Search is not available for this dataset
pipeline_tag
stringclasses 48
values | library_name
stringclasses 205
values | text
stringlengths 0
18.3M
| metadata
stringlengths 2
1.07B
| id
stringlengths 5
122
| last_modified
null | tags
listlengths 1
1.84k
| sha
null | created_at
stringlengths 25
25
|
|---|---|---|---|---|---|---|---|---|
text2text-generation
|
transformers
|
[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) for **Closed Book Question Answering**.
The model was pre-trained using T5's denoising objective on [C4](https://huggingface.co/datasets/c4), subsequently additionally pre-trained using [REALM](https://arxiv.org/pdf/2002.08909.pdf)'s salient span masking objective on [Wikipedia](https://huggingface.co/datasets/wikipedia), and finally fine-tuned on [Natural Questions (NQ)](https://huggingface.co/datasets/natural_questions).
**Note**: The model was fine-tuned on 100% of the train splits of [Natural Questions (NQ)](https://huggingface.co/datasets/natural_questions) for 10k steps.
Other community Checkpoints: [here](https://huggingface.co/models?search=ssm)
Paper: [How Much Knowledge Can You Pack
Into the Parameters of a Language Model?](https://arxiv.org/abs/1910.10683.pdf)
Authors: *Adam Roberts, Colin Raffel, Noam Shazeer*
## Results on Natural Questions - Test Set
|Id | link | Exact Match |
|---|---|---|
|T5-small|https://huggingface.co/google/t5-small-ssm-nq|25.5|
|T5-large|https://huggingface.co/google/t5-large-ssm-nq|30.4|
|T5-xl|https://huggingface.co/google/t5-xl-ssm-nq|35.6|
|T5-xxl|https://huggingface.co/google/t5-xxl-ssm-nq|37.9|
|**T5-3b**|**https://huggingface.co/google/t5-3b-ssm-nq**|**33.2**|
|T5-11b|https://huggingface.co/google/t5-11b-ssm-nq|36.6|
## Usage
The model can be used as follows for **closed book question answering**:
```python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
t5_qa_model = AutoModelForSeq2SeqLM.from_pretrained("google/t5-3b-ssm-nq")
t5_tok = AutoTokenizer.from_pretrained("google/t5-3b-ssm-nq")
input_ids = t5_tok("When was Franklin D. Roosevelt born?", return_tensors="pt").input_ids
gen_output = t5_qa_model.generate(input_ids)[0]
print(t5_tok.decode(gen_output, skip_special_tokens=True))
```
## Abstract
It has recently been observed that neural language models trained on unstructured text can implicitly store and retrieve knowledge using natural language queries. In this short paper, we measure the practical utility of this approach by fine-tuning pre-trained models to answer questions without access to any external context or knowledge. We show that this approach scales with model size and performs competitively with open-domain systems that explicitly retrieve answers from an external knowledge source when answering questions. To facilitate reproducibility and future work, we release our code and trained models at https://goo.gle/t5-cbqa.
|
{"language": "en", "license": "apache-2.0", "datasets": ["c4", "wikipedia", "natural_questions"], "pipeline_tag": "text2text-generation"}
|
google/t5-3b-ssm-nq
| null |
[
"transformers",
"pytorch",
"tf",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"dataset:wikipedia",
"dataset:natural_questions",
"arxiv:2002.08909",
"arxiv:1910.10683",
"license:apache-2.0",
"autotrain_compatible",
"endpoints_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) for **Closed Book Question Answering**.
The model was pre-trained using T5's denoising objective on [C4](https://huggingface.co/datasets/c4), subsequently additionally pre-trained using [REALM](https://arxiv.org/pdf/2002.08909.pdf)'s salient span masking objective on [Wikipedia](https://huggingface.co/datasets/wikipedia), and finally fine-tuned on [Natural Questions (NQ)](https://huggingface.co/datasets/natural_questions).
**Note**: The model was fine-tuned on 90% of the train splits of [Natural Questions (NQ)](https://huggingface.co/datasets/natural_questions) for 20k steps and validated on the held-out 10% of the train split.
Other community Checkpoints: [here](https://huggingface.co/models?search=ssm)
Paper: [How Much Knowledge Can You Pack
Into the Parameters of a Language Model?](https://arxiv.org/abs/1910.10683.pdf)
Authors: *Adam Roberts, Colin Raffel, Noam Shazeer*
## Results on Natural Questions - Test Set
|Id | link | Exact Match |
|---|---|---|
|T5-large|https://huggingface.co/google/t5-large-ssm-nqo|29.0|
|T5-xxl|https://huggingface.co/google/t5-xxl-ssm-nqo|35.2|
|**T5-3b**|**https://huggingface.co/google/t5-3b-ssm-nqo**|**31.7**|
|T5-11b|https://huggingface.co/google/t5-11b-ssm-nqo|34.8|
## Usage
The model can be used as follows for **closed book question answering**:
```python
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
t5_qa_model = AutoModelForSeq2SeqLM.from_pretrained("google/t5-3b-ssm-nqo")
t5_tok = AutoTokenizer.from_pretrained("google/t5-3b-ssm-nqo")
input_ids = t5_tok("When was Franklin D. Roosevelt born?", return_tensors="pt").input_ids
gen_output = t5_qa_model.generate(input_ids)[0]
print(t5_tok.decode(gen_output, skip_special_tokens=True))
```
## Abstract
It has recently been observed that neural language models trained on unstructured text can implicitly store and retrieve knowledge using natural language queries. In this short paper, we measure the practical utility of this approach by fine-tuning pre-trained models to answer questions without access to any external context or knowledge. We show that this approach scales with model size and performs competitively with open-domain systems that explicitly retrieve answers from an external knowledge source when answering questions. To facilitate reproducibility and future work, we release our code and trained models at https://goo.gle/t5-cbqa.
|
{"language": "en", "license": "apache-2.0", "datasets": ["c4", "wikipedia", "natural_questions"]}
|
google/t5-3b-ssm-nqo
| null |
[
"transformers",
"pytorch",
"tf",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"dataset:wikipedia",
"dataset:natural_questions",
"arxiv:2002.08909",
"arxiv:1910.10683",
"license:apache-2.0",
"autotrain_compatible",
"endpoints_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) for **Closed Book Question Answering**.
The model was pre-trained using T5's denoising objective on [C4](https://huggingface.co/datasets/c4) and subsequently additionally pre-trained using [REALM](https://arxiv.org/pdf/2002.08909.pdf)'s salient span masking objective on [Wikipedia](https://huggingface.co/datasets/wikipedia).
**Note**: This model should be fine-tuned on a question answering downstream task before it is useable for closed book question answering.
Other Community Checkpoints: [here](https://huggingface.co/models?search=ssm)
Paper: [How Much Knowledge Can You Pack
Into the Parameters of a Language Model?](https://arxiv.org/abs/1910.10683.pdf)
Authors: *Adam Roberts, Colin Raffel, Noam Shazeer*
## Abstract
It has recently been observed that neural language models trained on unstructured text can implicitly store and retrieve knowledge using natural language queries. In this short paper, we measure the practical utility of this approach by fine-tuning pre-trained models to answer questions without access to any external context or knowledge. We show that this approach scales with model size and performs competitively with open-domain systems that explicitly retrieve answers from an external knowledge source when answering questions. To facilitate reproducibility and future work, we release our code and trained models at https://goo.gle/t5-cbqa.
|
{"language": "en", "license": "apache-2.0", "datasets": ["c4", "wikipedia"]}
|
google/t5-3b-ssm
| null |
[
"transformers",
"pytorch",
"tf",
"t5",
"text2text-generation",
"en",
"dataset:c4",
"dataset:wikipedia",
"arxiv:2002.08909",
"arxiv:1910.10683",
"license:apache-2.0",
"autotrain_compatible",
"endpoints_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) Version 1.1 - LM-Adapted
## Version 1.1 - LM-Adapted
[T5 Version 1.1 - LM Adapted](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#lm-adapted-t511lm100k) includes the following improvements compared to the original [T5 model](https://huggingface.co/t5-base):
- GEGLU activation in feed-forward hidden layer, rather than ReLU - see [here](https://arxiv.org/abs/2002.05202).
- Dropout was turned off in pre-training (quality win). Dropout should be re-enabled during fine-tuning.
- Pre-trained on C4 only without mixing in the downstream tasks.
- no parameter sharing between embedding and classifier layer
- "xl" and "xxl" replace "3B" and "11B". The model shapes are a bit different - larger `d_model` and smaller `num_heads` and `d_ff`.
and is pretrained on both the denoising and language modeling objective.
More specifically, this checkpoint is initialized from [T5 Version 1.1 - Base](https://huggingface.co/google/https://huggingface.co/google/t5-v1_1-base)
and then trained for an additional 100K steps on the LM objective discussed in the [T5 paper](https://arxiv.org/pdf/1910.10683.pdf).
This adaptation improves the ability of the model to be used for prompt tuning.
**Note**: A popular fine-tuned version of the *T5 Version 1.1 - LM Adapted* model is [BigScience's T0pp](https://huggingface.co/bigscience/T0pp).
Pretraining Dataset: [C4](https://huggingface.co/datasets/c4)
Other Community Checkpoints: [here](https://huggingface.co/models?other=t5-lm-adapt)
Paper: [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)
Authors: *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu*
## Abstract
Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
|
{"language": "en", "license": "apache-2.0", "tags": ["t5-lm-adapt"], "datasets": ["c4"]}
|
google/t5-base-lm-adapt
| null |
[
"transformers",
"pytorch",
"tf",
"t5",
"text2text-generation",
"t5-lm-adapt",
"en",
"dataset:c4",
"arxiv:2002.05202",
"arxiv:1910.10683",
"license:apache-2.0",
"autotrain_compatible",
"endpoints_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DL2 (Deep-Narrow version)
T5-Efficient-BASE-DL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dl2** - is of model type **Base** with the following variations:
- **dl** is **2**
It has **128.52** million parameters and thus requires *ca.* **514.09 MB** of memory in full precision (*fp32*)
or **257.05 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dl2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DL4 (Deep-Narrow version)
T5-Efficient-BASE-DL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dl4** - is of model type **Base** with the following variations:
- **dl** is **4**
It has **147.4** million parameters and thus requires *ca.* **589.62 MB** of memory in full precision (*fp32*)
or **294.81 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dl4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DL6 (Deep-Narrow version)
T5-Efficient-BASE-DL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dl6** - is of model type **Base** with the following variations:
- **dl** is **6**
It has **166.29** million parameters and thus requires *ca.* **665.15 MB** of memory in full precision (*fp32*)
or **332.57 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dl6
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DL8 (Deep-Narrow version)
T5-Efficient-BASE-DL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dl8** - is of model type **Base** with the following variations:
- **dl** is **8**
It has **185.17** million parameters and thus requires *ca.* **740.67 MB** of memory in full precision (*fp32*)
or **370.34 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DM1000 (Deep-Narrow version)
T5-Efficient-BASE-DM1000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dm1000** - is of model type **Base** with the following variations:
- **dm** is **1000**
It has **297.23** million parameters and thus requires *ca.* **1188.93 MB** of memory in full precision (*fp32*)
or **594.47 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dm1000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DM2000 (Deep-Narrow version)
T5-Efficient-BASE-DM2000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dm2000** - is of model type **Base** with the following variations:
- **dm** is **2000**
It has **594.44** million parameters and thus requires *ca.* **2377.75 MB** of memory in full precision (*fp32*)
or **1188.87 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dm2000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DM256 (Deep-Narrow version)
T5-Efficient-BASE-DM256 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dm256** - is of model type **Base** with the following variations:
- **dm** is **256**
It has **74.33** million parameters and thus requires *ca.* **297.32 MB** of memory in full precision (*fp32*)
or **148.66 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dm256
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-DM512 (Deep-Narrow version)
T5-Efficient-BASE-DM512 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-dm512** - is of model type **Base** with the following variations:
- **dm** is **512**
It has **148.63** million parameters and thus requires *ca.* **594.52 MB** of memory in full precision (*fp32*)
or **297.26 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-dm512
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-EL16 (Deep-Narrow version)
T5-Efficient-BASE-EL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-el16** - is of model type **Base** with the following variations:
- **el** is **16**
It has **251.25** million parameters and thus requires *ca.* **1005.01 MB** of memory in full precision (*fp32*)
or **502.51 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-el16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-EL2 (Deep-Narrow version)
T5-Efficient-BASE-EL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-el2** - is of model type **Base** with the following variations:
- **el** is **2**
It has **152.13** million parameters and thus requires *ca.* **608.51 MB** of memory in full precision (*fp32*)
or **304.26 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-el2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-EL4 (Deep-Narrow version)
T5-Efficient-BASE-EL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-el4** - is of model type **Base** with the following variations:
- **el** is **4**
It has **166.29** million parameters and thus requires *ca.* **665.16 MB** of memory in full precision (*fp32*)
or **332.58 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-el4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-EL6 (Deep-Narrow version)
T5-Efficient-BASE-EL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-el6** - is of model type **Base** with the following variations:
- **el** is **6**
It has **180.45** million parameters and thus requires *ca.* **721.8 MB** of memory in full precision (*fp32*)
or **360.9 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-el6
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-EL8 (Deep-Narrow version)
T5-Efficient-BASE-EL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-el8** - is of model type **Base** with the following variations:
- **el** is **8**
It has **194.61** million parameters and thus requires *ca.* **778.44 MB** of memory in full precision (*fp32*)
or **389.22 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-el8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-FF1000 (Deep-Narrow version)
T5-Efficient-BASE-FF1000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-ff1000** - is of model type **Base** with the following variations:
- **ff** is **1000**
It has **147.43** million parameters and thus requires *ca.* **589.74 MB** of memory in full precision (*fp32*)
or **294.87 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-ff1000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-FF12000 (Deep-Narrow version)
T5-Efficient-BASE-FF12000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-ff12000** - is of model type **Base** with the following variations:
- **ff** is **12000**
It has **562.67** million parameters and thus requires *ca.* **2250.68 MB** of memory in full precision (*fp32*)
or **1125.34 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-ff12000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-FF2000 (Deep-Narrow version)
T5-Efficient-BASE-FF2000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-ff2000** - is of model type **Base** with the following variations:
- **ff** is **2000**
It has **185.18** million parameters and thus requires *ca.* **740.73 MB** of memory in full precision (*fp32*)
or **370.37 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-ff2000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-FF6000 (Deep-Narrow version)
T5-Efficient-BASE-FF6000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-ff6000** - is of model type **Base** with the following variations:
- **ff** is **6000**
It has **336.18** million parameters and thus requires *ca.* **1344.71 MB** of memory in full precision (*fp32*)
or **672.36 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-ff6000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-FF9000 (Deep-Narrow version)
T5-Efficient-BASE-FF9000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-ff9000** - is of model type **Base** with the following variations:
- **ff** is **9000**
It has **449.42** million parameters and thus requires *ca.* **1797.7 MB** of memory in full precision (*fp32*)
or **898.85 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-ff9000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-KV128 (Deep-Narrow version)
T5-Efficient-BASE-KV128 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-kv128** - is of model type **Base** with the following variations:
- **kv** is **128**
It has **307.87** million parameters and thus requires *ca.* **1231.47 MB** of memory in full precision (*fp32*)
or **615.73 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-kv128
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-KV16 (Deep-Narrow version)
T5-Efficient-BASE-KV16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-kv16** - is of model type **Base** with the following variations:
- **kv** is **16**
It has **159.23** million parameters and thus requires *ca.* **636.92 MB** of memory in full precision (*fp32*)
or **318.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-kv16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-KV256 (Deep-Narrow version)
T5-Efficient-BASE-KV256 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-kv256** - is of model type **Base** with the following variations:
- **kv** is **256**
It has **477.74** million parameters and thus requires *ca.* **1910.94 MB** of memory in full precision (*fp32*)
or **955.47 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-kv256
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-KV32 (Deep-Narrow version)
T5-Efficient-BASE-KV32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-kv32** - is of model type **Base** with the following variations:
- **kv** is **32**
It has **180.46** million parameters and thus requires *ca.* **721.86 MB** of memory in full precision (*fp32*)
or **360.93 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-kv32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NH16 (Deep-Narrow version)
T5-Efficient-BASE-NH16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nh16** - is of model type **Base** with the following variations:
- **nh** is **16**
It has **251.24** million parameters and thus requires *ca.* **1004.97 MB** of memory in full precision (*fp32*)
or **502.49 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nh16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NH24 (Deep-Narrow version)
T5-Efficient-BASE-NH24 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nh24** - is of model type **Base** with the following variations:
- **nh** is **24**
It has **307.87** million parameters and thus requires *ca.* **1231.47 MB** of memory in full precision (*fp32*)
or **615.73 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nh24
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NH32 (Deep-Narrow version)
T5-Efficient-BASE-NH32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nh32** - is of model type **Base** with the following variations:
- **nh** is **32**
It has **364.49** million parameters and thus requires *ca.* **1457.96 MB** of memory in full precision (*fp32*)
or **728.98 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nh32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NH8 (Deep-Narrow version)
T5-Efficient-BASE-NH8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nh8** - is of model type **Base** with the following variations:
- **nh** is **8**
It has **194.62** million parameters and thus requires *ca.* **778.48 MB** of memory in full precision (*fp32*)
or **389.24 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nh8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL16 (Deep-Narrow version)
T5-Efficient-BASE-NL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl16** - is of model type **Base** with the following variations:
- **nl** is **16**
It has **289.02** million parameters and thus requires *ca.* **1156.07 MB** of memory in full precision (*fp32*)
or **578.03 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL2 (Deep-Narrow version)
T5-Efficient-BASE-NL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl2** - is of model type **Base** with the following variations:
- **nl** is **2**
It has **57.72** million parameters and thus requires *ca.* **230.88 MB** of memory in full precision (*fp32*)
or **115.44 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL24 (Deep-Narrow version)
T5-Efficient-BASE-NL24 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl24** - is of model type **Base** with the following variations:
- **nl** is **24**
It has **421.19** million parameters and thus requires *ca.* **1684.75 MB** of memory in full precision (*fp32*)
or **842.37 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl24
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL32 (Deep-Narrow version)
T5-Efficient-BASE-NL32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl32** - is of model type **Base** with the following variations:
- **nl** is **32**
It has **553.36** million parameters and thus requires *ca.* **2213.43 MB** of memory in full precision (*fp32*)
or **1106.71 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL36 (Deep-Narrow version)
T5-Efficient-BASE-NL36 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl36** - is of model type **Base** with the following variations:
- **nl** is **36**
It has **619.44** million parameters and thus requires *ca.* **2477.77 MB** of memory in full precision (*fp32*)
or **1238.88 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl36
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL4 (Deep-Narrow version)
T5-Efficient-BASE-NL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl4** - is of model type **Base** with the following variations:
- **nl** is **4**
It has **90.76** million parameters and thus requires *ca.* **363.05 MB** of memory in full precision (*fp32*)
or **181.52 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL40 (Deep-Narrow version)
T5-Efficient-BASE-NL40 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl40** - is of model type **Base** with the following variations:
- **nl** is **40**
It has **685.53** million parameters and thus requires *ca.* **2742.11 MB** of memory in full precision (*fp32*)
or **1371.05 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl40
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL48 (Deep-Narrow version)
T5-Efficient-BASE-NL48 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl48** - is of model type **Base** with the following variations:
- **nl** is **48**
It has **817.7** million parameters and thus requires *ca.* **3270.79 MB** of memory in full precision (*fp32*)
or **1635.39 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl48
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE-NL8 (Deep-Narrow version)
T5-Efficient-BASE-NL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base-nl8** - is of model type **Base** with the following variations:
- **nl** is **8**
It has **156.85** million parameters and thus requires *ca.* **627.39 MB** of memory in full precision (*fp32*)
or **313.69 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base-nl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-BASE (Deep-Narrow version)
T5-Efficient-BASE is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-base** - is of model type **Base** with no variations.
It has **222.93** million parameters and thus requires *ca.* **891.73 MB** of memory in full precision (*fp32*)
or **445.86 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-base
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL12 (Deep-Narrow version)
T5-Efficient-LARGE-DL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl12** - is of model type **Large** with the following variations:
- **dl** is **12**
It has **536.34** million parameters and thus requires *ca.* **2145.37 MB** of memory in full precision (*fp32*)
or **1072.69 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL16 (Deep-Narrow version)
T5-Efficient-LARGE-DL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl16** - is of model type **Large** with the following variations:
- **dl** is **16**
It has **603.47** million parameters and thus requires *ca.* **2413.88 MB** of memory in full precision (*fp32*)
or **1206.94 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL2 (Deep-Narrow version)
T5-Efficient-LARGE-DL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl2** - is of model type **Large** with the following variations:
- **dl** is **2**
It has **368.53** million parameters and thus requires *ca.* **1474.11 MB** of memory in full precision (*fp32*)
or **737.05 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL32 (Deep-Narrow version)
T5-Efficient-LARGE-DL32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl32** - is of model type **Large** with the following variations:
- **dl** is **32**
It has **871.98** million parameters and thus requires *ca.* **3487.91 MB** of memory in full precision (*fp32*)
or **1743.96 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL4 (Deep-Narrow version)
T5-Efficient-LARGE-DL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl4** - is of model type **Large** with the following variations:
- **dl** is **4**
It has **402.09** million parameters and thus requires *ca.* **1608.36 MB** of memory in full precision (*fp32*)
or **804.18 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL6 (Deep-Narrow version)
T5-Efficient-LARGE-DL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl6** - is of model type **Large** with the following variations:
- **dl** is **6**
It has **435.65** million parameters and thus requires *ca.* **1742.61 MB** of memory in full precision (*fp32*)
or **871.31 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl6
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DL8 (Deep-Narrow version)
T5-Efficient-LARGE-DL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dl8** - is of model type **Large** with the following variations:
- **dl** is **8**
It has **469.22** million parameters and thus requires *ca.* **1876.87 MB** of memory in full precision (*fp32*)
or **938.43 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DM128 (Deep-Narrow version)
T5-Efficient-LARGE-DM128 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dm128** - is of model type **Large** with the following variations:
- **dm** is **128**
It has **92.27** million parameters and thus requires *ca.* **369.06 MB** of memory in full precision (*fp32*)
or **184.53 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dm128
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DM2000 (Deep-Narrow version)
T5-Efficient-LARGE-DM2000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dm2000** - is of model type **Large** with the following variations:
- **dm** is **2000**
It has **1475.39** million parameters and thus requires *ca.* **5901.57 MB** of memory in full precision (*fp32*)
or **2950.78 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dm2000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DM256 (Deep-Narrow version)
T5-Efficient-LARGE-DM256 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dm256** - is of model type **Large** with the following variations:
- **dm** is **256**
It has **184.47** million parameters and thus requires *ca.* **737.9 MB** of memory in full precision (*fp32*)
or **368.95 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dm256
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DM512 (Deep-Narrow version)
T5-Efficient-LARGE-DM512 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dm512** - is of model type **Large** with the following variations:
- **dm** is **512**
It has **368.89** million parameters and thus requires *ca.* **1475.56 MB** of memory in full precision (*fp32*)
or **737.78 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dm512
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-DM768 (Deep-Narrow version)
T5-Efficient-LARGE-DM768 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-dm768** - is of model type **Large** with the following variations:
- **dm** is **768**
It has **553.31** million parameters and thus requires *ca.* **2213.23 MB** of memory in full precision (*fp32*)
or **1106.62 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-dm768
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-EL12 (Deep-Narrow version)
T5-Efficient-LARGE-EL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el12** - is of model type **Large** with the following variations:
- **el** is **12**
It has **586.69** million parameters and thus requires *ca.* **2346.78 MB** of memory in full precision (*fp32*)
or **1173.39 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-el12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-EL2 (Deep-Narrow version)
T5-Efficient-LARGE-EL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el2** - is of model type **Large** with the following variations:
- **el** is **2**
It has **460.84** million parameters and thus requires *ca.* **1843.34 MB** of memory in full precision (*fp32*)
or **921.67 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-el2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-EL4 (Deep-Narrow version)
T5-Efficient-LARGE-EL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el4** - is of model type **Large** with the following variations:
- **el** is **4**
It has **486.01** million parameters and thus requires *ca.* **1944.03 MB** of memory in full precision (*fp32*)
or **972.01 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-el4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-EL6 (Deep-Narrow version)
T5-Efficient-LARGE-EL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el6** - is of model type **Large** with the following variations:
- **el** is **6**
It has **511.18** million parameters and thus requires *ca.* **2044.72 MB** of memory in full precision (*fp32*)
or **1022.36 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-el6
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-EL8 (Deep-Narrow version)
T5-Efficient-LARGE-EL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-el8** - is of model type **Large** with the following variations:
- **el** is **8**
It has **536.35** million parameters and thus requires *ca.* **2145.4 MB** of memory in full precision (*fp32*)
or **1072.7 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-el8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-KV128 (Deep-Narrow version)
T5-Efficient-LARGE-KV128 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-kv128** - is of model type **Large** with the following variations:
- **kv** is **128**
It has **1039.71** million parameters and thus requires *ca.* **4158.86 MB** of memory in full precision (*fp32*)
or **2079.43 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-kv128
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-KV16 (Deep-Narrow version)
T5-Efficient-LARGE-KV16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-kv16** - is of model type **Large** with the following variations:
- **kv** is **16**
It has **511.23** million parameters and thus requires *ca.* **2044.93 MB** of memory in full precision (*fp32*)
or **1022.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-kv16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-KV256 (Deep-Narrow version)
T5-Efficient-LARGE-KV256 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-kv256** - is of model type **Large** with the following variations:
- **kv** is **256**
It has **1643.69** million parameters and thus requires *ca.* **6574.78 MB** of memory in full precision (*fp32*)
or **3287.39 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-kv256
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-KV32 (Deep-Narrow version)
T5-Efficient-LARGE-KV32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-kv32** - is of model type **Large** with the following variations:
- **kv** is **32**
It has **586.73** million parameters and thus requires *ca.* **2346.92 MB** of memory in full precision (*fp32*)
or **1173.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-kv32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH12 (Deep-Narrow version)
T5-Efficient-LARGE-NH12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh12** - is of model type **Large** with the following variations:
- **nh** is **12**
It has **662.23** million parameters and thus requires *ca.* **2648.91 MB** of memory in full precision (*fp32*)
or **1324.45 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH2 (Deep-Narrow version)
T5-Efficient-LARGE-NH2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh2** - is of model type **Large** with the following variations:
- **nh** is **2**
It has **473.48** million parameters and thus requires *ca.* **1893.93 MB** of memory in full precision (*fp32*)
or **946.96 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH24 (Deep-Narrow version)
T5-Efficient-LARGE-NH24 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh24** - is of model type **Large** with the following variations:
- **nh** is **24**
It has **888.72** million parameters and thus requires *ca.* **3554.88 MB** of memory in full precision (*fp32*)
or **1777.44 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh24
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH32 (Deep-Narrow version)
T5-Efficient-LARGE-NH32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh32** - is of model type **Large** with the following variations:
- **nh** is **32**
It has **1039.72** million parameters and thus requires *ca.* **4158.86 MB** of memory in full precision (*fp32*)
or **2079.43 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH4 (Deep-Narrow version)
T5-Efficient-LARGE-NH4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh4** - is of model type **Large** with the following variations:
- **nh** is **4**
It has **511.23** million parameters and thus requires *ca.* **2044.92 MB** of memory in full precision (*fp32*)
or **1022.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH8-NL32 (Deep-Narrow version)
T5-Efficient-LARGE-NH8-NL32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh8-nl32** - is of model type **Large** with the following variations:
- **nh** is **8**
- **nl** is **32**
It has **771.34** million parameters and thus requires *ca.* **3085.35 MB** of memory in full precision (*fp32*)
or **1542.68 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh8-nl32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NH8 (Deep-Narrow version)
T5-Efficient-LARGE-NH8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nh8** - is of model type **Large** with the following variations:
- **nh** is **8**
It has **586.73** million parameters and thus requires *ca.* **2346.92 MB** of memory in full precision (*fp32*)
or **1173.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nh8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL10 (Deep-Narrow version)
T5-Efficient-LARGE-NL10 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl10** - is of model type **Large** with the following variations:
- **nl** is **10**
It has **326.58** million parameters and thus requires *ca.* **1306.31 MB** of memory in full precision (*fp32*)
or **653.16 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl10
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL12 (Deep-Narrow version)
T5-Efficient-LARGE-NL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl12** - is of model type **Large** with the following variations:
- **nl** is **12**
It has **385.31** million parameters and thus requires *ca.* **1541.25 MB** of memory in full precision (*fp32*)
or **770.63 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL16 (Deep-Narrow version)
T5-Efficient-LARGE-NL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl16** - is of model type **Large** with the following variations:
- **nl** is **16**
It has **502.78** million parameters and thus requires *ca.* **2011.13 MB** of memory in full precision (*fp32*)
or **1005.57 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL2 (Deep-Narrow version)
T5-Efficient-LARGE-NL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl2** - is of model type **Large** with the following variations:
- **nl** is **2**
It has **91.64** million parameters and thus requires *ca.* **366.55 MB** of memory in full precision (*fp32*)
or **183.28 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL20 (Deep-Narrow version)
T5-Efficient-LARGE-NL20 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl20** - is of model type **Large** with the following variations:
- **nl** is **20**
It has **620.25** million parameters and thus requires *ca.* **2481.02 MB** of memory in full precision (*fp32*)
or **1240.51 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl20
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL32 (Deep-Narrow version)
T5-Efficient-LARGE-NL32 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl32** - is of model type **Large** with the following variations:
- **nl** is **32**
It has **972.67** million parameters and thus requires *ca.* **3890.66 MB** of memory in full precision (*fp32*)
or **1945.33 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl32
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL36 (Deep-Narrow version)
T5-Efficient-LARGE-NL36 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl36** - is of model type **Large** with the following variations:
- **nl** is **36**
It has **1090.14** million parameters and thus requires *ca.* **4360.54 MB** of memory in full precision (*fp32*)
or **2180.27 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl36
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL4 (Deep-Narrow version)
T5-Efficient-LARGE-NL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl4** - is of model type **Large** with the following variations:
- **nl** is **4**
It has **150.37** million parameters and thus requires *ca.* **601.49 MB** of memory in full precision (*fp32*)
or **300.75 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE-NL8 (Deep-Narrow version)
T5-Efficient-LARGE-NL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large-nl8** - is of model type **Large** with the following variations:
- **nl** is **8**
It has **267.84** million parameters and thus requires *ca.* **1071.37 MB** of memory in full precision (*fp32*)
or **535.69 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large-nl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-LARGE (Deep-Narrow version)
T5-Efficient-LARGE is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-large** - is of model type **Large** with no variations.
It has **737.72** million parameters and thus requires *ca.* **2950.9 MB** of memory in full precision (*fp32*)
or **1475.45 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-large
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-MINI-NL12 (Deep-Narrow version)
T5-Efficient-MINI-NL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini-nl12** - is of model type **Mini** with the following variations:
- **nl** is **12**
It has **69.01** million parameters and thus requires *ca.* **276.05 MB** of memory in full precision (*fp32*)
or **138.03 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-mini-nl12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-MINI-NL24 (Deep-Narrow version)
T5-Efficient-MINI-NL24 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini-nl24** - is of model type **Mini** with the following variations:
- **nl** is **24**
It has **125.69** million parameters and thus requires *ca.* **502.75 MB** of memory in full precision (*fp32*)
or **251.37 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-mini-nl24
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-MINI-NL6 (Deep-Narrow version)
T5-Efficient-MINI-NL6 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini-nl6** - is of model type **Mini** with the following variations:
- **nl** is **6**
It has **40.68** million parameters and thus requires *ca.* **162.7 MB** of memory in full precision (*fp32*)
or **81.35 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-mini-nl6
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-MINI-NL8 (Deep-Narrow version)
T5-Efficient-MINI-NL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini-nl8** - is of model type **Mini** with the following variations:
- **nl** is **8**
It has **50.12** million parameters and thus requires *ca.* **200.49 MB** of memory in full precision (*fp32*)
or **100.24 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-mini-nl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-MINI (Deep-Narrow version)
T5-Efficient-MINI is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-mini** - is of model type **Mini** with no variations.
It has **31.23** million parameters and thus requires *ca.* **124.92 MB** of memory in full precision (*fp32*)
or **62.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-mini
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DL12 (Deep-Narrow version)
T5-Efficient-SMALL-DL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dl12** - is of model type **Small** with the following variations:
- **dl** is **12**
It has **85.7** million parameters and thus requires *ca.* **342.82 MB** of memory in full precision (*fp32*)
or **171.41 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dl12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DL16 (Deep-Narrow version)
T5-Efficient-SMALL-DL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dl16** - is of model type **Small** with the following variations:
- **dl** is **16**
It has **102.49** million parameters and thus requires *ca.* **409.97 MB** of memory in full precision (*fp32*)
or **204.99 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dl16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DL2 (Deep-Narrow version)
T5-Efficient-SMALL-DL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dl2** - is of model type **Small** with the following variations:
- **dl** is **2**
It has **43.73** million parameters and thus requires *ca.* **174.93 MB** of memory in full precision (*fp32*)
or **87.46 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dl2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DL4 (Deep-Narrow version)
T5-Efficient-SMALL-DL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dl4** - is of model type **Small** with the following variations:
- **dl** is **4**
It has **52.13** million parameters and thus requires *ca.* **208.51 MB** of memory in full precision (*fp32*)
or **104.25 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dl4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DL8 (Deep-Narrow version)
T5-Efficient-SMALL-DL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dl8** - is of model type **Small** with the following variations:
- **dl** is **8**
It has **68.92** million parameters and thus requires *ca.* **275.66 MB** of memory in full precision (*fp32*)
or **137.83 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DM1000 (Deep-Narrow version)
T5-Efficient-SMALL-DM1000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm1000** - is of model type **Small** with the following variations:
- **dm** is **1000**
It has **121.03** million parameters and thus requires *ca.* **484.11 MB** of memory in full precision (*fp32*)
or **242.05 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dm1000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DM128 (Deep-Narrow version)
T5-Efficient-SMALL-DM128 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm128** - is of model type **Small** with the following variations:
- **dm** is **128**
It has **15.14** million parameters and thus requires *ca.* **60.57 MB** of memory in full precision (*fp32*)
or **30.28 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dm128
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DM2000 (Deep-Narrow version)
T5-Efficient-SMALL-DM2000 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm2000** - is of model type **Small** with the following variations:
- **dm** is **2000**
It has **242.04** million parameters and thus requires *ca.* **968.16 MB** of memory in full precision (*fp32*)
or **484.08 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dm2000
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DM256 (Deep-Narrow version)
T5-Efficient-SMALL-DM256 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm256** - is of model type **Small** with the following variations:
- **dm** is **256**
It has **30.27** million parameters and thus requires *ca.* **121.07 MB** of memory in full precision (*fp32*)
or **60.54 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dm256
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-DM768 (Deep-Narrow version)
T5-Efficient-SMALL-DM768 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-dm768** - is of model type **Small** with the following variations:
- **dm** is **768**
It has **90.77** million parameters and thus requires *ca.* **363.1 MB** of memory in full precision (*fp32*)
or **181.55 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-dm768
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL12 (Deep-Narrow version)
T5-Efficient-SMALL-EL12 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el12** - is of model type **Small** with the following variations:
- **el** is **12**
It has **79.41** million parameters and thus requires *ca.* **317.63 MB** of memory in full precision (*fp32*)
or **158.81 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el12
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL16-DL1 (Deep-Narrow version)
T5-Efficient-SMALL-EL16-DL1 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el16-dl1** - is of model type **Small** with the following variations:
- **el** is **16**
- **dl** is **1**
It has **71.01** million parameters and thus requires *ca.* **284.04 MB** of memory in full precision (*fp32*)
or **142.02 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el16-dl1
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL16-DL2 (Deep-Narrow version)
T5-Efficient-SMALL-EL16-DL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el16-dl2** - is of model type **Small** with the following variations:
- **el** is **16**
- **dl** is **2**
It has **75.21** million parameters and thus requires *ca.* **300.83 MB** of memory in full precision (*fp32*)
or **150.42 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el16-dl2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL16-DL4 (Deep-Narrow version)
T5-Efficient-SMALL-EL16-DL4 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el16-dl4** - is of model type **Small** with the following variations:
- **el** is **16**
- **dl** is **4**
It has **83.6** million parameters and thus requires *ca.* **334.41 MB** of memory in full precision (*fp32*)
or **167.21 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el16-dl4
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL16-DL8 (Deep-Narrow version)
T5-Efficient-SMALL-EL16-DL8 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el16-dl8** - is of model type **Small** with the following variations:
- **el** is **16**
- **dl** is **8**
It has **100.39** million parameters and thus requires *ca.* **401.57 MB** of memory in full precision (*fp32*)
or **200.78 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el16-dl8
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL16 (Deep-Narrow version)
T5-Efficient-SMALL-EL16 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el16** - is of model type **Small** with the following variations:
- **el** is **16**
It has **92.0** million parameters and thus requires *ca.* **367.99 MB** of memory in full precision (*fp32*)
or **183.99 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el16
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
text2text-generation
|
transformers
|
# T5-Efficient-SMALL-EL2 (Deep-Narrow version)
T5-Efficient-SMALL-EL2 is a variation of [Google's original T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) following the [T5 model architecture](https://huggingface.co/docs/transformers/model_doc/t5).
It is a *pretrained-only* checkpoint and was released with the
paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)**
by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.
In a nutshell, the paper indicates that a **Deep-Narrow** model architecture is favorable for **downstream** performance compared to other model architectures
of similar parameter count.
To quote the paper:
> We generally recommend a DeepNarrow strategy where the model’s depth is preferentially increased
> before considering any other forms of uniform scaling across other dimensions. This is largely due to
> how much depth influences the Pareto-frontier as shown in earlier sections of the paper. Specifically, a
> tall small (deep and narrow) model is generally more efficient compared to the base model. Likewise,
> a tall base model might also generally more efficient compared to a large model. We generally find
> that, regardless of size, even if absolute performance might increase as we continue to stack layers,
> the relative gain of Pareto-efficiency diminishes as we increase the layers, converging at 32 to 36
> layers. Finally, we note that our notion of efficiency here relates to any one compute dimension, i.e.,
> params, FLOPs or throughput (speed). We report all three key efficiency metrics (number of params,
> FLOPS and speed) and leave this decision to the practitioner to decide which compute dimension to
> consider.
To be more precise, *model depth* is defined as the number of transformer blocks that are stacked sequentially.
A sequence of word embeddings is therefore processed sequentially by each transformer block.
## Details model architecture
This model checkpoint - **t5-efficient-small-el2** - is of model type **Small** with the following variations:
- **el** is **2**
It has **47.93** million parameters and thus requires *ca.* **191.72 MB** of memory in full precision (*fp32*)
or **95.86 MB** of memory in half precision (*fp16* or *bf16*).
A summary of the *original* T5 model architectures can be seen here:
| Model | nl (el/dl) | ff | dm | kv | nh | #Params|
| ----| ---- | ---- | ---- | ---- | ---- | ----|
| Tiny | 4/4 | 1024 | 256 | 32 | 4 | 16M|
| Mini | 4/4 | 1536 | 384 | 32 | 8 | 31M|
| Small | 6/6 | 2048 | 512 | 32 | 8 | 60M|
| Base | 12/12 | 3072 | 768 | 64 | 12 | 220M|
| Large | 24/24 | 4096 | 1024 | 64 | 16 | 738M|
| Xl | 24/24 | 16384 | 1024 | 128 | 32 | 3B|
| XXl | 24/24 | 65536 | 1024 | 128 | 128 | 11B|
whereas the following abbreviations are used:
| Abbreviation | Definition |
| ----| ---- |
| nl | Number of transformer blocks (depth) |
| dm | Dimension of embedding vector (output vector of transformers block) |
| kv | Dimension of key/value projection matrix |
| nh | Number of attention heads |
| ff | Dimension of intermediate vector within transformer block (size of feed-forward projection matrix) |
| el | Number of transformer blocks in the encoder (encoder depth) |
| dl | Number of transformer blocks in the decoder (decoder depth) |
| sh | Signifies that attention heads are shared |
| skv | Signifies that key-values projection matrices are tied |
If a model checkpoint has no specific, *el* or *dl* than both the number of encoder- and decoder layers correspond to *nl*.
## Pre-Training
The checkpoint was pretrained on the [Colossal, Cleaned version of Common Crawl (C4)](https://huggingface.co/datasets/c4) for 524288 steps using
the span-based masked language modeling (MLM) objective.
## Fine-Tuning
**Note**: This model is a **pretrained** checkpoint and has to be fine-tuned for practical usage.
The checkpoint was pretrained in English and is therefore only useful for English NLP tasks.
You can follow on of the following examples on how to fine-tune the model:
*PyTorch*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization)
- [Question Answering](https://github.com/huggingface/transformers/blob/master/examples/pytorch/question-answering/run_seq2seq_qa.py)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/pytorch/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*Tensorflow*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/tensorflow/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
*JAX/Flax*:
- [Summarization](https://github.com/huggingface/transformers/tree/master/examples/flax/summarization)
- [Text Classification](https://github.com/huggingface/transformers/tree/master/examples/flax/text-classification) - *Note*: You will have to slightly adapt the training example here to make it work with an encoder-decoder model.
## Downstream Performance
TODO: Add table if available
## Computational Complexity
TODO: Add table if available
## More information
We strongly recommend the reader to go carefully through the original paper **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** to get a more nuanced understanding of this model checkpoint.
As explained in the following [issue](https://github.com/google-research/google-research/issues/986#issuecomment-1035051145), checkpoints including the *sh* or *skv*
model architecture variations have *not* been ported to Transformers as they are probably of limited practical usage and are lacking a more detailed description. Those checkpoints are kept [here](https://huggingface.co/NewT5SharedHeadsSharedKeyValues) as they might be ported potentially in the future.
|
{"language": ["en"], "license": "apache-2.0", "tags": ["deep-narrow"], "datasets": ["c4"], "inference": false}
|
google/t5-efficient-small-el2
| null |
[
"transformers",
"pytorch",
"tf",
"jax",
"t5",
"text2text-generation",
"deep-narrow",
"en",
"dataset:c4",
"arxiv:2109.10686",
"license:apache-2.0",
"autotrain_compatible",
"has_space",
"text-generation-inference",
"region:us"
] | null |
2022-03-02T23:29:05+00:00
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.