modelId
stringlengths 4
112
| lastModified
stringlengths 24
24
| tags
list | pipeline_tag
stringclasses 21
values | files
list | publishedBy
stringlengths 2
37
| downloads_last_month
int32 0
9.44M
| library
stringclasses 15
values | modelCard
large_stringlengths 0
100k
|
|---|---|---|---|---|---|---|---|---|
facebook/blenderbot_small-90M | 2021-01-12T10:12:52.000Z | [
"pytorch",
"tf",
"blenderbot-small",
"seq2seq",
"en",
"dataset:blended_skill_talk",
"arxiv:1907.06616",
"transformers",
"convAI",
"conversational",
"facebook",
"license:apache-2.0",
"text2text-generation"
]
| conversational | [
".gitattributes",
"README.md",
"config.json",
"merges.txt",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 3,128 | transformers | ---
language:
- en
thumbnail:
tags:
- convAI
- conversational
- facebook
license: apache-2.0
datasets:
- blended_skill_talk
metrics:
- perplexity
---
## Model description
+ Paper: [Recipes for building an open-domain chatbot](https://arxiv.org/abs/1907.06616)
+ [Original PARLAI Code](https://parl.ai/projects/recipes/)
### Abstract
Building open-domain chatbots is a challenging area for machine learning research. While prior work has shown that scaling neural models in the number of parameters and the size of the data they are trained on gives improved results, we show that other ingredients are important for a high-performing chatbot. Good conversation requires a number of skills that an expert conversationalist blends in a seamless way: providing engaging talking points and listening to their partners, both asking and answering questions, and displaying knowledge, empathy and personality appropriately, depending on the situation. We show that large scale models can learn these skills when given appropriate training data and choice of generation strategy. We build variants of these recipes with 90M, 2.7B and 9.4B parameter neural models, and make our models and code publicly available. Human evaluations show our best models are superior to existing approaches in multi-turn dialogue in terms of engagingness and humanness measurements. We then discuss the limitations of this work by analyzing failure cases of our models.
|
facebook/deit-base-distilled-patch16-224 | 2021-04-09T18:26:08.000Z | [
"pytorch",
"deit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 804 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Distilled Data-efficient Image Transformer (base-sized model)
Distilled data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is a distilled Vision Transformer (ViT). It uses a distillation token, besides the class token, to effectively learn from a teacher (CNN) during both pre-training and fine-tuning. The distillation token is learned through backpropagation, by interacting with the class ([CLS]) and patch tokens through the self-attention layers.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a distilled ViT model, you can plug it into DeiTModel, DeiTForImageClassification or DeiTForImageClassificationWithTeacher. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, DeiTForImageClassificationWithTeacher
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-base-distilled-patch16-224')
model = DeiTForImageClassificationWithTeacher.from_pretrained('facebook/deit-base-distilled-patch16-224')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
This model was pretrained and fine-tuned with distillation on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (256x256), center-cropped at 224x224 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| **DeiT-base distilled** | **83.4** | **96.5** | **87M** | **https://huggingface.co/facebook/deit-base-distilled-patch16-224** |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-base-distilled-patch16-384 | 2021-04-09T18:23:31.000Z | [
"pytorch",
"deit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 173 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Distilled Data-efficient Image Transformer (base-sized model)
Distilled data-efficient Image Transformer (DeiT) model pre-trained at resolution 224x224 and fine-tuned at resolution 384x384 on ImageNet-1k (1 million images, 1,000 classes). It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is a distilled Vision Transformer (ViT). It uses a distillation token, besides the class token, to effectively learn from a teacher (CNN) during both pre-training and fine-tuning. The distillation token is learned through backpropagation, by interacting with the class ([CLS]) and patch tokens through the self-attention layers.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a distilled ViT model, you can plug it into DeiTModel, DeiTForImageClassification or DeiTForImageClassificationWithTeacher. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, DeiTForImageClassificationWithTeacher
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-base-distilled-patch16-384')
model = DeiTForImageClassificationWithTeacher.from_pretrained('facebook/deit-base-distilled-patch16-384')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
This model was pretrained and fine-tuned with distillation on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (438x438), center-cropped at 384x384 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Pre-training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|-------------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| **DeiT-base distilled 384 (1000 epochs)** | **85.2** | **97.2** | **88M** | **https://huggingface.co/facebook/deit-base-distilled-patch16-384** |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-base-patch16-224 | 2021-04-09T18:23:55.000Z | [
"pytorch",
"vit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 57 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Data-efficient Image Transformer (base-sized model)
Data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is actually a more efficiently trained Vision Transformer (ViT).
The Vision Transformer (ViT) is a transformer encoder model (BERT-like) pre-trained and fine-tuned on a large collection of images in a supervised fashion, namely ImageNet-1k, at a resolution of 224x224 pixels.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds a [CLS] token to the beginning of a sequence to use it for classification tasks. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder.
By pre-training the model, it learns an inner representation of images that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled images for instance, you can train a standard classifier by placing a linear layer on top of the pre-trained encoder. One typically places a linear layer on top of the [CLS] token, as the last hidden state of this token can be seen as a representation of an entire image.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a more efficiently trained ViT model, you can plug it into ViTModel or ViTForImageClassification. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, ViTForImageClassification
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-base-patch16-224')
model = ViTForImageClassification.from_pretrained('facebook/deit-base-patch16-224')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
The ViT model was pretrained on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (256x256), center-cropped at 224x224 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| **DeiT-base** | **81.8** | **95.6** | **86M** | **https://huggingface.co/facebook/deit-base-patch16-224** |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-base-patch16-384 | 2021-04-09T18:26:49.000Z | [
"pytorch",
"vit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 64 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Data-efficient Image Transformer (base-sized model)
Data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 384x384. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is actually a more efficiently trained Vision Transformer (ViT).
The Vision Transformer (ViT) is a transformer encoder model (BERT-like) pre-trained at resolution 224 and fine-tuned at resolution 384 on a large collection of images in a supervised fashion, namely ImageNet-1k.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds a [CLS] token to the beginning of a sequence to use it for classification tasks. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder.
By pre-training the model, it learns an inner representation of images that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled images for instance, you can train a standard classifier by placing a linear layer on top of the pre-trained encoder. One typically places a linear layer on top of the [CLS] token, as the last hidden state of this token can be seen as a representation of an entire image.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a more efficiently trained ViT model, you can plug it into ViTModel or ViTForImageClassification. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, ViTForImageClassification
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-base-patch16-384')
model = ViTForImageClassification.from_pretrained('facebook/deit-base-patch16-384')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
The ViT model was pretrained on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (438x438), center-cropped at 384x384 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Pre-training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| **DeiT-base 384** | **82.9** | **96.2** | **87M** | **https://huggingface.co/facebook/deit-base-patch16-384** |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-small-distilled-patch16-224 | 2021-04-09T18:23:10.000Z | [
"pytorch",
"deit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 12 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Distilled Data-efficient Image Transformer (small-sized model)
Distilled data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is a distilled Vision Transformer (ViT). It uses a distillation token, besides the class token, to effectively learn from a teacher (CNN) during both pre-training and fine-tuning. The distillation token is learned through backpropagation, by interacting with the class ([CLS]) and patch tokens through the self-attention layers.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a distilled ViT model, you can plug it into DeiTModel, DeiTForImageClassification or DeiTForImageClassificationWithTeacher. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, DeiTForImageClassificationWithTeacher
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-small-distilled-patch16-224')
model = DeiTForImageClassificationWithTeacher.from_pretrained('facebook/deit-small-distilled-patch16-224')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
This model was pretrained and fine-tuned with distillation on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (256x256), center-cropped at 224x224 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| **DeiT-small distilled** | **81.2** | **95.4** | **22M** | **https://huggingface.co/facebook/deit-small-distilled-patch16-224** |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-small-patch16-224 | 2021-04-09T18:27:20.000Z | [
"pytorch",
"vit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 72 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Data-efficient Image Transformer (small-sized model)
Data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is actually a more efficiently trained Vision Transformer (ViT).
The Vision Transformer (ViT) is a transformer encoder model (BERT-like) pre-trained and fine-tuned on a large collection of images in a supervised fashion, namely ImageNet-1k, at a resolution of 224x224 pixels.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds a [CLS] token to the beginning of a sequence to use it for classification tasks. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder.
By pre-training the model, it learns an inner representation of images that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled images for instance, you can train a standard classifier by placing a linear layer on top of the pre-trained encoder. One typically places a linear layer on top of the [CLS] token, as the last hidden state of this token can be seen as a representation of an entire image.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a more efficiently trained ViT model, you can plug it into ViTModel or ViTForImageClassification. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, ViTForImageClassification
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-small-patch16-224')
model = ViTForImageClassification.from_pretrained('facebook/deit-small-patch16-224')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
The ViT model was pretrained on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (256x256), center-cropped at 224x224 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| **DeiT-small** | **79.9** | **95.0** | **22M** | **https://huggingface.co/facebook/deit-small-patch16-224** |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-tiny-distilled-patch16-224 | 2021-04-09T18:27:35.000Z | [
"pytorch",
"deit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 36 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Distilled Data-efficient Image Transformer (tiny-sized model)
Distilled data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is a distilled Vision Transformer (ViT). It uses a distillation token, besides the class token, to effectively learn from a teacher (CNN) during both pre-training and fine-tuning. The distillation token is learned through backpropagation, by interacting with the class ([CLS]) and patch tokens through the self-attention layers.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a distilled ViT model, you can plug it into DeiTModel, DeiTForImageClassification or DeiTForImageClassificationWithTeacher. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, DeiTForImageClassificationWithTeacher
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-tiny-distilled-patch16-224')
model = DeiTForImageClassificationWithTeacher.from_pretrained('facebook/deit-tiny-distilled-patch16-224')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
This model was pretrained and fine-tuned with distillation on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (256x256), center-cropped at 224x224 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| DeiT-tiny | 72.2 | 91.1 | 5M | https://huggingface.co/facebook/deit-tiny-patch16-224 |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| **DeiT-tiny distilled** | **74.5** | **91.9** | **6M** | **https://huggingface.co/facebook/deit-tiny-distilled-patch16-224** |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/deit-tiny-patch16-224 | 2021-04-09T18:27:55.000Z | [
"pytorch",
"vit",
"dataset:imagenet",
"arxiv:2012.12877",
"arxiv:2006.03677",
"transformers",
"license:apache-2.0",
"image-classification"
]
| image-classification | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 33 | transformers | ---
license: apache-2.0
tags:
- image-classification
datasets:
- imagenet
---
# Data-efficient Image Transformer (tiny-sized model)
Data-efficient Image Transformer (DeiT) model pre-trained and fine-tuned on ImageNet-1k (1 million images, 1,000 classes) at resolution 224x224. It was first introduced in the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Touvron et al. and first released in [this repository](https://github.com/facebookresearch/deit). However, the weights were converted from the [timm repository](https://github.com/rwightman/pytorch-image-models) by Ross Wightman.
Disclaimer: The team releasing DeiT did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
This model is actually a more efficiently trained Vision Transformer (ViT).
The Vision Transformer (ViT) is a transformer encoder model (BERT-like) pre-trained and fine-tuned on a large collection of images in a supervised fashion, namely ImageNet-1k, at a resolution of 224x224 pixels.
Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which are linearly embedded. One also adds a [CLS] token to the beginning of a sequence to use it for classification tasks. One also adds absolute position embeddings before feeding the sequence to the layers of the Transformer encoder.
By pre-training the model, it learns an inner representation of images that can then be used to extract features useful for downstream tasks: if you have a dataset of labeled images for instance, you can train a standard classifier by placing a linear layer on top of the pre-trained encoder. One typically places a linear layer on top of the [CLS] token, as the last hidden state of this token can be seen as a representation of an entire image.
## Intended uses & limitations
You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=facebook/deit) to look for
fine-tuned versions on a task that interests you.
### How to use
Since this model is a more efficiently trained ViT model, you can plug it into ViTModel or ViTForImageClassification. Note that the model expects the data to be prepared using DeiTFeatureExtractor. Here we use AutoFeatureExtractor, which will automatically use the appropriate feature extractor given the model name.
Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes:
```python
from transformers import AutoFeatureExtractor, ViTForImageClassification
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = AutoFeatureExtractor.from_pretrained('facebook/deit-tiny-patch16-224')
model = ViTForImageClassification.from_pretrained('facebook/deit-tiny-patch16-224')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
# model predicts one of the 1000 ImageNet classes
predicted_class_idx = logits.argmax(-1).item()
print("Predicted class:", model.config.id2label[predicted_class_idx])
```
Currently, both the feature extractor and model support PyTorch. Tensorflow and JAX/FLAX are coming soon.
## Training data
The ViT model was pretrained on [ImageNet-1k](http://www.image-net.org/challenges/LSVRC/2012/), a dataset consisting of 1 million images and 1k classes.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/deit/blob/ab5715372db8c6cad5740714b2216d55aeae052e/datasets.py#L78).
At inference time, images are resized/rescaled to the same resolution (256x256), center-cropped at 224x224 and normalized across the RGB channels with the ImageNet mean and standard deviation.
### Pretraining
The model was trained on a single 8-GPU node for 3 days. Training resolution is 224. For all hyperparameters (such as batch size and learning rate) we refer to table 9 of the original paper.
## Evaluation results
| Model | ImageNet top-1 accuracy | ImageNet top-5 accuracy | # params | URL |
|---------------------------------------|-------------------------|-------------------------|----------|------------------------------------------------------------------|
| **DeiT-tiny** | **72.2** | **91.1** | **5M** | **https://huggingface.co/facebook/deit-tiny-patch16-224** |
| DeiT-small | 79.9 | 95.0 | 22M | https://huggingface.co/facebook/deit-small-patch16-224 |
| DeiT-base | 81.8 | 95.6 | 86M | https://huggingface.co/facebook/deit-base-patch16-224 |
| DeiT-tiny distilled | 74.5 | 91.9 | 6M | https://huggingface.co/facebook/deit-tiny-distilled-patch16-224 |
| DeiT-small distilled | 81.2 | 95.4 | 22M | https://huggingface.co/facebook/deit-small-distilled-patch16-224 |
| DeiT-base distilled | 83.4 | 96.5 | 87M | https://huggingface.co/facebook/deit-base-distilled-patch16-224 |
| DeiT-base 384 | 82.9 | 96.2 | 87M | https://huggingface.co/facebook/deit-base-patch16-384 |
| DeiT-base distilled 384 (1000 epochs) | 85.2 | 97.2 | 88M | https://huggingface.co/facebook/deit-base-distilled-patch16-384 |
Note that for fine-tuning, the best results are obtained with a higher resolution (384x384). Of course, increasing the model size will result in better performance.
### BibTeX entry and citation info
```bibtex
@misc{touvron2021training,
title={Training data-efficient image transformers & distillation through attention},
author={Hugo Touvron and Matthieu Cord and Matthijs Douze and Francisco Massa and Alexandre Sablayrolles and Hervé Jégou},
year={2021},
eprint={2012.12877},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@misc{wu2020visual,
title={Visual Transformers: Token-based Image Representation and Processing for Computer Vision},
author={Bichen Wu and Chenfeng Xu and Xiaoliang Dai and Alvin Wan and Peizhao Zhang and Zhicheng Yan and Masayoshi Tomizuka and Joseph Gonzalez and Kurt Keutzer and Peter Vajda},
year={2020},
eprint={2006.03677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
```bibtex
@inproceedings{deng2009imagenet,
title={Imagenet: A large-scale hierarchical image database},
author={Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li},
booktitle={2009 IEEE conference on computer vision and pattern recognition},
pages={248--255},
year={2009},
organization={Ieee}
}
``` |
facebook/detr-resnet-101-dc5 | 2021-06-08T14:20:25.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"object-detection"
]
| object-detection | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 4 | transformers | ---
license: apache-2.0
tags:
- object-detection
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-101 backbone (dilated C5 stage)
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 object detection (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
## Intended uses & limitations
You can use the raw model for object detection. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForObjectDetection
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-101-dc5')
model = DetrForObjectDetection.from_pretrained('facebook/detr-resnet-101-dc5')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts bounding boxes and corresponding COCO classes
logits = outputs.logits
bboxes = outputs.pred_boxes
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 object detection](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/google-research/vision_transformer/blob/master/vit_jax/input_pipeline.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves an AP (average precision) of **44.9** on COCO 2017 validation. For more details regarding evaluation results, we refer to table 1 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/detr-resnet-101-panoptic | 2021-06-08T14:22:54.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"image-segmentation"
]
| image-segmentation | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 6 | transformers | ---
license: apache-2.0
tags:
- image-segmentation
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-101 backbone
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 panoptic (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
DETR can be naturally extended to perform panoptic segmentation, by adding a mask head on top of the decoder outputs.
## Intended uses & limitations
You can use the raw model for panoptic segmentation. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForSegmentation
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-101-panoptic')
model = DetrForSegmentation.from_pretrained('facebook/detr-resnet-101-panoptic')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts COCO classes, bounding boxes, and masks
logits = outputs.logits
bboxes = outputs.pred_boxes
masks = outputs.pred_masks
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 panoptic](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/detr/blob/master/datasets/coco_panoptic.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves the following results on COCO 2017 validation: a box AP (average precision) of **40.1**, a segmentation AP (average precision) of **33** and a PQ (panoptic quality) of **45.1**.
For more details regarding evaluation results, we refer to table 5 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/detr-resnet-101 | 2021-06-08T14:17:01.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"object-detection"
]
| object-detection | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 15 | transformers | ---
license: apache-2.0
tags:
- object-detection
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-101 backbone
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 object detection (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
## Intended uses & limitations
You can use the raw model for object detection. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForObjectDetection
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-101')
model = DetrForObjectDetection.from_pretrained('facebook/detr-resnet-101')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts bounding boxes and corresponding COCO classes
logits = outputs.logits
bboxes = outputs.pred_boxes
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 object detection](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/google-research/vision_transformer/blob/master/vit_jax/input_pipeline.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves an AP (average precision) of **43.5** on COCO 2017 validation. For more details regarding evaluation results, we refer to table 1 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/detr-resnet-50-dc5-panoptic | 2021-06-08T14:11:56.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"image-segmentation"
]
| image-segmentation | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 45 | transformers | ---
license: apache-2.0
tags:
- image-segmentation
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-50 backbone (dilated C5 stage)
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 panoptic (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
DETR can be naturally extended to perform panoptic segmentation, by adding a mask head on top of the decoder outputs.
## Intended uses & limitations
You can use the raw model for panoptic segmentation. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForSegmentation
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-50-dc5-panoptic')
model = DetrForSegmentation.from_pretrained('facebook/detr-resnet-50-dc5-panoptic')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts COCO classes, bounding boxes, and masks
logits = outputs.logits
bboxes = outputs.pred_boxes
masks = outputs.pred_masks
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 panoptic](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/detr/blob/master/datasets/coco_panoptic.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves the following results on COCO 2017 validation: a box AP (average precision) of **40.2**, a segmentation AP (average precision) of **31.9** and a PQ (panoptic quality) of **44.6**.
For more details regarding evaluation results, we refer to table 5 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/detr-resnet-50-dc5 | 2021-06-08T14:09:02.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"object-detection"
]
| object-detection | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 7 | transformers | ---
license: apache-2.0
tags:
- object-detection
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-50 backbone (dilated C5 stage)
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 object detection (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
## Intended uses & limitations
You can use the raw model for object detection. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForObjectDetection
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-50-dc5')
model = DetrForObjectDetection.from_pretrained('facebook/detr-resnet-50-dc5')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts bounding boxes and corresponding COCO classes
logits = outputs.logits
bboxes = outputs.pred_boxes
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 object detection](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/google-research/vision_transformer/blob/master/vit_jax/input_pipeline.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves an AP (average precision) of **43.3** on COCO 2017 validation. For more details regarding evaluation results, we refer to table 1 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/detr-resnet-50-panoptic | 2021-06-08T14:14:38.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"image-segmentation"
]
| image-segmentation | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 204 | transformers | ---
license: apache-2.0
tags:
- image-segmentation
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-50 backbone
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 panoptic (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
DETR can be naturally extended to perform panoptic segmentation, by adding a mask head on top of the decoder outputs.
## Intended uses & limitations
You can use the raw model for panoptic segmentation. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForSegmentation
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-50-panoptic')
model = DetrForSegmentation.from_pretrained('facebook/detr-resnet-50-panoptic')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts COCO classes, bounding boxes, and masks
logits = outputs.logits
bboxes = outputs.pred_boxes
masks = outputs.pred_masks
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 panoptic](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/facebookresearch/detr/blob/master/datasets/coco_panoptic.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves the following results on COCO 2017 validation: a box AP (average precision) of **38.8**, a segmentation AP (average precision) of **31.1** and a PQ (panoptic quality) of **43.4**.
For more details regarding evaluation results, we refer to table 5 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/detr-resnet-50 | 2021-06-08T14:04:44.000Z | [
"pytorch",
"detr",
"dataset:coco",
"arxiv:2005.12872",
"transformers",
"license:apache-2.0",
"object-detection"
]
| object-detection | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 1,162 | transformers | ---
license: apache-2.0
tags:
- object-detection
datasets:
- coco
---
# DETR (End-to-End Object Detection) model with ResNet-50 backbone
DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 object detection (118k annotated images). It was introduced in the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Carion et al. and first released in [this repository](https://github.com/facebookresearch/detr).
Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.
## Model description
The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.
The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.
## Intended uses & limitations
You can use the raw model for object detection. See the [model hub](https://huggingface.co/models?search=facebook/detr) to look for all available DETR models.
### How to use
Here is how to use this model:
```python
from transformers import DetrFeatureExtractor, DetrForObjectDetection
from PIL import Image
import requests
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
image = Image.open(requests.get(url, stream=True).raw)
feature_extractor = DetrFeatureExtractor.from_pretrained('facebook/detr-resnet-50')
model = DetrForObjectDetection.from_pretrained('facebook/detr-resnet-50')
inputs = feature_extractor(images=image, return_tensors="pt")
outputs = model(**inputs)
# model predicts bounding boxes and corresponding COCO classes
logits = outputs.logits
bboxes = outputs.pred_boxes
```
Currently, both the feature extractor and model support PyTorch.
## Training data
The DETR model was trained on [COCO 2017 object detection](https://cocodataset.org/#download), a dataset consisting of 118k/5k annotated images for training/validation respectively.
## Training procedure
### Preprocessing
The exact details of preprocessing of images during training/validation can be found [here](https://github.com/google-research/vision_transformer/blob/master/vit_jax/input_pipeline.py).
Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).
### Training
The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).
## Evaluation results
This model achieves an AP (average precision) of **42.0** on COCO 2017 validation. For more details regarding evaluation results, we refer to table 1 of the original paper.
### BibTeX entry and citation info
```bibtex
@article{DBLP:journals/corr/abs-2005-12872,
author = {Nicolas Carion and
Francisco Massa and
Gabriel Synnaeve and
Nicolas Usunier and
Alexander Kirillov and
Sergey Zagoruyko},
title = {End-to-End Object Detection with Transformers},
journal = {CoRR},
volume = {abs/2005.12872},
year = {2020},
url = {https://arxiv.org/abs/2005.12872},
archivePrefix = {arXiv},
eprint = {2005.12872},
timestamp = {Thu, 28 May 2020 17:38:09 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2005-12872.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
``` |
facebook/dpr-ctx_encoder-multiset-base | 2020-11-25T16:58:57.000Z | [
"pytorch",
"tf",
"dpr",
"transformers"
]
| [
".gitattributes",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
]
| facebook | 4,430 | transformers | ||
facebook/dpr-ctx_encoder-single-nq-base | 2020-11-25T16:58:35.000Z | [
"pytorch",
"tf",
"dpr",
"transformers"
]
| [
".DS_Store",
".gitattributes",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
]
| facebook | 115,732 | transformers | ||
facebook/dpr-question_encoder-multiset-base | 2020-11-25T16:59:33.000Z | [
"pytorch",
"tf",
"dpr",
"transformers"
]
| [
".gitattributes",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
]
| facebook | 4,636 | transformers | ||
facebook/dpr-question_encoder-single-nq-base | 2020-11-25T16:59:20.000Z | [
"pytorch",
"tf",
"dpr",
"transformers"
]
| [
".DS_Store",
".gitattributes",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
]
| facebook | 87,115 | transformers | ||
facebook/dpr-reader-multiset-base | 2020-11-25T17:00:07.000Z | [
"pytorch",
"tf",
"dpr",
"transformers"
]
| [
".gitattributes",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
]
| facebook | 211 | transformers | ||
facebook/dpr-reader-single-nq-base | 2020-11-25T16:59:53.000Z | [
"pytorch",
"tf",
"dpr",
"transformers"
]
| [
".DS_Store",
".gitattributes",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"tokenizer.json",
"tokenizer_config.json",
"vocab.txt"
]
| facebook | 3,519 | transformers | ||
facebook/hubert-base-ls960 | 2021-06-16T10:03:18.000Z | [
"pytorch",
"hubert",
"en",
"dataset:librispeech_asr",
"arxiv:2106.07447",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 42 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
license: apache-2.0
---
# Hubert-Base
[Facebook's Hubert](https://ai.facebook.com/blog/hubert-self-supervised-representation-learning-for-speech-recognition-generation-and-compression)
The base model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note that this model should be fine-tuned on a downstream task, like Automatic Speech Recognition, Speaker Identification, Intent Classification, Emotion Recognition, etc...
[Paper](https://arxiv.org/abs/2106.07447)
Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed
**Abstract**
Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/hubert .
# Usage
See [this blog](https://huggingface.co/blog/fine-tune-wav2vec2-english) for more information on how to fine-tune the model. Note that the class `Wav2Vec2ForCTC` has to be replaced by `HubertForCTC`. |
|
facebook/hubert-large-ll60k | 2021-06-16T10:06:34.000Z | [
"pytorch",
"hubert",
"en",
"dataset:libri-light",
"arxiv:2106.07447",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 30 | transformers | ---
language: en
datasets:
- libri-light
tags:
- speech
license: apache-2.0
---
# Hubert-Large
[Facebook's Hubert](https://ai.facebook.com/blog/hubert-self-supervised-representation-learning-for-speech-recognition-generation-and-compression)
The large model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note that this model should be fine-tuned on a downstream task, like Automatic Speech Recognition, Speaker Identification, Intent Classification, Emotion Recognition, etc...
The model was pretrained on [Libri-Light](https://github.com/facebookresearch/libri-light).
[Paper](https://arxiv.org/abs/2106.07447)
Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed
**Abstract**
Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/hubert .
# Usage
See [this blog](https://huggingface.co/blog/fine-tune-wav2vec2-english) for more information on how to fine-tune the model. Note that the class `Wav2Vec2ForCTC` has to be replaced by `HubertForCTC`. |
|
facebook/hubert-large-ls960-ft | 2021-06-16T11:23:13.000Z | [
"pytorch",
"hubert",
"en",
"dataset:libri-light",
"dataset:librispeech_asr",
"arxiv:2106.07447",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 646 | transformers | ---
language: en
datasets:
- libri-light
- librispeech_asr
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
# Hubert-Large-Finetuned
[Facebook's Hubert](https://ai.facebook.com/blog/hubert-self-supervised-representation-learning-for-speech-recognition-generation-and-compression)
The large model fine-tuned on 960h of Librispeech on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz.
The model is a fine-tuned version of [hubert-large-ll60k](https://huggingface.co/facebook/hubert-large-ll60k).
[Paper](https://arxiv.org/abs/2106.07447)
Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed
**Abstract**
Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/hubert .
# Usage
The model can be used for automatic-speech-recognition as follows:
```python
import torch
from transformers import Wav2Vec2Processor, HubertForCTC
from datasets import load_dataset
import soundfile as sf
processor = Wav2Vec2Processor.from_pretrained("facebook/hubert-large-ls960-ft")
model = HubertForCTC.from_pretrained("facebook/hubert-large-ls960-ft")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
input_values = processor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.decode(predicted_ids[0])
# ->"A MAN SAID TO THE UNIVERSE SIR I EXIST"
``` |
facebook/hubert-xlarge-ll60k | 2021-06-16T10:06:17.000Z | [
"pytorch",
"hubert",
"en",
"dataset:libri-light",
"arxiv:2106.07447",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 10 | transformers | ---
language: en
datasets:
- libri-light
tags:
- speech
license: apache-2.0
---
# Hubert-Extra-Large
[Facebook's Hubert](https://ai.facebook.com/blog/hubert-self-supervised-representation-learning-for-speech-recognition-generation-and-compression)
The extra large model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note that this model should be fine-tuned on a downstream task, like Automatic Speech Recognition, Speaker Identification, Intent Classification, Emotion Recognition, etc...
The model was pretrained on [Libri-Light](https://github.com/facebookresearch/libri-light).
[Paper](https://arxiv.org/abs/2106.07447)
Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed
**Abstract**
Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/hubert .
# Usage
See [this blog](https://huggingface.co/blog/fine-tune-wav2vec2-english) for more information on how to fine-tune the model. Note that the class `Wav2Vec2ForCTC` has to be replaced by `HubertForCTC`. |
|
facebook/hubert-xlarge-ls960-ft | 2021-06-16T11:23:31.000Z | [
"pytorch",
"hubert",
"en",
"dataset:libri-light",
"dataset:librispeech_asr",
"arxiv:2106.07447",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 175 | transformers | ---
language: en
datasets:
- libri-light
- librispeech_asr
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
# Hubert-Extra-Large-Finetuned
[Facebook's Hubert](https://ai.facebook.com/blog/hubert-self-supervised-representation-learning-for-speech-recognition-generation-and-compression)
The extra large model fine-tuned on 960h of Librispeech on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz.
The model is a fine-tuned version of [hubert-xlarge-ll60k](https://huggingface.co/facebook/hubert-xlarge-ll60k).
[Paper](https://arxiv.org/abs/2106.07447)
Authors: Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed
**Abstract**
Self-supervised approaches for speech representation learning are challenged by three unique problems: (1) there are multiple sound units in each input utterance, (2) there is no lexicon of input sound units during the pre-training phase, and (3) sound units have variable lengths with no explicit segmentation. To deal with these three problems, we propose the Hidden-Unit BERT (HuBERT) approach for self-supervised speech representation learning, which utilizes an offline clustering step to provide aligned target labels for a BERT-like prediction loss. A key ingredient of our approach is applying the prediction loss over the masked regions only, which forces the model to learn a combined acoustic and language model over the continuous inputs. HuBERT relies primarily on the consistency of the unsupervised clustering step rather than the intrinsic quality of the assigned cluster labels. Starting with a simple k-means teacher of 100 clusters, and using two iterations of clustering, the HuBERT model either matches or improves upon the state-of-the-art wav2vec 2.0 performance on the Librispeech (960h) and Libri-light (60,000h) benchmarks with 10min, 1h, 10h, 100h, and 960h fine-tuning subsets. Using a 1B parameter model, HuBERT shows up to 19% and 13% relative WER reduction on the more challenging dev-other and test-other evaluation subsets.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/hubert .
# Usage
The model can be used for automatic-speech-recognition as follows:
```python
import torch
from transformers import Wav2Vec2Processor, HubertForCTC
from datasets import load_dataset
import soundfile as sf
processor = Wav2Vec2Processor.from_pretrained("facebook/hubert-xlarge-ls960-ft")
model = HubertForCTC.from_pretrained("facebook/hubert-xlarge-ls960-ft")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
input_values = processor(ds["speech"][0], return_tensors="pt").input_values # Batch size 1
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = processor.decode(predicted_ids[0])
# ->"A MAN SAID TO THE UNIVERSE SIR I EXIST"
``` |
facebook/m2m100_1.2B | 2021-03-18T17:01:21.000Z | [
"pytorch",
"m2m_100",
"seq2seq",
"multilingual",
"arxiv:2010.11125",
"transformers",
"license:mit",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 7,474 | transformers | ---
language: multilingual
license: MIT
tags:
---
# M2M100 1.2B
M2M100 is a multilingual encoder-decoder (seq-to-seq) model trained for Many-to-Many multilingual translation.
It was introduced in this [paper](https://arxiv.org/abs/2010.11125) and first released in [this](https://github.com/pytorch/fairseq/tree/master/examples/m2m_100) repository.
The model that can directly translate between the 9,900 directions of 100 languages.
To translate into a target language, the target language id is forced as the first generated token.
To force the target language id as the first generated token, pass the `forced_bos_token_id` parameter to the `generate` method.
*Note: `M2M100Tokenizer` depends on `sentencepiece`, so make sure to install it before running the example.*
To install `sentencepiece` run `pip install sentencepiece`
```python
from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
hi_text = "जीवन एक चॉकलेट बॉक्स की तरह है।"
chinese_text = "生活就像一盒巧克力。"
model = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M")
tokenizer = M2M100Tokenizer.from_pretrained("facebook/m2m100_418M")
# translate Hindi to French
tokenizer.src_lang = "hi"
encoded_hi = tokenizer(hi_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_hi, forced_bos_token_id=tokenizer.get_lang_id("fr"))
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "La vie est comme une boîte de chocolat."
# translate Chinese to English
tokenizer.src_lang = "zh"
encoded_zh = tokenizer(chinese_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_zh, forced_bos_token_id=tokenizer.get_lang_id("en"))
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "Life is like a box of chocolate."
```
See the [model hub](https://huggingface.co/models?filter=m2m_100) to look for more fine-tuned versions.
## Languages covered
Afrikaans (af), Amharic (am), Arabic (ar), Asturian (ast), Azerbaijani (az), Bashkir (ba), Belarusian (be), Bulgarian (bg), Bengali (bn), Breton (br), Bosnian (bs), Catalan; Valencian (ca), Cebuano (ceb), Czech (cs), Welsh (cy), Danish (da), German (de), Greeek (el), English (en), Spanish (es), Estonian (et), Persian (fa), Fulah (ff), Finnish (fi), French (fr), Western Frisian (fy), Irish (ga), Gaelic; Scottish Gaelic (gd), Galician (gl), Gujarati (gu), Hausa (ha), Hebrew (he), Hindi (hi), Croatian (hr), Haitian; Haitian Creole (ht), Hungarian (hu), Armenian (hy), Indonesian (id), Igbo (ig), Iloko (ilo), Icelandic (is), Italian (it), Japanese (ja), Javanese (jv), Georgian (ka), Kazakh (kk), Central Khmer (km), Kannada (kn), Korean (ko), Luxembourgish; Letzeburgesch (lb), Ganda (lg), Lingala (ln), Lao (lo), Lithuanian (lt), Latvian (lv), Malagasy (mg), Macedonian (mk), Malayalam (ml), Mongolian (mn), Marathi (mr), Malay (ms), Burmese (my), Nepali (ne), Dutch; Flemish (nl), Norwegian (no), Northern Sotho (ns), Occitan (post 1500) (oc), Oriya (or), Panjabi; Punjabi (pa), Polish (pl), Pushto; Pashto (ps), Portuguese (pt), Romanian; Moldavian; Moldovan (ro), Russian (ru), Sindhi (sd), Sinhala; Sinhalese (si), Slovak (sk), Slovenian (sl), Somali (so), Albanian (sq), Serbian (sr), Swati (ss), Sundanese (su), Swedish (sv), Swahili (sw), Tamil (ta), Thai (th), Tagalog (tl), Tswana (tn), Turkish (tr), Ukrainian (uk), Urdu (ur), Uzbek (uz), Vietnamese (vi), Wolof (wo), Xhosa (xh), Yiddish (yi), Yoruba (yo), Chinese (zh), Zulu (zu)
## BibTeX entry and citation info
```
@misc{fan2020englishcentric,
title={Beyond English-Centric Multilingual Machine Translation},
author={Angela Fan and Shruti Bhosale and Holger Schwenk and Zhiyi Ma and Ahmed El-Kishky and Siddharth Goyal and Mandeep Baines and Onur Celebi and Guillaume Wenzek and Vishrav Chaudhary and Naman Goyal and Tom Birch and Vitaliy Liptchinsky and Sergey Edunov and Edouard Grave and Michael Auli and Armand Joulin},
year={2020},
eprint={2010.11125},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
``` |
facebook/m2m100_418M | 2021-06-12T09:29:34.000Z | [
"pytorch",
"rust",
"m2m_100",
"seq2seq",
"multilingual",
"arxiv:2010.11125",
"transformers",
"license:mit",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"rust_model.ot",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 23,333 | transformers | ---
language: multilingual
license: MIT
tags:
---
# M2M100 418M
M2M100 is a multilingual encoder-decoder (seq-to-seq) model trained for Many-to-Many multilingual translation.
It was introduced in this [paper](https://arxiv.org/abs/2010.11125) and first released in [this](https://github.com/pytorch/fairseq/tree/master/examples/m2m_100) repository.
The model that can directly translate between the 9,900 directions of 100 languages.
To translate into a target language, the target language id is forced as the first generated token.
To force the target language id as the first generated token, pass the `forced_bos_token_id` parameter to the `generate` method.
*Note: `M2M100Tokenizer` depends on `sentencepiece`, so make sure to install it before running the example.*
To install `sentencepiece` run `pip install sentencepiece`
```python
from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
hi_text = "जीवन एक चॉकलेट बॉक्स की तरह है।"
chinese_text = "生活就像一盒巧克力。"
model = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M")
tokenizer = M2M100Tokenizer.from_pretrained("facebook/m2m100_418M")
# translate Hindi to French
tokenizer.src_lang = "hi"
encoded_hi = tokenizer(hi_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_hi, forced_bos_token_id=tokenizer.get_lang_id("fr"))
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "La vie est comme une boîte de chocolat."
# translate Chinese to English
tokenizer.src_lang = "zh"
encoded_zh = tokenizer(chinese_text, return_tensors="pt")
generated_tokens = model.generate(**encoded_zh, forced_bos_token_id=tokenizer.get_lang_id("en"))
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "Life is like a box of chocolate."
```
See the [model hub](https://huggingface.co/models?filter=m2m_100) to look for more fine-tuned versions.
## Languages covered
Afrikaans (af), Amharic (am), Arabic (ar), Asturian (ast), Azerbaijani (az), Bashkir (ba), Belarusian (be), Bulgarian (bg), Bengali (bn), Breton (br), Bosnian (bs), Catalan; Valencian (ca), Cebuano (ceb), Czech (cs), Welsh (cy), Danish (da), German (de), Greeek (el), English (en), Spanish (es), Estonian (et), Persian (fa), Fulah (ff), Finnish (fi), French (fr), Western Frisian (fy), Irish (ga), Gaelic; Scottish Gaelic (gd), Galician (gl), Gujarati (gu), Hausa (ha), Hebrew (he), Hindi (hi), Croatian (hr), Haitian; Haitian Creole (ht), Hungarian (hu), Armenian (hy), Indonesian (id), Igbo (ig), Iloko (ilo), Icelandic (is), Italian (it), Japanese (ja), Javanese (jv), Georgian (ka), Kazakh (kk), Central Khmer (km), Kannada (kn), Korean (ko), Luxembourgish; Letzeburgesch (lb), Ganda (lg), Lingala (ln), Lao (lo), Lithuanian (lt), Latvian (lv), Malagasy (mg), Macedonian (mk), Malayalam (ml), Mongolian (mn), Marathi (mr), Malay (ms), Burmese (my), Nepali (ne), Dutch; Flemish (nl), Norwegian (no), Northern Sotho (ns), Occitan (post 1500) (oc), Oriya (or), Panjabi; Punjabi (pa), Polish (pl), Pushto; Pashto (ps), Portuguese (pt), Romanian; Moldavian; Moldovan (ro), Russian (ru), Sindhi (sd), Sinhala; Sinhalese (si), Slovak (sk), Slovenian (sl), Somali (so), Albanian (sq), Serbian (sr), Swati (ss), Sundanese (su), Swedish (sv), Swahili (sw), Tamil (ta), Thai (th), Tagalog (tl), Tswana (tn), Turkish (tr), Ukrainian (uk), Urdu (ur), Uzbek (uz), Vietnamese (vi), Wolof (wo), Xhosa (xh), Yiddish (yi), Yoruba (yo), Chinese (zh), Zulu (zu)
## BibTeX entry and citation info
```
@misc{fan2020englishcentric,
title={Beyond English-Centric Multilingual Machine Translation},
author={Angela Fan and Shruti Bhosale and Holger Schwenk and Zhiyi Ma and Ahmed El-Kishky and Siddharth Goyal and Mandeep Baines and Onur Celebi and Guillaume Wenzek and Vishrav Chaudhary and Naman Goyal and Tom Birch and Vitaliy Liptchinsky and Sergey Edunov and Edouard Grave and Michael Auli and Armand Joulin},
year={2020},
eprint={2010.11125},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
``` |
facebook/mbart-large-50-many-to-many-mmt | 2021-06-05T11:18:58.000Z | [
"pytorch",
"rust",
"mbart",
"seq2seq",
"multilingual",
"arxiv:2008.00401",
"transformers",
"mbart-50",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"rust_model.ot",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json"
]
| facebook | 7,863 | transformers | ---
language: multilingual
tags:
- mbart-50
---
# mBART-50 many to many multilingual machine translation
This model is a fine-tuned checkpoint of [mBART-large-50](https://huggingface.co/facebook/mbart-large-50). `mbart-large-50-many-to-many-mmt` is fine-tuned for multilingual machine translation. It was introduced in [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) paper.
The model can translate directly between any pair of 50 languages. To translate into a target language, the target language id is forced as the first generated token. To force the target language id as the first generated token, pass the `forced_bos_token_id` parameter to the `generate` method.
```python
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
article_hi = "संयुक्त राष्ट्र के प्रमुख का कहना है कि सीरिया में कोई सैन्य समाधान नहीं है"
article_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."
model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
# translate Hindi to French
tokenizer.src_lang = "hi_IN"
encoded_hi = tokenizer(article_hi, return_tensors="pt")
generated_tokens = model.generate(
**encoded_hi,
forced_bos_token_id=tokenizer.lang_code_to_id["fr_XX"]
)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "Le chef de l 'ONU affirme qu 'il n 'y a pas de solution militaire dans la Syrie."
# translate Arabic to English
tokenizer.src_lang = "ar_AR"
encoded_ar = tokenizer(article_ar, return_tensors="pt")
generated_tokens = model.generate(
**encoded_ar,
forced_bos_token_id=tokenizer.lang_code_to_id["en_XX"]
)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "The Secretary-General of the United Nations says there is no military solution in Syria."
```
See the [model hub](https://huggingface.co/models?filter=mbart-50) to look for more fine-tuned versions.
## Languages covered
Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)
## BibTeX entry and citation info
```
@article{tang2020multilingual,
title={Multilingual Translation with Extensible Multilingual Pretraining and Finetuning},
author={Yuqing Tang and Chau Tran and Xian Li and Peng-Jen Chen and Naman Goyal and Vishrav Chaudhary and Jiatao Gu and Angela Fan},
year={2020},
eprint={2008.00401},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
``` |
facebook/mbart-large-50-many-to-one-mmt | 2021-04-07T16:25:12.000Z | [
"pytorch",
"mbart",
"seq2seq",
"multilingual",
"arxiv:2008.00401",
"transformers",
"mbart-50",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json"
]
| facebook | 1,815 | transformers | ---
language: multilingual
tags:
- mbart-50
---
# mBART-50 many to one multilingual machine translation
This model is a fine-tuned checkpoint of [mBART-large-50](https://huggingface.co/facebook/mbart-large-50). `mbart-large-50-many-to-many-mmt` is fine-tuned for multilingual machine translation. It was introduced in [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) paper.
The model can translate directly between any pair of 50 languages.
```python
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
article_hi = "संयुक्त राष्ट्र के प्रमुख का कहना है कि सीरिया में कोई सैन्य समाधान नहीं है"
article_ar = "الأمين العام للأمم المتحدة يقول إنه لا يوجد حل عسكري في سوريا."
model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-one-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-one-mmt")
# translate Hindi to English
tokenizer.src_lang = "hi_IN"
encoded_hi = tokenizer(article_hi, return_tensors="pt")
generated_tokens = model.generate(**encoded_hi)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "The head of the UN says there is no military solution in Syria."
# translate Arabic to English
tokenizer.src_lang = "ar_AR"
encoded_ar = tokenizer(article_ar, return_tensors="pt")
generated_tokens = model.generate(**encoded_ar)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => "The Secretary-General of the United Nations says there is no military solution in Syria."
```
See the [model hub](https://huggingface.co/models?filter=mbart-50) to look for more fine-tuned versions.
## Languages covered
Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)
## BibTeX entry and citation info
```
@article{tang2020multilingual,
title={Multilingual Translation with Extensible Multilingual Pretraining and Finetuning},
author={Yuqing Tang and Chau Tran and Xian Li and Peng-Jen Chen and Naman Goyal and Vishrav Chaudhary and Jiatao Gu and Angela Fan},
year={2020},
eprint={2008.00401},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
``` |
facebook/mbart-large-50-one-to-many-mmt | 2021-03-12T17:21:29.000Z | [
"pytorch",
"mbart",
"seq2seq",
"multilingual",
"arxiv:2008.00401",
"transformers",
"mbart-50",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json"
]
| facebook | 2,352 | transformers | ---
language: multilingual
tags:
- mbart-50
---
# mBART-50 one to many multilingual machine translation
This model is a fine-tuned checkpoint of [mBART-large-50](https://huggingface.co/facebook/mbart-large-50). `mbart-large-50-one-to-many-mmt` is fine-tuned for multilingual machine translation. It was introduced in [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) paper.
The model can translate English to other 49 languages mentioned below.
To translate into a target language, the target language id is forced as the first generated token. To force the
target language id as the first generated token, pass the `forced_bos_token_id` parameter to the `generate` method.
```python
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
article_en = "The head of the United Nations says there is no military solution in Syria"
model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-one-to-many-mmt")
tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-one-to-many-mmt", src_lang="en_XX")
model_inputs = tokenizer(article_en, return_tensors="pt")
# translate from English to Hindi
generated_tokens = model.generate(
**model_inputs,
forced_bos_token_id=tokenizer.lang_code_to_id["hi_IN"]
)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => 'संयुक्त राष्ट्र के नेता कहते हैं कि सीरिया में कोई सैन्य समाधान नहीं है'
# translate from English to Chinese
generated_tokens = model.generate(
**model_inputs,
forced_bos_token_id=tokenizer.lang_code_to_id["zh_CN"]
)
tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
# => '联合国首脑说,叙利亚没有军事解决办法'
```
See the [model hub](https://huggingface.co/models?filter=mbart-50) to look for more fine-tuned versions.
## Languages covered
Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)
## BibTeX entry and citation info
```
@article{tang2020multilingual,
title={Multilingual Translation with Extensible Multilingual Pretraining and Finetuning},
author={Yuqing Tang and Chau Tran and Xian Li and Peng-Jen Chen and Naman Goyal and Vishrav Chaudhary and Jiatao Gu and Angela Fan},
year={2020},
eprint={2008.00401},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
``` |
facebook/mbart-large-50 | 2021-03-12T17:21:46.000Z | [
"pytorch",
"mbart",
"seq2seq",
"multilingual",
"arxiv:2008.00401",
"transformers",
"mbart-50",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json"
]
| facebook | 4,298 | transformers | ---
language: multilingual
tags:
- mbart-50
---
# mBART-50
mBART-50 is a multilingual Sequence-to-Sequence model pre-trained using the "Multilingual Denoising Pretraining" objective. It was introduced in [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) paper.
## Model description
mBART-50 is a multilingual Sequence-to-Sequence model. It was introduced to show that multilingual translation models can be created through multilingual fine-tuning.
Instead of fine-tuning on one direction, a pre-trained model is fine-tuned on many directions simultaneously. mBART-50 is created using the original mBART model and extended to add extra 25 languages to support multilingual machine translation models of 50 languages. The pre-training objective is explained below.
**Multilingual Denoising Pretraining**: The model incorporates N languages by concatenating data:
`D = {D1, ..., DN }` where each Di is a collection of monolingual documents in language `i`. The source documents are noised using two schemes,
first randomly shuffling the original sentences' order, and second a novel in-filling scheme,
where spans of text are replaced with a single mask token. The model is then tasked to reconstruct the original text.
35% of each instance's words are masked by random sampling a span length according to a Poisson distribution `(λ = 3.5)`.
The decoder input is the original text with one position offset. A language id symbol `LID` is used as the initial token to predict the sentence.
## Intended uses & limitations
`mbart-large-50` is pre-trained model and primarily aimed at being fine-tuned on translation tasks. It can also be fine-tuned on other multilingual sequence-to-sequence tasks.
See the [model hub](https://huggingface.co/models?filter=mbart-50) to look for fine-tuned versions.
## Training
As the model is multilingual, it expects the sequences in a different format. A special language id token is used as a prefix in both the source and target text. The text format is `[lang_code] X [eos]` with `X` being the source or target text respectively and `lang_code` is `source_lang_code` for source text and `tgt_lang_code` for target text. `bos` is never used. Once the examples are prepared in this format, it can be trained as any other sequence-to-sequence model.
```python
from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50")
tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50", src_lang="en_XX", tgt_lang="ro_RO")
src_text = " UN Chief Says There Is No Military Solution in Syria"
tgt_text = "Şeful ONU declară că nu există o soluţie militară în Siria"
model_inputs = tokenizer(src_text, return_tensors="pt")
with tokenizer.as_target_tokenizer():
labels = tokenizer(tgt_text, return_tensors="pt").input_ids
model(**model_inputs, labels=labels) # forward pass
```
## Languages covered
Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)
## BibTeX entry and citation info
```
@article{tang2020multilingual,
title={Multilingual Translation with Extensible Multilingual Pretraining and Finetuning},
author={Yuqing Tang and Chau Tran and Xian Li and Peng-Jen Chen and Naman Goyal and Vishrav Chaudhary and Jiatao Gu and Angela Fan},
year={2020},
eprint={2008.00401},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
``` |
facebook/mbart-large-cc25 | 2021-03-10T03:48:19.000Z | [
"pytorch",
"mbart",
"seq2seq",
"en",
"ar",
"cs",
"de",
"et",
"fi",
"fr",
"gu",
"hi",
"it",
"ja",
"kk",
"ko",
"lt",
"lv",
"my",
"ne",
"nl",
"ro",
"ru",
"si",
"tr",
"vi",
"zh",
"multilingual",
"transformers",
"translation",
"text2text-generation"
]
| translation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentence.bpe.model",
"sentencepiece.bpe.model",
"tokenizer.json"
]
| facebook | 21,830 | transformers | ---
tags:
- translation
language:
- en
- ar
- cs
- de
- et
- fi
- fr
- gu
- hi
- it
- ja
- kk
- ko
- lt
- lv
- my
- ne
- nl
- ro
- ru
- si
- tr
- vi
- zh
- multilingual
---
#### mbart-large-cc25
Pretrained (not finetuned) multilingual mbart model.
Original Languages
```
export langs=ar_AR,cs_CZ,de_DE,en_XX,es_XX,et_EE,fi_FI,fr_XX,gu_IN,hi_IN,it_IT,ja_XX,kk_KZ,ko_KR,lt_LT,lv_LV,my_MM,ne_NP,nl_XX,ro_RO,ru_RU,si_LK,tr_TR,vi_VN,zh_CN
```
Original Code: https://github.com/pytorch/fairseq/tree/master/examples/mbart
Docs: https://huggingface.co/transformers/master/model_doc/mbart.html
Finetuning Code: examples/seq2seq/finetune.py (as of Aug 20, 2020)
Can also be finetuned for summarization. |
facebook/mbart-large-en-ro | 2021-03-10T03:46:29.000Z | [
"pytorch",
"tf",
"mbart",
"en",
"ro",
"transformers",
"translation",
"license:mit"
]
| translation | [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"sentence.bpe.model",
"sentencepiece.bpe.model",
"tf_model.h5",
"tokenizer.json"
]
| facebook | 6,592 | transformers | ---
tags:
- translation
language:
- en
- ro
license: mit
---
### mbart-large-en-ro
This is mbart-large-cc25, finetuned on wmt_en_ro.
It scores BLEU 28.1 without post processing and BLEU 38 with postprocessing. Instructions in `romanian_postprocessing.md`
Original Code: https://github.com/pytorch/fairseq/tree/master/examples/mbart
Docs: https://huggingface.co/transformers/master/model_doc/mbart.html
Finetuning Code: examples/seq2seq/finetune.py (as of Aug 20, 2020)
|
facebook/rag-sequence-base | 2020-12-11T21:39:37.000Z | [
"pytorch",
"rag",
"arxiv:2005.11401",
"transformers",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"generator_tokenizer/merges.txt",
"generator_tokenizer/special_tokens_map.json",
"generator_tokenizer/tokenizer_config.json",
"generator_tokenizer/vocab.json",
"question_encoder_tokenizer/special_tokens_map.json",
"question_encoder_tokenizer/tokenizer_config.json",
"question_encoder_tokenizer/vocab.txt"
]
| facebook | 554 | transformers | ---
license: apache-2.0
thumbnail: https://huggingface.co/front/thumbnails/facebook.png
---
## RAG
This is a non-finetuned version of the RAG-Sequence model of the the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/pdf/2005.11401.pdf)
by Patrick Lewis, Ethan Perez, Aleksandara Piktus et al.
Rag consits of a *question encoder*, *retriever* and a *generator*. The retriever should be a `RagRetriever` instance. The *question encoder* can be any model that can be loaded with `AutoModel` and the *generator* can be any model that can be loaded with `AutoModelForSeq2SeqLM`.
This model is a non-finetuned RAG-Sequence model and was created as follows:
```python
from transformers import RagTokenizer, RagRetriever, RagSequenceForGeneration, AutoTokenizer
model = RagSequenceForGeneration.from_pretrained_question_encoder_generator("facebook/dpr-question_encoder-single-nq-base", "facebook/bart-large")
question_encoder_tokenizer = AutoTokenizer.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
generator_tokenizer = AutoTokenizer.from_pretrained("facebook/bart-large")
tokenizer = RagTokenizer(question_encoder_tokenizer, generator_tokenizer)
model.config.use_dummy_dataset = True
model.config.index_name = "exact"
retriever = RagRetriever(model.config, question_encoder_tokenizer, generator_tokenizer)
model.save_pretrained("./")
tokenizer.save_pretrained("./")
retriever.save_pretrained("./")
```
Note that the model is *uncased* so that all capital input letters are converted to lower-case.
## Usage:
*Note*: the model uses the *dummy* retriever as a default. Better results are obtained by using the full retriever,
by setting `config.index_name="legacy"` and `config.use_dummy_dataset=False`.
The model can be fine-tuned as follows:
```python
from transformers import RagTokenizer, RagRetriever, RagTokenForGeneration
tokenizer = RagTokenizer.from_pretrained("facebook/rag-sequence-base")
retriever = RagRetriever.from_pretrained("facebook/rag-sequence-base")
model = RagTokenForGeneration.from_pretrained("facebook/rag-sequence-base", retriever=retriever)
input_dict = tokenizer.prepare_seq2seq_batch("who holds the record in 100m freestyle", "michael phelps", return_tensors="pt")
outputs = model(input_dict["input_ids"], labels=input_dict["labels"])
loss = outputs.loss
# train on loss
```
|
|
facebook/rag-sequence-nq | 2021-03-12T11:04:28.000Z | [
"pytorch",
"tf",
"rag",
"en",
"dataset:wiki_dpr",
"arxiv:2005.11401",
"transformers",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"generator_tokenizer/merges.txt",
"generator_tokenizer/special_tokens_map.json",
"generator_tokenizer/tokenizer_config.json",
"generator_tokenizer/vocab.json",
"question_encoder_tokenizer/special_tokens_map.json",
"question_encoder_tokenizer/tokenizer_config.json",
"question_encoder_tokenizer/vocab.txt"
]
| facebook | 2,104 | transformers | ---
language: en
license: apache-2.0
datasets:
- wiki_dpr
thumbnail: https://huggingface.co/front/thumbnails/facebook.png
---
## RAG
This is the RAG-Sequence Model of the the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/pdf/2005.11401.pdf)
by Patrick Lewis, Ethan Perez, Aleksandara Piktus et al.
The model is a *uncased* model, which means that capital letters are simply converted to lower-case letters.
The model consits of a *question_encoder*, *retriever* and a *generator*. The retriever extracts relevant passages from the *wiki_dpr* `train` datasets, which is linked above.
The question_encoder and retriever are based on `facebook/dpr-question_encoder-single-nq-base` and `facebook/bart-large`, which were jointly finetuned on
on the *wiki_dpr* QA dataset in an end-to-end fashion.
## Usage:
**Note**: In the usage example below only the *dummy* retriever of *wiki_dpr* is used because the complete *lecagy* index requires over 75 GB of RAM.
The model can generate answers to any factoid question as follows:
```python
from transformers import RagTokenizer, RagRetriever, RagSequenceForGeneration
tokenizer = RagTokenizer.from_pretrained("facebook/rag-sequence-nq")
retriever = RagRetriever.from_pretrained("facebook/rag-sequence-nq", index_name="exact", use_dummy_dataset=True)
model = RagSequenceForGeneration.from_pretrained("facebook/rag-sequence-nq", retriever=retriever)
input_dict = tokenizer.prepare_seq2seq_batch("how many countries are in europe", return_tensors="pt")
generated = model.generate(input_ids=input_dict["input_ids"])
print(tokenizer.batch_decode(generated, skip_special_tokens=True)[0])
# should give 54 => google says either 44 or 51
```
|
|
facebook/rag-token-base | 2020-12-11T21:39:44.000Z | [
"pytorch",
"rag",
"en",
"dataset:wiki_dpr",
"arxiv:2005.11401",
"transformers",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"generator_tokenizer/merges.txt",
"generator_tokenizer/special_tokens_map.json",
"generator_tokenizer/tokenizer_config.json",
"generator_tokenizer/vocab.json",
"question_encoder_tokenizer/special_tokens_map.json",
"question_encoder_tokenizer/tokenizer_config.json",
"question_encoder_tokenizer/vocab.txt"
]
| facebook | 2,022 | transformers | ---
language: en
license: apache-2.0
datasets:
- wiki_dpr
thumbnail: https://huggingface.co/front/thumbnails/facebook.png
---
## RAG
This is a non-finetuned version of the RAG-Token model of the the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/pdf/2005.11401.pdf)
by Patrick Lewis, Ethan Perez, Aleksandara Piktus et al.
Rag consits of a *question encoder*, *retriever* and a *generator*. The retriever should be a `RagRetriever` instance. The *question encoder* can be any model that can be loaded with `AutoModel` and the *generator* can be any model that can be loaded with `AutoModelForSeq2SeqLM`.
This model is a non-finetuned RAG-Token model and was created as follows:
```python
from transformers import RagTokenizer, RagRetriever, RagTokenForGeneration, AutoTokenizer
model = RagTokenForGeneration.from_pretrained_question_encoder_generator("facebook/dpr-question_encoder-single-nq-base", "facebook/bart-large")
question_encoder_tokenizer = AutoTokenizer.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
generator_tokenizer = AutoTokenizer.from_pretrained("facebook/bart-large")
tokenizer = RagTokenizer(question_encoder_tokenizer, generator_tokenizer)
model.config.use_dummy_dataset = True
model.config.index_name = "exact"
retriever = RagRetriever(model.config, question_encoder_tokenizer, generator_tokenizer)
model.save_pretrained("./")
tokenizer.save_pretrained("./")
retriever.save_pretrained("./")
```
Note that the model is *uncased* so that all capital input letters are converted to lower-case.
## Usage:
*Note*: the model uses the *dummy* retriever as a default. Better results are obtained by using the full retriever,
by setting `config.index_name="legacy"` and `config.use_dummy_dataset=False`.
The model can be fine-tuned as follows:
```python
from transformers import RagTokenizer, RagRetriever, RagTokenForGeneration
tokenizer = RagTokenizer.from_pretrained("facebook/rag-token-base")
retriever = RagRetriever.from_pretrained("facebook/rag-token-base")
model = RagTokenForGeneration.from_pretrained("facebook/rag-token-base", retriever=retriever)
input_dict = tokenizer.prepare_seq2seq_batch("who holds the record in 100m freestyle", "michael phelps", return_tensors="pt")
outputs = model(input_dict["input_ids"], labels=input_dict["labels"])
loss = outputs.loss
# train on loss
```
|
|
facebook/rag-token-nq | 2021-03-12T10:55:22.000Z | [
"pytorch",
"tf",
"rag",
"en",
"dataset:wiki_dpr",
"arxiv:2005.11401",
"transformers",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"pytorch_model.bin",
"tf_model.h5",
"generator_tokenizer/merges.txt",
"generator_tokenizer/special_tokens_map.json",
"generator_tokenizer/tokenizer_config.json",
"generator_tokenizer/vocab.json",
"question_encoder_tokenizer/special_tokens_map.json",
"question_encoder_tokenizer/tokenizer_config.json",
"question_encoder_tokenizer/vocab.txt"
]
| facebook | 27,796 | transformers | ---
language: en
license: apache-2.0
datasets:
- wiki_dpr
thumbnail: https://huggingface.co/front/thumbnails/facebook.png
---
## RAG
This is the RAG-Token Model of the the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/pdf/2005.11401.pdf)
by Patrick Lewis, Ethan Perez, Aleksandara Piktus et al.
The model is a *uncased* model, which means that capital letters are simply converted to lower-case letters.
The model consits of a *question_encoder*, *retriever* and a *generator*. The retriever extracts relevant passages from the *wiki_dpr* `train` datasets, which is linked above.
The question_encoder and retriever are based on `facebook/dpr-question_encoder-single-nq-base` and `facebook/bart-large`, which were jointly finetuned on
on the *wiki_dpr* QA dataset in an end-to-end fashion.
## Usage:
**Note**: In the usage example below only the *dummy* retriever of *wiki_dpr* is used because the complete *lecagy* index requires over 75 GB of RAM.
The model can generate answers to any factoid question as follows:
```python
from transformers import RagTokenizer, RagRetriever, RagTokenForGeneration
tokenizer = RagTokenizer.from_pretrained("facebook/rag-token-nq")
retriever = RagRetriever.from_pretrained("facebook/rag-token-nq", index_name="exact", use_dummy_dataset=True)
model = RagTokenForGeneration.from_pretrained("facebook/rag-token-nq", retriever=retriever)
input_dict = tokenizer.prepare_seq2seq_batch("who holds the record in 100m freestyle", return_tensors="pt")
generated = model.generate(input_ids=input_dict["input_ids"])
print(tokenizer.batch_decode(generated, skip_special_tokens=True)[0])
# should give michael phelps => sounds reasonable
```
|
|
facebook/s2t-large-librispeech-asr | 2021-03-16T16:59:59.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"dataset:librispeech_asr",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"automatic-speech-recognition",
"license:mit",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 189 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- audio
- automatic-speech-recognition
license: MIT
---
# S2T-LARGE-LIBRISPEECH-ASR
`s2t-large-librispeech-asr` is a Speech to Text Transformer (S2T) model trained for automatic speech recognition (ASR).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is an end-to-end sequence-to-sequence transformer model. It is trained with standard
autoregressive cross-entropy loss and generates the transcripts autoregressively.
## Intended uses & limitations
This model can be used for end-to-end speech recognition (ASR).
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-large-librispeech-asr")
processor = Speech2Textprocessor.from_pretrained("facebook/s2t-large-librispeech-asr")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
input_features = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
).input_features # Batch size 1
generated_ids = model.generate(input_ids=input_features)
transcription = processor.batch_decode(generated_ids)
```
#### Evaluation on LibriSpeech Test
The following script shows how to evaluate this model on the [LibriSpeech](https://huggingface.co/datasets/librispeech_asr)
*"clean"* and *"other"* test dataset.
```python
from datasets import load_dataset, load_metric
from transformers import Speech2TextForConditionalGeneration, Speech2TextProcessor
import soundfile as sf
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test") # change to "other" for other test dataset
wer = load_metric("wer")
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-large-librispeech-asr").to("cuda")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-large-librispeech-asr", do_upper_case=True)
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=16000, padding=True, return_tensors="pt")
input_features = features.input_features.to("cuda")
attention_mask = features.attention_mask.to("cuda")
gen_tokens = model.generate(input_ids=input_features, attention_mask=attention_mask)
batch["transcription"] = processor.batch_decode(gen_tokens, skip_special_tokens=True)
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=8, remove_columns=["speech"])
print("WER:", wer(predictions=result["transcription"], references=result["text"]))
```
*Result (WER)*:
| "clean" | "other" |
|:-------:|:-------:|
| 3.3 | 7.5 |
## Training data
The S2T-LARGE-LIBRISPEECH-ASR is trained on [LibriSpeech ASR Corpus](https://www.openslr.org/12), a dataset consisting of
approximately 1000 hours of 16kHz read English speech.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 10,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively.
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
``` |
facebook/s2t-medium-librispeech-asr | 2021-03-17T05:37:24.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"dataset:librispeech_asr",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 460 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- audio
- automatic-speech-recognition
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
license: MIT
---
# S2T-MEDIUM-LIBRISPEECH-ASR
`s2t-medium-librispeech-asr` is a Speech to Text Transformer (S2T) model trained for automatic speech recognition (ASR).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is an end-to-end sequence-to-sequence transformer model. It is trained with standard
autoregressive cross-entropy loss and generates the transcripts autoregressively.
## Intended uses & limitations
This model can be used for end-to-end speech recognition (ASR).
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-medium-librispeech-asr")
processor = Speech2Textprocessor.from_pretrained("facebook/s2t-medium-librispeech-asr")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
input_features = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
).input_features # Batch size 1
generated_ids = model.generate(input_ids=input_features)
transcription = processor.batch_decode(generated_ids)
```
#### Evaluation on LibriSpeech Test
The following script shows how to evaluate this model on the [LibriSpeech](https://huggingface.co/datasets/librispeech_asr)
*"clean"* and *"other"* test dataset.
```python
from datasets import load_dataset, load_metric
from transformers import Speech2TextForConditionalGeneration, Speech2TextProcessor
import soundfile as sf
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test") # change to "other" for other test dataset
wer = load_metric("wer")
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-medium-librispeech-asr").to("cuda")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-medium-librispeech-asr", do_upper_case=True)
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=16000, padding=True, return_tensors="pt")
input_features = features.input_features.to("cuda")
attention_mask = features.attention_mask.to("cuda")
gen_tokens = model.generate(input_ids=input_features, attention_mask=attention_mask)
batch["transcription"] = processor.batch_decode(gen_tokens, skip_special_tokens=True)
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=8, remove_columns=["speech"])
print("WER:", wer(predictions=result["transcription"], references=result["text"]))
```
*Result (WER)*:
| "clean" | "other" |
|:-------:|:-------:|
| 3.5 | 7.8 |
## Training data
The S2T-MEDIUM-LIBRISPEECH-ASR is trained on [LibriSpeech ASR Corpus](https://www.openslr.org/12), a dataset consisting of
approximately 1000 hours of 16kHz read English speech.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 10,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively.
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
``` |
facebook/s2t-medium-mustc-multilingual-st | 2021-03-16T16:58:19.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"de",
"nl",
"es",
"fr",
"it",
"pt",
"ro",
"ru",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"text2text-generation"
]
| text2text-generation | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 286 | transformers | ---
language:
- en
- de
- nl
- es
- fr
- it
- pt
- ro
- ru
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
---
# S2T-MEDIUM-MUSTC-MULTILINGUAL-ST
`s2t-medium-mustc-multilingual-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Multilingual Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to French text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
For multilingual speech translation models, `eos_token_id` is used as the `decoder_start_token_id` and
the target language id is forced as the first generated token. To force the target language id as the first
generated token, pass the `forced_bos_token_id` parameter to the `generate()` method. The following
example shows how to transate English speech to French and German text using the `facebook/s2t-medium-mustc-multilingual-st`
checkpoint.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-medium-mustc-multilingual-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-medium-mustc-multilingual-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
inputs = processor(ds["speech"][0], sampling_rate=16_000, return_tensors="pt")
# translate English Speech To French Text
generated_ids = model.generate(
input_ids=inputs["input_features"],
attention_mask=inputs["attention_mask"],
forced_bos_token_id=processor.tokenizer.lang_code_to_id["fr"]
)
translation_fr = processor.batch_decode(generated_ids)
# translate English Speech To German Text
generated_ids = model.generate(
input_ids=inputs["input_features"],
attention_mask=inputs["attention_mask"],
forced_bos_token_id=processor.tokenizer.lang_code_to_id["de"]
)
translation_de = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-medium-mustc-multilingual-st is trained on [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 10,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for multilingual ASR. For multilingual models, target language ID token
is used as target BOS.
## Evaluation results
MuST-C test results (BLEU score):
| En-De | En-Nl | En-Es | En-Fr | En-It | En-Pt | En-Ro | En-Ru |
|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|
| 24.5 | 28.6 | 28.2 | 34.9 | 24.6 | 31.1 | 23.8 | 16.0 |
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
``` |
facebook/s2t-small-covost2-ca-en-st | 2021-04-15T08:45:48.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"ca",
"en",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 23 | transformers | ---
language:
- ca
- en
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-CA-EN-ST
`s2t-small-covost2-ca-en-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end Catalan speech to English text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-ca-en-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-ca-en-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-ca-en-st is trained on Catalan-English subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for ca-en (BLEU score): 17.85
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-de-en-st | 2021-04-15T08:45:54.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"de",
"en",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 19 | transformers | ---
language:
- de
- en
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-DE-EN-ST
`s2t-small-covost2-de-en-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end German speech to English text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-de-en-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-de-en-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-de-en-st is trained on German-English subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for de-en (BLEU score): 17.58
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-en-ca-st | 2021-04-15T08:46:01.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"ca",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 27 | transformers | ---
language:
- en
- ca
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-EN-CA-ST
`s2t-small-covost2-en-ca-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Catlan text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-en-ca-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-en-ca-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-en-ca-st is trained on English-Catlan subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for en-ca (BLEU score): 21.68
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-en-de-st | 2021-04-15T08:46:07.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"de",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 7 | transformers | ---
language:
- en
- de
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-EN-DE-ST
`s2t-small-covost2-en-de-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to German text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-en-de-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-en-de-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-en-de-st is trained on English-German subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for en-de (BLEU score): 16.29
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-en-et-st | 2021-04-15T08:46:13.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"et",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 24 | transformers | ---
language:
- en
- et
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-EN-ET-ST
`s2t-small-covost2-en-et-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Estonian text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-en-et-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-en-et-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-en-et-st is trained on English-Estonian subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for en-et (BLEU score): 13.01
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-en-fa-st | 2021-04-15T08:46:19.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"fa",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 56 | transformers | ---
language:
- en
- fa
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-EN-FA-ST
`s2t-small-covost2-en-fa-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Farsi text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-en-fa-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-en-fa-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-en-fa-st is trained on English-Farsi subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for en-fa (BLEU score): 11.43
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-es-en-st | 2021-04-15T08:46:27.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"es",
"en",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 39 | transformers | ---
language:
- es
- en
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-ES-EN-ST
`s2t-small-covost2-es-en-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end Spanish speech to English text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-es-en-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-es-en-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-es-en-st is trained on Spanish-English subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for es-en (BLEU score): 22.31
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-covost2-fr-en-st | 2021-04-15T08:46:33.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"fr",
"en",
"dataset:covost2",
"arxiv:2010.05171",
"arxiv:1912.06670",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 40 | transformers | ---
language:
- fr
- en
datasets:
- covost2
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-COVOST2-FR-EN-ST
`s2t-small-covost2-fr-en-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end French speech to English text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-covost2-fr-en-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-covost2-fr-en-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=48_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-covost2-fr-en-st is trained on French-English subset of [CoVoST2](https://github.com/facebookresearch/covost).
CoVoST is a large-scale multilingual ST corpus based on [Common Voice](https://arxiv.org/abs/1912.06670), created to to foster
ST research with the largest ever open dataset
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using character based SentencePiece vocab.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
CoVOST2 test results for fr-en (BLEU score): 26.25
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-librispeech-asr | 2021-03-17T05:29:55.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"dataset:librispeech_asr",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 14,508 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
- audio
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-LIBRISPEECH-ASR
`s2t-small-librispeech-asr` is a Speech to Text Transformer (S2T) model trained for automatic speech recognition (ASR).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is an end-to-end sequence-to-sequence transformer model. It is trained with standard
autoregressive cross-entropy loss and generates the transcripts autoregressively.
## Intended uses & limitations
This model can be used for end-to-end speech recognition (ASR).
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
*Note: The feature extractor depends on [torchaudio](https://github.com/pytorch/audio) and the tokenizer depends on [sentencepiece](https://github.com/google/sentencepiece)
so be sure to install those packages before running the examples.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-librispeech-asr")
processor = Speech2Textprocessor.from_pretrained("facebook/s2t-small-librispeech-asr")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
input_features = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
).input_features # Batch size 1
generated_ids = model.generate(input_ids=input_features)
transcription = processor.batch_decode(generated_ids)
```
#### Evaluation on LibriSpeech Test
The following script shows how to evaluate this model on the [LibriSpeech](https://huggingface.co/datasets/librispeech_asr)
*"clean"* and *"other"* test dataset.
```python
from datasets import load_dataset, load_metric
from transformers import Speech2TextForConditionalGeneration, Speech2TextProcessor
import soundfile as sf
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test") # change to "other" for other test dataset
wer = load_metric("wer")
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-librispeech-asr").to("cuda")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-librispeech-asr", do_upper_case=True)
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=16000, padding=True, return_tensors="pt")
input_features = features.input_features.to("cuda")
attention_mask = features.attention_mask.to("cuda")
gen_tokens = model.generate(input_ids=input_features, attention_mask=attention_mask)
batch["transcription"] = processor.batch_decode(gen_tokens, skip_special_tokens=True)
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=8, remove_columns=["speech"])
print("WER:", wer(predictions=result["transcription"], references=result["text"]))
```
*Result (WER)*:
| "clean" | "other" |
|:-------:|:-------:|
| 4.3 | 9.0 |
## Training data
The S2T-SMALL-LIBRISPEECH-ASR is trained on [LibriSpeech ASR Corpus](https://www.openslr.org/12), a dataset consisting of
approximately 1000 hours of 16kHz read English speech.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 10,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively.
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
``` |
facebook/s2t-small-mustc-en-de-st | 2021-04-15T08:52:07.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"de",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 14 | transformers | ---
language:
- en
- de
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-DE-ST
`s2t-small-mustc-en-de-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to German text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-de-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-de-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-de-st is trained on English-German subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-de (BLEU score): 22.7
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-es-st | 2021-04-15T08:52:13.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"es",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 16 | transformers | ---
language:
- en
- es
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-ES-ST
`s2t-small-mustc-en-es-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Spanish text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-es-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-es-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-es-st is trained on English-Spanish subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-es (BLEU score): 27.2
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-fr-st | 2021-04-15T08:52:19.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"fr",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 91 | transformers | ---
language:
- en
- fr
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-FR-ST
`s2t-small-mustc-en-fr-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to French text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-fr-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-fr-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-fr-st is trained on English-French subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-fr (BLEU score): 32.9
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-it-st | 2021-04-15T08:52:24.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"it",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 236 | transformers | ---
language:
- en
- it
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-IT-ST
`s2t-small-mustc-en-it-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Italian text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-it-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-it-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-it-st is trained on English-Italian subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-it (BLEU score): 22.7
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-nl-st | 2021-04-15T08:52:30.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"nl",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 22 | transformers | ---
language:
- en
- nl
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-NL-ST
`s2t-small-mustc-en-nl-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Dutch text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-nl-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-nl-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-nl-st is trained on English-Dutch subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-nl (BLEU score): 27.3
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-pt-st | 2021-04-15T08:52:35.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"pt",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 23 | transformers | ---
language:
- en
- pt
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-PT-ST
`s2t-small-mustc-en-pt-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Portuguese text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-pt-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-pt-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-pt-st is trained on English-Portuguese subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-pt (BLEU score): 28.1
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-ro-st | 2021-04-15T08:52:41.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"ro",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 15 | transformers | ---
language:
- en
- ro
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-RO-ST
`s2t-small-mustc-en-ro-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Romanian text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-ro-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-ro-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-ro-st is trained on English-Romanian subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-ro (BLEU score): 21.9
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/s2t-small-mustc-en-ru-st | 2021-04-15T08:52:47.000Z | [
"pytorch",
"speech_to_text",
"seq2seq",
"en",
"ru",
"dataset:mustc",
"arxiv:2010.05171",
"arxiv:1904.08779",
"transformers",
"audio",
"speech-translation",
"automatic-speech-recognition",
"license:mit",
"pipeline_tag:automatic-speech-recognition",
"text2text-generation"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"sentencepiece.bpe.model",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 68 | transformers | ---
language:
- en
- ru
datasets:
- mustc
tags:
- audio
- speech-translation
- automatic-speech-recognition
license: MIT
pipeline_tag: automatic-speech-recognition
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# S2T-SMALL-MUSTC-EN-RU-ST
`s2t-small-mustc-en-ru-st` is a Speech to Text Transformer (S2T) model trained for end-to-end Speech Translation (ST).
The S2T model was proposed in [this paper](https://arxiv.org/abs/2010.05171) and released in
[this repository](https://github.com/pytorch/fairseq/tree/master/examples/speech_to_text)
## Model description
S2T is a transformer-based seq2seq (encoder-decoder) model designed for end-to-end Automatic Speech Recognition (ASR) and Speech
Translation (ST). It uses a convolutional downsampler to reduce the length of speech inputs by 3/4th before they are
fed into the encoder. The model is trained with standard autoregressive cross-entropy loss and generates the
transcripts/translations autoregressively.
## Intended uses & limitations
This model can be used for end-to-end English speech to Russian text translation.
See the [model hub](https://huggingface.co/models?filter=speech_to_text) to look for other S2T checkpoints.
### How to use
As this a standard sequence to sequence transformer model, you can use the `generate` method to generate the
transcripts by passing the speech features to the model.
*Note: The `Speech2TextProcessor` object uses [torchaudio](https://github.com/pytorch/audio) to extract the
filter bank features. Make sure to install the `torchaudio` package before running this example.*
You could either install those as extra speech dependancies with
`pip install transformers"[speech, sentencepiece]"` or install the packages seperatly
with `pip install torchaudio sentencepiece`.
```python
import torch
from transformers import Speech2TextProcessor, Speech2TextForConditionalGeneration
from datasets import load_dataset
import soundfile as sf
model = Speech2TextForConditionalGeneration.from_pretrained("facebook/s2t-small-mustc-en-ru-st")
processor = Speech2TextProcessor.from_pretrained("facebook/s2t-small-mustc-en-ru-st")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
ds = load_dataset(
"patrickvonplaten/librispeech_asr_dummy",
"clean",
split="validation"
)
ds = ds.map(map_to_array)
inputs = processor(
ds["speech"][0],
sampling_rate=16_000,
return_tensors="pt"
)
generated_ids = model.generate(input_ids=inputs["input_features"], attention_mask=inputs["attention_mask"])
translation = processor.batch_decode(generated_ids, skip_special_tokens=True)
```
## Training data
The s2t-small-mustc-en-ru-st is trained on English-Russian subset of [MuST-C](https://ict.fbk.eu/must-c/).
MuST-C is a multilingual speech translation corpus whose size and quality facilitates the training of end-to-end systems
for speech translation from English into several languages. For each target language, MuST-C comprises several hundred
hours of audio recordings from English TED Talks, which are automatically aligned at the sentence level with their manual
transcriptions and translations.
## Training procedure
### Preprocessing
The speech data is pre-processed by extracting Kaldi-compliant 80-channel log mel-filter bank features automatically from
WAV/FLAC audio files via PyKaldi or torchaudio. Further utterance-level CMVN (cepstral mean and variance normalization)
is applied to each example.
The texts are lowercased and tokenized using SentencePiece and a vocabulary size of 8,000.
### Training
The model is trained with standard autoregressive cross-entropy loss and using [SpecAugment](https://arxiv.org/abs/1904.08779).
The encoder receives speech features, and the decoder generates the transcripts autoregressively. To accelerate
model training and for better performance the encoder is pre-trained for English ASR.
## Evaluation results
MuST-C test results for en-ru (BLEU score): 15.3
### BibTeX entry and citation info
```bibtex
@inproceedings{wang2020fairseqs2t,
title = {fairseq S2T: Fast Speech-to-Text Modeling with fairseq},
author = {Changhan Wang and Yun Tang and Xutai Ma and Anne Wu and Dmytro Okhonko and Juan Pino},
booktitle = {Proceedings of the 2020 Conference of the Asian Chapter of the Association for Computational Linguistics (AACL): System Demonstrations},
year = {2020},
}
```
|
facebook/wav2vec2-base-100h | 2021-02-25T14:57:27.000Z | [
"pytorch",
"wav2vec2",
"en",
"dataset:librispeech_asr",
"arxiv:2006.11477",
"transformers",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 365 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- audio
- automatic-speech-recognition
license: apache-2.0
---
# Wav2Vec2-Base-100h
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The base model pretrained and fine-tuned on 100 hours of Librispeech on 16kHz sampled speech audio. When using the model
make sure that your speech input is also sampled at 16Khz.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
To transcribe audio files the model can be used as a standalone acoustic model as follows:
```python
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
from datasets import load_dataset
import soundfile as sf
import torch
# load model and tokenizer
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-100h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-100h")
# define function to read in sound file
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
# load dummy dataset and read soundfiles
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
# tokenize
input_values = tokenizer(ds["speech"][:2], return_tensors="pt", padding="longest").input_values # Batch size 1
# retrieve logits
logits = model(input_values).logits
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
```
## Evaluation
This code snippet shows how to evaluate **facebook/wav2vec2-base-100h** on LibriSpeech's "clean" and "other" test data.
```python
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import soundfile as sf
import torch
from jiwer import wer
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-100h").to("cuda")
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-100h")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
input_values = tokenizer(batch["speech"], return_tensors="pt", padding="longest").input_values
with torch.no_grad():
logits = model(input_values.to("cuda")).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=1, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))
```
*Result (WER)*:
| "clean" | "other" |
|---|---|
| 6.1 | 13.5 |
|
facebook/wav2vec2-base-100k-voxpopuli | 2021-06-09T19:10:29.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"multilingual",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 3,352 | transformers | ---
language: multilingual
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 100k unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base-10k-voxpopuli-ft-cs | 2021-05-05T16:24:24.000Z | [
"pytorch",
"wav2vec2",
"cs",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 97 | transformers | ---
language: cs
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in cs (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-cs")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-cs")
# load dataset
ds = load_dataset("common_voice", "cs", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-de | 2021-05-05T16:24:24.000Z | [
"pytorch",
"wav2vec2",
"de",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 56 | transformers | ---
language: de
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in de (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-de")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-de")
# load dataset
ds = load_dataset("common_voice", "de", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-en | 2021-05-05T16:24:25.000Z | [
"pytorch",
"wav2vec2",
"en",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 131 | transformers | ---
language: en
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in en (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-en")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-en")
# load dataset
ds = load_dataset("common_voice", "en", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-es | 2021-05-06T20:15:48.000Z | [
"pytorch",
"wav2vec2",
"es",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 39 | transformers | ---
language: es
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in es (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-es")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-es")
# load dataset
ds = load_dataset("common_voice", "es", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-et | 2021-05-05T16:24:26.000Z | [
"et",
"arxiv:2101.00390",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"preprocessor_config.json",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 0 | ---
language: et
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in et (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-et")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-et")
# load dataset
ds = load_dataset("common_voice", "et", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
|
facebook/wav2vec2-base-10k-voxpopuli-ft-fi | 2021-05-05T16:24:27.000Z | [
"pytorch",
"wav2vec2",
"fi",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 207 | transformers | ---
language: fi
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in fi (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-fi")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-fi")
# load dataset
ds = load_dataset("common_voice", "fi", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-fr | 2021-05-05T16:24:27.000Z | [
"pytorch",
"wav2vec2",
"fr",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 56 | transformers | ---
language: fr
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in fr (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-fr")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-fr")
# load dataset
ds = load_dataset("common_voice", "fr", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-hr | 2021-05-05T16:24:28.000Z | [
"pytorch",
"wav2vec2",
"hr",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 13 | transformers | ---
language: hr
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in hr (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-hr")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-hr")
# load dataset
ds = load_dataset("common_voice", "hr", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-hu | 2021-05-05T16:24:28.000Z | [
"pytorch",
"wav2vec2",
"hu",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 27 | transformers | ---
language: hu
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in hu (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-hu")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-hu")
# load dataset
ds = load_dataset("common_voice", "hu", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-it | 2021-05-05T16:24:29.000Z | [
"pytorch",
"wav2vec2",
"it",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 12 | transformers | ---
language: it
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in it (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-it")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-it")
# load dataset
ds = load_dataset("common_voice", "it", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-lt | 2021-05-05T16:24:29.000Z | [
"lt",
"arxiv:2101.00390",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"preprocessor_config.json",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 0 | ---
language: lt
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in lt (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-lt")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-lt")
# load dataset
ds = load_dataset("common_voice", "lt", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
|
facebook/wav2vec2-base-10k-voxpopuli-ft-nl | 2021-05-05T16:24:30.000Z | [
"pytorch",
"wav2vec2",
"nl",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 48 | transformers | ---
language: nl
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in nl (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-nl")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-nl")
# load dataset
ds = load_dataset("common_voice", "nl", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-pl | 2021-05-05T16:24:30.000Z | [
"pytorch",
"wav2vec2",
"pl",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 27 | transformers | ---
language: pl
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in pl (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-pl")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-pl")
# load dataset
ds = load_dataset("common_voice", "pl", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-ro | 2021-05-05T16:24:31.000Z | [
"pytorch",
"wav2vec2",
"ro",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 13 | transformers | ---
language: ro
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in ro (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-ro")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-ro")
# load dataset
ds = load_dataset("common_voice", "ro", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-sk | 2021-05-05T16:24:32.000Z | [
"pytorch",
"wav2vec2",
"sk",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 14 | transformers | ---
language: sk
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in sk (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-sk")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-sk")
# load dataset
ds = load_dataset("common_voice", "sk", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli-ft-sl | 2021-05-05T16:24:32.000Z | [
"pytorch",
"wav2vec2",
"sl",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 8 | transformers | ---
language: sl
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli-Finetuned
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10K unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390) and fine-tuned on the transcribed data in sl (refer to Table 1 of paper for more information).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Usage for inference
In the following it is shown how the model can be used in inference on a sample of the [Common Voice dataset](https://commonvoice.mozilla.org/en/datasets)
```python
#!/usr/bin/env python3
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
from datasets import load_dataset
import torchaudio
import torch
# resample audio
# load model & processor
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-sl")
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-10k-voxpopuli-ft-sl")
# load dataset
ds = load_dataset("common_voice", "sl", split="validation[:1%]")
# common voice does not match target sampling rate
common_voice_sample_rate = 48000
target_sample_rate = 16000
resampler = torchaudio.transforms.Resample(common_voice_sample_rate, target_sample_rate)
# define mapping fn to read in sound file and resample
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
speech = resampler(speech)
batch["speech"] = speech[0]
return batch
# load all audio files
ds = ds.map(map_to_array)
# run inference on the first 5 data samples
inputs = processor(ds[:5]["speech"], sampling_rate=target_sample_rate, return_tensors="pt", padding=True)
# inference
logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, axis=-1)
print(processor.batch_decode(predicted_ids))
```
|
facebook/wav2vec2-base-10k-voxpopuli | 2021-06-09T19:07:53.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"multilingual",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 41 | transformers | ---
language: multilingual
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the 10k unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base-960h | 2021-06-13T19:41:37.000Z | [
"pytorch",
"tf",
"wav2vec2",
"en",
"dataset:librispeech_asr",
"arxiv:2006.11477",
"transformers",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"feature_extractor_config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 101,311 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- audio
- automatic-speech-recognition
license: apache-2.0
widget:
- label: Librispeech sample 1
src: https://cdn-media.huggingface.co/speech_samples/sample1.flac
- label: Librispeech sample 2
src: https://cdn-media.huggingface.co/speech_samples/sample2.flac
---
# Wav2Vec2-Base-960h
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The base model pretrained and fine-tuned on 960 hours of Librispeech on 16kHz sampled speech audio. When using the model
make sure that your speech input is also sampled at 16Khz.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
To transcribe audio files the model can be used as a standalone acoustic model as follows:
```python
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
from datasets import load_dataset
import soundfile as sf
import torch
# load model and tokenizer
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
# define function to read in sound file
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
# load dummy dataset and read soundfiles
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
# tokenize
input_values = tokenizer(ds["speech"][:2], return_tensors="pt", padding="longest").input_values # Batch size 1
# retrieve logits
logits = model(input_values).logits
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
```
## Evaluation
This code snippet shows how to evaluate **facebook/wav2vec2-base-960h** on LibriSpeech's "clean" and "other" test data.
```python
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import soundfile as sf
import torch
from jiwer import wer
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h").to("cuda")
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-base-960h")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
input_values = tokenizer(batch["speech"], return_tensors="pt", padding="longest").input_values
with torch.no_grad():
logits = model(input_values.to("cuda")).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=1, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))
```
*Result (WER)*:
| "clean" | "other" |
|---|---|
| 3.4 | 8.6 | |
facebook/wav2vec2-base-es-voxpopuli | 2021-06-09T18:54:54.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"es",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 54 | transformers | ---
language: es
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the es unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base-fr-voxpopuli | 2021-06-09T18:58:11.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"fr",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 81 | transformers | ---
language: fr
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the fr unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base-it-voxpopuli | 2021-06-09T19:00:31.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"it",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 24 | transformers | ---
language: it
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the it unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base-nl-voxpopuli | 2021-06-09T19:02:59.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"nl",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 60 | transformers | ---
language: nl
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the nl unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base-sv-voxpopuli | 2021-06-09T19:05:27.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"sv",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 7 | transformers | ---
language: sv
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Base-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) base model pretrained on the sv unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-base | 2021-06-09T18:15:58.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"en",
"dataset:librispeech_asr",
"arxiv:2006.11477",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 6,988 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
license: apache-2.0
---
# Wav2Vec2-Base
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The base model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note that this model should be fine-tuned on a downstream task, like Automatic Speech Recognition. Check out [this blog](https://huggingface.co/blog/fine-tune-wav2vec2-english) for more information.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
See [this notebook](https://colab.research.google.com/drive/1FjTsqbYKphl9kL-eILgUc-bl4zVThL8F?usp=sharing) for more information on how to fine-tune the model. |
|
facebook/wav2vec2-large-100k-voxpopuli | 2021-06-09T20:08:04.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"multilingual",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 444 | transformers | ---
language: multilingual
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the 100k unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-10k-voxpopuli | 2021-06-09T20:00:07.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"multilingual",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 13 | transformers | ---
language: multilingual
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the 10k unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-960h-lv60-self | 2021-06-13T19:53:32.000Z | [
"pytorch",
"tf",
"wav2vec2",
"en",
"dataset:librispeech_asr",
"arxiv:2010.11430",
"arxiv:2006.11477",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"feature_extractor_config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tf_model.h5",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 11,261 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
# Wav2Vec2-Large-960h-Lv60 + Self-Training
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The large model pretrained and fine-tuned on 960 hours of Libri-Light and Librispeech on 16kHz sampled speech audio. Model was trained with [Self-Training objective](https://arxiv.org/abs/2010.11430). When using the model make sure that your speech input is also sampled at 16Khz.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
To transcribe audio files the model can be used as a standalone acoustic model as follows:
```python
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
from datasets import load_dataset
import soundfile as sf
import torch
# load model and tokenizer
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
# define function to read in sound file
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
# load dummy dataset and read soundfiles
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
# tokenize
input_values = tokenizer(ds["speech"][:2], return_tensors="pt", padding="longest").input_values # Batch size 1
# retrieve logits
logits = model(input_values).logits
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
```
## Evaluation
This code snippet shows how to evaluate **facebook/wav2vec2-large-960h-lv60-self** on LibriSpeech's "clean" and "other" test data.
```python
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import soundfile as sf
import torch
from jiwer import wer
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60-self").to("cuda")
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h-lv60-self")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
inputs = tokenizer(batch["speech"], return_tensors="pt", padding="longest")
input_values = inputs.input_values.to("cuda")
attention_mask = inputs.attention_mask.to("cuda")
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=16, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))
```
*Result (WER)*:
| "clean" | "other" |
|---|---|
| 1.9 | 3.9 | |
facebook/wav2vec2-large-960h-lv60 | 2021-02-25T14:58:20.000Z | [
"pytorch",
"wav2vec2",
"en",
"dataset:librispeech_asr",
"arxiv:2006.11477",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 1,458 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
license: apache-2.0
---
# Wav2Vec2-Large-960h-Lv60
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The large model pretrained and fine-tuned on 960 hours of Libri-Light and Librispeech on 16kHz sampled speech audio. When using the model
make sure that your speech input is also sampled at 16Khz.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
To transcribe audio files the model can be used as a standalone acoustic model as follows:
```python
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
from datasets import load_dataset
import soundfile as sf
import torch
# load model and tokenizer
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h-lv60")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60")
# define function to read in sound file
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
# load dummy dataset and read soundfiles
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
# tokenize
input_values = tokenizer(ds["speech"][:2], return_tensors="pt", padding="longest").input_values # Batch size 1
# retrieve logits
logits = model(input_values).logits
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
```
## Evaluation
This code snippet shows how to evaluate **facebook/wav2vec2-large-960h-lv60** on LibriSpeech's "clean" and "other" test data.
```python
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import soundfile as sf
import torch
from jiwer import wer
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h-lv60").to("cuda")
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h-lv60")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
inputs = tokenizer(batch["speech"], return_tensors="pt", padding="longest")
input_values = inputs.input_values.to("cuda")
attention_mask = inputs.attention_mask.to("cuda")
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=16, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))
```
*Result (WER)*:
| "clean" | "other" |
|---|---|
| 2.2 | 4.5 | |
|
facebook/wav2vec2-large-960h | 2021-02-25T14:58:16.000Z | [
"pytorch",
"wav2vec2",
"en",
"dataset:librispeech_asr",
"arxiv:2006.11477",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 2,654 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
license: apache-2.0
---
# Wav2Vec2-Large-960h
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The large model pretrained and fine-tuned on 960 hours of Librispeech on 16kHz sampled speech audio. When using the model
make sure that your speech input is also sampled at 16Khz.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
To transcribe audio files the model can be used as a standalone acoustic model as follows:
```python
from transformers import Wav2Vec2Tokenizer, Wav2Vec2ForCTC
from datasets import load_dataset
import soundfile as sf
import torch
# load model and tokenizer
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h")
# define function to read in sound file
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
# load dummy dataset and read soundfiles
ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")
ds = ds.map(map_to_array)
# tokenize
input_values = tokenizer(ds["speech"][:2], return_tensors="pt", padding="longest").input_values # Batch size 1
# retrieve logits
logits = model(input_values).logits
# take argmax and decode
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
```
## Evaluation
This code snippet shows how to evaluate **facebook/wav2vec2-large-960h** on LibriSpeech's "clean" and "other" test data.
```python
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import soundfile as sf
import torch
from jiwer import wer
librispeech_eval = load_dataset("librispeech_asr", "clean", split="test")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h").to("cuda")
tokenizer = Wav2Vec2Tokenizer.from_pretrained("facebook/wav2vec2-large-960h")
def map_to_array(batch):
speech, _ = sf.read(batch["file"])
batch["speech"] = speech
return batch
librispeech_eval = librispeech_eval.map(map_to_array)
def map_to_pred(batch):
input_values = tokenizer(batch["speech"], return_tensors="pt", padding="longest").input_values
with torch.no_grad():
logits = model(input_values.to("cuda")).logits
predicted_ids = torch.argmax(logits, dim=-1)
transcription = tokenizer.batch_decode(predicted_ids)
batch["transcription"] = transcription
return batch
result = librispeech_eval.map(map_to_pred, batched=True, batch_size=1, remove_columns=["speech"])
print("WER:", wer(result["text"], result["transcription"]))
```
*Result (WER)*:
| "clean" | "other" |
|---|---|
| 2.8 | 6.3 | |
|
facebook/wav2vec2-large-es-voxpopuli | 2021-06-09T19:17:31.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"es",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 81 | transformers | ---
language: es
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the es unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-fr-voxpopuli | 2021-06-09T19:25:11.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"fr",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 731 | transformers | ---
language: fr
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the fr unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-it-voxpopuli | 2021-06-09T19:32:52.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"it",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 180 | transformers | ---
language: it
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the it unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-lv60 | 2021-06-09T18:50:02.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"en",
"dataset:librispeech_asr",
"arxiv:2006.11477",
"transformers",
"speech",
"license:apache-2.0"
]
| [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 4,488 | transformers | ---
language: en
datasets:
- librispeech_asr
tags:
- speech
license: apache-2.0
---
# Wav2Vec2-Large-LV60
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
The base model pretrained on 16kHz sampled speech audio. When using the model make sure that your speech input is also sampled at 16Khz. Note that this model should be fine-tuned on a downstream task, like Automatic Speech Recognition. Check out [this blog](https://huggingface.co/blog/fine-tune-wav2vec2-english) for more information.
[Paper](https://arxiv.org/abs/2006.11477)
Authors: Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli
**Abstract**
We show for the first time that learning powerful representations from speech audio alone followed by fine-tuning on transcribed speech can outperform the best semi-supervised methods while being conceptually simpler. wav2vec 2.0 masks the speech input in the latent space and solves a contrastive task defined over a quantization of the latent representations which are jointly learned. Experiments using all labeled data of Librispeech achieve 1.8/3.3 WER on the clean/other test sets. When lowering the amount of labeled data to one hour, wav2vec 2.0 outperforms the previous state of the art on the 100 hour subset while using 100 times less labeled data. Using just ten minutes of labeled data and pre-training on 53k hours of unlabeled data still achieves 4.8/8.2 WER. This demonstrates the feasibility of speech recognition with limited amounts of labeled data.
The original model can be found under https://github.com/pytorch/fairseq/tree/master/examples/wav2vec#wav2vec-20.
# Usage
See [this notebook](https://colab.research.google.com/drive/1FjTsqbYKphl9kL-eILgUc-bl4zVThL8F?usp=sharing) for more information on how to fine-tune the model. |
|
facebook/wav2vec2-large-nl-voxpopuli | 2021-06-09T19:40:45.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"nl",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 144 | transformers | ---
language: nl
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the nl unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-sv-voxpopuli | 2021-06-09T19:52:19.000Z | [
"pytorch",
"wav2vec2",
"pretraining",
"sv",
"arxiv:2101.00390",
"transformers",
"audio",
"automatic-speech-recognition",
"voxpopuli",
"license:cc-by-nc-4.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin"
]
| facebook | 10 | transformers | ---
language: sv
tags:
- audio
- automatic-speech-recognition
- voxpopuli
license: cc-by-nc-4.0
---
# Wav2Vec2-Large-VoxPopuli
[Facebook's Wav2Vec2](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/) large model pretrained on the sv unlabeled subset of [VoxPopuli corpus](https://arxiv.org/abs/2101.00390).
**Paper**: *[VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation
Learning, Semi-Supervised Learning and Interpretation](https://arxiv.org/abs/2101.00390)*
**Authors**: *Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, Emmanuel Dupoux* from *Facebook AI*
See the official website for more information, [here](https://github.com/facebookresearch/voxpopuli/)
# Fine-Tuning
Please refer to [this blog](https://huggingface.co/blog/fine-tune-xlsr-wav2vec2) on how to fine-tune this model on a specific language. Note that you should replace `"facebook/wav2vec2-large-xlsr-53"` with this checkpoint for fine-tuning.
|
facebook/wav2vec2-large-xlsr-53-dutch | 2021-03-11T11:33:41.000Z | [
"pytorch",
"wav2vec2",
"nl",
"dataset:common_voice",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 249 | transformers | ---
language: nl
datasets:
- common_voice
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
## Evaluation on Common Voice NL Test
```python
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
Wav2Vec2ForCTC,
Wav2Vec2Processor,
)
import torch
import re
import sys
model_name = "facebook/wav2vec2-large-xlsr-53-dutch"
device = "cuda"
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"]' # noqa: W605
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2Processor.from_pretrained(model_name)
ds = load_dataset("common_voice", "nl", split="test", data_dir="./cv-corpus-6.1-2020-12-11")
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
batch["sampling_rate"] = resampler.new_freq
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
return batch
ds = ds.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
input_values = features.input_values.to(device)
attention_mask = features.attention_mask.to(device)
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["predicted"] = processor.batch_decode(pred_ids)
batch["target"] = batch["sentence"]
return batch
result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
wer = load_metric("wer")
print(wer.compute(predictions=result["predicted"], references=result["target"]))
```
**Result**: 21.1 % |
facebook/wav2vec2-large-xlsr-53-french | 2021-03-11T12:22:56.000Z | [
"pytorch",
"wav2vec2",
"fr",
"dataset:common_voice",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 1,431 | transformers | ---
language: fr
datasets:
- common_voice
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
## Evaluation on Common Voice FR Test
```python
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
Wav2Vec2ForCTC,
Wav2Vec2Processor,
)
import torch
import re
import sys
model_name = "facebook/wav2vec2-large-xlsr-53-french"
device = "cuda"
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"]' # noqa: W605
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2Processor.from_pretrained(model_name)
ds = load_dataset("common_voice", "fr", split="test", data_dir="./cv-corpus-6.1-2020-12-11")
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
batch["sampling_rate"] = resampler.new_freq
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
return batch
ds = ds.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
input_values = features.input_values.to(device)
attention_mask = features.attention_mask.to(device)
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["predicted"] = processor.batch_decode(pred_ids)
batch["target"] = batch["sentence"]
return batch
result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
wer = load_metric("wer")
print(wer.compute(predictions=result["predicted"], references=result["target"]))
```
**Result**: 25.2 % |
facebook/wav2vec2-large-xlsr-53-german | 2021-03-11T11:43:52.000Z | [
"pytorch",
"wav2vec2",
"de",
"dataset:common_voice",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 1,521 | transformers | ---
language: de
datasets:
- common_voice
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
## Evaluation on Common Voice DE Test
```python
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
Wav2Vec2ForCTC,
Wav2Vec2Processor,
)
import torch
import re
import sys
model_name = "facebook/wav2vec2-large-xlsr-53-german"
device = "cuda"
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"]' # noqa: W605
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2Processor.from_pretrained(model_name)
ds = load_dataset("common_voice", "de", split="test", data_dir="./cv-corpus-6.1-2020-12-11")
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
batch["sampling_rate"] = resampler.new_freq
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
return batch
ds = ds.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
input_values = features.input_values.to(device)
attention_mask = features.attention_mask.to(device)
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["predicted"] = processor.batch_decode(pred_ids)
batch["target"] = batch["sentence"]
return batch
result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
wer = load_metric("wer")
print(wer.compute(predictions=result["predicted"], references=result["target"]))
```
**Result**: 18.5 % |
facebook/wav2vec2-large-xlsr-53-italian | 2021-03-11T11:41:55.000Z | [
"pytorch",
"wav2vec2",
"it",
"dataset:common_voice",
"transformers",
"speech",
"audio",
"automatic-speech-recognition",
"license:apache-2.0"
]
| automatic-speech-recognition | [
".gitattributes",
"README.md",
"config.json",
"preprocessor_config.json",
"pytorch_model.bin",
"special_tokens_map.json",
"tokenizer_config.json",
"vocab.json"
]
| facebook | 1,792 | transformers | ---
language: it
datasets:
- common_voice
tags:
- speech
- audio
- automatic-speech-recognition
license: apache-2.0
---
## Evaluation on Common Voice IT Test
```python
import torchaudio
from datasets import load_dataset, load_metric
from transformers import (
Wav2Vec2ForCTC,
Wav2Vec2Processor,
)
import torch
import re
import sys
model_name = "facebook/wav2vec2-large-xlsr-53-italian"
device = "cuda"
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"]' # noqa: W605
model = Wav2Vec2ForCTC.from_pretrained(model_name).to(device)
processor = Wav2Vec2Processor.from_pretrained(model_name)
ds = load_dataset("common_voice", "it", split="test", data_dir="./cv-corpus-6.1-2020-12-11")
resampler = torchaudio.transforms.Resample(orig_freq=48_000, new_freq=16_000)
def map_to_array(batch):
speech, _ = torchaudio.load(batch["path"])
batch["speech"] = resampler.forward(speech.squeeze(0)).numpy()
batch["sampling_rate"] = resampler.new_freq
batch["sentence"] = re.sub(chars_to_ignore_regex, '', batch["sentence"]).lower().replace("’", "'")
return batch
ds = ds.map(map_to_array)
def map_to_pred(batch):
features = processor(batch["speech"], sampling_rate=batch["sampling_rate"][0], padding=True, return_tensors="pt")
input_values = features.input_values.to(device)
attention_mask = features.attention_mask.to(device)
with torch.no_grad():
logits = model(input_values, attention_mask=attention_mask).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["predicted"] = processor.batch_decode(pred_ids)
batch["target"] = batch["sentence"]
return batch
result = ds.map(map_to_pred, batched=True, batch_size=16, remove_columns=list(ds.features.keys()))
wer = load_metric("wer")
print(wer.compute(predictions=result["predicted"], references=result["target"]))
```
**Result**: 22.1 % |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.