Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[]

[]

[]

[ "- `transformers` version: 2.9.1\r\n- Platform: Linux-3.10.0-1062.9.1.el7.x86_64-x86_64-with-centos-7.7.1908-Core\r\n- Python version: 3.6.10\r\n- PyTorch version (GPU?): not installed (NA)\r\n- Tensorflow version (GPU?): 2.2.0 (True)\r\n- Using GPU in script?: Yes both single gpu and multi-gpu training\r\n- Using distributed or parallel set-up in script?: Yes", "Error log:\r\n ValueError: Variable <tf.Variable 'tf_bert_for_token_classification/bert/pooler/dense/kernel:0' shape=(768, 768) dtype=float32> has `None` for gradient. Please make sure that all of your ops have\r\n a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.", "Tried with TF 2.0:\r\nError Message: AttributeError: module 'tensorflow_core._api.v2.config' has no attribute 'list_physical_devices'\r\nTF 2.1: Same as TF 2.2 i.e. ValueError: Variable <tf.Variable 'tf_bert_for_token_classification/bert/pooler/dense/kernel:0' shape=(768, 768) dtype=float32> has `None` for gradient. Please make sure that all of your ops have\r\n a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.", "I had same issue.\r\nadd ```--mode token-classification``` to the command\r\nreference code https://github.com/huggingface/transformers/blob/18d233d52588b4e08dc785fbfecd77529e9effa6/src/transformers/trainer_tf.py#L380", "Thanks @linhx13 ..Works fine.." ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "Yes having a \"Conditional Generation\" pipeline makes sense given that variety of tasks can be solved using it. We can use T5, BART for these tasks as well as the new Encoder-Decoder. I would like to call it `TextToTextPipeline` though, since we can solve non-generative tasks also as demonstrated in the T5 paper. I think this pipeline will be really useful.", "Technically, any task using Text-To-Text is generative in nature right? But yeah, agree `TextToTextPipeline` will make the use case clearer :smile:\r\n\r\nHoping to get feedback from @patrickvonplaten before attempting this", "Yeah. To be honest, I'm not sure whether this is a good idea. The pipelines are supposed to be directly related to a task such as `translation`, `summarization` which are specific cases of `text2text` applications. \r\n\r\nI think for every task we should introduce a new `pipeline` before starting to have different levels of abstractions in `pipelines`. A `TextToTextPipeline could become quite a mess regarding different possible input formats, different prefixes (for T5), etc...For general tasks such as these ones I'd prefer to just implement your own code using the `.generate()` function. \r\n\r\n@LysandreJik - what do you think? ", "I think from a high level, more than just thinking about `text2text`, I'm foreseeing the future where multi-task learning becomes a standard way of deploying ML models. Having a pipeline to introduce this can be one step to accelerating that future.\r\n\r\nAlthough, I do understand that `text2text` is just one approach to doing this, but in my opinion, it's the most promising one at the moment, so it's a good interface to start with for a multi task model pipeline.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "I'm not sure that T5 is the most promising place to do a multi-task pipeline, since their results in that paper suggested it was hard to significantly beat the baseline of just fine tuning on the target task. \r\n\r\nThe recent AdapterHub library built off of HuggingFace seems a better place for building out multitask systems/pipelines imo. But of course the library designers have more intuition on this.", "I'm don't think anyone is arguing for the T5 model specifically, just that there is a trend towards `text2text` as a common method of doing multitask learning for NLP (GPT-3 frames tasks like this too for example).", "> I'm don't think anyone is arguing for the T5 model specifically, just that there is a trend towards `text2text` as a common method of doing multitask learning for NLP (GPT-3 frames tasks like this too for example).\r\n\r\nFair enough. I'm not one to argue against a feature, even if I wouldn't use it much myself. I've been using `text2text` myself for multiple tasks. \r\n\r\nMostly I just meant the multitask part of `text2text` is going to be a little tricky to abstract away conveniently into a pipeline. The main complexity there is mixing the proportion of each task / batch correctly. The T5 paper suggests performance and weights are very specific to the multitask learning, and if its not tuned properly the performance will be hurt by using multitasks. Uniform mixing for example performs quite poorly. I suspect that problem would apply to most `text2text` paradigms.\r\n\r\n What I've been doing myself is using a custom DataLoader class that handles the mixing of batch proportions of each task. A pipeline that can integrate something like that would be terrific to have.", "Hey everybody, after thinking a bit more about it, I think it does make sense to add a `ConditionalTextGeneration` pipeline which will be the equivalent of `TextGenerationPipeline` for all models in `AutoModelForSeq2Seq`. It should look very similar to the `TextGenerationPipeline` (probably we more or less the same at the moment), but it will give us more freedom in the future (for example when we add `decoder_input_ids` to the generation). \r\n@sshleifer , @yjernite , @LysandreJik - what are your thoughts on this?", "@patrickvonplaten happy to work on a PR for this if team agrees it makes sense :smile:", "I think we definitely need something like that.\r\n\r\nI'd probably go with a more explicit name though: e.g. `TextToTextPipeline` or `Text2TextGenerationPipeline`. `ConditionalTextGeneration` might cover other uses in the future (e.g. multiple input texts or multimodal inputs)", "Such a pipeline would be very welcome, indeed!", "Awesome, will send a PR in the next week or so :smile:", "I also want to work on this, @enzoampil let me know if you want to collaborate on the PR :)", "Sure thing, maybe we can collab on the same fork? :)" ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "Sorry I don't really understand how this PR removes pytorch code in `TFXLNetMainLayer` - can you explain a bit more in-detail?", "@patrickvonplaten \r\n\r\nThe codes in file modeling_tf_xlnet.py is the XLNet Model of tensorflow implementation. So the parameter **head_mask** should have a type of tf.Tensor or Numpy array (see Line 780 ).\r\n\r\nBut in Lines 643-650, the parameter **head_mask** uses the methods of the pytorch, such as \"expand\" or \"unsqueeze\" which can't apply to tf.Tensor or Numpy array. \r\n\r\nActually these codes are copied from modeling_xlnet.py by mistake and should be removed.", "Perfect thanks a lot! \r\n\r\n@LysandreJik - the RUN_SLOW=1 tests all poss for TFXLNetMainLayer.\r\n\r\nGood to merge for me!" ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Nice! cc @patrickvonplaten " ]

[ { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "Duplicate of #4185" ]

[]

[]

[]

[ { "id": 1260952223, "node_id": "MDU6TGFiZWwxMjYwOTUyMjIz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Discussion", "name": "Discussion", "color": "22870e", "default": false, "description": "Discussion on a topic (keep it focused or open a new issue though)" }, { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "First thing you can try is fine-tune T5/BART for summarization on your corpus and see how it performs.", "@patil-suraj where can I find a guide to this? I'm a bit confused by the documentation. ", "[Here's](https://github.com/huggingface/transformers/tree/master/examples/summarization/bart) the official example which fine-tunes BART on CNN/DM, you can just replace the cnn/dm dataset with your own summerization dataset.", "@patil-suraj Thanks for the example. I'm wondering if there is any simpler way to get started since I'm planning on training it in a Kaggle notebook due to GPU constraints, because otherwise I may need to copy paste entire folder into a Kaggle notebook.", "@kevinlu1248 \r\nThis [colab](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) shows how to fine-tune T5 with lightening. This is just the self-contained version of official example. You should be able to use the same `Trainer`, just replace the model with BART and use you own dataset.", "@patil-suraj Thanks, I'll look into it.", "> [Here's](https://github.com/huggingface/transformers/tree/master/examples/summarization/bart) the official example which fine-tunes BART on CNN/DM, you can just replace the cnn/dm dataset with your own summerization dataset.\r\n\r\nHi @patil-suraj, I am following that example and have my data in that format, and I can see the process using GPU/CPU, but I can't get tensorboard working. Do you have any hints? I am happy to contribute to documentation once I get it working.", "@sam-qordoba lightning handles logging itself and by default the tensorboard logs are saved in lightning_logs directory. So you should be able see the logs by passing lightning_logs as the logdir to tensorboard command.", "Thanks @patil-suraj ", "Hey @patil-suraj, I had OOM issues on Colab, so moved to a VM with 56GB RAM, and the behaviour is the same as on Colab: memory usage grows, until it uses up everything available (I even added 32GB of swap, so, it's a really impressive amount of memory usage), until I get locked out of the machine... and the only time it writes to `lightning_logs` is right when it starts. \r\n\r\n```sh\r\njupyter@pytorch-20200529-155153:~/lightning_logs$ tree\r\n.\r\n└── version_0\r\n ├── events.out.tfevents.1590794134.pytorch-20200529-155753.8733.0\r\n └── hparams.yaml\r\n\r\n1 directory, 2 files\r\n```\r\n\r\n`nvidia-smi` looks like this:\r\n\r\n```\r\njupyter@pytorch-20200529-155753:~$ nvidia-smi \r\nSat May 30 00:07:12 2020 \r\n+-----------------------------------------------------------------------------+\r\n| NVIDIA-SMI 418.87.01 Driver Version: 418.87.01 CUDA Version: 10.1 |\r\n|-------------------------------+----------------------+----------------------+\r\n| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |\r\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\r\n|===============================+======================+======================|\r\n| 0 Tesla T4 Off | 00000000:00:04.0 Off | 0 |\r\n| N/A 77C P0 35W / 70W | 2579MiB / 15079MiB | 0% Default |\r\n+-------------------------------+----------------------+----------------------+\r\n \r\n+-----------------------------------------------------------------------------+\r\n| Processes: GPU Memory |\r\n| GPU PID Type Process name Usage |\r\n|=============================================================================|\r\n| 0 8733 C /opt/conda/bin/python 2569MiB |\r\n+-----------------------------------------------------------------------------+\r\n```\r\n\r\nThe cell `trainer.fit(model)` outputs the model definition, but no progress bar on anything,\r\n\r\n``` \r\n\r\n | Name | Type | Params\r\n-----------------------------------------------------------------------------------------------------------------\r\n0 | model | T5ForConditionalGeneration | 222 M \r\n1 | model.shared | Embedding | 24 M \r\n2 | model.encoder | T5Stack | 109 M \r\n...\r\n514 | model.decoder.block.11.layer.2.dropout | Dropout | 0 \r\n515 | model.decoder.final_layer_norm | T5LayerNorm | 768 \r\n516 | model.decoder.dropout | Dropout | 0 \r\n517 | model.lm_head | Linear | 24 M \r\nSelected optimization level O1: Insert automatic casts around Pytorch functions and Tensor methods.\r\n\r\nDefaults for this optimization level are:\r\nenabled : True\r\nopt_level : O1\r\ncast_model_type : None\r\npatch_torch_functions : True\r\nkeep_batchnorm_fp32 : None\r\nmaster_weights : None\r\nloss_scale : dynamic\r\nProcessing user overrides (additional kwargs that are not None)...\r\nAfter processing overrides, optimization options are:\r\nenabled : True\r\nopt_level : O1\r\ncast_model_type : None\r\npatch_torch_functions : True\r\nkeep_batchnorm_fp32 : None\r\nmaster_weights : None\r\nloss_scale : dynamic\r\n```\r\n\r\nSorry to keep bothering you, but do you have any hints? It's hard to know what's going on because it doesn't seem to log", "It shouldn't take that much memory, did you try reducing the batch size ?\r\n\r\nAlso seems that you are using fp16 here. I haven't tried it with fp16 yet.\r\n\r\ntagging @sshleifer ", "Ok, I tried fp16 as a \"maybe this will use less memory\" experiment, I will try without. I tried batch size of 4, could go lower I guess. Should I just double the learning rate each time I halve the batch size, or are other changes needed?", "Could somebody who has fine-tuned BART give me an estimate of how long it takes / how many epochs until convergence? Also any tricks to speed it up (weight freezing etc)?\r\n\r\n1 epoch takes c. 150 hrs for my dataset so wondering how many I need...", "Sounds like you have a huge dataset?\r\nIt's tough to know exactly how many you will need, but for xsum and cnn most of the model's I've need have required 4-6 to converged.\r\nThe [original authors] https://github.com/pytorch/fairseq/blob/master/examples/bart/README.summarization.md#4-fine-tuning-on-cnn-dm-summarization-task say 15-20K Steps.\r\n\r\nI have had to go down to batch size=1 or 2 on some occasions. \r\nYou can use `--gradient_accumulation_steps` to keep the \"effective\" batch size (how many examples your model processes per backward pass) consistent.\r\n\r\n@sam-qordoba is your `Dataset/DataLoader` putting all the examples in memory before training? That could be an issue on a large dataset.\r\n\r\n", "You can also freeze the `BartForConditionalGeneration.model.encoder` using the function below to reduce memory cost.\r\n```\r\ndef freeze_part(model: nn.Module):\r\n for par in model.parameters():\r\n par.requires_grad = False\r\n```\r\n\r\nYou can also use `val_check_interval` in lightning to check validation statistics more frequently, but unfortunately your checkpoints will still be saved at the end of every epoch.", "@sshleifer thanks for coming back with this- all very helpful.\r\n\r\nYes- essentially I am just trying out using BART to for longer docs (arXiv/PubMed) as a baseline to compare more sophisticated methods against. This means training set has 300k samples and only 1 sample fits on the GPU at once (12Gb- using 1,024 input length).\r\n\r\nLots for me to play around with and see what works well. Thanks for your help.", "> Yes- essentially I am just trying out using BART to for longer docs (arXiv/PubMed) as a baseline to compare more sophisticated methods against\r\n\r\n@alexgaskell10 If you are interested in using BART for long documents then keep an eye here.\r\nhttps://github.com/patil-suraj/longbart\r\n\r\nI'm trying to convert BART to it's long version using longformer's sliding-window attention.\r\n\r\nI've been able to replace BART encoder's `SelfAttention` with `LongformerSelfAttention` with 4096 max length. Now I'm working on adding gradient checkpointing to allow it to train on smaller GPU's. Hope to finish it soon. \r\n\r\ngradient checkpointing and fp16 with '02' opt level should allow to use larger batch size", "@patil-suraj thanks for this- adapting BART for LongformerSelfAttention was actually something I was going to start looking into over the next couple of weeks. Thanks for sharing- I'll be sure to give it a go soon.", "Hey @patil-suraj, any updates on your latest progress on LongBART? Thinking about diving into a similar block of work: expanding BART via Longformer", "Hi @virattt , I've been able to replace bart encoder's self attention with sliding window attention. Also added gradient checkpoiting in the encoder. \r\n\r\nGradient checkpoiting in decoder is not working so going to remove it for now. Will update the repo this weekend and will put some instructions in the readme.", "Sounds great, thanks @patil-suraj ", "Would love to hear `LongBart` experimental results whenever they are available!", "@sshleifer I have been playing around with `LongBart` recently and have some preliminary experimental results. This is using @patil-suraj 's longbart repo fine-tuned on the PubMed dataset using the hf summarization finetune.py script.\r\n\r\nThe best result so far is ROUGE-1 = 36.8 (for comparison, fine-tuning vanilla `BART` on PubMed and truncating articles at 1024 tokens I got 42.3 ROUGE-1). I have only run a few configs so far and will be running many more so I expect this to improve. Next steps:\r\n- Have been only using a 12Gb GPU so far so have frozen the embeddings and encoder otherwise too large. I have a much larger cluster I can move to so will start running trials on this soon which will give more freedom to try different configs\r\n- I am only fine-tuning at the moment. Might explore doing some pre-training although this may be too expensive.\r\n\r\nLet me know if there is anything you would like to see and I'll try to schedule it in.", "Hi @alexgaskell10 , did you use the code as it is ? I think we'll need to train the embeddings for few epochs then we can freeze it.\r\nHowever without freezing the embeddings I ran into OOM halfway through the epoch even with bart-base with '02' fp16 on 16GB V100.\r\n\r\n@sshleifer do you have any ideas why this might be happening ? It went well till 60% of first epoch then OOM. Batch size was 1 and max_seq_len 4096 ?\r\n\r\n@alexgaskell10 can you share more details, how many epochs, batch size, fp16 or not ? ", "Yes, I used the code as is (minor changes to integrate with hf finetune.py script). I agree that the embeddings and encoder should not be frozen from the beginning but I couldn't fit it on my 12Gb GPU. Once I get setup on the cluster I'll try this.\r\n\r\nMore details on all my runs so far can be found in my [wandb project](https://app.wandb.ai/alexgaskell/Covid01-scripts_models/overview?workspace=user-alexgaskell). To answer your question, max a couple epochs so far, batch size between 4 and 16 depending on what fits, not fp16 so far (haven't set up yet but will do soon).", "Thanks @alexgaskell10 , I think you'll be able to use bart-base with fp16 and max 2048 seq len without frezzing embdddings on 12GB GPU ", "@patil-suraj: \r\n- 4096 is a very large max_seq_len, but I know that doesn't answer your question. I would guess that the answer is that you got a really big batch. The batches are not all the same size. We trim them to save padding computation. If you are on one GPU you can use `--sortish_sampler` which ensures that the first batch is the largest, so you get OOM at the beginning of the epoch at least. You also get hopefully a further reduction in padding computation. \r\n- I would be interested to know how much `--sortish_sampler` reduces the training cost of 1 epoch with other parameters fixed. \r\n\r\n\r\n@alexgaskell10 : \r\nThanks for sharing your wandb, it makes understanding what you're doing way easier.\r\n\r\n- From [pegasus](https://arxiv.org/pdf/1912.08777.pdf) Table 2, it seems like SOTA for PubMed is around `45.49/19.90/27.69`. (Rouge 1, Rouge 2, Rouge L) So there is still some headroom! (Note we will add pegasus in the library sometime in July).\r\n- From looking at your wandb logs, your models are still getting better when training stops. When you move to a beefier setup you might consider training for longer. \r\n- I think there is general consensus that Rouge2 and Rouge-L are better metrics than Rouge-1. \r\n\r\nSome questions I would love to know the answer to (for any dataset):\r\n\r\n1. which `--model_name_or_path` is the best starting point for finetuning: bart-base vs. bart-large vs. bart-large-xsum vs distilbart-xsum-12-6, for example.\r\n2. How does `LongBart` compare in performance to `BartForConditionalGeneration`?\r\n3. Does increasing `--adam_eps` improve performance? Jeremy Howard at fastai recommended this once, and the default 1e-8 seems to be a fairly low setting.\r\n4. What is the impact of `--freeze-encoder` and `--freeze_embeds` on time per epoch, max batch size, and performance.\r\n\r\n", "@sshleifer thanks for coming back to me. Several of your questions I can answer immediately, the others I will try to take a look at. If you're interested, I have a [separate wandb project](https://app.wandb.ai/alexgaskell/transformers-examples_summarization?workspace=user-alexgaskell) containing a bunch of fine-tuning runs for `BartForConditionalGeneration` on PubMed to act as a baseline for `Longformer`. All of these runs have frozen embs and enc because of size constraints- only batch size 1 or 2 fit on GPU and that didn't perform well. If I get a bigger setup I'll try with these unfrozen and a larger batch size.\r\n\r\nAddressing your questions:\r\n1. I have been using facebook/bart-large-cnn so far- will investigate if I get time\r\n2. This can be seen in the two wandb repos I've shared here and above. So far my best `BartForConditionalGeneration` is 0.426/0.177/0.263 and my best `Longformer` is 0.367/0.120/0.222 so BART is much better so far. However, both of these have frozen embs and enc (and presumably PEGASUS was fine-tuned without these frozen) so there are more experiments to run\r\n3. Haven't looked at this. Will give it a go\r\n4. Freezing both has a big impact (haven't looked at freezing each separately).\r\n- Time per epoch I think order of 3-4x quicker (8hrs vs 24+hrs per epoch using 12Gb GPU)\r\n- Batch size >8x improvement (2 vs 16)\r\n- Performance seemed much better when frozen. Probably due to small batch size training was unstable when using a small batch size. The img below shows a comparison between frozen (grey, bsz=16) and unfrozen (blue, bsz=2).\r\n\r\n<img width=\"1093\" alt=\"Screenshot 2020-07-01 at 22 06 57\" src=\"https://user-images.githubusercontent.com/51463426/86291868-96cbce80-bbe7-11ea-8427-52619710d2fa.png\">\r\n", "> BartForConditionalGeneration is 0.426/0.177/0.263 and my best Longformer is 0.367/0.120/0.222 \r\n\r\nThere's a bug related to masking that could be the reason for the performance drop. I started working on `LongformerEncoderDecoder` and have a fix [here](https://github.com/allenai/longformer/blob/encoderdecoder/longformer/longformer_encoder_decoder.py#L66). \r\n", "Thanks for flagging, I will take a look. \r\n\r\nIn any case, I have much better `LongBart` results now (0.433, 0.189, 0.273). I found that fine-tuning `LongBart` without freezing the embs or enc worked much better, whereas `Bart` performed better when embs and enc were frozen. This probably makes sense given that `LongBart` is using weight transfer so needs more comprehensive training to be effective. Hopefully the bug fix will improve these results even more. " ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Hi @howardhsu, the file path is not correct, it should be something like `model_cards/activebus/BERT_Review/README.md` ", "I updated paths as suggested, thanks.", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=h1) Report\n> Merging [#4405](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/3e0f06210646a440509efa718b30d18322d6a830&el=desc) will **increase** coverage by `0.02%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4405 +/- ##\n==========================================\n+ Coverage 78.16% 78.19% +0.02% \n==========================================\n Files 120 120 \n Lines 20058 20058 \n==========================================\n+ Hits 15679 15684 +5 \n+ Misses 4379 4374 -5 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4405/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.93% <0.00%> (+0.82%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=footer). Last update [3e0f062...466c62e](https://codecov.io/gh/huggingface/transformers/pull/4405?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Great: [model pages](https://huggingface.co/activebus)" ]

[]

[]

[ "It don't think it should really be a warning if you don't use wandb.\r\n\r\nThe root issue here is probably discoverability of wandb for users who don't know it? Then it would probably be better solved in documentation.\r\n\r\nWe will start some documentation on `Trainer`/`TFTrainer` in the coming weeks (cc @LysandreJik) we'll mention wandb there. (if you want to help with this let us know)", "Makes sense, let me know when it's started and I can help writing the section related to wandb." ]

[]

[]

[ "Thank you!" ]

[]

[]

[ "@jcblaisecruz02 \r\nYes, there's a bug in version 2.91 which hangs the trainer. It's been fixed in master branch. Install from master branch for TPU training.\r\n\r\nSee this pull request #4339", "Is it possible to run `run_language_modeling.py` on more than 8 cores when using pytorch and `xls_spawn`?\r\nAnd what about tensorflow?" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=h1) Report\n> Merging [#4401](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/2d184cb553ee20943b03b253f44300e466357871&el=desc) will **increase** coverage by `0.85%`.\n> The diff coverage is `100.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4401 +/- ##\n==========================================\n+ Coverage 77.30% 78.15% +0.85% \n==========================================\n Files 120 120 \n Lines 20027 20027 \n==========================================\n+ Hits 15481 15652 +171 \n+ Misses 4546 4375 -171 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl90NS5weQ==) | `95.16% <100.00%> (ø)` | |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.77% <0.00%> (+0.16%)` | :arrow_up: |\n| [src/transformers/trainer.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90cmFpbmVyLnB5) | `43.25% <0.00%> (+1.10%)` | :arrow_up: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `81.78% <0.00%> (+1.37%)` | :arrow_up: |\n| [src/transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG5ldC5weQ==) | `75.73% <0.00%> (+2.29%)` | :arrow_up: |\n| [src/transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19jdHJsLnB5) | `97.81% <0.00%> (+2.62%)` | :arrow_up: |\n| [src/transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19yb2JlcnRhLnB5) | `95.71% <0.00%> (+10.00%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_pytorch\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4401/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9weXRvcmNoX3V0aWxzLnB5) | `89.93% <0.00%> (+81.20%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=footer). Last update [2d184cb...fa80cf3](https://codecov.io/gh/huggingface/transformers/pull/4401?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "If needed I can even provide the dataset (Did not want to clutter): \r\nthe error stacktrace : \r\n```\r\nTraceback (most recent call last):\r\n File \"error.py\", line 129, in <module>\r\n a = enumerate(train_data_loader)\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 279, in __iter__\r\n return _MultiProcessingDataLoaderIter(self)\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\site-packages\\torch\\utils\\data\\dataloader.py\", line 719, in __init__\r\n w.start()\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\process.py\", line 121, in start\r\n self._popen = self._Popen(self)\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\context.py\", line 224, in _Popen\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n return _default_context.get_context().Process._Popen(process_obj)\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\context.py\", line 326, in _Popen\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\spawn.py\", line 116, in spawn_main\r\n return Popen(process_obj)\r\nexitcode = _main(fd, parent_sentinel) File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\popen_spawn_win32.py\", line 93, in __init__\r\n\r\n File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\spawn.py\", line 126, in _main\r\n reduction.dump(process_obj, to_child)\r\nself = reduction.pickle.load(from_parent) File \"C:\\Users\\admin\\miniconda3\\envs\\machine_learning\\lib\\multiprocessing\\reduction.py\", line 60, in dump\r\n\r\nEOFError: Ran out of input\r\n ForkingPickler(file, protocol).dump(obj)\r\nTypeError: cannot pickle 'Tokenizer' object\r\n```\r\nBut when run in kaggle notebook this works perfectly well. (The same script having same tokenizer and transformers version)\r\n@julien-c @sshleifer any help here?", "~Yes, this is fixed by PR #4389 , so you could `pip install -e .` off of that branch.~", "Hi @sshleifer I did these steps : \r\n1. git fetch origin pull/4389/head:temp_fix\r\n2. git checkout temp_fix\r\n3. pip install -e .\r\nstill the above fix doesn't seem to work. \r\n\r\nBy looking into the PR , I am guessing that it is fixed for `MarianTokenizer ` and not for `BertWordPieceTokenizer` that I am using in the above script.\r\n ", "Tested with another environment with python = 3.7.7, same issue is observed.", "Just wanted to mention that providing `num_workers = 0` bypasses the problem. So it only fails when multiprocessing is involved. This issue is not only in `BertWordPieceTokenizer`, It also fails with `ByteLevelBPETokenizer` .", "And this probably should be moved to the tokenizer repo @sshleifer to confirm.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "That would be an interesting project.", "I have worked on question generation using T5. I've trained answer aware question generator on SQuAD 2.2 which achieves 41.4641 BLUE1 score and 41.0823 ROUGE_L on the dev set given gold answers.\r\n\r\nI've also trained T5 for extracting answers from the text, and written a simple pipeline where the answer generator generates answers and then the answer-aware que generator generates questions with those answers. You can check the demo [here](https://colab.research.google.com/drive/1_2_mS5l29QHI1pXaqa4YLzAO5xm-HmH9?usp=sharing)\r\n\r\nI've also trained T5 for direct question generation on Yahoo questions dataset. It generates single question given a context.\r\n\r\nI would be happy to contribute to this project.", "@julien-c \r\nAny update on this ?", "We don't have any immediate plan to work on this ourselves but feel free to take a stab", "@julien-c \r\nOkay, I just need your feedback on one thing. I'm not sure if adding a pipeline here will make sense since there are multiple ways to generate questions\r\n1. ans aware\r\n2. generate 1 question directly\r\n3. generate multiple questions simultaneously\r\n\r\nAlso ans aware model will need an ans extractor/generator or the user will need to supply answers. And different models could process inputs differently. \r\nSo does adding pipeline makes sense here ? If not I can just upload the models and provide a inference script and decide next steps with community feedback. \r\n\r\nThank you!", "@patil-suraj I think it makes sense to generate questions that are answer aware as this has more use cases. I also think that questions should not be so narrow that a single word from the context is the answer. Artit's [Text2Text ](https://github.com/artitw/text2text) did a pretty good job. I used this to generate 1,000 random questions from a random context and plan to have them judged by human raters. One of the main issues I saw in these questions is that sometimes the answer is in the question text. \r\n\r\nI'll take a look at your demo.\r\n\r\nI also think it should generate multiple questions that can be so that some evaluation metric (BLUE1, ROUGE_L and METEOR) can be produced and have them ranked. Then the user can decide what to do with the highest metric questions. There's some research that shows the METEOR correlates better with human judgement, but this was for evaluating machine translation tasks and might not apply here.", "@danielduckworth \r\nYes answer aware seems to be the best chose right now. And we can generate multiple question if we have multiple answers. If you find the above demo interesting then I'll share the models so you can play with it and then we can decide how to proceed from there. \r\n\r\nAlso the METEOR score on dev set for demo model is 26.0676", "@patil-suraj Yes that would be great. How do you want to share the models?", "@danielduckworth \r\nI've setup everything in the same [colab](https://colab.research.google.com/drive/1_2_mS5l29QHI1pXaqa4YLzAO5xm-HmH9?usp=sharing). Please have a look.", "Great, thanks. I'll have a play over the next few days and get back to you.\r\n\r\nSo just to confirm, where do the models come from? Is the base model T5 and it has been tuned using the SQUAD data reference questions, contexts and answers?", "@patil-suraj I've had a quick look, it's very impressive! I've only tried two passages, but the questions are sensible and the answers are more than one word which is better than the Text2Text pipeline. I definitely want to do some more work with this.\r\n\r\nFirst, I'll generate a large set of questions I can get human raters to score to validate whether the quantitative metrics (METEOR etc) correlate with human judgement.\r\n\r\nThen I would like to work on some architecture experiments to investigate the following:\r\n\r\n1. Can questions be generated where the corresponding generated answer is not explicitly stated in the text? This would require the reader to make connections between information in the text (implied information) and make inferences. I think this could be achieved with some NLTK work but is not ideal as I think the semantics of these questions are best learned from text data rather than expert-system rules.\r\n\r\n2. Can additional tuning of the que_gen model be done with other question/answer/context datasets that are of a different text type. For example, Wikipedia is primary factual information texts. But what about discursive texts? Or narrative texts?\r\n\r\nAnyway, I'll continue to explore what you have with a developer I work with and maybe we can form repository we can work in with the goal of creating a pipeline for inclusion in the Transformer package.\r\n\r\nWhat do you think?", "> Great, thanks. I'll have a play over the next few days and get back to you.\r\n> \r\n> So just to confirm, where do the models come from? Is the base model T5 and it has been tuned using the SQUAD data reference questions, contexts and answers?\r\n\r\nYes both of the models are t5-base trained on SQuAD", "> @patil-suraj I've had a quick look, it's very impressive! I've only tried two passages, but the questions are sensible and the answers are more than one word which is better than the Text2Text pipeline. I definitely want to do some more work with this.\r\n> \r\n> First, I'll generate a large set of questions I can get human raters to score to validate whether the quantitative metrics (METEOR etc) correlate with human judgement.\r\n> \r\n> Then I would like to work on some architecture experiments to investigate the following:\r\n> \r\n> 1. Can questions be generated where the corresponding generated answer is not explicitly stated in the text? This would require the reader to make connections between information in the text (implied information) and make inferences. I think this could be achieved with some NLTK work but is not ideal as I think the semantics of these questions are best learned from text data rather than expert-system rules.\r\n> 2. Can additional tuning of the que_gen model be done with other question/answer/context datasets that are of a different text type. For example, Wikipedia is primary factual information texts. But what about discursive texts? Or narrative texts?\r\n> \r\n> Anyway, I'll continue to explore what you have with a developer I work with and maybe we can form repository we can work in with the goal of creating a pipeline for inclusion in the Transformer package.\r\n> \r\n> What do you think?\r\n\r\n@danielduckworth \r\nI'm not sure about the first, we will need to run small experiments and see if it can be achieved.\r\n\r\n2) Yes, I do think additional fine-tuning on more diverse datasets should improve the results. My goal is to first get factual questions correct and then move to narrative texts.\r\n\r\nAnd sure we can create different repo and take this forward.", "@patil-suraj Thanks for the examples. Would you mind sharing your fine-tuning code used to train the model as well? ", "Sure. I'm planning to release model as well as the fine-tuning code. I'll comment here once I do that.", "@danai-antoniou Thanks for the wonderful suggestion. \r\n\r\n@patil-suraj, I also played your colab code, and it looks super great. I look forward to the release.", "@danielduckworth, I am looking into recent works in Question Generation especially using Transformer based architecture leveraging fine-tuning for small data sets. This thread is very interesting to me.\r\n\r\n> Can questions be generated where the corresponding generated answer is not explicitly stated in the text? This would require the reader to make connections between information in the text (implied information) and make inferences. I think this could be achieved with some NLTK work but is not ideal as I think the semantics of these questions are best learned from text data rather than expert-system rules.\r\n\r\nI have some previous experience in generating non-factoid questions from a text, with the goal of having descriptive answers rather than quiz-like. In my project, the data was not sufficient for DL models, even for fine-tuning.\r\n\r\nWe had done some user studies on generated questions as well, and found out METEOR is better correlated with human judgment on how reasonable or well-formed the questions are rather than BLEU or ROUGE scores.\r\n\r\nFor extracting relations in text, to capture more complex answers, Semantic Role Labeling (SRL) + Dependency parse tree of the text might be useful for extracting some descriptive answers. I used a tool called ClearNLP to do that.\r\n\r\n@patil-suraj, I also checked out your Collab demo and the questions generated are looking good, much better than other models I worked with so far. Great job. Definitely looking forward to the release of the models and knowing more about the fine-tuning process.\r\n\r\nFor non-factoid questions, there is room for improvement.", "\r\nHi @emadg, @hunkim Thank you for your interest! :)\r\n\r\n@emadg\r\n\r\n>For extracting relations in text, to capture more complex answers, Semantic Role Labeling (SRL) + Dependency parse tree of the text might be useful for extracting some descriptive answers. I used a tool called ClearNLP to do that.\r\n \r\nThis sure sounds like a good idea. My current goal is to have a end-to-end model for generating questions and answers.", "> For non-factoid questions, there is room for improvement.\r\n\r\nFor those looking for less factual questions, I was actually able to get some reasonable results with a T5 pre-trained for query prediction, but there's definitely room for improvement. \r\n\r\nhttps://github.com/castorini/docTTTTTquery", "Hey people, I've setup few experiments for question generation. Let me know if anyone wants to collaborate on this, I would really appreciate some help and maybe some multi-gpu compute. Everything will be open sourced after the experiments are finished.\r\nThank you! ", "This is very interesting to me. I'm writing a master's thesis over the summer, working on transfer learning for question generation. I don't have much experience with contributing to large pre-existing frameworks like this but would definitely be happy to contribute wherever I can. \r\n\r\n@patil-suraj \r\nI had a look at your notebook. This is very impressive! Looking forward to seeing the fine-tuning process to get an idea of how this can be done using the transformer framework. ", "@patil-suraj \r\nI would like to help. Let me know if we can collaborate. Although I have somewhat limited time to contribute.", "Hi @emadg and @vegarab, thank you for your interest,\r\nMy goal is to do open source study on que generation. Here's what I have planned \r\nFor ans aware que generation we usually need 3 models \r\nfirst which will extract ans like spans\r\nsecond model will generate question on that answer\r\nand third will be a QA model which will take the question and produce an answer, \r\nthen we can compare the two answers to see if the generated question is correct or not.\r\n\r\nHaving 3 models for single task is lot of complexity, so goal is to create a multi-task model which can do all of these 3 tasks\r\n1. extract ans like spans\r\n2. generate question based on the answer \r\n3. QA\r\n\r\nAlso I want to see if we can generate multiple questions end-to-end without answers.\r\n\r\nAnother experiment is generating non-factoid questions. First we need to find a right dataset for this.\r\n\r\nI've trained t5-small model in multi-task way and its giving really good results, so now I want to train more models (t5-base, t5-large, bart-base, bart-large, bert-2-bert) and see if they improve the results.\r\n\r\nI've also trained t5-small and t5-base for end-2-end QG and that too is giving interesting results. \r\n\r\nSo regarding help I'm looking for some compute to train large models, multitask t5-small took 10hrs on single V100 GPU. I also want someone to provide rigorous feedback on the work (find out mistakes, asses quality of questions etc) and help with creating a write-up for study.", "Hi all! @patil-suraj really great work! I am using your approach and seems to be working very well.\r\n\r\nI have one question though, when fine-tuning the models, did you use a case or an uncased model? Because when giving too much uppercase text as context, it's generating questions partially or totally in uppercase and are much worse. It seems to be a cased model because when lowercasing the text, the questions are much better.\r\n\r\nThanks in advance!", "@patil-suraj,\r\n\r\n> and third will be a QA model which will take the question and produce an answer\r\n\r\nI think using a QA model to evaluate the generated Question is an interesting approach. Why should do this instead of evaluating with BLEU or METEOR score?\r\n\r\n> So regarding help I'm looking for some compute to train large models, multitask t5-small took 10hrs on single V100 GPU.\r\n\r\nCan we use GCP for the compute? if it is not going to cost a lot. I don't have GPUs myself.\r\n\r\n> I also want someone to provide rigorous feedback on the work (find out mistakes, asses quality of questions etc) and help with creating a write-up for study.\r\n\r\nI think I can spend some time and provide feedback about the work. Although, I need to catch up with the details related to T5 model and the fine-tuning method used here.", "Interesting thread @patil-suraj I had the similar thoughts on multi task training \r\n\r\n- Finetune the model combining the data for both question generation & answering(one example is **context:c1 answer: a1 ---> question : q1** & another example context:c1 question : q1 ----> answer:a1)\r\n- Way to generate multiple questions is either using topk and topp sampling or using multiple beams.\r\n\r\n\r\n", "Hey everyone,\r\nhere's a sneak peek of whats coming, everything will be available by the end of this week. stay tuned !\r\n", "@santhoshkolloju \r\n> * Way to generate multiple questions is either using topk and topp sampling or using multiple beams.\r\n\r\nYes, this is what I have tried in another model.\r\n\r\n\r\nHi @emadg \r\n\r\n> Why should do this instead of evaluating with BLEU or METEOR score?\r\n\r\nBLEU and METEOR can be used to evaluate the model when you have the original reference questions, but at inference time how can we decide if the generated question is correct (makes sense or not, has answer or not) or not without original question ? Which is why the QA model.\r\n\r\n>I think I can spend some time and provide feedback about the work. Although, I need to catch up with the details related to T5 model and the fine-tuning method used here.\r\n\r\nYou can start your analysis once I make it available. Human feedbacks will be most valuable. Thanks ! \r\n\r\n", "Hi all, tagging everyone for notification \r\n@emadg , @vegarab , @gabisurita , @hunkim , @ugmSorcero , @santhoshkolloju .\r\n\r\nHappy to finally release the project. You can find everything in [this repo](https://github.com/patil-suraj/question_generation).\r\n\r\n[All models](https://huggingface.co/models?filter=question-generation) are available on hub with configured inference API. You can search using question-generation tag.\r\n\r\nHere’s a [colab](https://colab.research.google.com/github/patil-suraj/question_generation/blob/master/question_generation.ipynb) if anyone wants to play more with it.", "@patil-suraj Thanks!!" ]

[]

[]

[ "Good spot! I think @julien-c's latest PR (https://github.com/huggingface/transformers/pull/4243) for distributed eval will also take care of this! " ]

[]

[]

[ "I succeeded to do it somehow, but i'm sure it's not the way it should work, and it won't scale well for large datasets. I would be happy to know if there is a better way. \r\n\r\nWhat worked: \r\n1. squad_convert_examples_to_features ( return_dataset = False) - getting the features\r\n2. Creating a dictionary of features and labels, where each item is list of tensorflow vectors obtained by `tf.convert_to_tensor`\r\n3. Constructing the dataset with `tfdataset = tf.data.Dataset.from_tensor_slices((features, labels)).batch(8)`\r\n4. Training with `fit_generator` method (`fit` fails)\r\n\r\nFull code: \r\n```python\r\n tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n\r\n processor = SquadV1Processor()\r\n # processor = SquadV2Processor()\r\n examples = processor.get_train_examples(args.data_dir, filename=args.train_file)\r\n train_dataset = squad_convert_examples_to_features(\r\n examples=examples,\r\n tokenizer=tokenizer,\r\n max_seq_length=args.max_seq_length,\r\n doc_stride=args.doc_stride,\r\n max_query_length=args.max_query_length,\r\n is_training=True\r\n )\r\n\r\n def create_features_and_labels_tf_tensors_from_dataset(train_dataset):\r\n all_input_ids = []\r\n all_token_type_ids = []\r\n all_attention_mask = []\r\n all_start_pos = []\r\n all_end_pos = []\r\n ex: SquadFeatures\r\n for ex in train_dataset:\r\n all_input_ids.append(ex.input_ids)\r\n all_token_type_ids.append(ex.token_type_ids)\r\n all_attention_mask.append(ex.attention_mask)\r\n all_start_pos.append(ex.start_position)\r\n all_end_pos.append(ex.end_position)\r\n all_input_ids_tensor = tf.convert_to_tensor(all_input_ids)\r\n all_token_type_ids_tensor = tf.convert_to_tensor(all_token_type_ids)\r\n all_attention_mask_tensor = tf.convert_to_tensor(all_attention_mask)\r\n all_start_pos_tensor = tf.convert_to_tensor(all_start_pos)\r\n all_end_pos_tensor = tf.convert_to_tensor(all_end_pos)\r\n features = {'input_ids': all_input_ids_tensor, 'token_type_ids': all_token_type_ids_tensor,\r\n 'attention_mask': all_attention_mask_tensor}\r\n labels = {\"output_1\": all_start_pos_tensor, 'output_2': all_end_pos_tensor}\r\n return features, labels\r\n\r\n features, labels = create_features_and_labels_tf_tensors_from_dataset(train_dataset)\r\n tfdataset = tf.data.Dataset.from_tensor_slices((features, labels)).batch(8)\r\n\r\n model = TFBertForQuestionAnswering.from_pretrained(\"bert-base-cased\")\r\n loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE, from_logits=True)\r\n opt = tf.keras.optimizers.Adam(learning_rate=3e-5)\r\n\r\n model.compile(optimizer=opt,\r\n loss={'output_1': loss_fn, 'output_2': loss_fn},\r\n loss_weights={'output_1': 1., 'output_2': 1.},\r\n metrics=['accuracy'])\r\n\r\n # Now let's train our model\r\n try:\r\n history = model.fit(tfdataset, epochs=1, steps_per_epoch=3)\r\n print(f'Success with fit')\r\n except Exception as ex:\r\n traceback.print_exc()\r\n print(f\"Failed using fit, {ex}\")\r\n history = model.fit_generator(tfdataset, epochs=1, steps_per_epoch=3)\r\n print(f'Success with fit_generator')\r\n print(\"Done\")\r\n```\r\nError message for `fit`:\r\n ```python\r\nFile \"/home/ec2-user/yonatab/ZeroShot/transformers_experiments/src/minimal_example_for_git.py\", line 73, in main\r\n history = model.fit(tfdataset, epochs=1, steps_per_epoch=3)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 819, in fit\r\n use_multiprocessing=use_multiprocessing)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_v2.py\", line 235, in fit\r\n use_multiprocessing=use_multiprocessing)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_v2.py\", line 593, in _process_training_inputs\r\n use_multiprocessing=use_multiprocessing)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_v2.py\", line 706, in _process_inputs\r\n use_multiprocessing=use_multiprocessing)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/data_adapter.py\", line 702, in __init__\r\n x = standardize_function(x)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_v2.py\", line 660, in standardize_function\r\n standardize(dataset, extract_tensors_from_dataset=False)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 2360, in _standardize_user_data\r\n self._compile_from_inputs(all_inputs, y_input, x, y)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 2580, in _compile_from_inputs\r\n target, self.outputs)\r\n File \"/home/ec2-user/anaconda3/envs/yonatan_env_tf2/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_utils.py\", line 1341, in cast_if_floating_dtype_and_mismatch\r\n if target.dtype != out.dtype:\r\nAttributeError: 'str' object has no attribute 'dtype'\r\nFailed using fit, 'str' object has no attribute 'dtype'\r\nWARNING:tensorflow:From /home/ec2-user/yonatab/ZeroShot/transformers_experiments/src/minimal_example_for_git.py:78: Model.fit_generator (from tensorflow.python.keras.engine.training) is deprecated and will be removed in a future version.\r\nInstructions for updating:\r\nPlease use Model.fit, which supports generators.\r\n```\r\n\r\nIt also fails when trying to add `validation_data` to the `fit` function", "I think it's a bug, i'm closing & opening another bug issue. " ]

[]

[]

[ "Never mind, my mistake", "What was your mistake?\r\n\r\nIs your system Windows?\r\nI am trying to reproduce the colab tutorial from https://blogs.helsinki.fi/language-technology/2020/05/14/helsinkinlp-in-huggingface/ on Windows but I get errors.\r\n", "@R4ZZ3 Hello! I just appended `>>tag<<` for text in cycle, and my text quickly became a mess of tags. And no Windows, sorry, can't help.", "Can anyone help with this issue: #5040 ?" ]

[ { "id": 1834052129, "node_id": "MDU6TGFiZWwxODM0MDUyMTI5", "url": "https://api.github.com/repos/huggingface/transformers/labels/High-Level%20feature", "name": "High-Level feature", "color": "f7c9a3", "default": false, "description": "" }, { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "That is available on this page - https://marian-nmt.github.io/faq:\r\n\r\nCan I generate n-best lists?\r\n Yes. Just use --n-best and the set --beam-size 6 for an n-best list size of 6.\r\n\r\nI do not know how to apply it here.", "That feature is not supported in our implementation, unfortunately.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "You can try passing `num_return_sequences=5` to generate, but that behavior is untested, and I have never used it.", "> You can try passing `num_return_sequences=5` to generate, but that behavior is untested, and I have never used it.\r\n\r\nHowever, do not forget to set the beams accordingly. The beams must be defined at least as large as the desired number of alternative results." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "Hi,\r\nCan anyone from Huggingface team look into this?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Similar to #6025 ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "Nope. Often the reverse model exists, however, e.g. `en-fr` and `fr-en`.", "> Nope. Often the reverse model exists, however, e.g. `en-fr` and `fr-en`.\r\n\r\nWhat if reverse model does not exist for my langauge? ", "What's your language? \r\nMaybe it's part of a multilingual group (below) or has been [trained](http://opus.nlpl.eu/Opus-MT/) since we did the conversion?\r\nOtherwise, you're out of luck unfortunately.\r\n\r\n```\r\nGROUP_MEMBERS = {\r\n 'ZH': ['cmn', 'cn', 'yue', 'ze_zh', 'zh_cn', 'zh_CN', 'zh_HK', 'zh_tw', 'zh_TW', 'zh_yue', 'zhs', 'zht', 'zh'],\r\n 'ROMANCE': ['fr', 'fr_BE', 'fr_CA', 'fr_FR', 'wa', 'frp', 'oc', 'ca', 'rm', 'lld', 'fur', 'lij', 'lmo', 'es', 'es_AR', 'es_CL', 'es_CO', 'es_CR', 'es_DO', 'es_EC', 'es_ES', 'es_GT', 'es_HN', 'es_MX', 'es_NI', 'es_PA', 'es_PE', 'es_PR', 'es_SV', 'es_UY', 'es_VE', 'pt', 'pt_br', 'pt_BR', 'pt_PT', 'gl', 'lad', 'an', 'mwl', 'it', 'it_IT', 'co', 'nap', 'scn', 'vec', 'sc', 'ro', 'la'],\r\n 'NORTH_EU': ['de', 'nl', 'fy', 'af', 'da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],\r\n 'SCANDINAVIA': ['da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],\r\n 'SAMI': ['se', 'sma', 'smj', 'smn', 'sms'],\r\n 'NORWAY': ['nb_NO', 'nb', 'nn_NO', 'nn', 'nog', 'no_nb', 'no'],\r\n 'CELTIC': ['ga', 'cy', 'br', 'gd', 'kw', 'gv']\r\n}\r\n```", "> What's your language?\r\n> Maybe it's part of a multilingual group (below) or has been [trained](http://opus.nlpl.eu/Opus-MT/) since we did the conversion?\r\n> Otherwise, you're out of luck unfortunately.\r\n> \r\n> ```\r\n> GROUP_MEMBERS = {\r\n> 'ZH': ['cmn', 'cn', 'yue', 'ze_zh', 'zh_cn', 'zh_CN', 'zh_HK', 'zh_tw', 'zh_TW', 'zh_yue', 'zhs', 'zht', 'zh'],\r\n> 'ROMANCE': ['fr', 'fr_BE', 'fr_CA', 'fr_FR', 'wa', 'frp', 'oc', 'ca', 'rm', 'lld', 'fur', 'lij', 'lmo', 'es', 'es_AR', 'es_CL', 'es_CO', 'es_CR', 'es_DO', 'es_EC', 'es_ES', 'es_GT', 'es_HN', 'es_MX', 'es_NI', 'es_PA', 'es_PE', 'es_PR', 'es_SV', 'es_UY', 'es_VE', 'pt', 'pt_br', 'pt_BR', 'pt_PT', 'gl', 'lad', 'an', 'mwl', 'it', 'it_IT', 'co', 'nap', 'scn', 'vec', 'sc', 'ro', 'la'],\r\n> 'NORTH_EU': ['de', 'nl', 'fy', 'af', 'da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],\r\n> 'SCANDINAVIA': ['da', 'fo', 'is', 'no', 'nb', 'nn', 'sv'],\r\n> 'SAMI': ['se', 'sma', 'smj', 'smn', 'sms'],\r\n> 'NORWAY': ['nb_NO', 'nb', 'nn_NO', 'nn', 'nog', 'no_nb', 'no'],\r\n> 'CELTIC': ['ga', 'cy', 'br', 'gd', 'kw', 'gv']\r\n> }\r\n> ```\r\n\r\nSwahili, oooh!" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1834056635, "node_id": "MDU6TGFiZWwxODM0MDU2NjM1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Core:%20Tokenization", "name": "Core: Tokenization", "color": "FF4446", "default": false, "description": "Internals of the library; Tokenization." } ]

[]

[ "@mfuntowicz @LysandreJik might understand the best!", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[ { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=h1) Report\n> Merging [#4389](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/48c3a70b4eaedab1dd9ad49990cfaa4d6cb8f6a0&el=desc) will **decrease** coverage by `0.37%`.\n> The diff coverage is `96.22%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4389 +/- ##\n==========================================\n- Coverage 78.41% 78.03% -0.38% \n==========================================\n Files 123 123 \n Lines 20432 20477 +45 \n==========================================\n- Hits 16021 15980 -41 \n- Misses 4411 4497 +86 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_marian.py](https://codecov.io/gh/huggingface/transformers/pull/4389/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fbWFyaWFuLnB5) | `93.10% <96.22%> (+5.77%)` | :arrow_up: |\n| [src/transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/4389/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3F1YWQucHk=) | `28.66% <0.00%> (-28.03%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4389/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.85% <0.00%> (+0.41%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=footer). Last update [48c3a70...1bd2df5](https://codecov.io/gh/huggingface/transformers/pull/4389?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Can anyone help with this issue: #5040 ?" ]

[ { "id": 1834067346, "node_id": "MDU6TGFiZWwxODM0MDY3MzQ2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Documentation", "name": "Documentation", "color": "77cc3b", "default": false, "description": "" } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "unstale?", "This was fixed with https://github.com/huggingface/transformers/pull/5665!" ]

[ { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "this works for me:\r\n```python\r\nmodelname='Helsinki-NLP/opus-mt-en-de'\r\nmodel = MarianMTModel.from_pretrained(modelname,output_attentions=True, output_hidden_states=True)\r\noutput_tuple = model(**model.dummy_inputs)\r\nassert len(output_tuple) == 6\r\n```", "#4388 is the real issue." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "The (new) run_glue.py is just a short template of code that you can copy/paste/customize (e.g. to plug your own Dataset), thanks to the [Trainer](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py).", "Thanks! which class would replace `GlueDataset` in that case? and what argument should I pass instead of `task_name`? and how to set the number of labels, which in `run_glue.py` is defined by `glue_tasks_num_labels`? ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Are you restarting from your checkpoint with the exact same set of script parameters?", "@julien-c Actually, no. And I guess I undestood what the reason of that behavior is. Originally I used to train with number of epochs = 5. After that I selected 10 epochs and hoped that it'll continue pretraining from 5'th epoch up to 10'th. However as far as I understand lr-value is calculated based on the number of epochs. The current value of lr is the same as I got at 15k iteration (30k/2). So I have should selected 10 (or whatever number) epochs originally", "Yes. You can always implement this behaviour yourself if you need it." ]

[]

[]

[ "The Albert original model has a SentencePiece tokenizer, so the `AlbertTokenizer` can only handle SentencePiece vocabularies right now. The `tokenizers` library doesn't handle these vocabularies as of now.\r\n\r\nWe're still thinking of how we should proceed when training a model from scratch with a tokenizer different to the original, as right now the model-tokenizer pairs have a fixed tokenizer backend. cc @julien-c @thomwolf ", "@LysandreJik But when I added spiece.model it started training. So is it taking vocab created by tokenizer library along with spiece.model?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Does this issue have been solved? When can we train the lm from scratch?", "You can already train an LM from scratch, there are examples in `examples/language-modeling`.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Are there any updates on training Albert from scratch? Perhaps via the Trainer?", "Hi, a `DataCollatorForSOP` has been added to the library, which you can indeed use with the `Trainer` object. You can see it's implementation here:\r\n\r\nhttps://github.com/huggingface/transformers/blob/6303b5a7185fba43830db0cbb06c61861f57ddff/src/transformers/data/data_collator.py#L201-L206", "> Hi, a `DataCollatorForSOP` has been added to the library, which you can indeed use with the `Trainer` object. You can see it's implementation here:\r\n> \r\n> https://github.com/huggingface/transformers/blob/6303b5a7185fba43830db0cbb06c61861f57ddff/src/transformers/data/data_collator.py#L201-L206\r\n\r\nWhen i try to use this `DataCollatorForSOP` it says that it is deprecated and that i should use `DataCollatorForLanguageModeling`. It is however not clear whether this data collator also produces the objective sentence order prediction as in the original ALBERT paper. Can anyone verify whether it does this or not?" ]

[]

[]

[ "Can you upgrade your version of transformers?", "Same issue I am not able to run the command `transformer-cli env` even after installing from source @julien-c .\r\n\r\n`transformers-cli is not recognized as an internal or external command operable program or batch file` \r\n\r\nJust wondering if this is due to the fact of using `conda` as in latest versions it is not preferring to modify the PYTHONPATH. hence the non visibility of the command?\r\n\r\n", "Upgrading the transformers version to 2.9.1 works!" ]

[]

[]

[ "Indeed, you're totally correct. Thank you for reporting the issue, I'm fixing it in #4385." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Hi @W4ngatang did you see that we have a new `Trainer` class in the library? It's basically a refactor of the previous version of `run_glue.py` so should be pretty easy to adopt.\r\n\r\nLet us know if we can help, and very excited to have an easy path to SuperGLUE tasks in the repo :)", "Hi Julien,\r\n\r\nSaw it! I started writing this a while ago, but you all move very fast. We're preparing this as some reference code for a competition, so we may keep the training code as is for a bit and then switch to the trainer when we're ready to actually merge into the main repo.", "@thomwolf ", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=h1) Report\n> Merging [#4378](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/90f4b2452077ac3bac9453bdc63e0359aa4fe4d2&el=desc) will **decrease** coverage by `1.47%`.\n> The diff coverage is `22.83%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4378 +/- ##\n==========================================\n- Coverage 78.00% 76.52% -1.48% \n==========================================\n Files 138 139 +1 \n Lines 23766 24408 +642 \n==========================================\n+ Hits 18539 18679 +140 \n- Misses 5227 5729 +502 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `83.36% <12.12%> (-4.14%)` | :arrow_down: |\n| [src/transformers/data/metrics/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL21ldHJpY3MvX19pbml0X18ucHk=) | `16.66% <12.35%> (-10.00%)` | :arrow_down: |\n| [src/transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19yb2JlcnRhLnB5) | `83.80% <19.44%> (-10.99%)` | :arrow_down: |\n| [src/transformers/data/processors/superglue.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3VwZXJnbHVlLnB5) | `21.73% <21.73%> (ø)` | |\n| [src/transformers/data/processors/utils.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvdXRpbHMucHk=) | `35.59% <84.61%> (+7.96%)` | :arrow_up: |\n| [src/transformers/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9fX2luaXRfXy5weQ==) | `99.21% <100.00%> (ø)` | |\n| [src/transformers/data/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL19faW5pdF9fLnB5) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/data/processors/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvX19pbml0X18ucHk=) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/data/processors/glue.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvZ2x1ZS5weQ==) | `49.44% <100.00%> (ø)` | |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `79.51% <0.00%> (-1.39%)` | :arrow_down: |\n| ... and [2 more](https://codecov.io/gh/huggingface/transformers/pull/4378/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=footer). Last update [90f4b24...576ecaf](https://codecov.io/gh/huggingface/transformers/pull/4378?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Looks super-cool, especially with SustaiNLP just around the corner! Is there any help that could be given that would accelerate this PR?", "hi @W4ngatang @thomwolf \r\n\r\nany help you need for this PR? I think pushing this PR through can help many researchers with their experiments. ", "Hi Tianyi,\n\nYes, help would be great! The code is correct from my end, but uses an old\nversion of Transformers. At this point, it's just a matter of merging the\nmost recent version of Transformers in and adhering to the Transformers\ncontributor guidelines.\n\nBest,\nAlex\n\nOn Thu, Oct 8, 2020 at 5:17 PM Tianyi <[email protected]> wrote:\n\n> hi @W4ngatang <https://github.com/W4ngatang> @thomwolf\n> <https://github.com/thomwolf>\n>\n> any help you need for this PR? I think pushing this PR through can help\n> many researchers with their experiments.\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/pull/4378#issuecomment-705828560>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABKDWG7FSDKTCVL4MI6TXEDSJYT7DANCNFSM4NBH5EJA>\n> .\n>\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "I am also facing the same problem", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Doesn't work right now as links are relative links. We should update them to global links before, I can take care of it." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "You may also need to rebase to fix `check_code_quality`.\r\nAlso `run_examples_torch` will only pass if it doesn't add new dependencies.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[ { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[]

[ "Also, how can I transfer to cuda `BatchEncoding` object which return after call `prepare_translation_batch` method on `MarianTokenizer`", "This could be fixed by specifiying a `save_vocabulary` function, like it is used in:\r\n\r\nhttps://github.com/huggingface/transformers/blob/7defc6670fa76e857109e1b99f3e919da8d11f42/src/transformers/tokenization_xlm_roberta.py#L292-L311\r\n\r\n(both source and target vocab needs to be written),\r\n\r\nAdditionally, it would be awesome if the `MarianTokenizer` class would be `pickle`-able using `__getstate__` and `__setstate__` methods :)\r\n\r\n/cc @sshleifer ", "@NonaryR `BatchEncoding.to(device:str)` should work.\r\nWill fix vocab saving, great catch!\r\n", "Thank you, `.to(device)` helped!\r\nI can close an issue if you want it. \r\nAlso, many thanks for this new release with all these new awesome translation models, they will bring significant changes to the community!!", "But how can you load the saved MarianTokenizer from your local directory? \r\nCan anyone give a complete example about how to save and load MarianTokenizer?", "Can anyone help with this issue: #5040 ?" ]

[]

[]

[ "Hi, this would be something you would need to build yourself.\r\n\r\nAlternatively you could use our Layers as Keras Layers?" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1565794707, "node_id": "MDU6TGFiZWwxNTY1Nzk0NzA3", "url": "https://api.github.com/repos/huggingface/transformers/labels/Write%20With%20Transformer", "name": "Write With Transformer", "color": "a84bf4", "default": false, "description": "" } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi @rugk, open to any suggestion for a better documentation link – e.g. should we link to https://huggingface.co/blog/how-to-generate ?\r\n\r\nMaybe https://medium.com/huggingface/how-to-write-with-transformer-5ee58d6f51fa ?\r\n\r\nOr to https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313#79c5 like on https://convai.huggingface.co/\r\n\r\ncc @patrickvonplaten @sgugger ", "I dunno. My complaint was just an UX complaint from a user with less scientific background, so linking to very in-deep document is likely bad.\r\n\r\nHow you solve this hmm?\r\nMaybe just write a new doc/small wiki entry here on GitHub or so?\r\n\r\nThough if I check your links the \"Advanced settings\" section in https://medium.com/huggingface/how-to-write-with-transformer-5ee58d6f51fa It does what I want: explains the parameter.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "shh #badbot :no_bell::robot::no_bell: This is still a TODO.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "Hi @Damiox,\r\n\r\nI think I see the problem. That's quite an edge case here. Since `2.8.0` (here the PR: https://github.com/huggingface/transformers/pull/3734), the `input_ids` will always be cut to just the last token when using `past` to have a consistent API with other models and allow to keep putting in the whole `input_ids` when speeding up decoding. Also `past` was never supposed to be used together with `input_ids` longer than 1, so I did not think of this edge case :-/. I'm quite surprised that this even worked before, but I'm also not sure whether the `causal mask` in GPT2 was correct for this case before.\r\n\r\nI think a simple fix in your case would be to generate the output incrementally and only use `input_ids` of length 1 or not use the `past` variable at all here.\r\n\r\n@LysandreJik, looking at this script, I think I did introduce a breaking change in the PR: https://github.com/huggingface/transformers/pull/3734 for this edge case when:\r\n`past` is given to the model **and** `input_ids` is longer than 1. It's quite an edge case here and I'm not sure whether it's worth reversing the PR. What do you think? \r\n\r\n\r\n", "Hi @patrickvonplaten - I'm a bit worried about this... I cannot apply the simple fix because I'm unable to get rid of `past` variable or run more inferences at this point because the performance would be affected for my use case which is over a realtime bigdata dataset. The current results are accurate and are based on discussions in https://github.com/huggingface/transformers/issues/3095 - this is already deployed.\r\n\r\nJust to be clear... Will this not be fixed?", "That's an interesting use-case indeed. I don't believe we ever intended the `past` to be used with `input_ids` larger than one, and I'm quite surprised as well that it did work.\r\n\r\n@thomwolf, do you have any opinion on this?", "I'm really looking forward to knowing if this can be fixed so to continue working as it was working up to v2.8.0. Any new thoughts on this? Thanks", "Hi @Damiox,\r\n\r\nSorry to answer so late. We had some discussion internally about it. And you are correct, we should revert the PR I merged earlier. This is done in #4581 and should be included in the next version. @thomwolf @LysandreJik " ]

[]

[]

[]

[]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "These are numbers that I am getting when running the examples (in the article or here: https://github.com/nlpyang/PreSumm) they mention much higher numbers: \r\n\r\n****** ROUGE SCORES ******\r\n\r\n** ROUGE 1\r\nF1 >> 0.195\r\nPrecision >> 0.209\r\nRecall >> 0.186\r\n\r\n** ROUGE 2\r\nF1 >> 0.094\r\nPrecision >> 0.104\r\nRecall >> 0.089\r\n\r\n** ROUGE L\r\nF1 >> 0.221\r\nPrecision >> 0.233\r\nRecall >> 0.212", "same here. Also getting ROUGE scores in the same ball-park", "Well, I already switched to Bart model. I do not know what is the reasons for low numbers here, possibly some bugs in the porting, but in general BertAbs model does not seem the best approach for it.", "Alex, by the way I see that you are currently working on Summarization topic too. If you like you can write me directly to [email protected] (I am currently conversational AI researcher at Nvidia) and we can have more detailed discussion on this topic.", "Will do, thanks!", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Thanks for catching this, @patil-suraj!" ]

[]

[]

[ "By the way we're hoping to soon have TPU-specific CI in place, cc @LysandreJik " ]

[]

[]

[]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "Thanks for the issue dude! Feel free to open a PR if you have a branch that fixed that lying around haha ;-) " ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Added language metadata for the model to appear on https://huggingface.co/models?filter=turkish" ]

[]

[]

[ "I meeting the same issue with a custom training script (built using GLUE script as model), using ELECTRA for classification task.", "Hi! This should have been solved by https://github.com/huggingface/transformers/pull/4450.\r\nCould you install from source and let me know if it fixes your issue?", "The code is running, but I receive this warning :\r\n\r\n```\r\nUserWarning: This overload of addcdiv_ is deprecated:\r\n\taddcdiv_(Number value, Tensor tensor1, Tensor tensor2)\r\nConsider using one of the following signatures instead:\r\n\taddcdiv_(Tensor tensor1, Tensor tensor2, *, Number value) (Triggered internally at /pytorch/torch/csrc/utils/python_arg_parser.cpp:760.)\r\n p.data.addcdiv_(-step_size, exp_avg, denom)\r\n```\r\n\r\nand it's **very slow** (2h for 1 epoch). \r\n\r\nOn a K80 GPU, it takes 30 minutes...\r\n\r\n---\r\n\r\nRunning on 8 cores is not working but I think this is due to the RAM limitations.", "Can confirm that the training loss is now declining as expected in colab notebook and that one epoch takes around 2h per epoch with 1 TPU-core. I also tested with a K80 but unlike for @Colanim it took around 5h per epoch, which I find reasonable in the sense that the TPU is faster. The batch size was the default Trainer value for both tests (which is 5 I believe).\r\n\r\nI get @Colanim's UserWarning regardless of using GPU or TPU.\r\n\r\nSo imo the bug is fixed!" ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Great! I suspect Exbert won't work out of the box but we are looking into enabling it automatically" ]

[]

[ { "login": "LysandreJik", "id": 30755778, "node_id": "MDQ6VXNlcjMwNzU1Nzc4", "avatar_url": "https://avatars.githubusercontent.com/u/30755778?v=4", "gravatar_id": "", "url": "https://api.github.com/users/LysandreJik", "html_url": "https://github.com/LysandreJik", "followers_url": "https://api.github.com/users/LysandreJik/followers", "following_url": "https://api.github.com/users/LysandreJik/following{/other_user}", "gists_url": "https://api.github.com/users/LysandreJik/gists{/gist_id}", "starred_url": "https://api.github.com/users/LysandreJik/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/LysandreJik/subscriptions", "organizations_url": "https://api.github.com/users/LysandreJik/orgs", "repos_url": "https://api.github.com/users/LysandreJik/repos", "events_url": "https://api.github.com/users/LysandreJik/events{/privacy}", "received_events_url": "https://api.github.com/users/LysandreJik/received_events", "type": "User", "site_admin": false } ]

[ "Do you mind giving us a reproducible code example so that we may check on our end?", "I have the same issue, my trained models work on 2.8.0 but I get the same error when loading saved model on 2.9.1", "Hi, sorry for not getting back to this @LysandreJik .\r\n\r\nIt seems that in Transformers 2.8.0 and below, there is no parameter called `masked_bias`.\r\n\r\nHowever in 2.9+, there is now a parameter called `masked_bias` on line 109 in `modeling_gpt2.py`: https://github.com/huggingface/transformers/blob/v2.9.0/src/transformers/modeling_gpt2.py#L109\r\n\r\nThus, the error above is seen.\r\n\r\nCould anything be done about this? Otherwise it renders all GPT2 checkpoints before 2.8.0 unusable, unless manually tinkering with the checkpoint to add this parameter?", "Hi! I'm trying to reproduce, but I don't get an error like you do. These values indeed do not get loaded, but this does not raise a `RuntimeError`.\r\n\r\nDo you mind giving me a short code sample so that I may debug on my side? Thanks a lot.", "The weights are all proprietary unfortunately but let me try and do something.", "@LysandreJik \r\n\r\nFirst, `pip install transformers==2.8.0`\r\n\r\nThen do:\r\n```\r\nimport torch\r\nfrom transformers import GPT2Config, GPT2Model\r\ntorch.save(GPT2Model(GPT2Config(n_layer=2)), \"gpt2_state_dict.pt\")\r\n```\r\n\r\nThen, `pip install transformers==2.9.0`\r\nand do\r\n```\r\nimport torch\r\nfrom transformers import GPT2Config, GPT2Model\r\ngpt2 = GPT2Model(GPT2Config(n_layer=2))\r\nstate_dict = torch.load(\"gpt2_state_dict.pt\")\r\ngpt2.load_state_dict(state_dict)\r\n```\r\n\r\nAnd you get an error: \r\n```\r\nRuntimeError: Error(s) in loading state_dict for GPT2Model:\r\n\tMissing key(s) in state_dict: \"h.0.attn.masked_bias\", \"h.1.attn.masked_bias\".\r\n```", "I see! Indeed, there's not much we can do about that, unfortunately. We have our own loading/saving method for that purpose.\r\n\r\nOut of curiosity, why don't you use `from_pretrained`/`save_pretrained`?", "I'm wrapping the model in a Pytorch Lightning module, and the most convenient way to save that is to save the state dictionary.", "In that case, you could have a workaround like:\r\n\r\n```py\r\nimport torch\r\nfrom transformers import GPT2Config, GPT2Model\r\n\r\n# Use the from_pretrained method\r\ntemp = GPT2Model.from_pretrained(directory)\r\ntemp.save_pretrained(directory)\r\n\r\ngpt2 = GPT2Model(GPT2Config(n_layer=2))\r\nstate_dict = torch.load(f\"{directory}/pytorch_model.pt\")\r\ngpt2.load_state_dict(state_dict)\r\n```\r\n\r\nLoading/saving using `from_pretrained` shouldn't raise an error, and once it's saved you can then load the state dict like you did. Once you have saved it using this, you can always just load the state dict.\r\n", "@LysandreJik yes, that seems like a good workaround. Thanks!", "Glad I could help, sorry for taking so long to get back to you. Feel free to reopen if you face any other issues down the road." ]

[]

[]

[ "It should be automatic.\r\n\r\nWhat does torch.cuda.is_available say?", "torch.cuda.is_available return False.\r\nThis is what my setup is by running nvidia-smi\r\n```\r\nThu May 14 15:04:38 2020\r\n+-----------------------------------------------------------------------------+\r\n| NVIDIA-SMI 431.87 Driver Version: 431.87 CUDA Version: 10.1 |\r\n|-------------------------------+----------------------+----------------------+\r\n| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |\r\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\r\n|===============================+======================+======================|\r\n| 0 GeForce GTX 166... WDDM | 00000000:01:00.0 Off | N/A |\r\n| N/A 57C P8 3W / N/A | 153MiB / 6144MiB | 0% Default |\r\n+-------------------------------+----------------------+----------------------+\r\n\r\n+-----------------------------------------------------------------------------+\r\n| Processes: GPU Memory |\r\n| GPU PID Type Process name Usage |\r\n|=============================================================================|\r\n| No running processes found |\r\n+-----------------------------------------------------------------------------+\r\n```", "Yes just checked. torch is not able to find cuda", "You should upgrade your nvidia driver.\r\n\r\nYou probably installed pytorch 1.5.0 bumped up the required version of nvidia-driver.", "yes.\r\n`pip install torch==1.4.0`\r\nwill solve the issue." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false } ]

[ "Awesome this looks good to me! \r\n\r\n@LysandreJik - I tried this fix and it does fix the issue https://github.com/huggingface/transformers/issues/4021 . To me it seems like this line was just forgotten when implementing T5 tokenizer.", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=h1) Report\n> Merging [#4353](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/94cb73c2d2efeb188b522ff352f98b15124ba9f8&el=desc) will **increase** coverage by `0.00%`.\n> The diff coverage is `100.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4353 +/- ##\n=======================================\n Coverage 78.21% 78.21% \n=======================================\n Files 120 120 \n Lines 20038 20039 +1 \n=======================================\n+ Hits 15673 15674 +1 \n Misses 4365 4365 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4353/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `89.49% <100.00%> (+0.01%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4353/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.77% <0.00%> (-0.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4353/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `91.08% <0.00%> (+0.12%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=footer). Last update [94cb73c...c9824f4](https://codecov.io/gh/huggingface/transformers/pull/4353?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hi,\r\n\r\nIs the below expected behavior? I am initializing the tokenizer with `tokenizer = T5Tokenizer.from_pretrained('t5-large', extra_ids=0, additional_special_tokens=special_tokens)` where special_tokens is a list like ['_self_say', '_partner_say', '_self_persona', '_partner_persona']. Despite specifying extra_ids=0 I see a lot of sentinel tokens in the outputs with negative ids. For example:\r\n\r\n```\r\nInput: _context _setting_name Tower _setting_desc The inside tower is made from a combination of wood and brick. It also has some metal to hold all the doors in place. _object a door _object a door _self_name pet dog _self_persona I am mans best friend and I wouldn't have it any other way. I tend to my master and never leave his side. I sleep at his feet and guard the room at night from things that go bump in the night. _option hug knight _option hit knight _history \r\n\r\nResponse: <extra_id_-100>,<extra_id_-99> erreichen<extra_id_-98> hit knightăminterească a door<extra_id_-97>îî<extra_id_-96> a doorăm!<extra_id_-95> timpul<extra_id_-94> pentru<extra_id_-93>itățile Tower<extra_id_-92> I am Mans best friend.<extra_id_-91> doud<extra_id_-90> casă facem<extra_id_-89> a doorșteștehlen<extra_id_-88><extra_id_-87> tower is made from different materials. The outside\r\n```\r\n\r\n" ]

[]

[]

[ "This is awesome @ibeltagy . I´m very excited about longformer.\r\n\r\nIn the original implementation a good speedup was thanks to the personalized kernel in tvm. In this pull, this speedup is lost, or it was adapted somehow?", "Thanx @ibeltagy! \r\n\r\nIs it possible for you or someone else who is familiar with TF2 to submit also a PR for TF2? I tried to migrate your code in the past but I couldn’t migrate the new attention mechanism written for TVM, so I moved on. Your work alongside Reformer has special interest for people like me and my colleagues, who are usually doing research with long documents (>=1000words). \r\n\r\nMaybe @thomwolf and @julien-c know the right person...\r\n", "@ibeltagy - Looks awesome the PR :-) I added some minor comments. \r\n\r\nI think we can merge this very soon :-) \r\n\r\nTo pass the check_code_qualtiy test, you can run `make style` from the root folder. \r\n\r\nRegarding the unit test, I think you can copy-paste a lot of the code from `test_roberta_modeling.py`. In order to skip tests for `torch_script, pruning, ..` (features we can add in another PR) you can do the same as is done in the reformer tests here: https://github.com/huggingface/transformers/blob/94cb73c2d2efeb188b522ff352f98b15124ba9f8/tests/test_modeling_reformer.py#L510 . \r\n\r\nIf possible it would also be great to add some integration tests comparing the outputs (and maybe even gradients) to those of your original model code. I think you could use the same integration test structure as was done for Roberta or Reformer.", "@bratao , @iliaschalkidis, about the custom CUDA kernel. We [recently](https://github.com/allenai/longformer/pull/27/files#diff-04c6e90faac2675aa89e2176d2eec7d8R4) added an efficient PyTorch implementation of our sliding window attention (without dilation), which is faster and uses less memory (with fp16) and should be easy to port to TF. The TVM code is still needed for the dilated sliding window attention, but dilation is more important for char-lm than fine-tuning on downstream tasks (QA, classification .. etc). This PR only uses the PyTorch code.", "Thanks, @patrickvonplaten. I will address your comments and let you know. \r\n\r\nI also want to add `LongformerForSequenceClassification`, `LongformerForMultipleChoice`, `LongformerForTokenClassification`, `LongformerForQuestionAnswering` with automatic setting of global attention based on the task, so that the user doesn't need to worry about it. I will keep this to a different PR unless you think it is better to add it here. ", "> Thanks, @patrickvonplaten. I will address your comments and let you know.\r\n> \r\n> I also want to add `LongformerForSequenceClassification`, `LongformerForMultipleChoice`, `LongformerForTokenClassification`, `LongformerForQuestionAnswering` with automatic setting of global attention based on the task, so that the user doesn't need to worry about it. I will keep this to a different PR unless you think it is better to add it here.\r\n\r\nIt'd be great to add those as well :-) I think doing it in a second PR would be better as well", "@patrickvonplaten, @sshleifer, I am done addressing your reviews. I still need to upload model weights but waiting to get access to the AllenAI organization account. Also, can you check the failed tests? they seem like issues with CircleCI rather than the code. \r\n", "@sshleifer, addressed your comments. Now `ci/circleci: run_tests_torch_and_tf` is timing out.", "I'm sorry.\r\n\r\nRebase may help the timeout, but I agree that it does not appear to be caused by this code.\r\n\r\nI'd just verify that there aren't new tests that should be decorated `@slow` (append `--durations 0` to your local pytest command) and it should be fine.", "The slowest one is 0.10s. \r\n```\r\n0.10s call tests/test_modeling_longformer.py::LongformerModelTest::test_save_load\r\n0.10s call tests/test_modeling_longformer.py::LongformerModelTest::test_attention_outputs\r\n0.07s call tests/test_modeling_longformer.py::LongformerModelTest::test_resize_tokens_embeddings\r\n0.06s call tests/test_modeling_longformer.py::LongformerModelTest::test_determinism\r\n0.05s call tests/test_modeling_longformer.py::LongformerModelTest::test_inputs_embeds\r\n0.04s call tests/test_modeling_longformer.py::LongformerModelTest::test_correct_missing_keys\r\n0.04s call tests/test_modeling_longformer.py::LongformerModelTest::test_hidden_states_output\r\n0.04s call tests/test_modeling_longformer.py::LongformerModelTest::test_longformer_model\r\n0.03s call tests/test_modeling_longformer.py::LongformerModelTest::test_initialization\r\n0.03s call tests/test_modeling_longformer.py::LongformerModelTest::test_longformer_for_masked_lm\r\n0.02s call tests/test_modeling_longformer.py::LongformerModelTest::test_model_common_attributes\r\n```\r\n\r\nMaybe the integration tests are slower than average because of the long document? ", "@patrickvonplaten, @sshleifer, looks like rebasing fixed the problem. Thanks.", "> ## Adding the Longformer model\r\n> Paper: https://arxiv.org/abs/2004.05150\r\n> The code is mostly copied from https://github.com/allenai/longformer\r\n> Addressing issue: #3783\r\n> \r\n> **TODOs**\r\n> \r\n> * [x] Add unit tests\r\n> * [x] Address PR reviews\r\n> * [ ] Upload model to HF instead of AllenAI s3 buckets.\r\n> \r\n> **TODOs for other PRs**\r\n> \r\n> * Add `LongformerForSequenceClassification`, `LongformerForMultipleChoice`, `LongformerForTokenClassification`, `LongformerForQuestionAnswering` with automatic setting of global attention based on the task.\r\n> * Add `LongformerEncoderDecoder`\r\n> * A few small TODOs in the code\r\n> * Add an example to show how to convert existing pretrained models into their long version.\r\n\r\nIs there any particular timeframe by when we can have LongformerForQuestionAnswering? Would love to use it in a small project of mine.", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4352?src=pr&el=h1) Report\n> :exclamation: No coverage uploaded for pull request base (`master@18d233d`). [Click here to learn what that means](https://docs.codecov.io/docs/error-reference#section-missing-base-commit).\n> The diff coverage is `84.48%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4352?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4352 +/- ##\n=========================================\n Coverage ? 78.30% \n=========================================\n Files ? 123 \n Lines ? 20373 \n Branches ? 0 \n=========================================\n Hits ? 15953 \n Misses ? 4420 \n Partials ? 0 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4352?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19sb25nZm9ybWVyLnB5) | `82.75% <82.75%> (ø)` | |\n| [src/transformers/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9fX2luaXRfXy5weQ==) | `99.13% <100.00%> (ø)` | |\n| [src/transformers/configuration\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2F1dG8ucHk=) | `92.68% <100.00%> (ø)` | |\n| [src/transformers/configuration\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2xvbmdmb3JtZXIucHk=) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/modeling\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19hdXRvLnB5) | `78.57% <100.00%> (ø)` | |\n| [src/transformers/tokenization\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYXV0by5weQ==) | `97.61% <100.00%> (ø)` | |\n| [src/transformers/tokenization\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fbG9uZ2Zvcm1lci5weQ==) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/tokenization\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fb3BlbmFpLnB5) | `83.72% <0.00%> (ø)` | |\n| [src/transformers/configuration\\_bart.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2JhcnQucHk=) | `93.47% <0.00%> (ø)` | |\n| [src/transformers/modeling\\_tf\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9iZXJ0LnB5) | `96.54% <0.00%> (ø)` | |\n| ... and [120 more](https://codecov.io/gh/huggingface/transformers/pull/4352/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4352?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4352?src=pr&el=footer). Last update [18d233d...d1c4bbc](https://codecov.io/gh/huggingface/transformers/pull/4352?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=h1) Report\n> Merging [#4351](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/e8db8b845a971b0cf63a0896b9deb5b316028a8b&el=desc) will **increase** coverage by `0.05%`.\n> The diff coverage is `42.50%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4351 +/- ##\n==========================================\n+ Coverage 76.80% 76.86% +0.05% \n==========================================\n Files 128 128 \n Lines 21602 21669 +67 \n==========================================\n+ Hits 16591 16655 +64 \n- Misses 5011 5014 +3 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9yb2JlcnRhLnB5) | `74.48% <ø> (-0.26%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl90NS5weQ==) | `91.26% <13.33%> (-2.46%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9jdHJsLnB5) | `97.61% <33.33%> (-0.78%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9kaXN0aWxiZXJ0LnB5) | `82.43% <33.33%> (-0.45%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_electra.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9lbGVjdHJhLnB5) | `90.33% <33.33%> (-0.96%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9ncHQyLnB5) | `94.68% <33.33%> (-0.64%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9vcGVuYWkucHk=) | `94.87% <33.33%> (-0.71%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl94bG0ucHk=) | `75.54% <33.33%> (-0.31%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl94bG5ldC5weQ==) | `80.03% <33.33%> (-0.27%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9hbGJlcnQucHk=) | `75.11% <50.00%> (-0.23%)` | :arrow_down: |\n| ... and [7 more](https://codecov.io/gh/huggingface/transformers/pull/4351/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=footer). Last update [e8db8b8...8dc8f96](https://codecov.io/gh/huggingface/transformers/pull/4351?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "hi team- i know you guys are busy. do you mind taking a look when you have a chance & let me know w/ any comments? thanks!", "Hi- Thanks for the feedback. Sure, I can.\r\n\r\nSorry context switching is difficult at times :) I remember that `TFBertMainLayer` implementation was slightly different because it does not inherit from `TFPreTrainedModel` as T5 does. So it requires its own `_get_resized_embeddings` and test modules. But to the extent other architectures are similar to `T5`, it shouldn't be a problem. \r\n\r\nIs there a source of truth for all the TF architectures implemented? Is https://github.com/huggingface/transformers#model-architectures kept up to date?\r\n\r\nDeniz\r\n", "Yes, that list is kept up to date!", "@LysandreJik please take a look. thanks!", "This looks good! Pinging @jplu for review as he's the TF master around here.", "Thank you @jplu @LysandreJik ! Looking forward to contributing more. " ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "love it, thanks Manuel! 👍 ", "You are welcome. From now on I will add it to all my models." ]