Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "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." ]

[]

[]

[ "> Thanks for working on it!\r\n> \r\n> It is very specific to S2T and Wav2Vec2 but I don't think that's too much of an issue, we can adapt later.\r\n> \r\n> Could you add this pipeline to:\r\n> \r\n> * the documentation\r\n\r\nYes !\r\n> \r\n> * the main init\r\n\r\nYes\r\n\r\n> \r\n> * the `pipeline` factory method\r\n\r\nI wanted to defer this into a follow-up PR, because of the AutoModel quirckness and to avoid making super long PRs.\r\nIf you feel we can't have it as a separate PR, I'll start including the work directly here.\r\n> \r\n> \r\n> We will also probably need a new auto model.\r\n\r\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." ]

[]

[]

[]

[]

[]

[ "Ah, indeed, I think the identifier is incorrect. It needs to be updated to `microsoft/DialogRPT-updown`. Do you want to open a PR to fix this?\r\n\r\nThe issue you mention regarding the random weights is because the `gpt2` checkpoint doesn't have a sequence classification head. This is not the case for the checkpoint mentioned above (`microsoft/DialogRPT-updown`) which does have a sequence classification head, so it will not be reinitialized every different run.", "@LysandreJik Thanks a lot! Indeed, the identifier `microsoft/DialogRPT-updown` works! But, can I change the code example on hugging face site through a pull request? Since I won't really change some source code, is it a change that must be made in the Transformers repo? The (wrong) code example is as following\r\n\r\n\r\n", "Yes, the example is created here: https://github.com/huggingface/transformers/blob/5e04d7086803ae4a3892f4082f2835a756592c2c/src/transformers/models/gpt2/modeling_gpt2.py#L1239\r\n\r\nIf you can update this, then it will update it in the docs as soon as we merge it.\r\n\r\nAll the docs visible on the website are in the repository :)", "Ok, thanks! I'll deal with it then :)", "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." ]

[]

[]

[ "The `is_split_into_words` should be set to skip pre-tokenization (splitting on whitespace), not tokenization. This flag should be set to `True` if you have split your text into individual words, and you're now looking to have each word split into tokens and converted to IDs.\r\n\r\nThis seems to be unclear from the documentation, we'll work on improving this.", "Makes sense, thank you for the clarification! I'd be happy to work on this if needed.", "Yes, we would welcome such a contribution! I guess we would need to find all occurrences of that `is_split_into_words` parameter and clarify that pre-tokenization isn't tokenization as one could expect. \r\n\r\nWe would gladly welcome a PR!", "Great, I will work on this next week! Thanks again for the help!" ]

[]

[]

[ "Sorted by length?\r\nHow this helps you?", "> Sorted by length?\r\n> How this helps you?\r\n\r\nfor next lstm layer", "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." ]

[]

[]

[ "Merging (cc @LysandreJik, @sgugger)" ]

[]

[]

[ "Hi, @talkhaldi thanks for correcting the other comment!\r\n\r\nThere is one code quality check failing, could run `make style && make quality` and push the code again?", "Hi!\r\n\r\nThanks for your comment! The command you mentioned was requiring a lot of dependencies, some of which conflicted with the venv I was using, so instead of creating a new venv to run it, I removed the extra trailing space which technically was the only extra thing, and hoped it would work, but seems not ^^' \r\n\r\nI created a new venv and running the command now...", "Hi @patil-suraj,\r\n\r\nSo it seems make style works well, but I get this error with make quality:\r\n\r\n```\r\n> make quality\r\nblack --check� �amples tests src utils\r\n839 files would be left unchanged.\r\nisort --check-only examples tests src utils\r\npython utils/custom_init_isort.py --check_only\r\nflake8 examples tests src utils\r\nmake extra_quality_checks\r\nmake[1]: Entering directory '/mnt/berry/home/alkhaldi/contribcode/transformers'\r\npython utils/check_copies.py\r\npython utils/check_table.py\r\npython utils/check_dummies.py\r\npython utils/check_repo.py\r\nTraceback (most recent call last):\r\n File \"utils/check_repo.py\", line 22, in <module>\r\n from transformers.models.auto import get_values\r\nImportError: cannot import name 'get_values' from 'transformers.models.auto' (unknown location)\r\nmake[1]: *** [Makefile:33: extra_quality_checks] Error 1\r\nmake[1]: Leaving directory '/mnt/berry/home/alkhaldi/contribcode/transformers'\r\nmake: *** [Makefile:42: quality] Error 2\r\n```\r\n\r\nDo you have an idea of what's wrong?", "hi @talkhaldi\r\n\r\nplease make sure you have all the dev deps before running make\r\n\r\nto do that you could run `pip install -e \".[dev]\"` from the root of the repo.", "Hi @patil-suraj,\r\n\r\nI have applied that command successfully, but it seems the tests are still failing :o What could be the problem?\r\n\r\nNote: That command changed many files, but I only added modeling_t5.py to the commit. Should I include the changes of all files even though I haven't changed them manually?", "Hi @talkhaldi \r\n\r\nI think this is because we upgraded the version of `black`. Could you rebase your branch with master and then push again ?", "Hi @patil-suraj,\r\n\r\nThe check_code_quality test passes now, but another test fails. Can you please tell me what's wrong?" ]

[]

[]

[ "Hi, thanks for this update!\r\nI guess my huggingface username is [`camembert`](https://huggingface.co/camembert) ...?", "> # What does this PR do?\r\n> This PR mentions by HF username the person who added a given model in each doc file, and updates the template so this keeps being consistently done.\r\n> \r\n> A few persons are missing, tagging there below. If you are one of the persons tagged, it would be great if you could create a HF account and report your HF username here so we can properly attribute the model addition to you :-)\r\n> \r\n> * Bert japanese: @singletongue\r\n> * CamemBert: @louismartin\r\n> * DeBerta (v1 and v2): @BigBird01\r\n => Pengcheng He: @BigBird01\r\n> * ProphetNet and XLM-ProphetNet: @qiweizhen\r\n\r\nThanks I just replied my name and account inline.", "Hi @BigBird01, there is no huggingface account with the user name BigBird01 (I'm not talking about the GitHub account but the account on [huggingface.co](https://huggingface.co/)).", "Thanks! It’s DeBERTa😊, DeBERTa (Pengcheng He) (huggingface.co)<https://huggingface.co/DeBERTa>\r\n\r\nThanks!\r\nPengcheng\r\n\r\n", "Thank you for the mentioning!\r\nThe HF account for `bert-japanese` models is [`cl-tohoku`](https://huggingface.co/cl-tohoku).", "Failure is spurious, so merging. @qiweizhen if you create (or already have) a HF account, I can add you in a followup PR." ]

[]

[]

[ "Side note, the torchhub test uses the master branch to check which dependency to install (see [here](https://github.com/huggingface/transformers/blob/c0328a6c263494fff527fac7288faa627e3267e0/.github/workflows/github-torch-hub.yml#L36)), which means it can't pass until this PR is merged. Unless I missed something!", "This is awesome - very much looking forward to have this feature! \r\n\r\nOne thing I'd love to discuss are that they are 4 input arguments to to the `push_to_hub(...)` method - do we really need all of those? \r\n\r\n1.) I actually wouldn't put `save_directory` as an argument to `push_to_hub` because I think only the object on which `push_to_hub(...)` is called should be uploaded and not files that are unrelated to the object on which `push_to_hub()` is called. \r\n\r\n*E.g.*: If I call `model.save_pretrained(\"name/of/dir\")` the model and config is saved in this dir => I would therefore expect `model.push_to_hub(\"name_of_repo\")` to upload only model to the hub. IMO, this is more intuitive and more concise with how we design `save_pretrained(...)`. I don't really understand why a model object should be responsible to upload files to the hub that are not related to the model itself. Similar to how we call `model.save_pretrained(...)` and `tokenizer.save_pretrained(...)` to save everything we need, we should have to call `model.push_to_hub(...)` and `tokenizer.push_to_hub(...)` IMO.\r\n\r\n2.) Also, I think it would be a bit nicer if we could reduce the args: `repo_name`, `repo_url`, `organization` to just `repo_name` and `organization`. I would make `repo_name` mandatory (by default it will be pushed under one's namespace) and `organization` optionally. Do we really need three args to define the repo URL? What do you think?\r\n\r\n\r\nI feel quite strongly about 1.) to have consistency in the library, less strongly about 2.)" ]

[]

[]

[ "It looks like mobileBERT was pretrained on TPUs using bfloat16, which then often result in NaNs when using FP16 for further fine-tuning (see #11076 or #10956). You'll be best off training in FP32 or use another model compatible with FP16.", "Makes sense! That's interesting that affects the training on GPUs! I will pass this info on to my colleague who deals with reproducibility! And for now I shall stick with FP32 when fine-tuning the MobileBERT model!\r\n\r\nMany thanks for the reply!", "> You'll be best off training in FP32 or use another model compatible with FP16.\r\n\r\nAnd at some point we should also add `--bf16` mode to Trainer, for those who want to do finetuning and inference on hardware that supports it . e.g. high-end Ampere RTX-3090 and A100 should already support it, and of course TPU v2+. \r\n\r\nDoes it make sense?\r\n\r\nFYI, `bf16` AMP is being discussed here: https://github.com/pytorch/pytorch/issues/55374" ]

[]

[]

[ "Could you upgrade to the latest version of Transformers and see if the problem persists? I have tried to reproduce but it all works fine on my side.", "I have the same issue.", "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.", "I'm facing the same problem with `4.10.0.dev0`. @ameet-1997 could you find a solution for this?", "Looks like you already found the solution, thanks for that!\r\nI wasn't able to fix it earlier." ]

[]

[]

[ "@sgugger, please merge if you're happy with the updated PR :)" ]

[]

[]

[]

[]

[]

[ "You will only have perfectly reproducible results using checkpointing if the only randomness comes from the shuffling in your data (this is enforced by the CI). The way this is programmed inside the Trainer is to go through each epoch before the current one (which triggers the random shuffling) and then each batch (which puts you in the same position as before the checkpoint).\r\n\r\nSince your results differ slightly, it looks like there are other random calls in your training code, which you did not share. There is no way to have the exact same results while resuming from a checkpoint if this is the case.", "Hi @sgugger thanks for the reply, I do not have any other randomness in my codes, and I am using run_seq2seq.py codes to train t5 models on mrpc dataset, without modifications, I really appreciate your help on this issue as this is really crucial for me to have this working thanks a lot \r\n\r\nI initialize only the weights randomly, but I assume huggnigface well taking care of setting seeds, and there is really no other randomness ", "@sgugger I confirm also training the vanilla t5 have the same issue exists:\r\nHere is the run for t5-base for 100 steps:\r\n\r\n```\r\n{'loss': 6.1045, 'learning_rate': 6e-07, 'epoch': 0.02} \r\n 0%| | 10/60000 [00:06<10:25:12, 1.60it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:01<00:00, 2.44it/s]\r\n{'mrpc_en_eval_loss': 6.924696445465088, 'mrpc_en_eval_f1': 0.0, 'mrpc_en_eval_accuracy': 0.0, 'mrpc_en_eval_gen_len': 3.3137254901960786, 'mrpc_en_eval_runtime': 1.9287, 'mrpc_en_eval_samples_per_second': 105.771, 'epoch': 0.22} \r\n{'mrpc_en_eval_loss': 6.924696445465088, 'mrpc_en_eval_f1': 0.0, 'mrpc_en_eval_accuracy': 0.0, 'mrpc_en_eval_gen_len': 3.3137254901960786, 'mrpc_en_eval_runtime': 1.9287, 'mrpc_en_eval_samples_per_second': 105.771, 'epoch': 0.22, 'eval_average_metrics': 0.0} \r\n 0%| | 20/60000 [00:27<13:37:00, 1.22it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:01<00:00, 2.49it/s]\r\n{'mrpc_en_eval_loss': 5.22016716003418, 'mrpc_en_eval_f1': 0.0, 'mrpc_en_eval_accuracy': 0.0, 'mrpc_en_eval_gen_len': 3.764705882352941, 'mrpc_en_eval_runtime': 1.8761, 'mrpc_en_eval_samples_per_second': 108.737, 'epoch': 0.43} \r\n{'mrpc_en_eval_loss': 5.22016716003418, 'mrpc_en_eval_f1': 0.0, 'mrpc_en_eval_accuracy': 0.0, 'mrpc_en_eval_gen_len': 3.764705882352941, 'mrpc_en_eval_runtime': 1.8761, 'mrpc_en_eval_samples_per_second': 108.737, 'epoch': 0.43, 'eval_average_metrics': 0.0} \r\n 0%| | 30/60000 [00:47<12:58:53, 1.28it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:01<00:00, 2.37it/s]\r\n{'mrpc_en_eval_loss': 1.3517154455184937, 'mrpc_en_eval_f1': 0.0, 'mrpc_en_eval_accuracy': 18.137254901960784, 'mrpc_en_eval_gen_len': 3.2205882352941178, 'mrpc_en_eval_runtime': 1.9678, 'mrpc_en_eval_samples_per_second': 103.67, 'epoch': 0.65} \r\n{'mrpc_en_eval_loss': 1.3517154455184937, 'mrpc_en_eval_f1': 0.0, 'mrpc_en_eval_accuracy': 18.137254901960784, 'mrpc_en_eval_gen_len': 3.2205882352941178, 'mrpc_en_eval_runtime': 1.9678, 'mrpc_en_eval_samples_per_second': 103.67, 'epoch': 0.65, 'eval_average_metrics': 9.068627450980392} \r\n 0%| | 40/60000 [01:08<13:00:06, 1.28it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 4.62it/s]\r\n{'mrpc_en_eval_loss': 0.4487058222293854, 'mrpc_en_eval_f1': 81.3953488372093, 'mrpc_en_eval_accuracy': 68.62745098039215, 'mrpc_en_eval_gen_len': 2.0, 'mrpc_en_eval_runtime': 1.0261, 'mrpc_en_eval_samples_per_second': 198.811, 'epoch': 0.87} \r\n{'mrpc_en_eval_loss': 0.4487058222293854, 'mrpc_en_eval_f1': 81.3953488372093, 'mrpc_en_eval_accuracy': 68.62745098039215, 'mrpc_en_eval_gen_len': 2.0, 'mrpc_en_eval_runtime': 1.0261, 'mrpc_en_eval_samples_per_second': 198.811, 'epoch': 0.87, 'eval_average_metrics': 75.01139990880073} \r\n 0%| | 50/60000 [01:27<12:31:06, 1.33it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.72it/s]\r\n{'mrpc_en_eval_loss': 0.25695744156837463, 'mrpc_en_eval_f1': 83.79204892966361, 'mrpc_en_eval_accuracy': 74.01960784313727, 'mrpc_en_eval_gen_len': 2.0833333333333335, 'mrpc_en_eval_runtime': 1.2653, 'mrpc_en_eval_samples_per_second': 161.228, 'epoch': 1.09} \r\n{'mrpc_en_eval_loss': 0.25695744156837463, 'mrpc_en_eval_f1': 83.79204892966361, 'mrpc_en_eval_accuracy': 74.01960784313727, 'mrpc_en_eval_gen_len': 2.0833333333333335, 'mrpc_en_eval_runtime': 1.2653, 'mrpc_en_eval_samples_per_second': 161.228, 'epoch': 1.09, 'eval_average_metrics': 78.90582838640043} \r\n 0%|▏ | 60/60000 [01:47<12:36:18, 1.32it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 4.29it/s]\r\n{'mrpc_en_eval_loss': 0.27573078870773315, 'mrpc_en_eval_f1': 82.11143695014663, 'mrpc_en_eval_accuracy': 70.09803921568627, 'mrpc_en_eval_gen_len': 2.014705882352941, 'mrpc_en_eval_runtime': 1.1521, 'mrpc_en_eval_samples_per_second': 177.063, 'epoch': 1.3} \r\n{'mrpc_en_eval_loss': 0.27573078870773315, 'mrpc_en_eval_f1': 82.11143695014663, 'mrpc_en_eval_accuracy': 70.09803921568627, 'mrpc_en_eval_gen_len': 2.014705882352941, 'mrpc_en_eval_runtime': 1.1521, 'mrpc_en_eval_samples_per_second': 177.063, 'epoch': 1.3, 'eval_average_metrics': 76.10473808291644} \r\n 0%|▏ | 70/60000 [02:09<13:15:00, 1.26it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.75it/s]\r\n{'mrpc_en_eval_loss': 0.16758881509304047, 'mrpc_en_eval_f1': 87.04318936877075, 'mrpc_en_eval_accuracy': 80.88235294117648, 'mrpc_en_eval_gen_len': 2.2107843137254903, 'mrpc_en_eval_runtime': 1.2665, 'mrpc_en_eval_samples_per_second': 161.075, 'epoch': 1.52} \r\n{'mrpc_en_eval_loss': 0.16758881509304047, 'mrpc_en_eval_f1': 87.04318936877075, 'mrpc_en_eval_accuracy': 80.88235294117648, 'mrpc_en_eval_gen_len': 2.2107843137254903, 'mrpc_en_eval_runtime': 1.2665, 'mrpc_en_eval_samples_per_second': 161.075, 'epoch': 1.52, 'eval_average_metrics': 83.96277115497361} \r\n 0%|▏ | 80/60000 [02:30<13:18:49, 1.25it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.64it/s]\r\n{'mrpc_en_eval_loss': 0.1627584546804428, 'mrpc_en_eval_f1': 89.86486486486486, 'mrpc_en_eval_accuracy': 85.29411764705883, 'mrpc_en_eval_gen_len': 2.235294117647059, 'mrpc_en_eval_runtime': 1.2734, 'mrpc_en_eval_samples_per_second': 160.198, 'epoch': 1.74} \r\n{'mrpc_en_eval_loss': 0.1627584546804428, 'mrpc_en_eval_f1': 89.86486486486486, 'mrpc_en_eval_accuracy': 85.29411764705883, 'mrpc_en_eval_gen_len': 2.235294117647059, 'mrpc_en_eval_runtime': 1.2734, 'mrpc_en_eval_samples_per_second': 160.198, 'epoch': 1.74, 'eval_average_metrics': 87.57949125596184} \r\n 0%|▏ | 90/60000 [02:50<12:35:38, 1.32it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.71it/s]\r\n{'mrpc_en_eval_loss': 0.178583025932312, 'mrpc_en_eval_f1': 90.78014184397163, 'mrpc_en_eval_accuracy': 87.25490196078431, 'mrpc_en_eval_gen_len': 2.303921568627451, 'mrpc_en_eval_runtime': 1.2507, 'mrpc_en_eval_samples_per_second': 163.108, 'epoch': 1.96} \r\n{'mrpc_en_eval_loss': 0.178583025932312, 'mrpc_en_eval_f1': 90.78014184397163, 'mrpc_en_eval_accuracy': 87.25490196078431, 'mrpc_en_eval_gen_len': 2.303921568627451, 'mrpc_en_eval_runtime': 1.2507, 'mrpc_en_eval_samples_per_second': 163.108, 'epoch': 1.96, 'eval_average_metrics': 89.01752190237798} \r\n 0%|▏ | 100/60000 [03:09<12:29:36, 1.33it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.70it/s]\r\n{'mrpc_en_eval_loss': 0.18296584486961365, 'mrpc_en_eval_f1': 88.72727272727272, 'mrpc_en_eval_accuracy': 84.80392156862744, 'mrpc_en_eval_gen_len': 2.338235294117647, 'mrpc_en_eval_runtime': 1.2762, 'mrpc_en_eval_samples_per_second': 159.845, 'epoch': 2.17} \r\n{'mrpc_en_eval_loss': 0.18296584486961365, 'mrpc_en_eval_f1': 88.72727272727272, 'mrpc_en_eval_accuracy': 84.80392156862744, 'mrpc_en_eval_gen_len': 2.338235294117647, 'mrpc_en_eval_runtime': 1.2762, 'mrpc_en_eval_samples_per_second': 159.845, 'epoch': 2.17, 'eval_average_metrics': 86.76559714795007} \r\n\r\n```\r\n\r\nNow lets see the results of t5-base after resuming from step = 60\r\n```\r\n 0%|▏ | 60/60000 [00:06<9:21:55, 1.78it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 4.00it/s]\r\n{'mrpc_en_eval_loss': 0.2794328033924103, 'mrpc_en_eval_f1': 82.11143695014663, 'mrpc_en_eval_accuracy': 70.09803921568627, 'mrpc_en_eval_gen_len': 2.014705882352941, 'mrpc_en_eval_runtime': 1.2224, 'mrpc_en_eval_samples_per_second': 166.887, 'epoch': 1.3} \r\n{'mrpc_en_eval_loss': 0.2794328033924103, 'mrpc_en_eval_f1': 82.11143695014663, 'mrpc_en_eval_accuracy': 70.09803921568627, 'mrpc_en_eval_gen_len': 2.014705882352941, 'mrpc_en_eval_runtime': 1.2224, 'mrpc_en_eval_samples_per_second': 166.887, 'epoch': 1.3, 'eval_average_metrics': 76.10473808291644} \r\n 0%|▏ | 70/60000 [00:28<13:22:56, 1.24it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.59it/s]\r\n{'mrpc_en_eval_loss': 0.16057834029197693, 'mrpc_en_eval_f1': 88.43537414965986, 'mrpc_en_eval_accuracy': 83.33333333333334, 'mrpc_en_eval_gen_len': 2.2450980392156863, 'mrpc_en_eval_runtime': 1.3058, 'mrpc_en_eval_samples_per_second': 156.222, 'epoch': 1.52} \r\n{'mrpc_en_eval_loss': 0.16057834029197693, 'mrpc_en_eval_f1': 88.43537414965986, 'mrpc_en_eval_accuracy': 83.33333333333334, 'mrpc_en_eval_gen_len': 2.2450980392156863, 'mrpc_en_eval_runtime': 1.3058, 'mrpc_en_eval_samples_per_second': 156.222, 'epoch': 1.52, 'eval_average_metrics': 85.8843537414966} \r\n 0%|▏ | 80/60000 [00:48<12:55:04, 1.29it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.69it/s]\r\n{'mrpc_en_eval_loss': 0.15957750380039215, 'mrpc_en_eval_f1': 88.81118881118881, 'mrpc_en_eval_accuracy': 84.31372549019608, 'mrpc_en_eval_gen_len': 2.284313725490196, 'mrpc_en_eval_runtime': 1.291, 'mrpc_en_eval_samples_per_second': 158.021, 'epoch': 1.74} \r\n{'mrpc_en_eval_loss': 0.15957750380039215, 'mrpc_en_eval_f1': 88.81118881118881, 'mrpc_en_eval_accuracy': 84.31372549019608, 'mrpc_en_eval_gen_len': 2.284313725490196, 'mrpc_en_eval_runtime': 1.291, 'mrpc_en_eval_samples_per_second': 158.021, 'epoch': 1.74, 'eval_average_metrics': 86.56245715069244} \r\n 0%|▏ | 90/60000 [01:11<13:47:58, 1.21it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.67it/s]\r\n{'mrpc_en_eval_loss': 0.19618992507457733, 'mrpc_en_eval_f1': 87.17948717948718, 'mrpc_en_eval_accuracy': 82.84313725490196, 'mrpc_en_eval_gen_len': 2.3480392156862746, 'mrpc_en_eval_runtime': 1.2811, 'mrpc_en_eval_samples_per_second': 159.235, 'epoch': 1.96} \r\n{'mrpc_en_eval_loss': 0.19618992507457733, 'mrpc_en_eval_f1': 87.17948717948718, 'mrpc_en_eval_accuracy': 82.84313725490196, 'mrpc_en_eval_gen_len': 2.3480392156862746, 'mrpc_en_eval_runtime': 1.2811, 'mrpc_en_eval_samples_per_second': 159.235, 'epoch': 1.96, 'eval_average_metrics': 85.01131221719457} \r\n 0%|▏ | 100/60000 [01:33<12:55:11, 1.29it/s]***** Running Evaluation *****\r\n Num examples = 204\r\n Batch size = 80\r\n ### n_samples 204███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 3.75it/s]\r\n{'mrpc_en_eval_loss': 0.21464459598064423, 'mrpc_en_eval_f1': 87.96992481203009, 'mrpc_en_eval_accuracy': 84.31372549019608, 'mrpc_en_eval_gen_len': 2.3823529411764706, 'mrpc_en_eval_runtime': 1.2654, 'mrpc_en_eval_samples_per_second': 161.214, 'epoch': 2.17} \r\n{'mrpc_en_eval_loss': 0.21464459598064423, 'mrpc_en_eval_f1': 87.96992481203009, 'mrpc_en_eval_accuracy': 84.31372549019608, 'mrpc_en_eval_gen_len': 2.3823529411764706, 'mrpc_en_eval_runtime': 1.2654, 'mrpc_en_eval_samples_per_second': 161.214, 'epoch': 2.17, 'eval_average_metrics': 86.14182515111308} \r\n 0%|▏ \r\n\r\n```\r\n", "Dear @sgugger @patrickvonplaten @patil-suraj \r\nCould you kindly have a look into this issue, this is really important to have the checkpointing workings, as in many cases one cannot train the models for larger periods, thnaks ", "Following up on @sgugger's suggestion, if I understand the methodology correctly it doesn't quite apply to the generic checkpointing method, but one could subclass the Trainer to save the RNG state at the moment of saving the checkpoint, and then restore the same RNG state on resume. You'd probably need to do that for at least python and pytorch (and numpy and other libraries if you use those).\r\n\r\n@dorooddorood606, look into:\r\n```\r\n# before saving\r\npy_rng_state = random.getstate()\r\npt_rng_state = torch.get_rng_state()\r\nnp_rng_state = numpy.random.get_state()\r\n\r\n# post resume\r\nrandom.setstate(py_rng_state)\r\ntorch.set_rng_state(pt_rng_state)\r\nnumpy.random.set_state(np_rng_state)\r\n```", "Dear @stas00 \r\nThank you very much for following up on this, I implemented this suggestion, and I still see the discrepancies after resuming the checkpoints. I emphasize I tried with \"vanilla t5-base\" so no changes from huggingface codes. In my own codes, I have some initialization which is the only part with randomness, I would be grateful if you could tell me if there might be an issue with these lines:\r\n```\r\nnn.init.normal_(linear_layer.weight, std=std)\r\nnn.init.zeros_(linear_layer.bias)\r\n```\r\nbut still since vanillat t5-base also has this issue, I was wondering if you might think this might be relevant to the trainer code as a general issue? I greatly appreciate it if you could kindly consider this issue.\r\n\r\nthanks a lot in advance for the great work you do and your hard efforts.\r\n", "> Thank you very much for following up on this, I implemented this suggestion, \r\n\r\nCould we first validate that this was done correctly?\r\n\r\nTo test you can debug print some random number generated **immediately after saving the checkpoint** and RNG state and doing the same right **after the checkpoint and RNG states were restored** when you run the program 2nd time with resume. If you get the same number generated then we know you restored the RNG state. You probably want to check one for torch and one for python.\r\n\r\n> I have some initialization which is the only part with randomness, I would be grateful if you could tell me if there might be an issue with these lines:\r\n> \r\n> ```\r\n> nn.init.normal_(linear_layer.weight, std=std)\r\n\r\nThis line would definitely impact the RNG state. If you're uncertain you can always debug and generate a random number with that line of code and w/o it and see if it's the same.\r\n \r\nSo for example one workaround you could do is to restore the RNG state after your custom code above.\r\n\r\nOr better don't re-run this line, but save the outcome with the checkpoint and then restore it on subsequent runs, rather the needing to fiddle with RNG states.\r\n", "Dear @stas00 \r\nFirst, I would like to thank you very much for taking your precious time and answering to my question. \r\nI observe that between different runs my codes generate different results. I was assuming since HuggingFace run_glue.py codes set the seeds initially, then it is well taking care of randomness. All my code has is some initialization, like what I sent, coming all after the \"set_seed()\" function. Considering only one run, putting check-pointing aside, could you kindly tell me if one needs to set seeds before each initialization? shall I bring them all in init_weights function of BERT? I appreciate your response a lot.\r\nThank you. ", "First a few requests, @dorooddorood606 \r\n- please don't re-post the same question on Issues and forums, once is plenty - honestly I'm lost at what we are trying to solve here.\r\n- we all appreciate your appreciations, you're clearly a very nice person, but it becomes overbearing when we get copious amounts of it in every post.\r\n- let's focus on the problem only so that the sound-to-noise ratio is manageable.\r\n\r\nThank you!\r\n\r\n------------\r\n\r\nNow, let's try to summarize what doesn't work.\r\n\r\n1. From what I understand you extended the library with your own modifications. And now you're experiencing inconsistent randomness issues when you resume the model, correct?\r\n\r\n Does the library produce the expected results if you remove your modifications?\r\n\r\n2. Is there an easy way to provide a reproducible example that shows how the main library works correctly and then it breaks when with your modification? Perhaps a simple google colab notebook? If you do that please make sure that it's very easy to quickly see what the problem is and where it comes from. So no production-level hundreds of lines of code, but toy examples if possible.\r\n\r\n", "Dear @stas00 \r\nThank you for the remind, I will follow the points you mentioned. I was thinking there is also a bug in the trainer as I was also observing it for the Bert-base model unchanged, but the randomness issue resolved with upgrading to 4.6.0 version of transformers.\r\n\r\n\r\n", "Dear @stas00 \r\n\r\nI appreciate your input on the issue of reproducibility from resuming from checkpoints a lot. I tried to follow your points to state it in a clearer way. \r\n\r\n### Problem statement \r\nIf a user train a model till some steps and then reload the model from a checkpoint, the results differs from the training the model without breaks. \r\n\r\n### How to reproduce the issue \r\nTransformer version: I am using 4.6.0dev version of transformer \r\n```\r\nhttps://github.com/huggingface/transformers/commit/04ab2ca639ee6fd1002ce0a498d892245f0c9093\r\n```\r\n\r\nPlease kindly clone this repository with a minimal example\r\n```\r\ngit clone [email protected]:dorooddorood606/reproducibility.git\r\n```\r\nTo run the codes, please kindly run this command, between the runs in every 50 steps after save of the model, kill the model for like 2-3 times. Please then compare the final results of running for the full iterations with resuming, with ```raining without any breaks```. The results would differ. \r\n```\r\nTASK_NAME=mrpc\r\npython run_glue.py --model_name_or_path bert-base-cased --task_name $TASK_NAME --do_train --do_eval --max_seq_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 2 --output_dir /temp/$TASK_NAME/ --eval_steps 50 --evaluation_strategy steps --load_best_model_at_end --fp16 --do_predict\r\n```\r\n\r\nPlease let me know if you need any further information on this. \r\n\r\n### Which modifications done on Trainer class to make it reproducible:\r\nI apply the following modifications to the trainer class:\r\n1) Following your suggestions. I save the random states and I reload them before reloading the checkpoint in the trainer class. Please see https://github.com/dorooddorood606/reproducibility/blob/f5902af4669bba8aaee326efdb0cd459e25be675/trainer.py#L126 \r\nand https://github.com/dorooddorood606/reproducibility/blob/f5902af4669bba8aaee326efdb0cd459e25be675/trainer.py#L200\r\n\r\n2) In each saving of checkpoints, I also save a copy of checkpoint in the output_dir, this is because I personally believe we need to also keep the last checkpoint to resume from in addition to keeping only checkpoint of the best model so far, to be able to continue training from the last state. Please see https://github.com/dorooddorood606/reproducibility/blob/f5902af4669bba8aaee326efdb0cd459e25be675/trainer.py#L87\r\n\r\n3) I get the last checkpoint in run_glue.py based on the checkpoint saved in the main output_dir, please see https://github.com/dorooddorood606/reproducibility/blob/f5902af4669bba8aaee326efdb0cd459e25be675/run_glue.py#L46\r\n\r\n### Larger impact of this issue\r\nTo me this issue with resuming from checkpoint, can also help other users and would be beneficial to all users who need to use this option. I appreciate a lot if you could sparse me some time from your precious time and help on this issue.\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "Thank you for your detailed followup, @dorooddorood606. And sharing what experiments you have tried. \r\n\r\nI agree that it'd be awesome to be able to resume as if there was no stopping.\r\n\r\nPlease give us some time, we are going discuss whether it is feasible to make it happen as there are so many moving parts to consider and if so will build this ground up.\r\n\r\nWe will keep you posted.", "Dear @stas00 \r\nthank you. Sure, meanwhile if you might have some ideas and suggestions for me to try, I greatly appreciate your help. I searched for this issue a lot, and apart from the things HuggingFace repo has already implemented I could not find more tricks to do to solve the issue. \r\nThanks a lot in advance for your time and assistance. ", "@sgugger is working on it in https://github.com/huggingface/transformers/pull/11582", "Hi \r\nI cannot really express how much I appreciate this. Thank you very much both for working on this. This would be wonderful to have resuming fixed in trainer. Thanks for your efforts. ", "I totally agree!\r\n\r\nAll kudos go to @sgugger , who has a much better understanding of the nooks and crannies of the HF Trainer.\r\n\r\n", "Dear @sgugger \r\n\r\nThanks for the hard work. I tested it but the issue is not resolved, specially for small datasets it can make large changes in final results, I appreciate if you could share with me some suggestions on how to resolve the issue:\r\n\r\nThe original one:\r\n```\r\ncheckpoint: 200\r\n{'eval_loss': 0.44332757592201233, 'eval_accuracy': 0.7941176470588235, 'eval_f1': 0.8521126760563381, 'eval_combined_score': 0.8231151615575808, 'eval_runtime': 1.5259, 'eval_samples_per_second': 133.692, 'eval_average_metrics': 0.8231151615575808, 'epoch': 1.74}\r\n```\r\n\r\nThe resumed one:\r\n\r\n```\r\ncheckpoint: 200\r\n{'eval_loss': 0.4352119266986847, 'eval_accuracy': 0.7941176470588235, 'eval_f1': 0.85, 'eval_combined_score': 0.8220588235294117, 'eval_runtime': 1.4451, 'eval_samples_per_second': 141.165, 'eval_average_metrics': 0.8220588235294117, 'epoch': 1.74} \r\n```\r\nThe differences accumulate a lot over time \r\n\r\n# To reproduce please run:\r\n```\r\nTASK_NAME=mrpc \r\npython run_glue.py --model_name_or_path bert-base-cased --task_name $TASK_NAME --do_train --do_eval --max_seq_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 3 --output_dir /temp/results --eval_steps 50 --evaluation_strategy steps --load_best_model_at_end --fp16 --do_test --save_total_limit 1 \r\n``` \r\n\r\nHere are the final results without drop:\r\n```\r\n[INFO|trainer_pt_utils.py:907] 2021-05-09 17:35:14,973 >> ***** eval metrics *****\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,973 >> epoch = 3.0\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,973 >> eval_accuracy = 0.701\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_average_metrics = 0.7605196946035051\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_combined_score = 0.7605\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_f1 = 0.8201\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_loss = 0.604\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_mem_cpu_alloc_delta = 2MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_mem_cpu_peaked_delta = 2MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_mem_gpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_mem_gpu_peaked_delta = 33MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_runtime = 0:00:01.95\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_samples = 204\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:14,974 >> eval_samples_per_second = 104.502\r\n05/09/2021 17:35:14 - INFO - __main__ - *** Test ***\r\n[INFO|trainer.py:515] 2021-05-09 17:35:15,036 >> The following columns in the evaluation set don't have a corresponding argument in `BertForSequenceClassification.forward` and have been ignored: idx, sentence1, sentence2.\r\n[INFO|trainer.py:2089] 2021-05-09 17:35:15,040 >> ***** Running Evaluation *****\r\n[INFO|trainer.py:2091] 2021-05-09 17:35:15,041 >> Num examples = 204\r\n[INFO|trainer.py:2094] 2021-05-09 17:35:15,041 >> Batch size = 8\r\n100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 26/26 [00:01<00:00, 13.77it/s]\r\n[INFO|trainer_pt_utils.py:907] 2021-05-09 17:35:17,070 >> ***** test metrics *****\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> epoch = 3.0\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_accuracy = 0.6863\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_average_metrics = 0.7490196078431373\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_combined_score = 0.749\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_f1 = 0.8118\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_loss = 0.6198\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_mem_cpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_mem_cpu_peaked_delta = 2MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_mem_gpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_mem_gpu_peaked_delta = 33MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_runtime = 0:00:01.95\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> eval_samples_per_second = 104.281\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:35:17,070 >> test_samples = 204\r\n\r\n```\r\n\r\nwith breaking in between:\r\n```\r\n[INFO|trainer_pt_utils.py:907] 2021-05-09 17:41:22,953 >> ***** eval metrics *****\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> epoch = 3.0\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_accuracy = 0.6863\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_average_metrics = 0.7467517127332861\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_combined_score = 0.7468\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_f1 = 0.8072\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_loss = 0.6106\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_mem_cpu_alloc_delta = 2MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_mem_cpu_peaked_delta = 1MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_mem_gpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_mem_gpu_peaked_delta = 33MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_runtime = 0:00:01.82\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,953 >> eval_samples = 204\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:22,954 >> eval_samples_per_second = 111.603\r\n05/09/2021 17:41:22 - INFO - __main__ - *** Test ***\r\n[INFO|trainer.py:515] 2021-05-09 17:41:23,014 >> The following columns in the evaluation set don't have a corresponding argument in `BertForSequenceClassification.forward` and have been ignored: idx, sentence2, sentence1.\r\n[INFO|trainer.py:2089] 2021-05-09 17:41:23,018 >> ***** Running Evaluation *****\r\n[INFO|trainer.py:2091] 2021-05-09 17:41:23,019 >> Num examples = 204\r\n[INFO|trainer.py:2094] 2021-05-09 17:41:23,019 >> Batch size = 8\r\n100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 26/26 [00:01<00:00, 14.71it/s]\r\n[INFO|trainer_pt_utils.py:907] 2021-05-09 17:41:24,916 >> ***** test metrics *****\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,916 >> epoch = 3.0\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,916 >> eval_accuracy = 0.701\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,916 >> eval_average_metrics = 0.7572180248246088\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,916 >> eval_combined_score = 0.7572\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,916 >> eval_f1 = 0.8135\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,916 >> eval_loss = 0.6068\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> eval_mem_cpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> eval_mem_cpu_peaked_delta = 1MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> eval_mem_gpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> eval_mem_gpu_peaked_delta = 33MB\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> eval_runtime = 0:00:01.83\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> eval_samples_per_second = 111.455\r\n[INFO|trainer_pt_utils.py:912] 2021-05-09 17:41:24,917 >> test_samples = 204\r\n```\r\n\r\nThis is that different that still does not allow using checkpointing, I only have access to gpus which are interruptable and really appreciate your help \r\n\r\nI also have added `CUBLAS_WORKSPACE_CONFIG=:16:8` as described in `https://discuss.pytorch.org/t/random-seed-with-external-gpu/102260/3` to make torch deterministic, still does not work, ", "Are you sure you are running on a source install of Transformers? The command produces the exact same results on my end.", "Dear Sylvain,\r\nThanks for the response. Yes, I install transformers as pip install git+https://github.com/huggingface/transformers.git\r\n\r\nbut the results differs a lot. Please kindly run this command and break it after first checkpoint (iterations = 50)\r\n```\r\nTASK_NAME=mrpc\r\npython run_glue.py --model_name_or_path bert-base-cased --task_name $TASK_NAME --do_train --do_eval --max_seq_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 3 --output_dir /tmp/ --eval_steps 50 --evaluation_strategy steps --load_best_model_at_end --fp16 --do_test \r\n``` \r\n", "This might be due to the FP16 parameter. Could you check if you get the same result without FP16?\r\nThe reason is due to the fact we don't save the state of the gradient scaler in mixed precision training, which is another thing to restore to its state. Can make a PR to fix that tomorrow.", "Dear Sylvain\r\n\r\nThank you for taking your precious time and answering this issue. you are absolutely right. I checked it without fp16 and I confirm this works fine without fp16, it would be wonderful to have the fp16 mode also working when you have time.\r\n\r\nThank you for your hard work and great job you do :) ", "Problem was fixed on my side with the PR above. Let me know if this is not the case for you.", "Dear @sgugger \r\n\r\nThank you for the PR, I checked it with the last version of transformers now, and the issue still exists, please kindly run this command and break this after first 50 steps:\r\n```\r\nTASK_NAME=mrpc\r\npython run_glue.py --model_name_or_path bert-base-cased --task_name $TASK_NAME --do_train --do_eval --max_seq_length 128 --per_device_train_batch_size 32 --learning_rate 2e-5 --num_train_epochs 3 --output_dir /tmp/$TASK_NAME/ --eval_steps 50 --evaluation_strategy steps --load_best_model_at_end --fp16 --do_test \r\n```\r\nHere are the results:\r\nIf you do not break:\r\n```\r\nAfter 50 steps:\r\n\r\n{'eval_loss': 0.6383711695671082, 'eval_accuracy': 0.6764705882352942, 'eval_f1': 0.8070175438596491, 'eval_combined_score': 0.7417440660474717, 'eval_runtime': 2.1914, 'eval_samples_per_second': 93.091, 'eval_average_metrics': 0.7417440660474717, 'epoch': 0.43}\r\n\r\nAfter 100 steps:\r\n{'eval_loss': 0.6184656023979187, 'eval_accuracy': 0.6862745098039216, 'eval_f1': 0.813953488372093, 'eval_combined_score': 0.7501139990880072, 'eval_runtime': 2.1089, 'eval_samples_per_second': 96.731, 'eval_average_metrics': 0.7501139990880072, 'epoch': 0.87}\r\n ```\r\n\r\nif you break after 50 steps:\r\n\r\n```\r\nAfter 100 steps\r\n{'eval_loss': 0.6308265328407288, 'eval_accuracy': 0.6862745098039216, 'eval_f1': 0.813953488372093, 'eval_combined_score': 0.7501139990880072, 'eval_runtime': 2.1549, 'eval_samples_per_second': 94.668, 'eval_average_metrics': 0.7501139990880072, 'epoch': 0.87} \r\n```\r\n\r\nThe differences accumulates and the results at the end varies a lot that resumed results are not usable.\r\nI really appreciate if you could kindly have another look. Could you kindly reopen this issue as well?\r\n\r\nthanks. \r\n", "I sadly cannot reproduce (get the exact same results with the command you indicated using a source install on current master) so this comes from something in your particular setup at this stage." ]

[]

[]

[ "We probably need to re-think the \"placement on device\" logic. And to do it explicitly in each stage and in `__init__` only in those special cases where it's absolutely required." ]

[]

[]

[ "Hi @biggoron \r\n\r\n`BertForMaskedLM` can't be used as a decoder, it's intended for masked LM. The `BertLMHeadModel` model should be used if you wan to use bert as a decoder. Also the right way to initialize the bert as decoder is as follows\r\n\r\n```python\r\ndec_config = BertConfig.from_pretrained(\"bert-base-uncased\")\r\ndec_config.add_cross_attentions = True # add cross attention if you want to use it in EncoderDecoderModel\r\ndec_config.is_decoder = True\r\ndec = BertLMHeadModel.from_pretrained(\"bert-base-uncased\", config=dec_config)\r\n```\r\n\r\nor simply use the `from_encoder_decoder_pretrained` method which takes care of this.", "Thanks a lot for your help, it is much clearer!" ]

[ { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[ { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "After asking about this on the EleutherAI discord, it was pointed out to me that 256 tokens corresponds to the local attention span of the model. Looking at the plot above, the first allocation peaks appear after about 256 tokens. After about 512 tokens, another shorter set of spikes occur, with more and shorter spikes being added every 256. This could indicate that there is an issue related to the implementation of local attention.", "One more comment with additional information. During another run of the of the test script, I added some [logging](https://gist.github.com/finetuneanon/5b8b5cdaf4c27836ebbbf1ed0d238c5b) to modeling_gpt_neo.py in an exception handler. For another run where an OOM crash occured with sequence length 1871, before converting query and key to float32, 5270MB are used. Afterwards, 9984MB are in use. query.shape is [1, 1871, 20, 257, 128] and key.shape is [1, 1871, 20, 1, 128]. The transpose makes no additional allocation, but the matmul attempts to allocate another 4.6GB. The main culprit seems to be the dimension of size 257, which greatly increases the size of the tensor.", "Hi @finetuneanon \r\n\r\nThanks for the detailed issue!\r\n\r\nSo what is happening here is, the way local attention is designed is a bit weird (not the implementation), in that it splits the `seq_length` dim into `(num_blocks, block_length)` but here `block_length` is actually dynamic.\r\n\r\nIt's equal to `window_size` by default which is 256. But when the `seq_length` is not evenly divisible by `block_length` then it's adjusted as follows\r\n\r\n```python\r\ndef _get_block_length_and_num_blocks(seq_length, window_size):\r\n \"\"\"\r\n Computes ``block_length`` and ``num_blocks`` such that ``seq_length`` becomes evenly divisible by\r\n ``block_length``.\r\n \"\"\"\r\n block_length = window_size\r\n while seq_length % block_length != 0:\r\n block_length -= 1\r\n num_blocks = seq_length // block_length\r\n return block_length, num_blocks\r\n\r\n```\r\nsuch that, the `seq_length` becomes evenly divisible by `block_length`. \r\n\r\nSo the shape of `query` becomes `(batch, num_blocks, block_length, hidden_dim)`\r\nand then the `keys` and `values` are padded and the `seq_length` dim is split such that their shape becomes \r\n`(batch, num_blocks, window_size + block_length, hidden_dim`).\r\n\r\nHere's a simple function to get the shape of `query` and `key` for given `seq_length`\r\n```python\r\ndef get_query_key_shape(seq_len, window_size, hidden_dim):\r\n block_length, num_blocks = _get_block_length_and_num_blocks(seq_len, window_size)\r\n query_shape = (1, num_blocks, block_length, hidden_dim)\r\n key_shape = (1, num_blocks, window_size + block_length, hidden_dim)\r\n return query_shape, key_shape\r\n```\r\n\r\nLet's print the shapes for few lengths\r\n```python\r\nwindow_size = 256\r\nhidden_dim = 2560\r\nfor seq_len in range(256, 266):\r\n query_shape, key_shape = get_query_key_shape(seq_len, window_size, hidden_dim)\r\n print(f\"seq_len: {seq_len}, query_shape: {query_shape}, key_shape: {key_shape}\"\r\n```\r\nwhich gives \r\n\r\n```\r\nseq_len: 256, query_shape: (1, 1, 256, 2560), key_shape: (1, 1, 512, 2560)\r\nseq_len: 257, query_shape: (1, 257, 1, 2560), key_shape: (1, 257, 257, 2560)\r\nseq_len: 258, query_shape: (1, 2, 129, 2560), key_shape: (1, 2, 385, 2560)\r\nseq_len: 259, query_shape: (1, 7, 37, 2560), key_shape: (1, 7, 293, 2560)\r\nseq_len: 260, query_shape: (1, 2, 130, 2560), key_shape: (1, 2, 386, 2560)\r\nseq_len: 261, query_shape: (1, 3, 87, 2560), key_shape: (1, 3, 343, 2560)\r\nseq_len: 262, query_shape: (1, 2, 131, 2560), key_shape: (1, 2, 387, 2560)\r\nseq_len: 263, query_shape: (1, 263, 1, 2560), key_shape: (1, 263, 257, 2560)\r\nseq_len: 264, query_shape: (1, 2, 132, 2560), key_shape: (1, 2, 388, 2560)\r\nseq_len: 265, query_shape: (1, 5, 53, 2560), key_shape: (1, 5, 309, 2560))\r\n```\r\n\r\nas you can see, because of the dynamic `block_length` the dimensions are very different for different `seq_length` which explains the irregular VRAM usage.\r\n\r\nif you set the seq_length to 1871 you'll get \r\n```\r\nseq_len: 1871, query_shape: (1, 1871, 1, 2560), key_shape: (1, 1871, 257, 2560)\r\n```\r\nas you posted above.\r\n\r\nSo I wouldn't say this is an implementation issue, that's how the local attention algorithm is designed in mesh-tf.", "Thank you for taking the time and walking me through the calculations. It makes sense and certainly explains the irregular pattern. However, I wonder if it is possible to reach the same end result in a way that is less memory intensive. A bit earlier I was looking for more information about local self-attention and I found this [implementation](https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py). Running it for a [1, 1871, 2560] tensor results in a peak allocation of just about 253MB:\r\n\r\n```\r\n>>> import torch\r\n>>> q = torch.rand(1,1871,128*20).to(torch.float32).cuda()\r\n>>> k = torch.rand(1,1871,128*20).to(torch.float32).cuda()\r\n>>> v = torch.rand(1,1871,128*20).to(torch.float32).cuda()\r\n>>> from local_attention import LocalAttention\r\n>>> local_attention = LocalAttention(256, causal=True, look_forward=0, autopad=True, dim=2560).cuda()\r\n>>> torch.cuda.memory_allocated(), torch.cuda.max_memory_allocated()\r\n(57482240, 57482240)\r\n>>> result = local_attention(q, k, v)\r\n>>> torch.cuda.memory_allocated(), torch.cuda.max_memory_allocated()\r\n(78453760, 252826624)\r\n```\r\n\r\nSimply running this implementation and GPTNeoLocalSelfAttention on the same input does seem to give different results however, so I think there may also be some difference between the algorithms.\r\n\r\nEdit: Experimenting with it a bit, I think my best bet is to just limit the sequence length to 1750. The padding approach of that approach is very different.", "I have thought more about it and think I have found a solution to reduce memory use.\r\n\r\n```\r\n--- src/transformers/models/gpt_neo/modeling_gpt_neo.py.backup\t2021-04-07 22:28:43.049493417 +0200\r\n+++ src/transformers/models/gpt_neo/modeling_gpt_neo.py\t2021-04-22 10:53:41.274276535 +0200\r\n@@ -413,4 +413,18 @@\r\n batch_size, seq_length = hidden_states.shape[:2]\r\n full_seq_length = seq_length + past_length\r\n+\r\n+ padding = None\r\n+ if layer_past is None and full_seq_length % self.window_size != 0 and full_seq_length > self.window_size:\r\n+ padding = self.window_size-(full_seq_length%self.window_size)\r\n+ if attention_mask is None:\r\n+ attention_mask = torch.zeros(query.shape[0], query.shape[1] + padding).to(query.device)\r\n+ attention_mask[:, padding:] = 1\r\n+ else:\r\n+ attention_mask = torch.cat([torch.zeros(attention_mask.shape[0], padding).to(attention_mask.device), attention_mask], axis=1)\r\n+ pad = lambda x: torch.cat([torch.zeros(x.shape[0],padding,x.shape[2]).to(x.device), x], axis=1)\r\n+ query, key, value = map(pad, (query, key, value))\r\n+ seq_length += padding\r\n+ full_seq_length += padding\r\n+\r\n block_length, num_blocks = self._get_block_length_and_num_blocks(full_seq_length, self.window_size)\r\n \r\n@@ -454,5 +468,9 @@\r\n attn_output = attn_output.reshape(batch_size, seq_length, self.embed_dim)\r\n \r\n- attn_output = self.out_proj(attn_output)\r\n+ if padding is not None:\r\n+ attn_output = attn_output[:,padding:]\r\n+ attn_weights = attn_weights[:,padding:]\r\n+\r\n+ attn_output = self.out_proj(attn_output.to(hidden_states.dtype))\r\n attn_output = self.resid_dropout(attn_output)\r\n\r\n```\r\n\r\nBy padding q, k and v and adding a mask to mask out the padding, the it becomes unnecessary to split things in a way that leads to a very large dimension. From how I see it, this should not change the result of the _attn function due to masking. For some reason I needed to add an extra .to at the end for running the model in fp16. First results of doing inference with this change look okay, but I am still testing it more. It's not updated with cfd2eaa8cf82da8581825c6592b66d2789c5bc53 yet.\r\n\r\nThe purple line here is a run with the patch applied:\r\n\r\n\r\n", "EricHallahan from EleutherAI was so kind to run the lambada evaluation task with the patch applied and found no degradation in accuracy and negligible differences in speed over multiple runs.", "(It is worth noting that LAMBADA task contexts are significantly shorter than 256 tokens. I believe Eric is currently running QA4MRE, which has much longer contexts)", "Hi @finetunenon\r\n\r\nThanks a lot for working on this. Let me run a few experiments to verify this and get back to you.", "Great, I ran a small test and it seems to be working! (sorry about the earlier comment). Here's the script\r\n\r\n```python\r\nimport torch\r\nfrom torch import nn\r\nfrom transformers.models.gpt_neo.modeling_gpt_neo import GPTNeoAttentionMixin\r\nfrom transformers import GPTNeoConfig\r\n\r\nclass GPTNeoLocalSelfAttention(nn.Module, GPTNeoAttentionMixin):\r\n def __init__(self, config):\r\n super().__init__()\r\n\r\n self.register_buffer(\"masked_bias\", torch.tensor(-1e9))\r\n\r\n self.attn_dropout = nn.Dropout(config.attention_dropout)\r\n self.resid_dropout = nn.Dropout(config.resid_dropout)\r\n\r\n self.embed_dim = config.hidden_size\r\n self.num_heads = config.num_heads\r\n self.head_dim = self.embed_dim // self.num_heads\r\n if self.head_dim * self.num_heads != self.embed_dim:\r\n raise ValueError(\r\n f\"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`: {self.num_heads}).\"\r\n )\r\n\r\n self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)\r\n self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)\r\n self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)\r\n self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=True)\r\n\r\n self.window_size = config.window_size\r\n\r\n def forward(\r\n self,\r\n hidden_states,\r\n attention_mask=None,\r\n layer_past=None,\r\n head_mask=None,\r\n use_cache=False,\r\n output_attentions=False,\r\n pad_qkv=False\r\n ):\r\n query = self.q_proj(hidden_states)\r\n\r\n if layer_past is not None:\r\n past = layer_past[0]\r\n key_value_hidden_states = torch.cat([past, hidden_states], dim=1)\r\n past_length = past.size()[1]\r\n else:\r\n key_value_hidden_states = hidden_states\r\n past_length = 0\r\n\r\n key = self.k_proj(key_value_hidden_states)\r\n value = self.v_proj(key_value_hidden_states)\r\n \r\n \r\n # compute block length and num_blocks\r\n batch_size, seq_length = hidden_states.shape[:2]\r\n full_seq_length = seq_length + past_length\r\n \r\n padding = None\r\n if pad_qkv:\r\n if layer_past is None and full_seq_length % self.window_size != 0 and full_seq_length > self.window_size:\r\n padding = self.window_size-(full_seq_length%self.window_size)\r\n if attention_mask is None:\r\n attention_mask = torch.zeros(query.shape[0], query.shape[1] + padding).to(query.device)\r\n attention_mask[:, padding:] = 1\r\n else:\r\n attention_mask = torch.cat([torch.zeros(attention_mask.shape[0], padding).to(attention_mask.device), attention_mask], axis=1)\r\n pad = lambda x: torch.cat([torch.zeros(x.shape[0],padding,x.shape[2]).to(x.device), x], axis=1)\r\n query, key, value = map(pad, (query, key, value))\r\n seq_length += padding\r\n full_seq_length += padding\r\n \r\n block_length, num_blocks = self._get_block_length_and_num_blocks(full_seq_length, self.window_size)\r\n \r\n # create buckets\r\n if layer_past is not None:\r\n # we just need 1 block with block_length 1 when caching is enabled\r\n query = self._split_seq_length_dim_to(query, 1, 1)\r\n else:\r\n query = self._split_seq_length_dim_to(query, num_blocks, block_length)\r\n\r\n key = self._look_back(key, block_length, self.window_size)\r\n value = self._look_back(value, block_length, self.window_size)\r\n\r\n # select key/value vectors only for the last block\r\n if layer_past is not None:\r\n key = key[:, -1:, ...]\r\n value = value[:, -1:, ...]\r\n\r\n query = self._split_heads(query, self.num_heads, self.head_dim)\r\n key = self._split_heads(key, self.num_heads, self.head_dim)\r\n value = self._split_heads(value, self.num_heads, self.head_dim)\r\n \r\n attention_mask = GPTNeoAttentionMixin.create_local_attention_mask(\r\n batch_size, full_seq_length, self.window_size, \"cpu\", attention_mask\r\n )\r\n\r\n if layer_past is not None:\r\n # only take the mask for the last block\r\n attention_mask = attention_mask[:, -1:, :, -1:, :]\r\n\r\n # attn\r\n attn_output, attn_weights = self._attn(\r\n query,\r\n key,\r\n value,\r\n causal_mask=attention_mask,\r\n masked_bias=self.masked_bias,\r\n attn_dropout=self.attn_dropout,\r\n head_mask=head_mask,\r\n )\r\n\r\n attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)\r\n attn_output = attn_output.reshape(batch_size, seq_length, self.embed_dim)\r\n \r\n if padding is not None:\r\n attn_output = attn_output[:,padding:]\r\n attn_weights = attn_weights[:,padding:]\r\n\r\n attn_output = self.out_proj(attn_output)\r\n attn_output = self.resid_dropout(attn_output)\r\n\r\n outputs = (attn_output,)\r\n if output_attentions:\r\n outputs += (attn_weights,)\r\n\r\n return outputs # a, (attentions)\r\n\r\nconfig = GPTNeoConfig(hidden_size=16, num_heads=4)\r\nattn_layer = GPTNeoLocalSelfAttention(config).eval()\r\n\r\nmatched = []\r\nwith torch.no_grad():\r\n for seq_len in range(1, 2049):\r\n hidden_states = torch.randn(1, seq_len, 16)\r\n out = attn_layer(hidden_states)[0]\r\n out_with_padding = attn_layer(hidden_states, pad_qkv=True)[0]\r\n matched.append(torch.allclose(out, out_with_padding, atol=1e-5))\r\n\r\nall(matched)\r\n# True\r\n```\r\nI will run a few tests with the actual model and will let you know. If it works, feel free to open a PR :)\r\n", "Thanks for testing. If it works with the actual model, how should cfd2eaa8cf82da8581825c6592b66d2789c5bc53 be handled? I tried adapting the patch, but the attention mask seems to work in a very different way and I haven't been able to figure it out yet.", "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." ]

[]

[]

[ "You should upgrade to the last version of transformers, which fully relies on the repository of a pretrained model instead of using special files like here.", "Thanks!" ]

[]

[]

[]

[]

[]

[ "Similar issue to https://github.com/huggingface/transformers/issues/11294", "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." ]

[]

[]

[ "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." ]

[]

[]

[ "This code seems to be using `simpletransformers`, sadly we won't dive into that. You could use the `run_translation.py` script [here](https://github.com/huggingface/transformers/tree/master/examples/seq2seq) which supports multi-gpu training. See this [doc](https://github.com/huggingface/transformers/tree/master/examples#distributed-training-and-mixed-precision) for distributed training using `Trainer`.\r\n\r\nAnd with `Trainer` you could also leverage `deepspeed` to get more efficiency, see this [doc](https://huggingface.co/transformers/main_classes/trainer.html#trainer-integrations) for `deepspeed` integration ", "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." ]

[]

[]

[ "All tests pass, the max_length was actually a bug hidden by `while cur_len < max_length` that was still in there.\r\nThe bart tests (at least) caught it automatically and enabled me to change it to the correct comparison ! ", "> All tests pass, the max_length was actually a bug hidden by `while cur_len < max_length` that was still in there.\r\n> The bart tests (at least) caught it automatically and enabled me to change it to the correct comparison !\r\n\r\nGreat! Feel free to merge then!" ]

[ { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[]

[]

[]

[]

[ "The I-BERT framework allows for easy fine-tuning in PyTorch to find the optimal parameters. Once those are found, the model can be deployed in a setup using int-8 capable operations such as TensorRT.\r\n\r\n@kssteven418 will be able to explain better than me.", "Yes, the current I-BERT implementation (both in HF and Fairseq in my personal repo) only searches for the optimal int8 parameters through quantization-aware training and leaves out the actual model deployment. That is to say, it simulates int8 inference using floating-point representations and operations. One reason we are not supporting int8 execution is that, PyTorch only supports int8 inference via its own quantization APIs. Therefore, the optimal parameters found in the I-BERT framework must be then exported to different frameworks that can support int8 deployment (TensorRT, TVM, etc. are some popular frameworks). We haven't yet open-sourced the code for model deployment.", "@kssteven418 Thanks for the answer. What I want is the final output in `int8`. Then will it be ok to just cast the output into `int8` even with pytorch? The whole execution doesn't have to be run in integer mode.", "Yes, if you look at the quantization modules, e.g., QuantLinear, there are additional attributes such as `weight_integer` that represent the int8 model parameters in the torch.float. You can cast those numbers to torch.int8, but just make sure that you don't round down the numbers - they must be rounded.", "> Yes, if you look at the quantization modules, e.g., QuantLinear, there are additional attributes such as `weight_integer` that represent the int8 model parameters in the torch.float. You can cast those numbers to torch.int8, but just make sure that you don't round down the numbers - they must be rounded.\r\n\r\nThank you very much. This will be the last question I expect :) \r\n\r\nThe following is how I defined my network and how the input passes flows. Briefly, I want the output of `IBertModel` to pass one more `QuantLinear` layer to get the final representation, like this. \r\n1. input -> `IBertModel`\r\n1. `IBertModel` -> `QuantAct` \r\n1. `QuantAct` -> `QuantLinear` -> `final_representation`\r\n\r\nBut the `QuantLinear` module outputs two tensors: `quant_x` and `scaling_factor`. \r\n\r\n**Do I have to deal with both, or can I just use `quant_x` as the final representation?** \r\n\r\n```python\r\nreturn (\r\n F.linear(x_int, weight=self.weight_integer, bias=self.bias_integer) * bias_scaling_factor, # use only this?\r\n bias_scaling_factor, # or do I also need this?\r\n)\r\n```\r\n\r\nThis is my code.\r\n```python\r\n# Define network\r\nself.pre_linear = QuantAct(self.act_bit, quant_mode=self.quant_mode)\r\n\r\nself.linear = QuantLinear(\r\n self.input_size,\r\n self.n,\r\n quant_mode=self.quant_mode,\r\n)\r\n\r\n...\r\n\r\n# Generate output\r\nx, pre_scaling_factor = self.pre_linear(x)\r\nx, scaling_factor = self.linear(x, pre_scaling_factor)\r\n\r\n# x = x * scaling_factor?\r\n```", "As you have also noticed, all the quant modules including QuantLinear return two tensors: `quant_x` and `scaling_factor`. Here, `quant_x` / `scaling_factor` represents the quantized (integer) value for the activation - in other words, `quant_x` is the dequantized value. Therefore, you do not have to multiply it with the `scaling_factor`. ", "Hi!\r\n\r\nI would like to deploy IBERT on a framework like TensorRT. I am a complete beginner in that field and I was wandering if someone could give me some tips on the main steps of how to quantize IBERT? \r\n\r\nThank you!", "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." ]

[]

[]

[ "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." ]

[]

[ { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "@patil-suraj Can you please check if this speed decrease of GPT2LMHeadModel model call is not caused by your PR #11225?", "Hi @LSinev \r\n\r\nThank you for posting the detailed issue. I will take a look.", "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." ]

[]

[]

[ "The notebook looks amazing!\r\n\r\nOut of curiosity, have you tried using `Trainer` to fine-tune `ViT`?", "Added a notebook that uses the Trainer. Includes a nice confusion matrix at the end :) " ]

[]

[]

[ "I think I solved the memory leakage issue. In my case, I simply wanted to update the index parameters in **self.retriever object**. So I used **def set_index function**. But I observe for some reason ray works completely cannot flush out the old index and its related objects.\r\n\r\nSo now when I want to update it, I delete the self.retriever object and re-initialize it. \r\n\r\n```\r\nclass RayRetriever:\r\n def __init__(self):\r\n self.initialized = False\r\n\r\n def create_rag_retriever(self, config, question_encoder_tokenizer, ctx_encoder_tokenizer,generator_tokenizer, index):\r\n if not self.initialized:\r\n self.retriever = RagRetriever(\r\n config,\r\n question_encoder_tokenizer=question_encoder_tokenizer,\r\n ctx_encoder_tokenizer=ctx_encoder_tokenizer,\r\n generator_tokenizer=generator_tokenizer,\r\n index=index,\r\n init_retrieval=False,\r\n )\r\n self.initialized = True\r\n\r\n def init_retrieval(self):\r\n self.retriever.index.init_index()\r\n\r\n def set_index(self,index):\r\n self.retriever.index=index #with this new index class all the paramters in HFindex class get updated\r\n \r\n\r\n def clear_object(self): #we can call this first delete all thing things and again call create_rag_retriever\r\n del self.retriever\r\n self.initialized = False\r\n \r\n\r\n def retrieve(self, question_hidden_states, n_docs):\r\n\r\n doc_ids, retrieved_doc_embeds = self.retriever._main_retrieve(question_hidden_states, n_docs)\r\n doc_dicts= self.retriever.index.get_doc_dicts(doc_ids)\r\n\r\n return doc_ids, retrieved_doc_embeds,doc_dicts\r\n``````" ]

[ { "id": 1990918270, "node_id": "MDU6TGFiZWwxOTkwOTE4Mjcw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20First%20Issue", "name": "Good First Issue", "color": "bbf794", "default": false, "description": "" }, { "id": 2392046359, "node_id": "MDU6TGFiZWwyMzkyMDQ2MzU5", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20Second%20Issue", "name": "Good Second Issue", "color": "dd935a", "default": false, "description": "Issues that are more difficult to do than \"Good First\" issues - give it a try if you want!" } ]

[]

[ "@patrickvonplaten @patil-suraj @sgugger ", "This sounds like a nice feature, but I sadly won't have time to work on it - let's see if someone in the community could be interested :-)", "There is something like this which may help : https://github.com/lumaku/espnet/blob/espnet2_ctc_segmentation/espnet2/bin/asr_align.py\r\n\r\nI need some help in integrating it to wav2vec2 in hugging face. ", "@theainerd are you working on this feature?", "I would also really like to see this feature.\r\n\r\n@theainerd I'd be happy to help in any way I can although I'm not too familiar with the Wav2Vec transformer.\r\n\r\n@patrickvonplaten do you think you could write out a brief outline of what you think the steps required would be?", "Hi there!\r\n\r\nI'm very very new to collaborating on open-source projects as well as on using huggingface/transformers in general therefore I'm not confident I can come up with a solution for this issue -- however I did some poking around with tutorials surrounding Wav2Vec2 and I was thinking of ways on how this might be implemented:\r\n\r\nIt seems like the Wav2Vec2FeatureExtractor does most of the heavylifting of converting the raw audio array to suitable input values\r\n\r\n-> These input values are then fed into the model to obtain the logits (Dimension of the output is observed to be dropped a considerable amount here)\r\n\r\n\r\n-> after applying argmax to obtain the IDs, these IDs are then fed back into the Wav2Vec2CTCTokenizer decode/batch_decode function to obtain the transcription.\r\n\r\n\r\nPerhaps information of the sampling rate should be stored within the Tokenizer class such that during decode it's able to make use of this information to determine the timestamp? Or it might be possible to store it within the Wav2Vec2Processor class and have some wrapper functions take care of determining the timestamp and including it during the decode step\r\n\r\nA relation of how the input values dimensions are mapped to the output logit's dimensions would be needed for this, which I don't have the expertise at the moment to figure out \r\n\r\nCC:\r\n@theainerd \r\n@MerryOscar \r\n@patrickvonplaten \r\n\r\nsources I've been referring to -- \r\nhttps://www.kdnuggets.com/2021/03/speech-text-wav2vec.html (I realise this is outdated with the old tokenizer class, which seems to perform feature extraction as well)\r\n\r\nhttps://huggingface.co/blog/fine-tune-wav2vec2-english\r\n", "+1 on this, i'd really appreciate timestamped words as well. the datasets like timit, etc. seem to have this info, but i guess that's part of their test set, not an output from the model itself. ", "Here's what i've found so far: \r\nif speech length is - 480,000\r\ninput_values lenth - 480,000\r\nlogits length - 1499\r\n\r\nthis was for a 30s audio file. \r\n`\r\nmodel = Wav2Vec2ForCTC\r\nprocessor = Wav2Vec2Processor\r\n\r\n input_values = processor(speech, return_tensors=\"pt\").input_values\r\n logits = model(input_values).logits\r\n`", "> Here's what i've found so far:\r\n> if speech length is - 480,000\r\n> input_values lenth - 480,000\r\n> logits length - 1499\r\n> \r\n> this was for a 30s audio file.\r\n> `\r\n> model = Wav2Vec2ForCTC\r\n> processor = Wav2Vec2Processor\r\n> \r\n> ```\r\n> input_values = processor(speech, return_tensors=\"pt\").input_values\r\n> logits = model(input_values).logits\r\n> ```\r\n> \r\n> `\r\n\r\nThanks for investigating on this -- while I think it may be possible to just use the ratio and sampling rate to derive the timestamp, what I'm afraid of is that this ratio might just be a \"magic number\" and might differ if there are variations in the configuration of the Wav2Vec2 model\r\n\r\nCurrent ratio from input_values size to logits seem to be around **320**\r\n\r\ne.g.:\r\nDoes the ratio change if the [hyperparameters](https://huggingface.co/transformers/model_doc/wav2vec2.html#transformers.Wav2Vec2Config) of the model are changed? \r\n\r\n\r\nIs this ratio constant for varying size of audio? (Experiment with different size WAV clips and check the ratio)\r\n", "> > Here's what i've found so far:\r\n> > if speech length is - 480,000\r\n> > input_values lenth - 480,000\r\n> > logits length - 1499\r\n> > this was for a 30s audio file.\r\n> > `\r\n> > model = Wav2Vec2ForCTC\r\n> > processor = Wav2Vec2Processor\r\n> > ```\r\n> > input_values = processor(speech, return_tensors=\"pt\").input_values\r\n> > logits = model(input_values).logits\r\n> > ```\r\n> > \r\n> > \r\n> > `\r\n> \r\n> Thanks for investigating on this -- while I think it may be possible to just use the ratio and sampling rate to derive the timestamp, what I'm afraid of is that this ratio might just be a \"magic number\" and might differ if there are variations in the configuration of the Wav2Vec2 model\r\n> \r\n> Current ratio from input_values size to logits seem to be around **320**\r\n> \r\n> e.g.:\r\n> Does the ratio change if the [hyperparameters](https://huggingface.co/transformers/model_doc/wav2vec2.html#transformers.Wav2Vec2Config) of the model are changed?\r\n> \r\n> Is this ratio constant for varying size of audio? (Experiment with different size WAV clips and check the ratio)\r\n\r\nMaybe @patrickvonplaten could shed some light of whether we are going in the right direction about this (if it's not too much trouble) 😓 🙏 ", "> > Here's what i've found so far:\r\n> > if speech length is - 480,000\r\n> > input_values lenth - 480,000\r\n> > logits length - 1499\r\n> > this was for a 30s audio file.\r\n> > `\r\n> > model = Wav2Vec2ForCTC\r\n> > processor = Wav2Vec2Processor\r\n> > ```\r\n> > input_values = processor(speech, return_tensors=\"pt\").input_values\r\n> > logits = model(input_values).logits\r\n> > ```\r\n> > \r\n> > \r\n> > `\r\n> \r\n> Thanks for investigating on this -- while I think it may be possible to just use the ratio and sampling rate to derive the timestamp, what I'm afraid of is that this ratio might just be a \"magic number\" and might differ if there are variations in the configuration of the Wav2Vec2 model\r\n> \r\n> Current ratio from input_values size to logits seem to be around **320**\r\n> \r\n> e.g.:\r\n> Does the ratio change if the [hyperparameters](https://huggingface.co/transformers/model_doc/wav2vec2.html#transformers.Wav2Vec2Config) of the model are changed?\r\n> \r\n> Is this ratio constant for varying size of audio? (Experiment with different size WAV clips and check the ratio)\r\n\r\nhey @yushao2, what ratio are you referring to here ? sorry, not too familiar with audio processing", "@patrickvonplaten @yushao2 following up on this", "> @patrickvonplaten @yushao2 following up on this\r\n\r\nHi there! Sorry for not being responsive here.\r\n\r\nThe ratio here refers to the number you get when you divide the size of ``input_values`` to the size of ``logits``\r\n\r\nin this case, you mentioned\r\n>input_values lenth - 480,000\r\n>logits length - 1499\r\n\r\nthe ratio would be 480000/1499 which is approximately 320", "Hello all,\r\n\r\nThere is something I have found which may serve as a good starting point. Basically this returns the time offsets and the textual data as well . \r\n\r\nhttps://github.com/lumaku/ctc-segmentation\r\n\r\n```python\r\n\r\nimport torch\r\nimport torchaudio\r\nfrom datasets import load_dataset, load_metric\r\nfrom transformers import Wav2Vec2ForCTC, Wav2Vec2Processor\r\nimport re\r\n\r\nfrom ctc_segmentation import ctc_segmentation\r\nfrom ctc_segmentation import CtcSegmentationParameters\r\nfrom ctc_segmentation import determine_utterance_segments\r\nfrom ctc_segmentation import prepare_text\r\n\r\n# Get the Wav2Vec2 model and the predicted text\r\ntest_dataset = load_dataset(\"common_voice\", \"hi\", split=\"test\")\r\nwer = load_metric(\"wer\")\r\n\r\nprocessor = Wav2Vec2Processor.from_pretrained(\"theainerd/Wav2Vec2-large-xlsr-hindi\")\r\nmodel = Wav2Vec2ForCTC.from_pretrained(\"theainerd/Wav2Vec2-large-xlsr-hindi\")\r\nmodel.to(\"cuda\")\r\n\r\nchars_to_ignore_regex = '[\\,\\?\\.\\!\\-\\;\\:\\\"\\“]'\r\n\r\nresampler = torchaudio.transforms.Resample(48_000, 16_000)\r\n\r\n# Preprocessing the datasets.\r\n# We need to read the aduio files as arrays\r\ndef speech_file_to_array_fn(batch):\r\n batch[\"sentence\"] = re.sub(chars_to_ignore_regex, '', batch[\"sentence\"]).lower()\r\n speech_array, sampling_rate = torchaudio.load(batch[\"path\"])\r\n batch[\"speech\"] = resampler(speech_array).squeeze().numpy()\r\n return batch\r\n\r\ntest_dataset = test_dataset.map(speech_file_to_array_fn)\r\n\r\ninput_values = processor(test_dataset[\"speech\"][0], return_tensors=\"pt\").input_values # Batch size 1\r\nlogits = model(input_values.to(\"cuda\")).logits\r\npredicted_ids = torch.argmax(logits, dim=-1)\r\ntranscription = processor.decode(predicted_ids[0])\r\n\r\nsoftmax = torch.nn.Softmax(dim = -1)\r\n\r\n# apply configuration\r\nconfig = CtcSegmentationParameters()\r\n\r\nwith torch.no_grad():\r\n # Apply ctc layer to obtain log character probabilities\r\n lpz = softmax(logits)[0].cpu().numpy()\r\n\r\nchar_dict = {\"न\": 0, \"च\": 1, \"थ\": 2, \"ी\": 3, \"ऐ\": 4, \"ृ\": 5, \"ध\": 6, \"य\": 7, \"ह\": 8, \"ऊ\": 9, \"म\": 10, \"ण\": 11, \"ै\": 13, \"ौ\": 14, \"ा\": 15, \"ल\": 16, \"त\": 17, \"इ\": 18, \"ढ़\": 19, \"ष\": 20, \"भ\": 21, \"ग़\": 22, \"ख\": 23, \"ड़\": 24, \"ए\": 25, \"व\": 26, \"ु\": 27, \"ओ\": 28, \"र\": 29, \"श\": 30, \"औ\": 31, \"ट\": 32, \"आ\": 33, \"ो\": 34, \"ढ\": 35, \"झ\": 36, \"ग\": 37, \"ज़\": 38, \"अ\": 39, \"े\": 40, \"प\": 41, \"घ\": 42, \"द\": 43, \"ई\": 44, \"फ़\": 45, \"ब\": 46, \"ड\": 47, \"ँ\": 48, \"छ\": 49, \"ू\": 50, \"फ\": 51, \"ि\": 52, \"स\": 53, \"्\": 54, \"क\": 55, \"उ\": 56, \"ठ\": 57, \"ं\": 58, \"़\": 59, \"ज\": 60, \"क़\": 61, \"|\": 12, \"[UNK]\": 62, \"[PAD]\": 63}\r\nchar_list = list(char_dict.keys())\r\n\r\n# Prepare the text for aligning\r\nground_truth_mat, utt_begin_indices = prepare_text(config, transcription,char_list)\r\n# Align using CTC segmentation\r\ntimings, char_probs, state_list = ctc_segmentation(config, lpz, ground_truth_mat)\r\n\r\n# Obtain list of utterances with time intervals and confidence score\r\nsegments = determine_utterance_segments(config, utt_begin_indices, char_probs, timings, transcription)\r\n# Sample Output : 0.26 1.73 -0.0154 THE SALE OF THE HOTELS * An example picked up from the ctc_segmentation \r\n```\r\nNow if I have the time offsets but how to get this for each and every word rather than the segments. _Please don't take this as an absolute solution_ as I am not sure that this is a good direction to go but still something is better than nothing. Please share your thoughts. \r\n", "Hi everyone, here is some sample code which I have created to get the word-level start and end timestamps.\r\nIt's surely a bit hacky, and I could imagine there being some special cases where it might break, but for the cases I have tried it with it worked great.\r\n\r\n```python\r\nfrom itertools import groupby\r\nimport torch\r\nfrom transformers import Wav2Vec2Processor, Wav2Vec2ForCTC\r\nimport soundfile as sf\r\n\r\n##############\r\n# load model & audio and run audio through model\r\n##############\r\nmodel_name = 'facebook/wav2vec2-large-960h-lv60-self'\r\nprocessor = Wav2Vec2Processor.from_pretrained(model_name)\r\nmodel = Wav2Vec2ForCTC.from_pretrained(model_name).cuda()\r\n\r\n\r\naudio_filepath = ''\r\nspeech, sample_rate = sf.read(audio_filepath)\r\ninput_values = processor(speech, sampling_rate=sample_rate, return_tensors=\"pt\").input_values.cuda()\r\n\r\nlogits = model(input_values).logits\r\n\r\npredicted_ids = torch.argmax(logits, dim=-1)\r\ntranscription = processor.decode(predicted_ids[0]).lower()\r\n\r\n##############\r\n# this is where the logic starts to get the start and end timestamp for each word\r\n##############\r\nwords = [w for w in transcription.split(' ') if len(w) > 0]\r\npredicted_ids = predicted_ids[0].tolist()\r\nduration_sec = input_values.shape[1] / sample_rate\r\n\r\n\r\nids_w_time = [(i / len(predicted_ids) * duration_sec, _id) for i, _id in enumerate(predicted_ids)]\r\n# remove entries which are just \"padding\" (i.e. no characers are recognized)\r\nids_w_time = [i for i in ids_w_time if i[1] != processor.tokenizer.pad_token_id]\r\n# now split the ids into groups of ids where each group represents a word\r\nsplit_ids_w_time = [list(group) for k, group\r\n in groupby(ids_w_time, lambda x: x[1] == processor.tokenizer.word_delimiter_token_id)\r\n if not k]\r\n\r\nassert len(split_ids_w_time) == len(words) # make sure that there are the same number of id-groups as words. Otherwise something is wrong\r\n\r\nword_start_times = []\r\nword_end_times = []\r\nfor cur_ids_w_time, cur_word in zip(split_ids_w_time, words):\r\n _times = [_time for _time, _id in cur_ids_w_time]\r\n word_start_times.append(min(_times))\r\n word_end_times.append(max(_times))\r\n \r\nwords, word_start_times, word_end_times\r\n```", "@KB-g \r\nCongrats!\r\nIs there a chance to also extract the \"per word probability\"?", "@KB-g \r\nThe `assert len() == len()` triggers. \r\nThis audio: [assert.zip](https://github.com/huggingface/transformers/files/6721402/assert.zip)\r\nTestcase:\r\n````python\r\nfrom itertools import groupby\r\nimport torch\r\nfrom transformers import Wav2Vec2Processor, Wav2Vec2ForCTC\r\nimport soundfile as sf\r\n\r\nmodel_name = 'DewiBrynJones/wav2vec2-large-xlsr-welsh'\r\nprocessor = Wav2Vec2Processor.from_pretrained(model_name)\r\nmodel = Wav2Vec2ForCTC.from_pretrained(model_name)\r\n\r\naudio_filepath = '/tmp/assert.wav'\r\nspeech, sample_rate = sf.read(audio_filepath)\r\ninput_values = processor(speech, sampling_rate=sample_rate, return_tensors=\"pt\").input_values\r\nlogits = model(input_values).logits\r\npredicted_ids = torch.argmax(logits, dim=-1)\r\ntranscription = processor.decode(predicted_ids[0]).lower()\r\n\r\n##############\r\n# this is where the logic starts to get the start and end timestamp for each word\r\n##############\r\nwords = [w for w in transcription.split(' ') if len(w) > 0]\r\npredicted_ids = predicted_ids[0].tolist()\r\nduration_sec = input_values.shape[1] / sample_rate\r\nids_w_time = [(i / len(predicted_ids) * duration_sec, _id) for i, _id in enumerate(predicted_ids)]\r\nids_w_time = [i for i in ids_w_time if i[1] != processor.tokenizer.pad_token_id]\r\nsplit_ids_w_time = [list(group) for k, group\r\n in groupby(ids_w_time, lambda x: x[1] == processor.tokenizer.word_delimiter_token_id)\r\n if not k]\r\n# make sure that there are the same number of id-groups as words. Otherwise something is wrong\r\nassert len(split_ids_w_time) == len(words), (len(split_ids_w_time), len(words))\r\n````", "> @KB-g Congrats! Is there a chance to also extract the \"per word probability\"?\r\n\r\nHey @KB-g \r\nAny success on this?", "Hi @doublex , @abhirooptalasila,\r\nI haven't tried to get the per-word probability. If you come up with a solution, it would be great if you could let me know. I'd also be interested in a solution :)", "Hi @KB-g, @doublex and @abhirooptalasila,\r\n\r\nmaybe this [tutorial](https://pytorch.org/audio/main/tutorials/forced_alignment_tutorial.html) helps to find out a way to calculate a \"per-word probability\". In the function `merge_words`, the author calculates scores for each word based on tokens probabilities and theirs duration. ", "We need to document the time stamp retrieval a bit better here I think", "@KB-g Thanks for the code snippet, really useful. Made a small addition (no_grad) for inference, would help people facing OOM error(s):\r\n\r\n```python\r\nfrom itertools import groupby\r\nimport torch\r\nfrom transformers import Wav2Vec2Processor, Wav2Vec2ForCTC\r\nimport soundfile as sf\r\n\r\n##############\r\n# load model & audio and run audio through model\r\n##############\r\nmodel_name = 'facebook/wav2vec2-large-960h-lv60-self'\r\nprocessor = Wav2Vec2Processor.from_pretrained(model_name)\r\nmodel = Wav2Vec2ForCTC.from_pretrained(model_name).cuda()\r\n\r\n\r\naudio_filepath = ''\r\nspeech, sample_rate = sf.read(audio_filepath)\r\ninput_values = processor(speech, sampling_rate=sample_rate, return_tensors=\"pt\").input_values.cuda()\r\n\r\nwith torch.no_grad():\r\n logits = model(input_values).logits\r\n\r\npredicted_ids = torch.argmax(logits, dim=-1)\r\ntranscription = processor.decode(predicted_ids[0]).lower()\r\n\r\n##############\r\n# this is where the logic starts to get the start and end timestamp for each word\r\n##############\r\nwords = [w for w in transcription.split(' ') if len(w) > 0]\r\npredicted_ids = predicted_ids[0].tolist()\r\nduration_sec = input_values.shape[1] / sample_rate\r\n\r\n\r\nids_w_time = [(i / len(predicted_ids) * duration_sec, _id) for i, _id in enumerate(predicted_ids)]\r\n# remove entries which are just \"padding\" (i.e. no characers are recognized)\r\nids_w_time = [i for i in ids_w_time if i[1] != processor.tokenizer.pad_token_id]\r\n# now split the ids into groups of ids where each group represents a word\r\nsplit_ids_w_time = [list(group) for k, group\r\n in groupby(ids_w_time, lambda x: x[1] == processor.tokenizer.word_delimiter_token_id)\r\n if not k]\r\n\r\nassert len(split_ids_w_time) == len(words) # make sure that there are the same number of id-groups as words. Otherwise something is wrong\r\n\r\nword_start_times = []\r\nword_end_times = []\r\nfor cur_ids_w_time, cur_word in zip(split_ids_w_time, words):\r\n _times = [_time for _time, _id in cur_ids_w_time]\r\n word_start_times.append(min(_times))\r\n word_end_times.append(max(_times))\r\n \r\nwords, word_start_times, word_end_times\r\n```", "@Ap1075, thank you for the example you provided above. I'm having a hard time figuring out where/how to pass in transcribed text so it can be aligned with the audio. Is passing in pre-transcribed text possible, or am I misunderstanding how it works?", "I'm trying to get word timing for karaoke I have the lyrics... Would this be possible? 🤔" ]

[]

[]

[ "I will take a deeper look tomorrow and comment here - looks great already :-)", "Looks great to me so far @anton-l - I see no breaking changes, the modularization looks good to me and the parameter naming is fine as well -> don't think `self.quantizer.quantizer` is too awkward => we also have `self.self.self_attention` somewhere in BERT ;-)", "I think an important next step would be to verify that the pretraining works more or less :-) ", "Integration tests are now passing. Can be verified by running:\r\n\r\n```python\r\n#!/usr/bin/env python3 \r\nimport datasets \r\nimport fairseq \r\nimport torch \r\n \r\nimport soundfile as sf \r\nimport sys \r\nfrom fairseq.criterions.wav2vec_criterion import Wav2VecCriterionConfig, Wav2vecCriterion \r\nfrom fairseq.tasks.audio_pretraining import AudioPretrainingConfig, AudioPretrainingTask \r\n \r\nfrom transformers import Wav2Vec2ForPreTraining, Wav2Vec2FeatureExtractor \r\n \r\nhf_path = str(sys.argv[1]) \r\nfairseq_wav2vec2_path = str(sys.argv[2])\r\n\r\nmodel, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([fairseq_wav2vec2_path])\r\n\r\nfeature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(hf_path, do_normalize=False)\r\nhf_model = Wav2Vec2ForPreTraining.from_pretrained(hf_path)\r\n\r\nmodel = model[0]\r\nmodel.eval()\r\n\r\n\r\ndummy_speech_data = datasets.load_dataset(\"patrickvonplaten/librispeech_asr_dummy\", \"clean\", split=\"validation\")\r\n\r\n\r\ndef map_to_array(batch):\r\n speech_array, _ = sf.read(batch[\"file\"])\r\n batch[\"speech\"] = speech_array\r\n return batch\r\n\r\n\r\ndummy_speech_data = dummy_speech_data.map(map_to_array, remove_columns=[\"file\"])\r\ninputs = feature_extractor(dummy_speech_data[:3][\"speech\"], return_tensors=\"pt\", padding=\"longest\", return_attention_mask=True)\r\n\r\ninput_values = inputs.input_values\r\nattention_mask = inputs.attention_mask\r\n\r\naudio_cfg = AudioPretrainingConfig(labels=\"ltr\", data=\"./data\")\r\ntask = AudioPretrainingTask.setup_task(audio_cfg) \r\ncriterion = Wav2vecCriterion(Wav2VecCriterionConfig(infonce=True, log_keys=[\"prob_perplexity\", \"code_perplexity\", \"temp\"], loss_weights=[0.1, 10]), task)\r\n\r\nsample = {\r\n \"net_input\": {\r\n \"source\": input_values,\r\n \"padding_mask\": attention_mask.ne(1),\r\n },\r\n \"id\": torch.zeros((1,)),\r\n}\r\n\r\ntorch.manual_seed(0)\r\nresult = model(**sample[\"net_input\"])\r\ntorch.manual_seed(0)\r\nhf_result = hf_model(input_values, attention_mask=attention_mask)\r\n\r\n\r\nassert torch.allclose(hf_result.logits, result['x'], atol=1e-3), \"wrong logits\"\r\n\r\nloss, sample_size, log = criterion(model, sample)\r\n\r\nprint(\"Loss diff %\", 100 * (loss.detach().item() - hf_result.loss.detach().item()) / hf_result.loss.detach())\r\n```\r\n\r\nand using [this](https://dl.fbaipublicfiles.com/fairseq/wav2vec/wav2vec_small.pt) as the fairseq checkpoint and [this](https://huggingface.co/patrickvonplaten/wav2vec2-base) model as the HF model." ]

[]

[]

[ "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.", "i have the same exact problem when i use `do_sample=True` can you re-open this issue?", "Maybe @gante has an idea!", "Hi there @Muennighoff @zeke-john 👋 \r\n\r\nI've run the script above for both models on `v4.5.1` (and on `v4.22.dev0`) and it works with no problems -- you can see a colab [here](https://colab.research.google.com/drive/1bg7v0mxZbFxJTjj28AriYBeERsAN264E?usp=sharing).\r\n\r\nA potential cause for errors may be GPU memory -- generation with `num_beams` is memory intensive. Let me know if you have more details about your problem :)" ]

[]

[]

[ "I created an env that cannot be copied", "Please do not create duplicates.\r\nDuplicate of #11301 ", "sorry, i close it now" ]

[]

[ { "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 }, { "login": "lhoestq", "id": 42851186, "node_id": "MDQ6VXNlcjQyODUxMTg2", "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhoestq", "html_url": "https://github.com/lhoestq", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "repos_url": "https://api.github.com/users/lhoestq/repos", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "type": "User", "site_admin": false } ]

[ "This indeed looks like a typo! Thanks for your issue :-)\r\n\r\nThe kwargs should not be trimmed by `\"question_question_encoder\"`, but by `\"question_encoder\"`.\r\n\r\nWould you like to open a PR to fix it? Otherwise I can do it as well :-)", "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.", "Unping" ]

[]

[]

[ "The probem only seems to happen on some models. This model here works ok: https://huggingface.co/T-Systems-onsite/cross-en-de-roberta-sentence-transformer", "Tagging @Pierrci for visibility", "Hi @PhilipMay, thank you for reporting this, we just deployed a fix for that, let us know if you still encounter the problem, otherwise feel free to close the issue :)", "Works now. Thanks. " ]

[]

[]

[ "What code are you running that leads to that error?", "# God, someone finally replied to me，thanks！\r\n## code\r\n```python\r\nfrom transformers import AutoModel, AutoTokenizer, pipeline\r\nimport torch\r\n\r\nmodel_name = 'pre-model/' + 'longformer-encdec-base-16384'\r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\nmodel = AutoModel.from_pretrained(model_name)\r\nclassifier = pipeline('feature-extraction', model=model, tokenizer=tokenizer)\r\n\r\n# encoded_inputs = tokenizer([\"ATGCATGCNACT\"], [\"ATGCATGCNACT\"], return_token_type_ids=True, return_tensors='pt')\r\nencoded_inputs = tokenizer([\"ATGCATGCNACT\", \"ATGCATG\", \"ACTGGTCATGCAC\"], return_tensors='pt',\r\n padding=True)\r\nprint(encoded_inputs)\r\n# feature = model(input_ids=encoded_inputs['input_ids'], attention_mask=encoded_inputs['attention_mask'],\r\n# return_netsors='pt')\r\nfeature = model(**encoded_inputs,\r\n return_netsors='pt')\r\nprint(feature[0])\r\nprint(type(feature[0]))\r\n# feature = torch.as_tensor(feature)\r\n# print(feature.shape)\r\nprint(\"***\" * 48)\r\n\r\nfeature = classifier([\"ATG\", \"ATGCATG\", \"ACTGGTCATGCAC\"])\r\nprint(type(feature))\r\nfeature = torch.as_tensor(feature)\r\nprint(feature)\r\nprint(feature.shape)\r\nprint(\"***\" * 48)\r\n\r\n```\r\n## env info\r\n\r\n### can work: env0\r\n```bash\r\n# Name Version Build Channel\r\n_libgcc_mutex 0.1 main defaults\r\nabsl-py 0.12.0 pypi_0 pypi\r\nastunparse 1.6.3 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nbiopython 1.78 pypi_0 pypi\r\nblas 1.0 mkl defaults\r\nboto3 1.17.48 pypi_0 pypi\r\nbotocore 1.20.48 pypi_0 pypi\r\nbrotlipy 0.7.0 py36h27cfd23_1003 defaults\r\nca-certificates 2021.1.19 h06a4308_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncachetools 4.2.1 pypi_0 pypi\r\ncertifi 2020.12.5 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncffi 1.14.5 py36h261ae71_0 defaults\r\nchardet 4.0.0 py36h06a4308_1003 defaults\r\nclick 7.1.2 pyhd3eb1b0_0 defaults\r\ncryptography 3.4.7 py36hd23ed53_0 defaults\r\ncudatoolkit 10.0.130 0 defaults\r\ndataclasses 0.8 pyh4f3eec9_6 defaults\r\ndill 0.3.3 pypi_0 pypi\r\nfilelock 3.0.12 pyhd3eb1b0_1 defaults\r\nfreetype 2.10.4 h5ab3b9f_0 defaults\r\nfuture 0.18.2 pypi_0 pypi\r\ngoogle-auth 1.28.1 pypi_0 pypi\r\ngoogle-auth-oauthlib 0.4.4 pypi_0 pypi\r\ngrpcio 1.37.0 pypi_0 pypi\r\nidna 2.10 pyhd3eb1b0_0 defaults\r\nimageio 2.9.0 pypi_0 pypi\r\nimportlib-metadata 3.10.0 pypi_0 pypi\r\nintel-openmp 2020.2 254 defaults\r\njmespath 0.10.0 pypi_0 pypi\r\njoblib 1.0.1 pyhd3eb1b0_0 defaults\r\njpeg 9b h024ee3a_2 defaults\r\nlcms2 2.12 h3be6417_0 defaults\r\nld_impl_linux-64 2.33.1 h53a641e_7 defaults\r\nlibffi 3.3 he6710b0_2 defaults\r\nlibgcc-ng 9.1.0 hdf63c60_0 defaults\r\nlibpng 1.6.37 hbc83047_0 defaults\r\nlibprotobuf 3.14.0 h8c45485_0 defaults\r\nlibstdcxx-ng 9.1.0 hdf63c60_0 defaults\r\nlibtiff 4.1.0 h2733197_1 defaults\r\nlongformer 0.1 pypi_0 pypi\r\nlz4-c 1.9.3 h2531618_0 defaults\r\nmarkdown 3.3.4 pypi_0 pypi\r\nmkl 2020.2 256 defaults\r\nmkl-service 2.3.0 py36he8ac12f_0 defaults\r\nmkl_fft 1.3.0 py36h54f3939_0 defaults\r\nmkl_random 1.1.1 py36h0573a6f_0 defaults\r\nncurses 6.2 he6710b0_1 defaults\r\nninja 1.10.2 py36hff7bd54_0 defaults\r\nnlp 0.4.0 pypi_0 pypi\r\nnltk 3.6.1 pypi_0 pypi\r\nnumpy 1.19.5 pypi_0 pypi\r\nnumpy-base 1.19.2 py36hfa32c7d_0 defaults\r\noauthlib 3.1.0 pypi_0 pypi\r\nolefile 0.46 py36_0 defaults\r\nopenssl 1.1.1k h27cfd23_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npackaging 20.9 pyhd3eb1b0_0 defaults\r\npandas 1.1.5 pypi_0 pypi\r\npatsy 0.5.1 pypi_0 pypi\r\npillow 8.2.0 py36he98fc37_0 defaults\r\npip 21.0.1 py36h06a4308_0 defaults\r\nprotobuf 3.15.8 pypi_0 pypi\r\npyahocorasick 1.4.2 pypi_0 pypi\r\npyarrow 3.0.0 pypi_0 pypi\r\npyasn1 0.4.8 pypi_0 pypi\r\npyasn1-modules 0.2.8 pypi_0 pypi\r\npybedtools 0.8.2 pypi_0 pypi\r\npycparser 2.20 py_2 defaults\r\npyopenssl 20.0.1 pyhd3eb1b0_1 defaults\r\npyparsing 2.4.7 pyhd3eb1b0_0 defaults\r\npysam 0.16.0.1 pypi_0 pypi\r\npysocks 1.7.1 py36h06a4308_0 defaults\r\npython 3.6.13 hdb3f193_0 defaults\r\npython-dateutil 2.8.1 pypi_0 pypi\r\npython_abi 3.6 1_cp36m huggingface\r\npytorch-lightning 0.8.5 pypi_0 pypi\r\npytorch-transformers 1.2.0 pypi_0 pypi\r\npytz 2021.1 pypi_0 pypi\r\npyyaml 5.4.1 pypi_0 pypi\r\nreadline 8.1 h27cfd23_0 defaults\r\nregex 2021.4.4 py36h27cfd23_0 defaults\r\nrequests 2.25.1 pyhd3eb1b0_0 defaults\r\nrequests-oauthlib 1.3.0 pypi_0 pypi\r\nrouge-score 0.0.4 pypi_0 pypi\r\nrsa 4.7.2 pypi_0 pypi\r\ns3transfer 0.3.6 pypi_0 pypi\r\nsacremoses 0.0.44 pypi_0 pypi\r\nscikit-learn 0.24.1 pypi_0 pypi\r\nscipy 1.5.4 pypi_0 pypi\r\nsentencepiece 0.1.91 pypi_0 pypi\r\nseqeval 1.2.2 pypi_0 pypi\r\nsetuptools 52.0.0 py36h06a4308_0 defaults\r\nsix 1.15.0 py36h06a4308_0 defaults\r\nsqlite 3.35.4 hdfb4753_0 defaults\r\nstatsmodels 0.12.2 pypi_0 pypi\r\ntensorboard 2.4.1 pypi_0 pypi\r\ntensorboard-plugin-wit 1.8.0 pypi_0 pypi\r\ntensorboardx 2.2 pypi_0 pypi\r\ntest-tube 0.7.5 pypi_0 pypi\r\nthreadpoolctl 2.1.0 pypi_0 pypi\r\ntk 8.6.10 hbc83047_0 defaults\r\ntokenizers 0.5.0 pypi_0 pypi\r\ntorch 1.6.0 pypi_0 pypi\r\ntorchvision 0.5.0 py36_cu100 pytorch\r\ntqdm 4.60.0 pypi_0 pypi\r\ntransformers 3.1.0 pypi_0 pypi\r\ntyping-extensions 3.7.4.3 pypi_0 pypi\r\nurllib3 1.26.4 pyhd3eb1b0_0 defaults\r\nwerkzeug 1.0.1 pypi_0 pypi\r\nwheel 0.36.2 pyhd3eb1b0_0 defaults\r\nxxhash 2.0.2 pypi_0 pypi\r\nxz 5.2.5 h7b6447c_0 defaults\r\nzipp 3.4.1 pypi_0 pypi\r\nzlib 1.2.11 h7b6447c_3 defaults\r\nzstd 1.4.9 haebb681_0 defaults\r\n```\r\n### can not work\r\n\r\n#### env1：tf2-pt-keras\r\n```bash\r\n# Name Version Build Channel\r\n_libgcc_mutex 0.1 main https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\n_tflow_select 2.1.0 gpu https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nabsl-py 0.11.0 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\naiohttp 3.6.3 py36h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\napex 0.1 pypi_0 pypi\r\nargon2-cffi 20.1.0 py36h7b6447c_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nastor 0.8.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nastunparse 1.6.3 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nasync-timeout 3.0.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nasync_generator 1.10 py36h28b3542_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nattrs 20.3.0 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nbackcall 0.2.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nbert-serving-client 1.10.0 pypi_0 pypi\r\nbert-serving-server 1.10.0 pypi_0 pypi\r\nblas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nbleach 3.2.1 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nblinker 1.4 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nbrotlipy 0.7.0 py36h27cfd23_1003 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nc-ares 1.16.1 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nca-certificates 2021.4.13 h06a4308_1 defaults\r\ncachetools 4.1.1 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncertifi 2020.12.5 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncffi 1.14.3 py36h261ae71_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nchardet 3.0.4 py36h06a4308_1003 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nclick 7.1.2 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncryptography 3.2.1 py36h3c74f83_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncudatoolkit 10.1.243 h6bb024c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncudnn 7.6.5 cuda10.1_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncupti 10.1.168 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncycler 0.10.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ndataclasses 0.7 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ndbus 1.13.18 hb2f20db_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ndecorator 4.4.2 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ndefusedxml 0.6.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nentrypoints 0.3 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nexpat 2.2.10 he6710b0_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nfilelock 3.0.12 pyhd3eb1b0_1 defaults\r\nfontconfig 2.13.0 h9420a91_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nfreetype 2.10.4 h5ab3b9f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngast 0.2.2 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nglib 2.66.1 h92f7085_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngoogle-auth 1.23.0 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngoogle-auth-oauthlib 0.4.2 pyhd3eb1b0_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngoogle-pasta 0.2.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngputil 1.4.0 pypi_0 pypi\r\ngrpcio 1.31.0 py36hf8bcb03_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngst-plugins-base 1.14.0 hbbd80ab_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ngstreamer 1.14.0 hb31296c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nh5py 2.10.0 py36hd6299e0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nhdf5 1.10.6 hb1b8bf9_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nicu 58.2 he6710b0_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nidna 2.10 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nidna_ssl 1.1.0 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nimportlib-metadata 2.0.0 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nimportlib_metadata 2.0.0 1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nintel-openmp 2020.2 254 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nipykernel 5.3.4 py36h5ca1d4c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nipython 7.12.0 py36h5ca1d4c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge\r\nipython_genutils 0.2.0 pyhd3eb1b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nipywidgets 7.6.0 pyhd3eb1b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njedi 0.10.2 py36_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free\r\njinja2 2.11.2 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njoblib 0.17.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njpeg 9b h024ee3a_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njsonschema 3.2.0 py_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njupyter 1.0.0 py36_7 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njupyter_client 6.1.7 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njupyter_console 6.2.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njupyter_core 4.7.0 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njupyterlab_pygments 0.1.2 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nkeras 2.3.1 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nkeras-applications 1.0.8 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nkeras-base 2.3.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nkeras-preprocessing 1.1.0 py_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nkiwisolver 1.3.0 py36h2531618_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nkrb5 1.18.2 h173b8e3_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlcms2 2.11 h396b838_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nld_impl_linux-64 2.33.1 h53a641e_7 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibcurl 7.71.1 h20c2e04_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibedit 3.1.20191231 h14c3975_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibffi 3.3 he6710b0_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibgcc-ng 9.1.0 hdf63c60_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibgfortran-ng 7.3.0 hdf63c60_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibpng 1.6.37 hbc83047_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibprotobuf 3.13.0.1 hd408876_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibsodium 1.0.18 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibssh2 1.9.0 h1ba5d50_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibstdcxx-ng 9.1.0 hdf63c60_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibtiff 4.1.0 h2733197_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibuuid 1.0.3 h1bed415_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibuv 1.40.0 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibxcb 1.14 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibxml2 2.9.10 hb55368b_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlz4-c 1.9.2 heb0550a_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmarkdown 3.3.3 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmarkupsafe 1.1.1 py36h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmatplotlib 3.3.2 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmatplotlib-base 3.3.2 py36h817c723_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmistune 0.8.4 py36h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl 2020.2 256 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl-service 2.3.0 py36he904b0f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl_fft 1.2.0 py36h23d657b_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl_random 1.1.1 py36h0573a6f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmultidict 4.7.6 py36h7b6447c_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnbclient 0.5.1 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnbconvert 6.0.7 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnbformat 5.0.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nncurses 6.2 he6710b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnest-asyncio 1.4.3 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nninja 1.10.1 py36hfd86e86_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnotebook 6.1.6 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnumpy 1.19.2 py36h54aff64_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnumpy-base 1.19.2 py36hfa32c7d_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\noauthlib 3.1.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nolefile 0.46 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nopenssl 1.1.1k h27cfd23_0 defaults\r\nopt_einsum 3.1.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npackaging 20.8 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npandas 1.1.3 py36he6710b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npandoc 2.11 hb0f4dca_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npandocfilters 1.4.3 py36h06a4308_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npcre 8.44 he6710b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npexpect 4.8.0 pyhd3eb1b0_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npickleshare 0.7.5 pyhd3eb1b0_1003 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npillow 8.0.1 py36he98fc37_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npip 20.2.4 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nprometheus_client 0.9.0 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nprompt-toolkit 3.0.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nprompt_toolkit 3.0.8 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nprotobuf 3.13.0.1 py36he6710b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nptyprocess 0.6.0 pyhd3eb1b0_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyasn1 0.4.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyasn1-modules 0.2.8 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npycparser 2.20 py_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npygments 2.7.3 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyjwt 1.7.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyopenssl 19.1.0 pyhd3eb1b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyparsing 2.4.7 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyqt 5.9.2 py36h05f1152_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyrsistent 0.17.3 py36h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npysocks 1.7.1 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npython 3.6.12 hcff3b4d_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npython-dateutil 2.8.1 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npython_abi 3.6 1_cp36m https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge\r\npytorch 1.7.0 py3.6_cuda10.1.243_cudnn7.6.3_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch\r\npytz 2020.1 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyyaml 5.3.1 py36h7b6447c_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyzmq 20.0.0 py36h2531618_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nqt 5.9.7 h5867ecd_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nqtconsole 4.7.7 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nqtpy 1.9.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nreadline 8.0 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nregex 2021.4.4 py36h27cfd23_0 defaults\r\nrequests 2.24.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nrequests-oauthlib 1.3.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nrsa 4.6 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsacremoses 0.0.44 pypi_0 pypi\r\nscikit-learn 0.23.2 py36h0573a6f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nscipy 1.5.2 py36h0b6359f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nseaborn 0.11.1 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsend2trash 1.5.0 pyhd3eb1b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsetuptools 50.3.1 py36h06a4308_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsip 4.19.8 py36hf484d3e_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsix 1.15.0 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsqlite 3.33.0 h62c20be_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorboard 2.3.0 pyh4dce500_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorboard-plugin-wit 1.6.0 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorflow 2.1.0 gpu_py36h2e5cdaa_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorflow-base 2.1.0 gpu_py36h6c5654b_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorflow-estimator 2.1.0 pyhd54b08b_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorflow-gpu 2.1.0 h0d30ee6_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntermcolor 1.1.0 py36_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nterminado 0.9.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntestpath 0.4.4 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nthreadpoolctl 2.1.0 pyh5ca1d4c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntk 8.6.10 hbc83047_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntokenizers 0.10.2 pypi_0 pypi\r\ntorchaudio 0.7.0 py36 https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch\r\ntorchvision 0.1.8 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free\r\ntornado 6.0.4 py36h7b6447c_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntqdm 4.60.0 pypi_0 pypi\r\ntraitlets 4.3.3 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntransformers 4.4.2 py_0 huggingface\r\ntyping_extensions 3.7.4.3 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nurllib3 1.25.11 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwcwidth 0.2.5 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwebencodings 0.5.1 py36_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwerkzeug 1.0.1 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwheel 0.35.1 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwidgetsnbextension 3.5.1 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwrapt 1.12.1 py36h7b6447c_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nxz 5.2.5 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nyaml 0.2.5 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nyarl 1.6.2 py36h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nzeromq 4.3.3 he6710b0_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nzipp 3.4.0 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nzlib 1.2.11 h7b6447c_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nzstd 1.4.5 h9ceee32_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\n\r\n```\r\n### env2: copied from env0 but not worked\r\n```bash\r\n# Name Version Build Channel\r\n_libgcc_mutex 0.1 main https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nabsl-py 0.12.0 pypi_0 pypi\r\nastunparse 1.6.3 py_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nblas 1.0 mkl https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nboto3 1.17.53 pypi_0 pypi\r\nbotocore 1.20.53 pypi_0 pypi\r\nbrotlipy 0.7.0 py36h27cfd23_1003 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nca-certificates 2021.4.13 h06a4308_1 \r\ncachetools 4.2.1 pypi_0 pypi\r\ncertifi 2020.12.5 py36h06a4308_0 \r\ncffi 1.14.5 py36h261ae71_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nchardet 4.0.0 py36h06a4308_1003 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nclick 7.1.2 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncryptography 3.4.7 py36hd23ed53_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ncudatoolkit 10.0.130 0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ndataclasses 0.8 pyh4f3eec9_6 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ndill 0.3.3 pypi_0 pypi\r\nfilelock 3.0.12 pyhd3eb1b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nfreetype 2.10.4 h5ab3b9f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nfuture 0.18.2 pypi_0 pypi\r\ngoogle-auth 1.29.0 pypi_0 pypi\r\ngoogle-auth-oauthlib 0.4.4 pypi_0 pypi\r\ngrpcio 1.37.0 pypi_0 pypi\r\nidna 2.10 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nimageio 2.9.0 pypi_0 pypi\r\nimportlib-metadata 2.0.0 py_1 anaconda\r\nintel-openmp 2020.2 254 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njmespath 0.10.0 pypi_0 pypi\r\njoblib 1.0.1 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\njpeg 9b h024ee3a_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlcms2 2.12 h3be6417_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nld_impl_linux-64 2.33.1 h53a641e_7 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibffi 3.3 he6710b0_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibgcc-ng 9.1.0 hdf63c60_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibpng 1.6.37 hbc83047_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibprotobuf 3.14.0 h8c45485_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibstdcxx-ng 9.1.0 hdf63c60_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlibtiff 4.1.0 h2733197_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nlongformer 0.1 pypi_0 pypi\r\nlz4-c 1.9.3 h2531618_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmarkdown 3.3.4 pypi_0 pypi\r\nmkl 2020.2 256 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl-service 2.3.0 py36he8ac12f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl_fft 1.3.0 py36h54f3939_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nmkl_random 1.1.1 py36h0573a6f_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nncurses 6.2 he6710b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nninja 1.10.2 py36hff7bd54_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnlp 0.4.0 pypi_0 pypi\r\nnltk 3.6.1 pypi_0 pypi\r\nnumpy 1.19.2 py36h54aff64_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nnumpy-base 1.19.2 py36hfa32c7d_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\noauthlib 3.1.0 pypi_0 pypi\r\nolefile 0.46 py36_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nopenssl 1.1.1k h27cfd23_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npackaging 20.9 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npandas 1.1.5 pypi_0 pypi\r\npillow 8.2.0 py36he98fc37_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npip 21.0.1 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nprotobuf 3.15.8 pypi_0 pypi\r\npyarrow 3.0.0 pypi_0 pypi\r\npyasn1 0.4.8 pypi_0 pypi\r\npyasn1-modules 0.2.8 pypi_0 pypi\r\npycparser 2.20 py_2 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyopenssl 20.0.1 pyhd3eb1b0_1 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npyparsing 2.4.7 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npysocks 1.7.1 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npython 3.6.13 hdb3f193_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\npython-dateutil 2.8.1 pypi_0 pypi\r\npython_abi 3.6 1_cp36m huggingface\r\npytorch-lightning 0.8.5 pypi_0 pypi\r\npytorch-transformers 1.2.0 pypi_0 pypi\r\npytz 2021.1 pypi_0 pypi\r\npyyaml 5.4.1 pypi_0 pypi\r\nreadline 8.1 h27cfd23_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nregex 2021.4.4 py36h27cfd23_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nrequests 2.25.1 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nrequests-oauthlib 1.3.0 pypi_0 pypi\r\nrouge-score 0.0.4 pypi_0 pypi\r\nrsa 4.7.2 pypi_0 pypi\r\ns3transfer 0.3.7 pypi_0 pypi\r\nsacremoses 0.0.44 pypi_0 pypi\r\nsentencepiece 0.1.95 pypi_0 pypi\r\nsetuptools 52.0.0 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsix 1.15.0 py36h06a4308_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nsqlite 3.35.4 hdfb4753_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntensorboard 2.4.1 pypi_0 pypi\r\ntensorboard-plugin-wit 1.8.0 pypi_0 pypi\r\ntensorboardx 2.2 pypi_0 pypi\r\ntest-tube 0.7.5 pypi_0 pypi\r\ntk 8.6.10 hbc83047_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\ntokenizers 0.8.1rc2 pypi_0 pypi\r\ntorch 1.6.0 pypi_0 pypi\r\ntorchvision 0.5.0 py36_cu100 pytorch\r\ntqdm 4.60.0 pypi_0 pypi\r\ntransformers 3.1.0 pypi_0 pypi\r\nurllib3 1.26.4 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nwerkzeug 1.0.1 pypi_0 pypi\r\nwheel 0.36.2 pyhd3eb1b0_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nxxhash 2.0.2 pypi_0 pypi\r\nxz 5.2.5 h7b6447c_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nzipp 3.4.1 pyhd3eb1b0_0 \r\nzlib 1.2.11 h7b6447c_3 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\nzstd 1.4.9 haebb681_0 https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main\r\n```", "next step i wan to use gene seqs to pretrain longformer, but i has been seemly dead in step 0 ....", "Can you please post the output of:\r\n```\r\ntype(model)\r\n```\r\nof your working environment?\r\nIn case it is showing something with `....BartModel`, can you please show us the definition of the class BertEncoder? You can locate it in the directory of:\r\n```\r\nimport transformers\r\nprint(transformers.__file__)\r\n```\r\n", "> Can you please post the output of:\r\n> \r\n> ```\r\n> type(model)\r\n> ```\r\n> \r\n> of your working environment?\r\n> In case it is showing something with `....BartModel`, can you please show us the definition of the class BertEncoder? You can locate it in the directory of:\r\n> \r\n> ```\r\n> import transformers\r\n> print(transformers.__file__)\r\n> ```\r\n\r\n# code\r\n```python\r\nfrom transformers import AutoModel, AutoTokenizer # , pipeline\r\nimport transformers\r\nprint(transformers.__file__)\r\n\r\n\r\nmodel_name = 'pre-model/' + 'longformer-encdec-base-16384'\r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\nmodel = AutoModel.from_pretrained(model_name)\r\n# classifier = pipeline('feature-extraction', model=model, tokenizer=tokenizer)\r\n\r\nprint(type(model))\r\n\r\n\r\n```\r\n\r\n# env0:\r\n```bash\r\n/home/pbc/anaconda3/envs/dnabert/lib/python3.6/site-packages/transformers/__init__.py\r\nSome weights of the model checkpoint at pre-model/longformer-encdec-base-16384 were not used when initializing BartModel: ['model.encoder.layers.0.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.0.self_attn.output.weight', 'model.encoder.layers.0.self_attn.output.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.1.self_attn.output.weight', 'model.encoder.layers.1.self_attn.output.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.2.self_attn.output.weight', 'model.encoder.layers.2.self_attn.output.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.3.self_attn.output.weight', 'model.encoder.layers.3.self_attn.output.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.4.self_attn.output.weight', 'model.encoder.layers.4.self_attn.output.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.5.self_attn.output.weight', 'model.encoder.layers.5.self_attn.output.bias']\r\n- This IS expected if you are initializing BartModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).\r\n- This IS NOT expected if you are initializing BartModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of BartModel were not initialized from the model checkpoint at pre-model/longformer-encdec-base-16384 and are newly initialized: ['model.encoder.layers.0.self_attn.k_proj.weight', 'model.encoder.layers.0.self_attn.k_proj.bias', 'model.encoder.layers.0.self_attn.v_proj.weight', 'model.encoder.layers.0.self_attn.v_proj.bias', 'model.encoder.layers.0.self_attn.q_proj.weight', 'model.encoder.layers.0.self_attn.q_proj.bias', 'model.encoder.layers.0.self_attn.out_proj.weight', 'model.encoder.layers.0.self_attn.out_proj.bias', 'model.encoder.layers.1.self_attn.k_proj.weight', 'model.encoder.layers.1.self_attn.k_proj.bias', 'model.encoder.layers.1.self_attn.v_proj.weight', 'model.encoder.layers.1.self_attn.v_proj.bias', 'model.encoder.layers.1.self_attn.q_proj.weight', 'model.encoder.layers.1.self_attn.q_proj.bias', 'model.encoder.layers.1.self_attn.out_proj.weight', 'model.encoder.layers.1.self_attn.out_proj.bias', 'model.encoder.layers.2.self_attn.k_proj.weight', 'model.encoder.layers.2.self_attn.k_proj.bias', 'model.encoder.layers.2.self_attn.v_proj.weight', 'model.encoder.layers.2.self_attn.v_proj.bias', 'model.encoder.layers.2.self_attn.q_proj.weight', 'model.encoder.layers.2.self_attn.q_proj.bias', 'model.encoder.layers.2.self_attn.out_proj.weight', 'model.encoder.layers.2.self_attn.out_proj.bias', 'model.encoder.layers.3.self_attn.k_proj.weight', 'model.encoder.layers.3.self_attn.k_proj.bias', 'model.encoder.layers.3.self_attn.v_proj.weight', 'model.encoder.layers.3.self_attn.v_proj.bias', 'model.encoder.layers.3.self_attn.q_proj.weight', 'model.encoder.layers.3.self_attn.q_proj.bias', 'model.encoder.layers.3.self_attn.out_proj.weight', 'model.encoder.layers.3.self_attn.out_proj.bias', 'model.encoder.layers.4.self_attn.k_proj.weight', 'model.encoder.layers.4.self_attn.k_proj.bias', 'model.encoder.layers.4.self_attn.v_proj.weight', 'model.encoder.layers.4.self_attn.v_proj.bias', 'model.encoder.layers.4.self_attn.q_proj.weight', 'model.encoder.layers.4.self_attn.q_proj.bias', 'model.encoder.layers.4.self_attn.out_proj.weight', 'model.encoder.layers.4.self_attn.out_proj.bias', 'model.encoder.layers.5.self_attn.k_proj.weight', 'model.encoder.layers.5.self_attn.k_proj.bias', 'model.encoder.layers.5.self_attn.v_proj.weight', 'model.encoder.layers.5.self_attn.v_proj.bias', 'model.encoder.layers.5.self_attn.q_proj.weight', 'model.encoder.layers.5.self_attn.q_proj.bias', 'model.encoder.layers.5.self_attn.out_proj.weight', 'model.encoder.layers.5.self_attn.out_proj.bias']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\n<class 'transformers.modeling_bart.BartModel'>\r\n```\r\n# env1:tf2-pt-keras\r\n```bash\r\n/home/pbc/anaconda3/envs/tf2_pt_kr2/lib/python3.6/site-packages/transformers-4.4.2-py3.8.egg/transformers/__init__.py\r\nSome weights of the model checkpoint at pre-model/longformer-encdec-base-16384 were not used when initializing BartModel: ['model.encoder.layers.0.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.0.self_attn.output.weight', 'model.encoder.layers.0.self_attn.output.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.1.self_attn.output.weight', 'model.encoder.layers.1.self_attn.output.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.2.self_attn.output.weight', 'model.encoder.layers.2.self_attn.output.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.3.self_attn.output.weight', 'model.encoder.layers.3.self_attn.output.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.4.self_attn.output.weight', 'model.encoder.layers.4.self_attn.output.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.5.self_attn.output.weight', 'model.encoder.layers.5.self_attn.output.bias']\r\n- This IS expected if you are initializing BartModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing BartModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of BartModel were not initialized from the model checkpoint at pre-model/longformer-encdec-base-16384 and are newly initialized: ['model.encoder.layers.0.self_attn.k_proj.weight', 'model.encoder.layers.0.self_attn.k_proj.bias', 'model.encoder.layers.0.self_attn.v_proj.weight', 'model.encoder.layers.0.self_attn.v_proj.bias', 'model.encoder.layers.0.self_attn.q_proj.weight', 'model.encoder.layers.0.self_attn.q_proj.bias', 'model.encoder.layers.0.self_attn.out_proj.weight', 'model.encoder.layers.0.self_attn.out_proj.bias', 'model.encoder.layers.1.self_attn.k_proj.weight', 'model.encoder.layers.1.self_attn.k_proj.bias', 'model.encoder.layers.1.self_attn.v_proj.weight', 'model.encoder.layers.1.self_attn.v_proj.bias', 'model.encoder.layers.1.self_attn.q_proj.weight', 'model.encoder.layers.1.self_attn.q_proj.bias', 'model.encoder.layers.1.self_attn.out_proj.weight', 'model.encoder.layers.1.self_attn.out_proj.bias', 'model.encoder.layers.2.self_attn.k_proj.weight', 'model.encoder.layers.2.self_attn.k_proj.bias', 'model.encoder.layers.2.self_attn.v_proj.weight', 'model.encoder.layers.2.self_attn.v_proj.bias', 'model.encoder.layers.2.self_attn.q_proj.weight', 'model.encoder.layers.2.self_attn.q_proj.bias', 'model.encoder.layers.2.self_attn.out_proj.weight', 'model.encoder.layers.2.self_attn.out_proj.bias', 'model.encoder.layers.3.self_attn.k_proj.weight', 'model.encoder.layers.3.self_attn.k_proj.bias', 'model.encoder.layers.3.self_attn.v_proj.weight', 'model.encoder.layers.3.self_attn.v_proj.bias', 'model.encoder.layers.3.self_attn.q_proj.weight', 'model.encoder.layers.3.self_attn.q_proj.bias', 'model.encoder.layers.3.self_attn.out_proj.weight', 'model.encoder.layers.3.self_attn.out_proj.bias', 'model.encoder.layers.4.self_attn.k_proj.weight', 'model.encoder.layers.4.self_attn.k_proj.bias', 'model.encoder.layers.4.self_attn.v_proj.weight', 'model.encoder.layers.4.self_attn.v_proj.bias', 'model.encoder.layers.4.self_attn.q_proj.weight', 'model.encoder.layers.4.self_attn.q_proj.bias', 'model.encoder.layers.4.self_attn.out_proj.weight', 'model.encoder.layers.4.self_attn.out_proj.bias', 'model.encoder.layers.5.self_attn.k_proj.weight', 'model.encoder.layers.5.self_attn.k_proj.bias', 'model.encoder.layers.5.self_attn.v_proj.weight', 'model.encoder.layers.5.self_attn.v_proj.bias', 'model.encoder.layers.5.self_attn.q_proj.weight', 'model.encoder.layers.5.self_attn.q_proj.bias', 'model.encoder.layers.5.self_attn.out_proj.weight', 'model.encoder.layers.5.self_attn.out_proj.bias']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\nTraceback (most recent call last):\r\n File \"/home/pbc/anaconda3/envs/tf2_pt_kr2/lib/python3.6/site-packages/IPython/core/interactiveshell.py\", line 3331, in run_code\r\n exec(code_obj, self.user_global_ns, self.user_ns)\r\n File \"<ipython-input-2-b1f8935f1cfa>\", line 1, in <module>\r\n runfile('/home/pbc/Documents/PycharmProjects/myEPI/src/github.py', wdir='/home/pbc/Documents/PycharmProjects/myEPI/src')\r\n File \"/home/pbc/.local/share/JetBrains/Toolbox/apps/PyCharm-P/ch-0/202.7660.27/plugins/python/helpers/pydev/_pydev_bundle/pydev_umd.py\", line 197, in runfile\r\n pydev_imports.execfile(filename, global_vars, local_vars) # execute the script\r\n File \"/home/pbc/.local/share/JetBrains/Toolbox/apps/PyCharm-P/ch-0/202.7660.27/plugins/python/helpers/pydev/_pydev_imps/_pydev_execfile.py\", line 18, in execfile\r\n exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\r\n File \"/home/pbc/Documents/PycharmProjects/myEPI/src/github.py\", line 8, in <module>\r\n model = AutoModel.from_pretrained(model_name)\r\n File \"/home/pbc/anaconda3/envs/tf2_pt_kr2/lib/python3.6/site-packages/transformers-4.4.2-py3.8.egg/transformers/models/auto/modeling_auto.py\", line 815, in from_pretrained\r\n pretrained_model_name_or_path, *model_args, config=config, **kwargs\r\n File \"/home/pbc/anaconda3/envs/tf2_pt_kr2/lib/python3.6/site-packages/transformers-4.4.2-py3.8.egg/transformers/modeling_utils.py\", line 1183, in from_pretrained\r\n model.__class__.__name__, \"\\n\\t\".join(error_msgs)\r\nRuntimeError: Error(s) in loading state_dict for BartModel:\r\n\tsize mismatch for model.encoder.embed_positions.weight: copying a param with shape torch.Size([16386, 768]) from checkpoint, the shape in current model is torch.Size([1026, 768]).\r\n```\r\n# env2: copied from env0 but not worked\r\n```bash\r\nhome/pbc/anaconda3/envs/dnabert/lib/python3.6/site-packages/transformers/__init__.py\r\nSome weights of the model checkpoint at pre-model/longformer-encdec-base-16384 were not used when initializing BartModel: ['model.encoder.layers.0.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.0.self_attn.output.weight', 'model.encoder.layers.0.self_attn.output.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.1.self_attn.output.weight', 'model.encoder.layers.1.self_attn.output.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.2.self_attn.output.weight', 'model.encoder.layers.2.self_attn.output.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.3.self_attn.output.weight', 'model.encoder.layers.3.self_attn.output.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.4.self_attn.output.weight', 'model.encoder.layers.4.self_attn.output.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.5.self_attn.output.weight', 'model.encoder.layers.5.self_attn.output.bias']\r\n- This IS expected if you are initializing BartModel from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).\r\n- This IS NOT expected if you are initializing BartModel from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of BartModel were not initialized from the model checkpoint at pre-model/longformer-encdec-base-16384 and are newly initialized: ['model.encoder.layers.0.self_attn.k_proj.weight', 'model.encoder.layers.0.self_attn.k_proj.bias', 'model.encoder.layers.0.self_attn.v_proj.weight', 'model.encoder.layers.0.self_attn.v_proj.bias', 'model.encoder.layers.0.self_attn.q_proj.weight', 'model.encoder.layers.0.self_attn.q_proj.bias', 'model.encoder.layers.0.self_attn.out_proj.weight', 'model.encoder.layers.0.self_attn.out_proj.bias', 'model.encoder.layers.1.self_attn.k_proj.weight', 'model.encoder.layers.1.self_attn.k_proj.bias', 'model.encoder.layers.1.self_attn.v_proj.weight', 'model.encoder.layers.1.self_attn.v_proj.bias', 'model.encoder.layers.1.self_attn.q_proj.weight', 'model.encoder.layers.1.self_attn.q_proj.bias', 'model.encoder.layers.1.self_attn.out_proj.weight', 'model.encoder.layers.1.self_attn.out_proj.bias', 'model.encoder.layers.2.self_attn.k_proj.weight', 'model.encoder.layers.2.self_attn.k_proj.bias', 'model.encoder.layers.2.self_attn.v_proj.weight', 'model.encoder.layers.2.self_attn.v_proj.bias', 'model.encoder.layers.2.self_attn.q_proj.weight', 'model.encoder.layers.2.self_attn.q_proj.bias', 'model.encoder.layers.2.self_attn.out_proj.weight', 'model.encoder.layers.2.self_attn.out_proj.bias', 'model.encoder.layers.3.self_attn.k_proj.weight', 'model.encoder.layers.3.self_attn.k_proj.bias', 'model.encoder.layers.3.self_attn.v_proj.weight', 'model.encoder.layers.3.self_attn.v_proj.bias', 'model.encoder.layers.3.self_attn.q_proj.weight', 'model.encoder.layers.3.self_attn.q_proj.bias', 'model.encoder.layers.3.self_attn.out_proj.weight', 'model.encoder.layers.3.self_attn.out_proj.bias', 'model.encoder.layers.4.self_attn.k_proj.weight', 'model.encoder.layers.4.self_attn.k_proj.bias', 'model.encoder.layers.4.self_attn.v_proj.weight', 'model.encoder.layers.4.self_attn.v_proj.bias', 'model.encoder.layers.4.self_attn.q_proj.weight', 'model.encoder.layers.4.self_attn.q_proj.bias', 'model.encoder.layers.4.self_attn.out_proj.weight', 'model.encoder.layers.4.self_attn.out_proj.bias', 'model.encoder.layers.5.self_attn.k_proj.weight', 'model.encoder.layers.5.self_attn.k_proj.bias', 'model.encoder.layers.5.self_attn.v_proj.weight', 'model.encoder.layers.5.self_attn.v_proj.bias', 'model.encoder.layers.5.self_attn.q_proj.weight', 'model.encoder.layers.5.self_attn.q_proj.bias', 'model.encoder.layers.5.self_attn.out_proj.weight', 'model.encoder.layers.5.self_attn.out_proj.bias']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\nTraceback (most recent call last):\r\n File \"<input>\", line 1, in <module>\r\n File \"/home/pbc/.local/share/JetBrains/Toolbox/apps/PyCharm-P/ch-0/203.7148.72/plugins/python/helpers/pydev/_pydev_bundle/pydev_umd.py\", line 197, in runfile\r\n pydev_imports.execfile(filename, global_vars, local_vars) # execute the script\r\n File \"/home/pbc/.local/share/JetBrains/Toolbox/apps/PyCharm-P/ch-0/203.7148.72/plugins/python/helpers/pydev/_pydev_imps/_pydev_execfile.py\", line 18, in execfile\r\n exec(compile(contents+\"\\n\", file, 'exec'), glob, loc)\r\n File \"/home/pbc/PycharmProjects/bert/github.py\", line 7, in <module>\r\n model = AutoModel.from_pretrained(model_name)\r\n File \"/home/pbc/anaconda3/envs/dnabert/lib/python3.6/site-packages/transformers/modeling_auto.py\", line 523, in from_pretrained\r\n return model_class.from_pretrained(pretrained_model_name_or_path, *model_args, config=config, **kwargs)\r\n File \"/home/pbc/anaconda3/envs/dnabert/lib/python3.6/site-packages/transformers/modeling_utils.py\", line 972, in from_pretrained\r\n model.__class__.__name__, \"\\n\\t\".join(error_msgs)\r\nRuntimeError: Error(s) in loading state_dict for BartModel:\r\n\tsize mismatch for model.encoder.embed_positions.weight: copying a param with shape torch.Size([16386, 768]) from checkpoint, the shape in current model is torch.Size([1026, 768]).\r\n ```\r\n\r\nI found transformers.__file__ all are different\r\n\r\n", "Now please check this directory `/home/pbc/anaconda3/envs/dnabert/lib/python3.6/site-packages/transformers/` and locate the file called `modeling_bart.py`. Post the BertEncoder class definition here.\r\n\r\nYou should also pay attention to the weights that were not used from the pre-trained weights:\r\n`['model.encoder.layers.0.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.0.self_attn.output.weight', 'model.encoder.layers.0.self_attn.output.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.1.self_attn.output.weight', 'model.encoder.layers.1.self_attn.output.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.2.self_attn.output.weight', 'model.encoder.layers.2.self_attn.output.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.3.self_attn.output.weight', 'model.encoder.layers.3.self_attn.output.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.4.self_attn.output.weight', 'model.encoder.layers.4.self_attn.output.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.5.self_attn.output.weight', 'model.encoder.layers.5.self_attn.output.bias']\r\n`\r\n\r\nAre you sure that this model (including config and weights) can be used with the transformers AutoModel class? Currently, it looks to me that someone has built his own model with the transformers library (which is not supposed to work with the AutoClasses). ", "> Now please check this directory `/home/pbc/anaconda3/envs/dnabert/lib/python3.6/site-packages/transformers/` and locate the file called `modeling_bart.py`. Post the BertEncoder class definition here.\r\n> \r\n> You should also pay attention to the weights that were not used from the pre-trained weights:\r\n> `['model.encoder.layers.0.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.0.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.0.self_attn.output.weight', 'model.encoder.layers.0.self_attn.output.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.1.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.1.self_attn.output.weight', 'model.encoder.layers.1.self_attn.output.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.2.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.2.self_attn.output.weight', 'model.encoder.layers.2.self_attn.output.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.3.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.3.self_attn.output.weight', 'model.encoder.layers.3.self_attn.output.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.4.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.4.self_attn.output.weight', 'model.encoder.layers.4.self_attn.output.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.query_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.key_global.bias', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.weight', 'model.encoder.layers.5.self_attn.longformer_self_attn.value_global.bias', 'model.encoder.layers.5.self_attn.output.weight', 'model.encoder.layers.5.self_attn.output.bias'] `\r\n> \r\n> Are you sure that this model (including config and weights) can be used with the transformers AutoModel class? Currently, it looks to me that someone has built his own model with the transformers library (which is not supposed to work with the AutoClasses).\r\n\r\nI am not sure, when I start to use bert model by the same way , it works, So I make it again in longformer, then many errors occured one by one", "That is a different thing. What you have here `longformer-encdec-base-16384` is something that is provided by someone that is not supposed to work with the provided AutoClasses by hugging face. Please check the code of this someone and see what this person did.\r\nI think this is the repository you should check out: https://github.com/allenai/ms2 \r\nor maybe this code snippet: https://github.com/allenai/longformer/issues/154 ", "> That is a different thing. What you have here `longformer-encdec-base-16384` is something that is provided by someone that is not supposed to work with the provided AutoClasses by hugging face. Please check the code of this someone and see what this person did.\r\n> I think this is the repository you should check out: https://github.com/allenai/ms2\r\n> or maybe this code snippet: [allenai/longformer#154](https://github.com/allenai/longformer/issues/154)\r\n\r\nyeah, you means that I should install env with allenai/longformer and not huggingface, at start I read allenai/longformer's readme, i just found that it may from huggingface and don't look for any things about how to use its longformer model by python code. \r\n\r\nI have seen [allenai/longformer#154](https://github.com/allenai/longformer/issues/154), and I will try it through Imitating her code. \r\n\r\nAnd another question is, if I want to use hugging face env to load model, that means I should download in https://huggingface.co/? \r\n\r\nAs for ms2, I will view it soon, Thanks!\r\n\r\n# Finally,thank you very much! you save me!Thanks!ORZ", "Yes, the `allenai/longformer` is the framework you should use for `longformer-encdec-base-16384`.\r\n\r\n> And another question is, if I want to use hugging face env to load model, that means I should download in https://huggingface.co/?\r\n\r\nYes, you can check the pre-trained models here: https://huggingface.co/models", "Okay, and I am curious that how do you find `allenai/longformer#154` and https://github.com/allenai/ms2. If I have the skill, I can save myself quickly,haha.", "Use a search engine of your choice and look for `longformer-encdec-base-16384` ;-)", "> longformer-encdec-base-16384\r\n\r\nOK，thank you very much!", "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." ]

[]

[]

[ "@patrickvonplaten @patil-suraj : Could please help me with the failed test? The error message is not that expressive:\r\n\r\n> Run cat reports/tests_templates_failures_short.txt\r\n> cat reports/tests_templates_failures_short.txt\r\n> shell: /usr/bin/bash -e {0}\r\n> env:\r\n> pythonLocation: /opt/hostedtoolcache/Python/3.6.13/x64\r\n> cat: reports/tests_templates_failures_short.txt: No such file or directory\r\n> Error: Process completed with exit code 1.", "> Instead of modifying the config, I think one alternate solution is to assign the shared config object to the encoder and decoder, after this line\r\n> \r\n> https://github.com/huggingface/transformers/blob/95dab34d5588fb155dfed8293ac2fbb1217a95a7/src/transformers/models/encoder_decoder/modeling_encoder_decoder.py#L345\r\n> \r\n> ```python\r\n> encoder.config = config.encoder\r\n> decoder.config = config.decoder\r\n> ```\r\nPlease correct me if I am wrong but what happens, in that case, is the following:\r\n1. Casting already existing config objects to dictionaries.\r\n2. Recreating config objects from those dictionaries.\r\n3. Initializing EncoderDecoderConfig with those new config objects.\r\n4. Throwing away the newly generating config objects by assigning the ones that were already present before step 1.\r\n\r\nI think this PR is a cleaner implementation by avoids steps 1-3 and executing 4 directly.\r\n\r\n", "Good point!\r\n\r\nBut this breaks backward compatibility. With this change, none of the previously trained models will be able to load because the config will now be in-compatible. For ex if you try\r\n```python\r\nconfig = EncoderDecoderConfig.from_pretrained(\"google/bert2bert_L-24_wmt_de_en\")\r\n```\r\non this PR, it raises an exception. So loading model fails.\r\n\r\nIn any case, backward compatibility is of utmost importance.", "Hi @patil-suraj \r\nI have pushed a new version that is now backward compatible and also covers a case I have previously overlooked. After checking the implementation of the parent classes `PreTrainedModel` and `PretrainedConfig` I came to the conclusion that your suggestion is the best because they all transfer dictionaries as parameters and not config objects. \r\nWe could of course implement a type check like;\r\n```\r\nif type(encoder) == dict:\r\n#.....\r\n```\r\nbut I think this makes the code less readable. Would be great if you could have a look again and thanks for the constructive review so far :+1:.", "Thanks a lot for taking care of this @cronoik :-) It's a nice fix. It would be awesome if you could check out the suggestions and then we can merge this IMO.", "Hi @patil-suraj @patrickvonplaten, \r\nthanks for all the suggestions. I think I am done. Could you please have a look?" ]

[]

[]

[]

[]

[]

[ "Circle CI is unrelated - merging! Thanks a lot @nicola-decao " ]

[]

[]

[ "Tbh, I've never really worked with `get_concrete_function()`, etc.... @Rocketknight1 - do you have an idea by any chance?", "> Tbh, I've never really worked with `get_concrete_function()`, etc.... @Rocketknight1 - do you have an idea by any chance?\n\nI just realized I should tag the authors of the post I read about in the issue. I have edited the issue.", "Hi, I'm the TF maintainer! There are two problems here. The first is that the first two arguments to `TFGPT2LMHeadModel` are not `input_ids` and `attention_mask`, they are `input_ids` and `past`, see [here](https://huggingface.co/transformers/model_doc/gpt2.html#tfgpt2lmheadmodel). Also, `TFGPT2LMHeadModel` returns a tuple/dict of Tensors. Concrete functions do not support that - you need to pick which one you want. Try something like this, which should work (if you want an output other than \"logits\", you can just change that bit):\r\n\r\n```\r\nimport tensorflow as tf\r\nfrom transformers import TFGPT2LMHeadModel\r\nfrom transformers import GPT2Tokenizer\r\n\r\[email protected]\r\ndef call_model(input_ids, attention_mask):\r\n return model(input_ids=input_ids, attention_mask=attention_mask)['logits']\r\n\r\ntokenizer = GPT2Tokenizer.from_pretrained(\"distilgpt2\")\r\nmodel = TFGPT2LMHeadModel.from_pretrained(\"distilgpt2\", pad_token_id=tokenizer.eos_token_id)\r\nconcrete_function = call_model.get_concrete_function(tf.TensorSpec([None, 384], tf.int32, name=\"input_ids\"), tf.TensorSpec([None, 384], tf.int32, name=\"attention_mask\"))\r\ntf.saved_model.save(model, 'distilgpt2_sig', signatures=concrete_function)\r\n```", "> Hi, I'm the TF maintainer! There are two problems here. The first is that the first two arguments to `TFGPT2LMHeadModel` are not `input_ids` and `attention_mask`, they are `input_ids` and `past`, see [here](https://huggingface.co/transformers/model_doc/gpt2.html#tfgpt2lmheadmodel). Also, `TFGPT2LMHeadModel` returns a tuple/dict of Tensors. Concrete functions do not support that - you need to pick which one you want. Try something like this, which should work (if you want an output other than \"logits\", you can just change that bit):\r\n> \r\n> ```\r\n> import tensorflow as tf\r\n> from transformers import TFGPT2LMHeadModel\r\n> from transformers import GPT2Tokenizer\r\n> \r\n> @tf.function\r\n> def call_model(input_ids, attention_mask):\r\n> return model(input_ids=input_ids, attention_mask=attention_mask)['logits']\r\n> \r\n> tokenizer = GPT2Tokenizer.from_pretrained(\"distilgpt2\")\r\n> model = TFGPT2LMHeadModel.from_pretrained(\"distilgpt2\", pad_token_id=tokenizer.eos_token_id)\r\n> concrete_function = call_model.get_concrete_function(tf.TensorSpec([None, 384], tf.int32, name=\"input_ids\"), tf.TensorSpec([None, 384], tf.int32, name=\"attention_mask\"))\r\n> tf.saved_model.save(model, 'distilgpt2_sig', signatures=concrete_function)\r\n> ```\r\n\r\nThat works! Thank you for the help. I am not familiar with TF especially things like `get_concrete_function`, I didn't know you can define a function outside the model and then save it. " ]

[]

[]

[ "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." ]

[]

[]

[ "You can also consider the modification in finetune_trainer.py\r\n\r\nhttps://github.com/huggingface/transformers/blob/1c06240e1b3477728129bb58e7b6c7734bb5074e/examples/legacy/seq2seq/finetune_trainer.py#L233\r\n\r\nIf you freeze some parameters as done in the line above, those would not be there when you load the model and restarting the training, this is really a serious issue, thanks for the help \r\n", "Here is the minimal code to generate this bug, we make a model, we freeze, then we save it (as done in trainer checkpoint), then we load it (as done in train() in trainer.py) and we see if the feezed parameters are there or not\r\n\r\n```\r\nfrom transformers import T5ForConditionalGeneration\r\nfrom typing import Optional \r\nimport torch\r\nimport os\r\n\r\n# This is copied from trainer.py\r\ndef _save(model, output_dir: Optional[str] = None):\r\n os.makedirs(output_dir, exist_ok=True)\r\n print(f\"Saving model checkpoint to {output_dir}\")\r\n # Save a trained model and configuration using `save_pretrained()`.\r\n # They can then be reloaded using `from_pretrained()`\r\n state_dict = model.state_dict()\r\n model.save_pretrained(output_dir, state_dict=state_dict)\r\n\r\ndef print_num_parameters(model):\r\n for n,p in model.named_parameters():\r\n if (p.requires_grad):\r\n print(\"n \", n) \r\n\r\ndef freeze_params(model):\r\n for n,p in model.named_parameters():\r\n p.requires_grad = False\r\n\r\n\r\nmodel = T5ForConditionalGeneration.from_pretrained(\"t5-base\")\r\nfreeze_params(model)\r\nprint(\"#### parameters before saving ####\")\r\nprint_num_parameters(model)\r\n_save(model, \"temp_model\")\r\n\r\n# Now lets load the model as done in trainer from the checkpoint.\r\nmodel = model.from_pretrained(\"temp_model\")\r\n\r\n# Now lets print the number of parameters\r\nprint(\"#### parameters after saving ####\")\r\nprint_num_parameters(model)\r\n```\r\n\r\nsurprisingly, no, the freezed parameters are not freezed anymore after loading the checkpoint:\r\n```\r\n#### parameters before saving ####\r\nSaving model checkpoint to temp_model\r\n#### parameters after saving ####\r\nn shared.weight\r\nn encoder.block.0.layer.0.SelfAttention.q.weight\r\nn encoder.block.0.layer.0.SelfAttention.k.weight\r\nn encoder.block.0.layer.0.SelfAttention.v.weight\r\nn encoder.block.0.layer.0.SelfAttention.o.weight\r\nn encoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight\r\nn encoder.block.0.layer.0.layer_norm.weight\r\nn encoder.block.0.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.0.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.0.layer.1.layer_norm.weight\r\nn encoder.block.1.layer.0.SelfAttention.q.weight\r\nn encoder.block.1.layer.0.SelfAttention.k.weight\r\nn encoder.block.1.layer.0.SelfAttention.v.weight\r\nn encoder.block.1.layer.0.SelfAttention.o.weight\r\nn encoder.block.1.layer.0.layer_norm.weight\r\nn encoder.block.1.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.1.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.1.layer.1.layer_norm.weight\r\nn encoder.block.2.layer.0.SelfAttention.q.weight\r\nn encoder.block.2.layer.0.SelfAttention.k.weight\r\nn encoder.block.2.layer.0.SelfAttention.v.weight\r\nn encoder.block.2.layer.0.SelfAttention.o.weight\r\nn encoder.block.2.layer.0.layer_norm.weight\r\nn encoder.block.2.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.2.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.2.layer.1.layer_norm.weight\r\nn encoder.block.3.layer.0.SelfAttention.q.weight\r\nn encoder.block.3.layer.0.SelfAttention.k.weight\r\nn encoder.block.3.layer.0.SelfAttention.v.weight\r\nn encoder.block.3.layer.0.SelfAttention.o.weight\r\nn encoder.block.3.layer.0.layer_norm.weight\r\nn encoder.block.3.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.3.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.3.layer.1.layer_norm.weight\r\nn encoder.block.4.layer.0.SelfAttention.q.weight\r\nn encoder.block.4.layer.0.SelfAttention.k.weight\r\nn encoder.block.4.layer.0.SelfAttention.v.weight\r\nn encoder.block.4.layer.0.SelfAttention.o.weight\r\nn encoder.block.4.layer.0.layer_norm.weight\r\nn encoder.block.4.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.4.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.4.layer.1.layer_norm.weight\r\nn encoder.block.5.layer.0.SelfAttention.q.weight\r\nn encoder.block.5.layer.0.SelfAttention.k.weight\r\nn encoder.block.5.layer.0.SelfAttention.v.weight\r\nn encoder.block.5.layer.0.SelfAttention.o.weight\r\nn encoder.block.5.layer.0.layer_norm.weight\r\nn encoder.block.5.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.5.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.5.layer.1.layer_norm.weight\r\nn encoder.block.6.layer.0.SelfAttention.q.weight\r\nn encoder.block.6.layer.0.SelfAttention.k.weight\r\nn encoder.block.6.layer.0.SelfAttention.v.weight\r\nn encoder.block.6.layer.0.SelfAttention.o.weight\r\nn encoder.block.6.layer.0.layer_norm.weight\r\nn encoder.block.6.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.6.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.6.layer.1.layer_norm.weight\r\nn encoder.block.7.layer.0.SelfAttention.q.weight\r\nn encoder.block.7.layer.0.SelfAttention.k.weight\r\nn encoder.block.7.layer.0.SelfAttention.v.weight\r\nn encoder.block.7.layer.0.SelfAttention.o.weight\r\nn encoder.block.7.layer.0.layer_norm.weight\r\nn encoder.block.7.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.7.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.7.layer.1.layer_norm.weight\r\nn encoder.block.8.layer.0.SelfAttention.q.weight\r\nn encoder.block.8.layer.0.SelfAttention.k.weight\r\nn encoder.block.8.layer.0.SelfAttention.v.weight\r\nn encoder.block.8.layer.0.SelfAttention.o.weight\r\nn encoder.block.8.layer.0.layer_norm.weight\r\nn encoder.block.8.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.8.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.8.layer.1.layer_norm.weight\r\nn encoder.block.9.layer.0.SelfAttention.q.weight\r\nn encoder.block.9.layer.0.SelfAttention.k.weight\r\nn encoder.block.9.layer.0.SelfAttention.v.weight\r\nn encoder.block.9.layer.0.SelfAttention.o.weight\r\nn encoder.block.9.layer.0.layer_norm.weight\r\nn encoder.block.9.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.9.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.9.layer.1.layer_norm.weight\r\nn encoder.block.10.layer.0.SelfAttention.q.weight\r\nn encoder.block.10.layer.0.SelfAttention.k.weight\r\nn encoder.block.10.layer.0.SelfAttention.v.weight\r\nn encoder.block.10.layer.0.SelfAttention.o.weight\r\nn encoder.block.10.layer.0.layer_norm.weight\r\nn encoder.block.10.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.10.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.10.layer.1.layer_norm.weight\r\nn encoder.block.11.layer.0.SelfAttention.q.weight\r\nn encoder.block.11.layer.0.SelfAttention.k.weight\r\nn encoder.block.11.layer.0.SelfAttention.v.weight\r\nn encoder.block.11.layer.0.SelfAttention.o.weight\r\nn encoder.block.11.layer.0.layer_norm.weight\r\nn encoder.block.11.layer.1.DenseReluDense.wi.weight\r\nn encoder.block.11.layer.1.DenseReluDense.wo.weight\r\nn encoder.block.11.layer.1.layer_norm.weight\r\nn encoder.final_layer_norm.weight\r\nn decoder.block.0.layer.0.SelfAttention.q.weight\r\nn decoder.block.0.layer.0.SelfAttention.k.weight\r\nn decoder.block.0.layer.0.SelfAttention.v.weight\r\nn decoder.block.0.layer.0.SelfAttention.o.weight\r\nn decoder.block.0.layer.0.SelfAttention.relative_attention_bias.weight\r\nn decoder.block.0.layer.0.layer_norm.weight\r\nn decoder.block.0.layer.1.EncDecAttention.q.weight\r\nn decoder.block.0.layer.1.EncDecAttention.k.weight\r\nn decoder.block.0.layer.1.EncDecAttention.v.weight\r\nn decoder.block.0.layer.1.EncDecAttention.o.weight\r\nn decoder.block.0.layer.1.layer_norm.weight\r\nn decoder.block.0.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.0.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.0.layer.2.layer_norm.weight\r\nn decoder.block.1.layer.0.SelfAttention.q.weight\r\nn decoder.block.1.layer.0.SelfAttention.k.weight\r\nn decoder.block.1.layer.0.SelfAttention.v.weight\r\nn decoder.block.1.layer.0.SelfAttention.o.weight\r\nn decoder.block.1.layer.0.layer_norm.weight\r\nn decoder.block.1.layer.1.EncDecAttention.q.weight\r\nn decoder.block.1.layer.1.EncDecAttention.k.weight\r\nn decoder.block.1.layer.1.EncDecAttention.v.weight\r\nn decoder.block.1.layer.1.EncDecAttention.o.weight\r\nn decoder.block.1.layer.1.layer_norm.weight\r\nn decoder.block.1.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.1.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.1.layer.2.layer_norm.weight\r\nn decoder.block.2.layer.0.SelfAttention.q.weight\r\nn decoder.block.2.layer.0.SelfAttention.k.weight\r\nn decoder.block.2.layer.0.SelfAttention.v.weight\r\nn decoder.block.2.layer.0.SelfAttention.o.weight\r\nn decoder.block.2.layer.0.layer_norm.weight\r\nn decoder.block.2.layer.1.EncDecAttention.q.weight\r\nn decoder.block.2.layer.1.EncDecAttention.k.weight\r\nn decoder.block.2.layer.1.EncDecAttention.v.weight\r\nn decoder.block.2.layer.1.EncDecAttention.o.weight\r\nn decoder.block.2.layer.1.layer_norm.weight\r\nn decoder.block.2.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.2.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.2.layer.2.layer_norm.weight\r\nn decoder.block.3.layer.0.SelfAttention.q.weight\r\nn decoder.block.3.layer.0.SelfAttention.k.weight\r\nn decoder.block.3.layer.0.SelfAttention.v.weight\r\nn decoder.block.3.layer.0.SelfAttention.o.weight\r\nn decoder.block.3.layer.0.layer_norm.weight\r\nn decoder.block.3.layer.1.EncDecAttention.q.weight\r\nn decoder.block.3.layer.1.EncDecAttention.k.weight\r\nn decoder.block.3.layer.1.EncDecAttention.v.weight\r\nn decoder.block.3.layer.1.EncDecAttention.o.weight\r\nn decoder.block.3.layer.1.layer_norm.weight\r\nn decoder.block.3.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.3.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.3.layer.2.layer_norm.weight\r\nn decoder.block.4.layer.0.SelfAttention.q.weight\r\nn decoder.block.4.layer.0.SelfAttention.k.weight\r\nn decoder.block.4.layer.0.SelfAttention.v.weight\r\nn decoder.block.4.layer.0.SelfAttention.o.weight\r\nn decoder.block.4.layer.0.layer_norm.weight\r\nn decoder.block.4.layer.1.EncDecAttention.q.weight\r\nn decoder.block.4.layer.1.EncDecAttention.k.weight\r\nn decoder.block.4.layer.1.EncDecAttention.v.weight\r\nn decoder.block.4.layer.1.EncDecAttention.o.weight\r\nn decoder.block.4.layer.1.layer_norm.weight\r\nn decoder.block.4.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.4.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.4.layer.2.layer_norm.weight\r\nn decoder.block.5.layer.0.SelfAttention.q.weight\r\nn decoder.block.5.layer.0.SelfAttention.k.weight\r\nn decoder.block.5.layer.0.SelfAttention.v.weight\r\nn decoder.block.5.layer.0.SelfAttention.o.weight\r\nn decoder.block.5.layer.0.layer_norm.weight\r\nn decoder.block.5.layer.1.EncDecAttention.q.weight\r\nn decoder.block.5.layer.1.EncDecAttention.k.weight\r\nn decoder.block.5.layer.1.EncDecAttention.v.weight\r\nn decoder.block.5.layer.1.EncDecAttention.o.weight\r\nn decoder.block.5.layer.1.layer_norm.weight\r\nn decoder.block.5.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.5.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.5.layer.2.layer_norm.weight\r\nn decoder.block.6.layer.0.SelfAttention.q.weight\r\nn decoder.block.6.layer.0.SelfAttention.k.weight\r\nn decoder.block.6.layer.0.SelfAttention.v.weight\r\nn decoder.block.6.layer.0.SelfAttention.o.weight\r\nn decoder.block.6.layer.0.layer_norm.weight\r\nn decoder.block.6.layer.1.EncDecAttention.q.weight\r\nn decoder.block.6.layer.1.EncDecAttention.k.weight\r\nn decoder.block.6.layer.1.EncDecAttention.v.weight\r\nn decoder.block.6.layer.1.EncDecAttention.o.weight\r\nn decoder.block.6.layer.1.layer_norm.weight\r\nn decoder.block.6.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.6.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.6.layer.2.layer_norm.weight\r\nn decoder.block.7.layer.0.SelfAttention.q.weight\r\nn decoder.block.7.layer.0.SelfAttention.k.weight\r\nn decoder.block.7.layer.0.SelfAttention.v.weight\r\nn decoder.block.7.layer.0.SelfAttention.o.weight\r\nn decoder.block.7.layer.0.layer_norm.weight\r\nn decoder.block.7.layer.1.EncDecAttention.q.weight\r\nn decoder.block.7.layer.1.EncDecAttention.k.weight\r\nn decoder.block.7.layer.1.EncDecAttention.v.weight\r\nn decoder.block.7.layer.1.EncDecAttention.o.weight\r\nn decoder.block.7.layer.1.layer_norm.weight\r\nn decoder.block.7.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.7.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.7.layer.2.layer_norm.weight\r\nn decoder.block.8.layer.0.SelfAttention.q.weight\r\nn decoder.block.8.layer.0.SelfAttention.k.weight\r\nn decoder.block.8.layer.0.SelfAttention.v.weight\r\nn decoder.block.8.layer.0.SelfAttention.o.weight\r\nn decoder.block.8.layer.0.layer_norm.weight\r\nn decoder.block.8.layer.1.EncDecAttention.q.weight\r\nn decoder.block.8.layer.1.EncDecAttention.k.weight\r\nn decoder.block.8.layer.1.EncDecAttention.v.weight\r\nn decoder.block.8.layer.1.EncDecAttention.o.weight\r\nn decoder.block.8.layer.1.layer_norm.weight\r\nn decoder.block.8.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.8.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.8.layer.2.layer_norm.weight\r\nn decoder.block.9.layer.0.SelfAttention.q.weight\r\nn decoder.block.9.layer.0.SelfAttention.k.weight\r\nn decoder.block.9.layer.0.SelfAttention.v.weight\r\nn decoder.block.9.layer.0.SelfAttention.o.weight\r\nn decoder.block.9.layer.0.layer_norm.weight\r\nn decoder.block.9.layer.1.EncDecAttention.q.weight\r\nn decoder.block.9.layer.1.EncDecAttention.k.weight\r\nn decoder.block.9.layer.1.EncDecAttention.v.weight\r\nn decoder.block.9.layer.1.EncDecAttention.o.weight\r\nn decoder.block.9.layer.1.layer_norm.weight\r\nn decoder.block.9.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.9.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.9.layer.2.layer_norm.weight\r\nn decoder.block.10.layer.0.SelfAttention.q.weight\r\nn decoder.block.10.layer.0.SelfAttention.k.weight\r\nn decoder.block.10.layer.0.SelfAttention.v.weight\r\nn decoder.block.10.layer.0.SelfAttention.o.weight\r\nn decoder.block.10.layer.0.layer_norm.weight\r\nn decoder.block.10.layer.1.EncDecAttention.q.weight\r\nn decoder.block.10.layer.1.EncDecAttention.k.weight\r\nn decoder.block.10.layer.1.EncDecAttention.v.weight\r\nn decoder.block.10.layer.1.EncDecAttention.o.weight\r\nn decoder.block.10.layer.1.layer_norm.weight\r\nn decoder.block.10.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.10.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.10.layer.2.layer_norm.weight\r\nn decoder.block.11.layer.0.SelfAttention.q.weight\r\nn decoder.block.11.layer.0.SelfAttention.k.weight\r\nn decoder.block.11.layer.0.SelfAttention.v.weight\r\nn decoder.block.11.layer.0.SelfAttention.o.weight\r\nn decoder.block.11.layer.0.layer_norm.weight\r\nn decoder.block.11.layer.1.EncDecAttention.q.weight\r\nn decoder.block.11.layer.1.EncDecAttention.k.weight\r\nn decoder.block.11.layer.1.EncDecAttention.v.weight\r\nn decoder.block.11.layer.1.EncDecAttention.o.weight\r\nn decoder.block.11.layer.1.layer_norm.weight\r\nn decoder.block.11.layer.2.DenseReluDense.wi.weight\r\nn decoder.block.11.layer.2.DenseReluDense.wo.weight\r\nn decoder.block.11.layer.2.layer_norm.weight\r\nn decoder.final_layer_norm.weight\r\n```\r\n", "Your end example is not surprising: you are re-loading a new model so of course modifications done on the original model are erased.\r\nWe'll look at how we can fix this inside the trainer, to load the weights differently instead of using `from_pretrained`.", "Hi @sgugger thank you for the response, my intention from the example was showing that the procedure happening inside the trainer if the user resume training from a modified model. thank you. ", "@sgugger I also see some more issues with the trainer.py when I load the model from the checkpoint, the model which previously was training fine on the GPU gets out of memory issue, there must be a leakge of memory during the loading from a checkpoint, shall I make a separate ticket for this issue? thanks ", "If it's because you add frozen parameters previously, that would explain the OOM error.", "Hi @LysandreJik @sgugger \r\nthanks for the respose, but I do not see how this is relevant. \r\nDuring training, I also load the model and then freeze some of the parameters, then this trains fine, only during loading from a checkpoint, this goes to memory issues, but the procedure remains the same, by loading the model and then freezing params, i personally think there must be a bug somewhere in checkpoint loading, resulting in extra usage of memory,. thanks for your help ", "One simple test to check this @sgugger would be get a model, and choose a batch size in a way that it just fit the memory but larger than that wont, then resume the training from a checkpoint, then I am sure you would also see the memory issue, even without any modification, just the baseline t5 I do see this issue with huggingface codes. thanks for your help ", "> We'll look at how we can fix this inside the trainer, to load the weights differently instead of using `from_pretrained`.\r\n\r\n@sgugger, this is definitely related to what I need for deepspeed checkpoint resume. Currently we first load the model `from_pretrained` and then it gets dropped and replaced by the deepspeed checkpoint, which for huge models is a huge slowdown.\r\nSo let's coordinate this work.\r\n\r\nMy preliminary idea was to pass a new flag to `from_pretrained` which will do everything except actually loading the weights. I was planning to work on this this week.\r\n\r\n(plus we need to stop random init the weights when they are replaced with pre-trained weights, so this is related too but not directly to this particular issue)", "Thanks @stas00 for your attention to this issue, this would be really awesome to have this fixed, thanks a lot for the great work you do ", "You're in good hands, @dorooddorood606 - @sgugger is taking care of it already, I was just commenting that something similar needs to be done for deepspeed, so once @sgugger's PR goes in I will work on doing the same for deepspeed.", "Dear @stas00 \r\nThank you very much both of you @sgugger to your great efforts and the great job you do, I also observe the vanilla t5-base checkpointing gets very different results after resume, I reported the bug here, this can be a relevant bug to this one:\r\n\r\nhttps://github.com/huggingface/transformers/issues/11323 \r\n\r\nSo this seems there are some randomness in t5-base model which is not considered in trainer.py resume from checkpointing part, if you also think these bugs can be related, I would greatly appreciate if you could also consider the losses when one resume from a checkpoint.\r\n I would like to thank you so much for your time and your efforts and the great and awesome job you do. " ]

[]

[]

[ "> Attempted to read a PyTorch file with version 3, but the maximum supported version for reading is 2. Your PyTorch installation may be too old. \r\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." ]

[]

[]

[]

[]

[]

[ "Is it possible you're running out of RAM (not necessarily GPU RAM)?", "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." ]

[]

[]

[ "Hi @chz816 \r\n\r\nI can reproduce the issue. This is because pegasus doesn't really work with `fp16`since its trained with `bfloat16`, so in most cases, it overflows and returns `nan` logits. The model works as expected in `fp32`, so if you run the above command without the `--fp16` arg, it should give the expected results.\r\n\r\ncc @stas00 ", "Thank you @patil-suraj!\r\n\r\nI have generated the summaries using ```pegasus-cnn_dailymail``` with the following performance: ```{'rouge1': 43.146, 'rouge2': 20.7292, 'rougeL': 30.4596, 'n_obs': 11490, 'seconds_per_sample': 0.2415, 'n_gpus': 3}```. It is lower than expected, but I think it can be explained by smaller batch size, which is caused by the memory limitation.\r\n\r\n```bash\r\npython -m torch.distributed.launch --nproc_per_node=3 run_distributed_eval.py \\\r\n --model_name google/pegasus-cnn_dailymail \\\r\n --save_dir $OUTPUT_DIR \\\r\n --data_dir $DATA_DIR \\\r\n --bs 16\r\n```\r\n\r\nCan you maybe explain why this problem does not exist for ```google/pegasus-xsum```? Thank you!", "As @patil-suraj pointed out many models trained in `bfloat16` can't be run under mixed precision `fp16` (albeit pytorch are discussing bfloat16 mixed precision)\r\n\r\n`pegasus-cnn_dailymail` has an issue of underflow under `fp16`:\r\n\r\nlet's take a single frame - Linear forward for `lm_head`:\r\n\r\nfp32:\r\n\r\n```\r\nabs min abs max metadata\r\n lm_head Linear\r\n4.66e-10 1.13e+01 weight\r\n6.29e-07 4.47e+00 input[0]\r\n1.63e-07 3.00e+01 output\r\n```\r\n\r\nfp16:\r\n\r\n```\r\n lm_head Linear\r\n0.00e+00 1.13e+01 weight\r\n6.76e-07 5.38e+00 input[0]\r\n0.00e+00 3.08e+01 output\r\n```\r\n\r\nAs you can see `4.66e-10` under fp16 underflows into `0.0`.\r\n\r\n**edit:** well, actually this would be the case if we did `model.half()` (which is what deepspeed does, and that's where it'd immediately underflow on the very first use), so here it's probably something else. I will need some time to try to understand what's going on here.\r\n\r\nThis is from WIP PR https://github.com/huggingface/transformers/pull/11274 - still polishing some nuances but should be ready soon.\r\n\r\nLet me check `google/pegasus-xsum`", "Regarding the cnn_dailymail scores, please see this issue #6844", "@chz816, meanwhile could you please give me a way to reproduce your case? Ideally with some public dataset and best with the current version of the examples (master or last release), which would be using `examples/seq2seq/run_summarization.py`\r\n\r\ne.g.:\r\n```\r\npython examples/seq2seq/run_summarization.py --model_name_or_path google/pegasus-cnn_dailymail \\\r\n--do_train --do_eval --dataset_name cnn_dailymail --dataset_config \"3.0.0\" --source_prefix \\\r\n\"summarize: \" --output_dir /tmp/tst-summarization --per_device_train_batch_size=1 \\\r\n--per_device_eval_batch_size=1 --overwrite_output_dir --predict_with_generate \r\n```", "Thank you, @patil-suraj. \r\n\r\nOh, this is the legacy script so it does do:\r\n\r\n```\r\n if fp16:\r\n model = model.half()\r\n```\r\n\r\n", "```\r\nwget https://cdn-datasets.huggingface.co/summarization/pegasus_data/cnn_dailymail.tar.gz\r\ntar -xvzf cnn_dailymail.tar.gz\r\npython -m torch.distributed.launch --nproc_per_node=1 run_distributed_eval.py \\\r\n--model_name google/pegasus-cnn_dailymail --save_dir output_dir --data_dir cnn_dailymail \\\r\n--bs 8 --fp16\r\n```\r\n\r\nSo the detection is quick (had to bolt it on manually, since this script isn't using the `Trainer`):\r\n```\r\nDetected inf/nan during batch_number=0\r\nLast 10 forward frames:\r\nabs min abs max metadata\r\n model.encoder.layers.14.fc1 Linear\r\n0.00e+00 1.88e+01 weight\r\n2.73e-05 2.54e+00 bias\r\n5.96e-08 9.05e+00 input[0]\r\n0.00e+00 3.16e+02 output\r\n model.encoder.layers.14.fc2 Linear\r\n5.96e-08 3.29e+01 weight\r\n5.40e-03 2.66e+01 bias\r\n0.00e+00 1.03e+02 input[0]\r\n0.00e+00 8.00e+03 output\r\n model.encoder.layers.14 PegasusEncoderLayer\r\n0.00e+00 6.45e+04 input[0]\r\n0.00e+00 0.00e+00 input[1]\r\n0.00e+00 6.45e+04 output[0]\r\n model.encoder.layers.15.self_attn_layer_norm LayerNorm\r\n5.63e-03 3.85e-01 weight\r\n1.69e-05 2.49e-01 bias\r\n0.00e+00 6.45e+04 input[0]\r\n0.00e+00 1.50e+00 output\r\n model.encoder.layers.15.self_attn.q_proj Linear\r\n8.34e-07 2.95e+00 weight\r\n0.00e+00 0.00e+00 bias\r\n0.00e+00 1.50e+00 input[0]\r\n5.96e-08 8.52e+00 output\r\n model.encoder.layers.15.self_attn.k_proj Linear\r\n2.38e-07 1.85e+00 weight\r\n0.00e+00 0.00e+00 bias\r\n0.00e+00 1.50e+00 input[0]\r\n1.19e-07 9.30e+00 output\r\n model.encoder.layers.15.self_attn.v_proj Linear\r\n5.96e-08 4.03e+00 weight\r\n0.00e+00 0.00e+00 bias\r\n0.00e+00 1.50e+00 input[0]\r\n6.56e-07 2.95e+01 output\r\n model.encoder.layers.15.self_attn.out_proj Linear\r\n5.96e-08 2.25e+01 weight\r\n0.00e+00 0.00e+00 bias\r\n5.96e-08 1.25e+01 input[0]\r\n3.58e-07 1.29e+03 output\r\n model.encoder.layers.15.self_attn PegasusAttention\r\n3.58e-07 1.29e+03 output[0]\r\n None output[1]\r\n None output[2]\r\n model.encoder.layers.15.final_layer_norm LayerNorm\r\n7.32e-02 2.69e+00 weight\r\n2.00e-05 1.02e+00 bias\r\n0.00e+00 inf input[0]\r\n nan nan output\r\n```", "I'm able to reproduce this with the \"modern\" version of the script:\r\n\r\n```\r\nrm -rf output_dir; USE_TF=0 PYTHONPATH=src python examples/seq2seq/run_summarization.py \\\r\n--model_name_or_path google/pegasus-cnn_dailymail --do_eval --dataset_name cnn_dailymail \\\r\n--dataset_config \"3.0.0\" --output_dir output_dir \\\r\n--per_device_eval_batch_size=16 --predict_with_generate --fp16_full_eval --max_val_samples 10\r\n\r\n[...]\r\n\r\n***** eval metrics *****\r\n eval_gen_len = 9.0\r\n eval_loss = nan\r\n eval_mem_cpu_alloc_delta = -55MB\r\n eval_mem_cpu_peaked_delta = 55MB\r\n eval_mem_gpu_alloc_delta = 1089MB\r\n eval_mem_gpu_peaked_delta = 7241MB\r\n eval_rouge1 = 0.0\r\n eval_rouge2 = 0.0\r\n eval_rougeL = 0.0\r\n eval_rougeLsum = 0.0\r\n eval_runtime = 0:00:07.71\r\n eval_samples = 10\r\n eval_samples_per_second = 1.295\r\n init_mem_cpu_alloc_delta = 0MB\r\n init_mem_cpu_peaked_delta = 0MB\r\n init_mem_gpu_alloc_delta = 0MB\r\n init_mem_gpu_peaked_delta = 0MB\r\n```", "Thank you for your response @stas00 ! Yeah I am able to resolve the issue without ```--fp16```, but I am still little confused why ```google/pegasus-xsum``` works well with ```---fp16``` argument, since they are from the same seq2seq model. Any ideas? Thank you!", "For some reason I can't even run `google/pegasus-xsum` https://github.com/huggingface/transformers/issues/11344, so I'm not able to look inside.\r\n\r\nI can only guess that perhaps `google/pegasus-xsum` was trained in mixed precision fp16?" ]

[]

[]

[]

[]

[]

[ "The ", "> The\r\n\r\nYes?", "Answering my question: pipeline returns the model as well. You just have to use it directly to extract the hidden states." ]

[]

[]

[]

[]

[]

[ "This is caused by the EncoderDecoderConfig which initializes independent objects ([link](https://github.com/huggingface/transformers/blob/d9c62047a8d75e18d2849d345ab3394875a712ef/src/transformers/models/encoder_decoder/configuration_encoder_decoder.py#L84)) instead of utilizing the already existing ones.\r\n\r\nYou can fix that for the moment by calling:\r\n```\r\nmodel.config.decoder = model.decoder.config\r\nmodel.config.encoder = model.encoder.config\r\n```\r\nPR will follow." ]

[]

[]

[ "The script is not hanging, it is skipping the first 660,000 batches since you are resuming training from there which takes a lot of time. If you don't mind continuing training with the same data, you can use the option `ignore_data_skip=True` in your training arguments.", "@eh-93 Do you remember how much time it took to train that checkpoint? \r\n@sgugger How about we add a progress bar to make the trainer more user-friendly? ", "@sgugger got it, thanks\r\n@cronoik - around 5 days to get to that checkpoint\r\n\r\nAfter 28 hours the training resumed", "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.", "That's not a good solution. A better solution that should be implemented is to pass some boolean to the CustomDataset, telling it that we are now in a 'skip' mode, so that the CustomDataset could prevent from doing expensive and unneeded steps in the skipping phase, like for example tokenize words. Does HF work on such solution?" ]

[]

[ { "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 }, { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "Hey @tanujjain, \r\n\r\nWe are very interested in adding beam search for Wav2Vec2 + LM support in general, but sadly don't find the time to do so at the moment. We would be really happy about a contribution if you want to give it a try.\r\n\r\nAs a start we could add the logic to `examples/research_projects/wav2vec2` and if it's clean then move to upstream to `src/transformers`", "@patrickvonplaten Sure, I'll give it a go.", "Hello @patrickvonplaten and @tanujjain,\r\n\r\nI have already worked with prefix beam search decoding with language models for wav2vec2 and would like to implement it for huggingface, if you guys are okay with it.", "PRs are very much welcome!", "Any update on this? Specifically any transformer based lm that one can use with wav2vec 2.0?", "As a quick solution, I used the code by original author of the algo which can be found [here](https://gist.github.com/awni/56369a90d03953e370f3964c826ed4b0).\r\n\r\n``` python\r\nimport numpy as np\r\nimport math\r\nimport collections\r\n\r\nNEG_INF = -float(\"inf\")\r\n\r\ndef make_new_beam():\r\n fn = lambda : (NEG_INF, NEG_INF)\r\n return collections.defaultdict(fn)\r\n\r\ndef logsumexp(*args):\r\n \"\"\"\r\n Stable log sum exp.\r\n \"\"\"\r\n if all(a == NEG_INF for a in args):\r\n return NEG_INF\r\n a_max = max(args)\r\n lsp = math.log(sum(math.exp(a - a_max)\r\n for a in args))\r\n return a_max + lsp\r\n\r\ndef decode(probs, beam_size=100, blank=0):\r\n \"\"\"\r\n Performs inference for the given output probabilities.\r\n Arguments:\r\n probs: The output probabilities (e.g. post-softmax) for each\r\n time step. Should be an array of shape (time x output dim).\r\n beam_size (int): Size of the beam to use during inference.\r\n blank (int): Index of the CTC blank label.\r\n Returns the output label sequence and the corresponding negative\r\n log-likelihood estimated by the decoder.\r\n \"\"\"\r\n T, S = probs.shape\r\n probs = np.log(probs)\r\n \r\n # Elements in the beam are (prefix, (p_blank, p_no_blank))\r\n # Initialize the beam with the empty sequence, a probability of\r\n # 1 for ending in blank and zero for ending in non-blank\r\n # (in log space).\r\n beam = [(tuple(), (0.0, NEG_INF))]\r\n \r\n for t in range(T): # Loop over time\r\n next_beam = make_new_beam() # A default dictionary to store the next step candidates.\r\n for s in range(S): # Loop over vocab\r\n p = probs[t, s]\r\n # The variables p_b and p_nb are respectively the\r\n # probabilities for the prefix given that it ends in a\r\n # blank and does not end in a blank at this time step.\r\n for prefix, (p_b, p_nb) in beam: # Loop over beam\r\n # If we propose a blank the prefix doesn't change.\r\n # Only the probability of ending in blank gets updated\r\n if s == blank:\r\n n_p_b, n_p_nb = next_beam[prefix]\r\n n_p_b = logsumexp(n_p_b, p_b + p, p_nb + p)\r\n next_beam[prefix] = (n_p_b, n_p_nb)\r\n continue\r\n # Extend the prefix by the new character s and add it to\r\n # the beam. Only the probability of not ending in blank\r\n # gets updated.\r\n end_t = prefix[-1] if prefix else None\r\n n_prefix = prefix + (s,)\r\n n_p_b, n_p_nb = next_beam[n_prefix]\r\n if s != end_t:\r\n n_p_nb = logsumexp(n_p_nb, p_b + p, p_nb + p)\r\n else:\r\n # We don't include the previous probability of not ending\r\n # in blank (p_nb) if s is repeated at the end. The CTC\r\n # algorithm merges characters not separated by a blank.\r\n n_p_nb = logsumexp(n_p_nb, p_b + p)\r\n \r\n # *NB* this would be a good place to include an LM score.\r\n next_beam[n_prefix] = (n_p_b, n_p_nb) ## add lm here\r\n # If s is repeated at the end we also update the unchanged\r\n # prefix. This is the merging case.\r\n if s == end_t:\r\n n_p_b, n_p_nb = next_beam[prefix]\r\n n_p_nb = logsumexp(n_p_nb, p_nb + p)\r\n next_beam[prefix] = (n_p_b, n_p_nb)\r\n # Sort and trim the beam before moving on to the\r\n # next time-step.\r\n beam = sorted(next_beam.items(),\r\n key=lambda x : logsumexp(*x[1]),\r\n reverse=True)\r\n beam = beam[:beam_size]\r\n best = beam[0]\r\n return best[0], -logsumexp(*best[1])\r\n\r\n# Try the algo on an example\r\ntime = 50\r\noutput_dim = 20\r\nbatch_size = 16\r\n\r\nbatch_probs = np.random.rand(batch_size, time, output_dim)\r\ndecoded_batch = []\r\nfor b in batch_probs:\r\n norm_b = b/np.sum(b, axis=1, keepdims=True)\r\n decoded_batch.append(decode(norm_b, beam_size=3)[0])\r\n```\r\n\r\nTrying to add a language model (for german) like so:\r\n``` python\r\n\r\nfrom transformers import AutoTokenizer, AutoModelWithLMHead\r\ntokenizer_de = AutoTokenizer.from_pretrained(\"dbmdz/german-gpt2\")\r\nmodel_de = AutoModelWithLMHead.from_pretrained(\"dbmdz/german-gpt2\", return_dict_in_generate=True)\r\n\r\n\r\ndef lm_prob(sentence):\r\n last_word_token = tokenizer_de.encode(sentence.split(' ')[-1])\r\n earlier_sentence = ' '.join(sentence.split(' ')[:-1])\r\n input_ids_earlier_sent = tokenizer_de.encode(earlier_sentence, return_tensors=\"pt\") # tokenize rest of the sentence\r\n generated_outputs_lm = model_de.generate(input_ids_earlier_sent,\r\n max_length=len(input_ids_earlier_sent[0]) + 1,\r\n do_sample=True, \r\n num_return_sequences=1,\r\n output_scores=True)\r\n sftmax_prob_lm = generated_outputs_lm.scores[0].softmax(-1)\r\n prob = sftmax_prob_lm[0, last_word_token]\r\n return prob\r\n```\r\n\r\nThe lm snippet should give the prob of having the last word in a beam given all the other preceding characters, but the probabilities for the words I expect are almost always close to zero, so still working on figuring out how better to use the LM. Hence, haven't integrated the LM with the above snippet.\r\n\r\nAs for a decent implementation for beamsearchforctc, I'm thinking on the lines of running the above algo (not the same code obviously) with each sequence in the batch running an independent beamsearch on a different thread/process. \r\n\r\n**Anyone with less complex implementational ideas?** \r\n\r\nFound another implementation [here](https://github.com/githubharald/CTCDecoder/blob/master/src/BeamSearch.py) (without consideration for batch inference).\r\n", "> As for a decent implementation for beamsearchforctc, I'm thinking on the lines of running the above algo (not the same code obviously) with each sequence in the batch running an independent beamsearch on a different thread/process.\r\n\r\nThere you go: https://github.com/mozilla/DeepSpeech/blob/master/native_client/ctcdecode/ctc_beam_search_decoder.cpp#L287\r\n\r\nI'd highly encourage to also consider returning the frames where the probability of the token spikes as it can be used for alignment. Mozilla did it in their implementation and it works quite nicely.\r\n\r\nIs there any restriction on the programming language? The computational complexity of the algorithm is quite high and ctc beam search decoding often the bottleneck.", "I think we can try to add a dependency to wav2letter: https://github.com/flashlight/wav2letter and add LM decoding as explained here on fairseq: https://github.com/pytorch/fairseq/blob/master/examples/wav2vec/README.md#evaluating-a-ctc-model . It would be awesome if we manage to create a nice `run_wav2vec2_eval_with_lm.py` script that people can use out of the box with every wav2vec2 model. We can also make a nice blog post out of this and publish it on our blog :-) \r\n\r\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.", "ping", "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.", "For future developers, you may find this implementation useful. I used the simplest code possible to develop it\r\nhttps://github.com/farisalasmary/wav2vec2-kenlm\r\n", "I'm now working on this topic full time. \r\n\r\nWe will most likely foster a closer collaboration between [pyctcdecode](https://github.com/kensho-technologies/pyctcdecode) and Transformers. [Here](https://github.com/patrickvonplaten/Wav2Vec2_PyCTCDecode) is a github repo that shows how to use `pyctcdecode` with Wav2Vec2 for LM supported decoding. It works quite well with KenLM." ]

[]

[]

[ "Thanks for sharing this issue!\r\n\r\nThe issue here indeed comes from 1.0 that is not from the same dtype than `extended_attention_mask`. A better fix would be to align this line such as the other models by replacing it with:\r\n```\r\nextended_attention_mask = tf.cast(extended_attention_mask, dtype=embedding_output.dtype)\r\none_cst = tf.constant(1.0, dtype=embedding_output.dtype)\r\nten_thousand_cst = tf.constant(-10000.0, dtype=embedding_output.dtype)\r\nextended_attention_mask = tf.multiply(tf.subtract(one_cst, extended_attention_mask), ten_thousand_cst)\r\n```\r\nExtracted from BERT. I will do a fix.", "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." ]

[]

[]

[ "Hey @doubleplusnice,\r\n\r\nI can't really reproduce the bug. \r\n\r\nWhen running your code the output I get is:\r\n\r\n```\r\nSetting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\r\n/home/patrick/python_bin/transformers/generation_utils.py:963: UserWarning: `max_length` is deprecated in this function, use `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.\r\n warnings.warn(\r\nI like cheese, but I don't like cheese. I like cheese because\r\n50257 50257\r\nEmbedding(50257, 768)\r\n\r\n\r\n\r\nSetting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\r\nI like cheese cheese cheese cheese cheese cheese cheese cheese cheese cheese cheese cheese cheese\r\n50258 50257\r\nEmbedding(50258, 768)\r\n\r\n\r\n\r\nSetting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\r\nI like<|endoftext|>The first time I saw the new \"The Walking Dead\"\r\n50257 50257\r\nEmbedding(50257, 768)\r\n```\r\n\r\nwhich seems correct to me.\r\n\r\nCan you maybe try to update on master? Also, I'm testing with the `gpt2` checkpoint.", "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": 1897896961, "node_id": "MDU6TGFiZWwxODk3ODk2OTYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Migration", "name": "Migration", "color": "e99695", "default": false, "description": "" } ]

[]

[ "I don't know what your transformers version is, but from your error message, you should install `sentencepiece`: `pip install sentencepiece` or `conda install sentencepiece` with the appropriate channel (probably `conda-forge`)", "Thank u very much, I have solved this problem. Your work is really remarkable.&nbsp;\r\n\r\n\r\n\r\n---Original---\r\nFrom: \"Lysandre ***@***.***&gt;\r\nDate: Fri, Apr 16, 2021 21:17 PM\r\nTo: ***@***.***&gt;;\r\nCc: ***@***.******@***.***&gt;;\r\nSubject: Re: [huggingface/transformers] failed to import BertModel (#11280)\r\n\r\n\r\n\r\n\r\n \r\nI don't know what your transformers version is, but from your error message, you should install sentencepiece: pip install sentencepiece or conda install sentencepiece with the appropriate channel (probably conda-forge)\r\n \r\n—\r\nYou are receiving this because you authored the thread.\r\nReply to this email directly, view it on GitHub, or unsubscribe.", "Happy to help!", "The " ]

[]

[]

[ "Please fill in the issue template for us to help you. You seem to be on an older transformers version, this was fixed in recent versions.", "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." ]

[]

[]

[]

[]

[]

[ "Seems out of scope." ]

[ { "id": 2648621985, "node_id": "MDU6TGFiZWwyNjQ4NjIxOTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Feature%20request", "name": "Feature request", "color": "FBCA04", "default": false, "description": "Request for a new feature" } ]

[]

[]

[]

[]

[]

[]

[]

[ "> I haven't commented on each print statement, but they should use the logger maybe? \r\n\r\nI started with it first and then replaced with print, because we need all the horizontal space and the really busy long pre-amble is just getting in the way, IMHO. I also don't see what useful information it'd contribute because the tool raises an exception when it detects the problem. Finally, what if someone disabled the logger - it'd not be able to do its work then. Please correct me if I'm missing something.\r\n\r\n> Also I think the `DebugActivationOverflow` should be documented in our internals API doc (since we tell people to use it in their own Trainers). `internal/trainer_utils` is probably the place for that.\r\n\r\nWill do. Thank you for suggesting where to put it.\r\n\r\nConverted to .rst - found this new tool https://github.com/miyakogi/m2r that did it well, just needed to clean up a weird quirk.\r\n", "All links have been added and tweaked the doc some more to improve readability.", "OK, the original functionality has been expanded to include a lot more useful information. Please see the updated copious documentation both in the docstring and the user docs.\r\n\r\nThe main changes are that:\r\n1. we now print each frame separately and include inputs/outputs/weights\r\n2. there is a tracing mode which can easily trace any number of batches at will\r\n\r\nI wasn't sure how I could integrate the new features into the limited `--debug underflow_overflow` interface as it now has 3 optional parameters. So for now these can be activated directly from the script. If you can think how I could make these work with ``TrainingArguments`` I'm all ears.\r\n\r\nAs this has changed a lot inside I'd appreciate another look from @sgugger and @LysandreJik - no rush please. And thank you!" ]

[]

[]

[]

[]

[]

[ "Please use `squad_convert_example_to_features_init(yourtokenizer)` to set the tokenizer." ]

[]

[]

[ "I would guess that's a memory issue indeed! That should be the only difference between the two checkpoints.", "Oof makes sense. I'll check back if that doesn't fix it. Thanks!", "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." ]