Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Sorry about that!" ]

[]

[]

[ "Just tested the script locally and it seems to work great, congrats! We are almost done! The last part would be top adapt the end of the README in the text-classification folder to reflect how to use the new script (since the arguments are a bit different).", "> Just tested the script locally and it seems to work great, congrats! We are almost done! The last part would be top adapt the end of the README in the text-classification folder to reflect how to use the new script (since the arguments are a bit different).\r\n\r\nI've changed the script in README from\r\n```\r\nexport XNLI_DIR=/path/to/XNLI\r\n\r\npython run_xnli.py \\\r\n --model_name_or_path bert-base-multilingual-cased \\\r\n --language de \\\r\n --train_language en \\\r\n --do_train \\\r\n --do_eval \\\r\n --data_dir $XNLI_DIR \\\r\n --per_device_train_batch_size 32 \\\r\n --learning_rate 5e-5 \\\r\n --num_train_epochs 2.0 \\\r\n --max_seq_length 128 \\\r\n --output_dir /tmp/debug_xnli/ \\\r\n --save_steps -1\r\n```\r\nto \r\n```\r\npython run_xnli.py \\\r\n --model_name_or_path bert-base-multilingual-cased \\\r\n --language de \\\r\n --train_language en \\\r\n --do_train \\\r\n --do_eval \\\r\n --per_device_train_batch_size 32 \\\r\n --learning_rate 5e-5 \\\r\n --num_train_epochs 2.0 \\\r\n --max_seq_length 128 \\\r\n --output_dir /tmp/debug_xnli/ \\\r\n --save_steps -1\r\n```\r\n\r\nI've also removed these sentences below from [Fine-tuning on XNLI](https://github.com/huggingface/transformers/blob/master/examples/text-classification/README.md#fine-tuning-on-xnli)\r\n```\r\nThe data for XNLI can be downloaded with the following links and should be both saved (and un-zipped) in a $XNLI_DIR directory.\r\n\r\n XNLI 1.0\r\n XNLI-MT 1.0\r\n```", "> This is in good shape to be merged, thanks a lot for your work! I just have a few comments on how to simplify things here and there since there is only one task to deal with in the new script.\r\n> \r\n> One question I have is, is the tokenizer the same for the training and evaluation datasets, even if the languages, can be different?\r\n\r\nI'm puzzled. Is `Trainer()` class doing the magic under the hood when the languages are different? or is it `AutoTokenizer.from_pretrained`?\r\n", "I missed this is a multinlingual checkpoint, so there is no need for different tokenizers.\r\n\r\n@patil-suraj it's good to merge IMO, I'll let you review one last time and merge if you approve.", " @sgugger Yay :)\r\n@patil-suraj let me know if there's anything you would like me to change further\r\n", "Thanks @sgugger @patil-suraj for your helpful comments and guidance. I was jumping in at the deep end when I attempted this PR to be honest, but yay it's merged 😀", "Great job adding this example and thanks a lot for your PR! Don't hesitate to brag a little bit on Twitter about your contribution ;-) " ]

[]

[]

[ "Thanks for fixing!" ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!" ]

[]

[]

[ "Thank you!\r\n" ]

[]

[]

[ "@sgugger \r\nI got the error message with ```make style```.\r\n\r\n```\r\nkiyoung@medical-ubuntu:~/transformers$ make style\r\nrunning deps_table_update\r\nupdating src/transformers/dependency_versions_table.py\r\nblack examples tests src utils\r\nmake: black: Command not found\r\nMakefile:42: recipe for target 'style' failed\r\nmake: *** [style] Error 127\r\n```", "You need to follow the steps of the [contributing guide](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) do be able to make PRs. In particular you didn't follow the installation part by running `pip install -e \".[dev]\"` since you don't have `black` installed.", "@sgugger \r\nThanks, I did it.", "Thanks for fixing! Now the problem comes from a new release of jax, which has been fixed in master so this is safe to merge.", "This PR introduced a `datasets` submodule. I'm removing it in #9868." ]

[ { "id": 2690307185, "node_id": "MDU6TGFiZWwyNjkwMzA3MTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Performance", "name": "Performance", "color": "207F32", "default": false, "description": "" }, { "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": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "The automatic mixed precision and performance tuning recipes may be helpful.\r\nhttps://pytorch.org/tutorials/recipes/recipes/amp_recipe.html\r\nhttps://pytorch.org/tutorials/recipes/recipes/tuning_guide.html\r\n", "thank you very much, @mcarilli - this is exactly what I was looking for!", "Going to make it into a real doc here: https://github.com/huggingface/transformers/pull/12258" ]

[]

[]

[]

[]

[]

[ "Oh you need to run `make style` on your branch for the styling test to pass. Let me know if you run into any issue doing that!", "Sorry! I'm pretty rusty on the software dev stuff - college was 16 years ago. I think I've fixed it now." ]

[]

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "I'd love to get feedback on that. Thank you!", "As it doesn't seem to resonate as a real need, I'm closing this one." ]

[]

[]

[]

[]

[]

[ "I know than one can add, for example, a line like this\r\n```\r\n# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->FSMT\r\n```\r\nbefore `Attention` module in FSMT. However, this does not copy only additions, but the whole module from BART, which is, in this case, undesirable, I guess, as these modules are a little bit different. But maybe there is another way I am not aware of.", "@LysandreJik, @patrickvonplaten - how can we make sure fsmt gets tracked and synced with all the bart-family changes? while the tokenizer is different, the model is ~95% identical.", "as @stancld said, we can do that with some statements of the following kind:\r\n\r\n```\r\n# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->FSMT\r\n```\r\n\r\nThe difference between the BART and FSMT implementation of the targeted object must only be the \"BART\" occurrences that change to \"FSMT\". @sgugger can tell you more about it.", "Thank you, @LysandreJik\r\n\r\nI think this is really a question to @patrickvonplaten - who I remember was planning to refactor FSMT to match what he did for Bart. So if this is still planned, Patrick perhaps you could add this item to the agenda - keeping FSMT in sync with the Bart-family (modeling only - tokenizer is similar to xlm).\r\n\r\nSo the currently proposed solution can't be used, since Bart diverged since FSMT forked it.\r\n\r\nIt might help to treat FSMT as Bart with the main difference of it having a dual vocab and no tied weights - and a few layers that are different - but identical otherwise. (again for the model only).", "> I think this is really a question to @patrickvonplaten - who I remember was planning to refactor FSMT to match what he did for Bart. So if this is still planned, Patrick perhaps you could add this item to the agenda - keeping FSMT in sync with the Bart-family (modeling only - tokenizer is similar to xlm).\r\n\r\nYes, the FSTM / ProphetNet refactor is still on my ToDo List (think next week is reasonable). After the refactor I'll try to add as many # Copied from statements to keep the models in sync. Nevertheless, this PR can be merged as it is now!\r\n\r\nGreat work @stancld" ]

[]

[]

[ "If I may add, under `def train():`, I think the initialisation of `self._globalstep_last_logged = 0` should be `self._globalstep_last_logged=self.state.global_step`, to ensure that the first logging of the loss is correct when you later divide by `self.state.global_step-self._globalstep_last_logged`? " ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thanks for the fix!", "Thank you for your quick response!" ]

[]

[]

[]

[ { "id": 2669577093, "node_id": "MDU6TGFiZWwyNjY5NTc3MDkz", "url": "https://api.github.com/repos/huggingface/transformers/labels/PR%20for%20Model%20Addition", "name": "PR for Model Addition", "color": "5319e7", "default": false, "description": "" } ]

[]

[ "> Thanks for adding this new model! When referencing other pages in the documentation, it's better to use `:doc:` instead of a hard link, as it will then work in all versions of the documentation (which don't have the same base url).\r\n\r\nSorry that was my bad! I copied it from DialoGPT -> Updated it there as well", "Ah sorry @patrickvonplaten, the `model_doc/` should be removed are the pages are in the same folder. That should resolve the build doc error.", "Great job @stefan-it " ]

[]

[]

[ "Yes, there should be a not here. Do you want to open a PR since you found the problem and its fix?", "Thanks, I'd love to open a PR! Please wait a minute.", "I've opened a PR to fix this issue, and all checks have passed.\r\nI would be grateful if you could check it when you have time. " ]

[]

[]

[ "I like the idea of using `LogitsProcessor`, my only concern is now each time the user wants to use a different `max_length` they would need to pass the `forced_pos_id_pairs= {max_length: eos_token_id}`.\r\n\r\nAlso IMO `prefix_token` sounds more intuitive than `forced_pos_id_pairs`. So I think we could add `ForceTokenProcessor ` and keep both `prefix_token` and `forced_pos_id_pairs` arguments. \r\n\r\nAnd if `prefix_token` is passed or `config.force_bos_token_to_be_generated` is `True` we set\r\n```python\r\nforced_pos_id_pairs = {2: prefix_token or bos_token_id, max_length: eos_token_id}\r\n```\r\n\r\nthis would avoid breaking change.\r\n", "Putting our discussion with @patil-suraj down here for everyone to see. @patil-suraj brought up a good point that `forced_pos_id_pairs` is not super user friendly and might be hard to read and people usually never force \"in the middle\" tokens to be generated. So I like the proposed approach making two LogitsProcessor better as well => we should therefore make \r\na `ForcedBosTokenLogitsProcessor` that takes a `token` as input and always forces the first token to be generated and a `ForcedEosTokenLogitsProcessor` that also takes a `token` as input and forces this token to be generated at `max_length`. \r\nAs discussed we should delete all `adjust_logits` functionality and also get rid of Bart's `config.force_eos_to_be_generated` parameter while keeping full backwards compatibility as discussed.", "Thanks a lot for making this work @patil-suraj! \r\n\r\nWill review the PR now. I saw that you changed the config for `facebook/bart-large-cnn` online which is nice, but should not delete: `force_bos_token_to_be_generated` from the config or it won't be backwards compatible (previous transformer versions need to still use this param). Also it seems like you accidently uploaded a `.ipynb_checkpoints` folder: force_bos_token_to_be_generated", "> We have to make sure that all slow tests for Bart, MBart, Pegasus, Marian, Blenderbot, BlenderbotSmall, FSMT, RAG pass for both PT and TF. I think some of the RAGTokenGeneration could have been broken here\r\n\r\nall slow tests are passing for PT, will run the RAG tests now. Also, there is no `force_bos_token_id_to_be_generated` parameter in `RagConfig`, it's in the `generator` config if the `generator` is BART, and `BartConfig` already handles this.\r\n\r\n> Not sure whether we want to do inheritance for the MBart50Tokenizer\r\n\r\nNo strong opinion here will remove the inheritance.\r\n\r\nRegarding the `prepare_seq2seq_batch` method:\r\nThese checkpoints are mostly intended for multilingual fine-tuning/translation, in this case, it's actually nice to be able to pass lang_id directly for encoding rather than setting `src_lang` and `tgt_lang` on tokenizer each time we have a new language pair.\r\n\r\nIf we have fixed src and target language then in that case `prepare_seq2seq_batch` definitely doesn't make much sense." ]

[]

[]

[ "I'd suggest to still use `allenai/led-base-16384` and just pad the input to a maximum length of only `2048`. Also you could think about reducing `config.attention_window` of the model's config to something like 512 or 256 to make your model more efficient for `2048` input", "E.g. in this notebook: https://colab.research.google.com/drive/12LjJazBl7Gam0XBPy_y0CTOJZeZ34c2v?usp=sharing, I just pad the max length of every input to `8192` and use `allenai/led-base-16384` which works very well!", "Sounds great, I'll look into setting`led.config.attention_window=512` instead of 1024 and the `max_input_length=2048` for the encoder. Thank you for your feedback!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "I don't think that the tests that fails are related to this PR." ]

[]

[]

[ "Would it make sense to move [those 2 lines](https://github.com/huggingface/transformers/blob/0dd939bf1e01594eadf21b40e5cdb07001233cbf/src/transformers/integrations.py#L573-L574) to the init method instead of setting `self._log_model` to False?", "This should be equivalent indeed, and probably easier to read", "Looks great!" ]

[]

[]

[ "Hey @alexvaca0,\r\n\r\nThanks for your issue. We have started to create a more general script called `run_seq2seq.py` with which fine-tuning ProphetNet should work rather easily. \r\n\r\nCould you try to pull current master and do:\r\n\r\n```\r\npython examples/seq2seq/run_seq2seq.py --learning_rate=3e-5 --task summarization --do_train --do_eval --evaluation_strategy steps --model_name_or_path microsoft/prophetnet-large-uncased --output_dir myoutputdir --per_device_train_batch_size 8 --per_device_eval_batch_size 16 --eval_accumulation_steps 8 --gradient_accumulation_steps 8 --num_train_epochs=20 --eval_beams=1 --load_best_model_at_end --save_steps 25 --logging_steps 25 --fp16 --overwrite_output_dir --dataset_name cnn_dailymail --dataset_config_name 3.0.0\r\n```\r\n\r\nfor the cnn/dailymail dataset *e.g.*.\r\n\r\nPlease let me know how it goes, I'm very interested in ProphetNet fine-tuning results.", "Thank you very much for your quick response! @patrickvonplaten As soon as I can, I'll try that command to check if the new script run_seq2seq.py works fine with ProphetNet. When I have results/errors I'll let you know.\r\n", "I've tried to run the script you said @patrickvonplaten , but it returns the following error when evaluating:\r\n\r\n```{python}\r\nAll the weights of ProphetNetForConditionalGeneration were initialized from the model checkpoint at microsoft/prophetnet-large-uncased.\r\nIf your task is similar to the task the model of the checkpoint was trained on, you can already use ProphetNetForConditionalGeneration for predictions without further training.\r\nLoading cached processed dataset at /root/.cache/huggingface/datasets/csv/default-2def39d5bd2a9c76/0.0.0/2960f95a26e85d40ca41a230ac88787f715ee3003edaacb8b1f0891e9f04dda2/cache-7e4959c336c61e5a.arrow\r\nLoading cached processed dataset at /root/.cache/huggingface/datasets/csv/default-2def39d5bd2a9c76/0.0.0/2960f95a26e85d40ca41a230ac88787f715ee3003edaacb8b1f0891e9f04dda2/cache-b898db3404de8043.arrow\r\nThe following columns in the training set don't have a corresponding argument in `ProphetNetForConditionalGeneration.forward` and have been ignored: token_type_ids.\r\nThe following columns in the evaluation set don't have a corresponding argument in `ProphetNetForConditionalGeneration.forward` and have been ignored: token_type_ids.\r\n***** Running training *****\r\n Num examples = 33451\r\n Num Epochs = 20\r\n Instantaneous batch size per device = 8\r\n Total train batch size (w. parallel, distributed & accumulation) = 128\r\n Gradient Accumulation steps = 16\r\n Total optimization steps = 5220\r\n{'loss': 5.5221, 'learning_rate': 4.760536398467433e-05, 'epoch': 0.96}\r\n 5% 250/5220 [16:57<5:41:20, 4.12s/it]***** Running Evaluation *****\r\n Num examples = 2697\r\n Batch size = 16\r\n\r\n 0% 0/169 [00:00<?, ?it/s]\r\n 1% 2/169 [00:00<00:10, 16.33it/s]\r\n 2% 3/169 [00:00<00:12, 13.45it/s]\r\n 3% 5/169 [00:00<00:14, 11.34it/s]\r\n 4% 6/169 [00:00<00:15, 10.62it/s]\r\n 4% 7/169 [00:00<00:21, 7.53it/s]\r\n 5% 8/169 [00:00<00:21, 7.48it/s]\r\n 5% 9/169 [00:01<00:21, 7.54it/s]\r\n 6% 10/169 [00:01<00:24, 6.46it/s]\r\n 7% 11/169 [00:01<00:22, 7.01it/s]\r\n 7% 12/169 [00:01<00:21, 7.40it/s]\r\n 8% 13/169 [00:01<00:20, 7.47it/s]\r\n 8% 14/169 [00:01<00:19, 8.06it/s]\r\n 9% 15/169 [00:01<00:18, 8.48it/s]\r\n 9% 16/169 [00:01<00:19, 7.72it/s]\r\n 10% 17/169 [00:02<00:18, 8.19it/s]\r\n 11% 18/169 [00:02<00:19, 7.60it/s]\r\n 11% 19/169 [00:02<00:19, 7.77it/s]\r\n 12% 20/169 [00:02<00:19, 7.60it/s]\r\n 12% 21/169 [00:02<00:20, 7.31it/s]\r\n 13% 22/169 [00:02<00:18, 7.79it/s]\r\n 14% 23/169 [00:02<00:19, 7.36it/s]\r\n 14% 24/169 [00:03<00:18, 7.76it/s]\r\n 15% 25/169 [00:03<00:18, 7.77it/s]\r\n 15% 26/169 [00:03<00:18, 7.93it/s]\r\n 16% 27/169 [00:03<00:17, 8.29it/s]\r\n 17% 28/169 [00:03<00:18, 7.82it/s]\r\n 17% 29/169 [00:03<00:22, 6.14it/s]\r\n 18% 30/169 [00:03<00:24, 5.79it/s]\r\n 18% 31/169 [00:04<00:22, 6.04it/s]\r\n 19% 32/169 [00:04<00:21, 6.28it/s]\r\n 20% 34/169 [00:04<00:18, 7.11it/s]\r\n 21% 35/169 [00:04<00:17, 7.67it/s]\r\n 21% 36/169 [00:04<00:16, 7.98it/s]\r\n 22% 37/169 [00:04<00:15, 8.27it/s]\r\n 22% 38/169 [00:04<00:17, 7.38it/s]\r\n 23% 39/169 [00:05<00:20, 6.40it/s]\r\n 24% 40/169 [00:05<00:18, 7.00it/s]\r\n 25% 42/169 [00:05<00:16, 7.65it/s]\r\n 26% 44/169 [00:05<00:15, 8.00it/s]\r\n 27% 45/169 [00:05<00:14, 8.46it/s]\r\n 28% 47/169 [00:06<00:14, 8.53it/s]\r\n 28% 48/169 [00:06<00:13, 8.65it/s]\r\n 29% 49/169 [00:06<00:15, 7.98it/s]\r\n 30% 50/169 [00:06<00:14, 8.33it/s]\r\n 30% 51/169 [00:06<00:15, 7.67it/s]\r\n 31% 52/169 [00:06<00:14, 7.95it/s]\r\n 31% 53/169 [00:06<00:16, 7.03it/s]\r\n 32% 54/169 [00:07<00:21, 5.43it/s]\r\n 33% 55/169 [00:07<00:18, 6.27it/s]\r\n 33% 56/169 [00:07<00:16, 6.98it/s]\r\n 34% 57/169 [00:07<00:14, 7.61it/s]\r\n 34% 58/169 [00:07<00:13, 7.94it/s]\r\n 35% 59/169 [00:07<00:13, 8.30it/s]\r\n 36% 60/169 [00:07<00:12, 8.71it/s]\r\n 36% 61/169 [00:07<00:12, 8.69it/s]\r\n 37% 62/169 [00:08<00:12, 8.65it/s]\r\n 37% 63/169 [00:08<00:13, 7.87it/s]\r\n 38% 64/169 [00:08<00:13, 7.93it/s]\r\n 39% 66/169 [00:08<00:12, 8.46it/s]\r\n 40% 67/169 [00:08<00:13, 7.43it/s]\r\n 40% 68/169 [00:08<00:13, 7.74it/s]\r\n 41% 69/169 [00:08<00:12, 8.10it/s]\r\n 41% 70/169 [00:08<00:11, 8.41it/s]\r\n 42% 71/169 [00:09<00:11, 8.79it/s]\r\n 43% 72/169 [00:09<00:10, 9.06it/s]\r\n 43% 73/169 [00:09<00:10, 9.22it/s]\r\n 44% 74/169 [00:09<00:10, 9.02it/s]\r\n 45% 76/169 [00:09<00:10, 8.95it/s]\r\n 46% 77/169 [00:09<00:11, 8.09it/s]\r\n 46% 78/169 [00:09<00:10, 8.39it/s]\r\n 47% 79/169 [00:10<00:10, 8.45it/s]\r\n 47% 80/169 [00:10<00:10, 8.63it/s]\r\n 48% 81/169 [00:10<00:10, 8.44it/s]\r\n 49% 82/169 [00:10<00:12, 7.12it/s]\r\n 49% 83/169 [00:10<00:11, 7.58it/s]\r\n 50% 84/169 [00:10<00:13, 6.09it/s]\r\n 50% 85/169 [00:10<00:12, 6.88it/s]\r\n 51% 86/169 [00:11<00:11, 7.25it/s]\r\n 51% 87/169 [00:11<00:10, 7.80it/s]\r\n 52% 88/169 [00:11<00:09, 8.32it/s]\r\n 53% 89/169 [00:11<00:09, 8.67it/s]\r\n 53% 90/169 [00:11<00:08, 8.86it/s]\r\n 54% 91/169 [00:11<00:08, 8.92it/s]\r\n 54% 92/169 [00:11<00:08, 8.57it/s]\r\n 55% 93/169 [00:11<00:08, 8.56it/s]\r\n 56% 94/169 [00:11<00:08, 8.76it/s]\r\n 56% 95/169 [00:12<00:08, 8.68it/s]\r\n 57% 96/169 [00:12<00:08, 8.58it/s]\r\n 57% 97/169 [00:12<00:09, 7.23it/s]\r\n 58% 98/169 [00:12<00:09, 7.33it/s]\r\n 59% 99/169 [00:12<00:08, 7.89it/s]\r\n 59% 100/169 [00:12<00:08, 8.21it/s]\r\n 60% 101/169 [00:12<00:09, 7.45it/s]\r\n 60% 102/169 [00:12<00:08, 7.87it/s]\r\n 61% 103/169 [00:13<00:08, 7.92it/s]\r\n 62% 104/169 [00:13<00:08, 8.08it/s]\r\n 62% 105/169 [00:13<00:08, 7.98it/s]\r\n 63% 106/169 [00:13<00:08, 7.09it/s]/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [32,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [33,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [34,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [35,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [36,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [37,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [38,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [39,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [40,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [41,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [42,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [43,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [44,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [45,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [46,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [47,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [48,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [49,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [50,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [51,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [52,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [53,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [54,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [55,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [56,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [57,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [58,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [59,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [60,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [61,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [62,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [63,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [0,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [1,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [2,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [3,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [4,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [5,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [6,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [7,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [8,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [9,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [10,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [11,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [12,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [13,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [14,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [15,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [16,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [17,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [18,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [19,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [20,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [21,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [22,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [23,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [24,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [25,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [26,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [27,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [28,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [29,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [30,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [117,0,0], thread: [31,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [96,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [97,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [98,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [99,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [100,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [101,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [102,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [103,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [104,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [105,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [106,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [107,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [108,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [109,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [110,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [111,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [112,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [113,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [114,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [115,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [116,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [117,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [118,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [119,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [120,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [121,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [122,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [123,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [124,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [125,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [126,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [77,0,0], thread: [127,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [64,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [65,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [66,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [67,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [68,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [69,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [70,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [71,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [72,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [73,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [74,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [75,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [76,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [77,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [78,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [79,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [80,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [81,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [82,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [83,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [84,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [85,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [86,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [87,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [88,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [89,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [90,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [91,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [92,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [93,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [94,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [157,0,0], thread: [95,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [64,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [65,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [66,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [67,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [68,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [69,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [70,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [71,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [72,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [73,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [74,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [75,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [76,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [77,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [78,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [79,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [80,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [81,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [82,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [83,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [84,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [85,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [86,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [87,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [88,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [89,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [90,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [91,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [92,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [93,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [94,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [37,0,0], thread: [95,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\nTraceback (most recent call last):\r\n File \"transformers/examples/seq2seq/run_seq2seq.py\", line 541, in <module>\r\n main()\r\n File \"transformers/examples/seq2seq/run_seq2seq.py\", line 503, in main\r\n train_result = trainer.train(model_path=model_path)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer.py\", line 924, in train\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer.py\", line 999, in _maybe_log_save_evaluate\r\n metrics = self.evaluate()\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer_seq2seq.py\", line 96, in evaluate\r\n return super().evaluate(eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer.py\", line 1447, in evaluate\r\n metric_key_prefix=metric_key_prefix,\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer.py\", line 1564, in prediction_loop\r\n loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer_seq2seq.py\", line 175, in prediction_step\r\n model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/trainer.py\", line 1670, in prediction_step\r\n outputs = model(**inputs)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/models/prophetnet/modeling_prophetnet.py\", line 1772, in forward\r\n return_dict=return_dict,\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/models/prophetnet/modeling_prophetnet.py\", line 1656, in forward\r\n return_dict=return_dict,\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/content/gdrive/MyDrive/GColab_folder/transformers/src/transformers/models/prophetnet/modeling_prophetnet.py\", line 1223, in forward\r\n hidden_states = inputs_embeds + position_embeddings\r\nRuntimeError: CUDA error: device-side assert triggered\r\n 5% 250/5220 [17:12<5:42:05, 4.13s/it]\r\n```\r\nI've run it with --no_cuda and there are no errors, it works properly in that setting. Therefore it must be a cuda-related issue. I've tried disabling fp16 and the error persists.", "I confirm that with t5 it works, therefore it's prophetnet-related.", "Who is in charge of developing ProphetNet code? @patrickvonplaten @sgugger ", "Hey @alexvaca0, thanks for trying out the script! I'm quite sure that this is an indexing error that occers because a data sample is too large for the model to handle. It should be easy to fix by simply adding:\r\n\r\n```\r\n--max_source_length 512\r\n```\r\n\r\nto the command above. Could you try this and let me know if it works? :-)", "@patrickvonplaten Great! That was it, the sequence length! \r\n\r\nActually, I'm trying to fine-tune ProphetNet in a Summarization task, in which models like T5, BART etc achieve eval losses of around 0.5-0.6 (approx), but with ProphetNet I'm not able to go below 5, and the eval loss doesn't actually decrease over training, it seems like it's diverging. I've tried using the same parameters as with BART and T5, and also with the parameters of the paper (https://arxiv.org/pdf/2001.04063.pdf) for CNN/DailyMail, that is batch size 512, learning rate 1e-04 with warmup steps 1000 (in my case I use less due to training data size). \r\n\r\nAny recommendations/suggestions? ProphetNet was expected to work similarly to BART but its performance is much worse until now...", "I don't know if this warning provides some extra info: The following columns in the training set don't have a corresponding argument in `ProphetNetForConditionalGeneration.forward` and have been ignored: token_type_ids.\r\n @patrickvonplaten ", "> @patrickvonplaten Great! That was it, the sequence length!\r\n> \r\n> Actually, I'm trying to fine-tune ProphetNet in a Summarization task, in which models like T5, BART etc achieve eval losses of around 0.5-0.6 (approx), but with ProphetNet I'm not able to go below 5, and the eval loss doesn't actually decrease over training, it seems like it's diverging. I've tried using the same parameters as with BART and T5, and also with the parameters of the paper (https://arxiv.org/pdf/2001.04063.pdf) for CNN/DailyMail, that is batch size 512, learning rate 1e-04 with warmup steps 1000 (in my case I use less due to training data size).\r\n> \r\n> Any recommendations/suggestions? ProphetNet was expected to work similarly to BART but its performance is much worse until now...\r\n\r\nInteresting! Could you share the exact command you used here? Also pinging @qiweizhen - do you know what could be a problem for this? Are we sure that the n-gram loss is correctly implemented?", "```{bash}\r\npython transformers/examples/seq2seq/run_seq2seq.py \\\r\n --model_name_or_path microsoft/prophetnet-large-uncased \\\r\n --do_eval --do_train \\\r\n --task summarization \\\r\n --train_file train_df.csv \\\r\n --validation_file val_df.csv \\\r\n --output_dir prophetnet_0201 \\\r\n --overwrite_output_dir \\\r\n --per_device_train_batch_size=8 \\\r\n --per_device_eval_batch_size=16 \\\r\n --eval_accumulation_steps=10 \\\r\n --text_column text \\\r\n --max_source_length 364 \\\r\n --summary_column summary \\\r\n --max_target_length 60 \\\r\n --val_max_target_length 60 --evaluation_strategy steps \\\r\n --gradient_accumulation_steps 64 --num_train_epochs=20 --eval_beams=1 \\\r\n --load_best_model_at_end --save_steps 75 --logging_steps 75 --learning_rate 1e-04 --warmup_steps 200 \r\n```\r\nThis is the command I'm using. After trying some modifications I observe the same: no progress is made in evaluation, and almost no progress in training (loss 5.1 after almost 7 epochs), so it seems there may be some issue with ProphetNet implementation...\r\n@patrickvonplaten @qiweizhen ", "I got the same problem that the training crashed in run_seq2seq.py. BTW, I guess that it is related to sequence lengths in configuration and training sets.\r\n\r\n```bash\r\npython ./run_seq2seq.py \\\r\n --model_name_or_path sshleifer/student_marian_en_ro_6_1 \\\r\n --do_train \\\r\n --do_eval \\\r\n --task translation_en_to_ro \\\r\n --dataset_name wmt16 \\\r\n --dataset_config_name ro-en \\\r\n --source_lang en_XX \\\r\n --target_lang ro_RO\\\r\n --output_dir ~/tmp/tst-translation \\\r\n --per_device_train_batch_size=4 \\\r\n --per_device_eval_batch_size=4 \\\r\n --overwrite_output_dir \\\r\n --predict_with_generate\r\n```\r\n\r\ntransformers version: 4.4.0.dev0\r\nPlatform: Ubuntu 16.04.7\r\nPython version: 3.8.5\r\nPyTorch version (GPU?): 1.7.1 (YES)\r\nTensorflow version (GPU?):\r\nUsing GPU in script?: YES\r\nUsing distributed or parallel set-up in script?: Yes, it detects 6 GPUs.\r\n\r\nError:\r\n\r\n> /opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu/opt/conda/conda-bld/pytorch_1607369981906/work/at en/src/ATen/native/cuda/Indexing.cu:658/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: index SelectLargeIndex:658: indexSelectLargeIndex: block: [264,0,0], thread: [0,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [264,0,0], thr ead: [1,0: block: [267,0,0: indexSelectLargeIndex], thread: [32,0: block: [263,0,0,0,0], thread: [96,0,0] Assertion `srcIndex < srcSele ctDimSize` failed.\r\n] Assertion `srcIndex < srcSelectDimSize] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [263,0/opt/con da/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu` failed.\r\n,0:658/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu], thread: [97: indexSelectLargeIndex:658,0: block: [264: indexSelectLargeIndex,0,0: block: [267] Assertion `srcIndex < srcSelectDimSize,0,0` failed.\r\n], thread: [2,0/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu,0], thread: [33:658,0,0: indexSele ctLargeIndex] Assertion `srcIndex < srcSelectDimSize,0: block: [263` failed.\r\n] Assertion `srcIndex < srcSelectDimSize,0` failed.\r\n/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [267,0,0], thr ead: [34,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [267,0,0], thr ead: [35,0,0/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu,0] Assertion `srcIndex < srcSelectDim Size:658], thread: [98` failed.\r\n: indexSelectLargeIndex,0/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu: block: [264,0:658,0] As sertion `srcIndex < srcSelectDimSize: indexSelectLargeIndex,0` failed.\r\n: block: [267], thread: [3/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIn dex: block: [263,0,0], thread: [99,0,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [263,0,0,0,0], thread: [100,0,0,0], thread: [36] Assertion `srcIndex < srcSelectDimSize,0,0` failed.\r\n/opt/conda/conda-bld/pytorch_1607369981906/work/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [264,0,0], thr ead: [4,0] Assertion `srcIndex < srcSelectDimSize,0] Assertion `srcIndex < srcSelectDimSize` failed.\r\n,0` failed.\r\n", "Any updates on ProphetNet loss?? @patrickvonplaten ", "I have some more information on this. After training for 20 epochs, it learns almost nothing. Most interestingly, its outputs doesn't change when inputs change, that is, it always predicts the same. Predictions are like a mix of different summaries, getting elements from different types of summarizable texts, but it's the same for all... This brings me to think that in some way the network is constructed so that the output layer must always output the same thing, as if it must improve on all batches at the same time, I don't know if I'm explaining myself. It's clear that it is learning \"something\", in the sense that the summaries are clearly taken from my corpus style, but it's kind of learning to make the same summary for all texts. Since I'm using the same script as for other models, I guess there is some error in the network implementation...", "Hey @alexvaca0,\r\n\r\nI think I can reproduce your error. My training loss is also not improving after quite some time - will look into it!", "Okay perfect! Please let me know when the issue is solved.", "The original author @qiweizhen of the model was so nice to say he'll take a look. @qiweizhen - feel free to directly post any potential bugs in this PR.", "Any updates on this? @qiweizhen @patrickvonplaten ", "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.", "If prophetnet is not going to be fixed, then I think it should be removed from the library, as it is worthless having it here without being able to use it.", "The model works fine in inference - it's the training that seems to be buggy. @qiweizhen - do you think we could take a look at ProphetNet together? ", "> The model works fine in inference - it's the training that seems to be buggy. @qiweizhen - do you think we could take a look at ProphetNet together?\r\n\r\n\r\n\r\n> If prophetnet is not going to be fixed, then I think it should be removed from the library, as it is worthless having it here without being able to use it.\r\n\r\nSorry. Will fix it as soon as possible.\r\n", "It's strange that I can get correct inference / forward results with beam search, but as you pointed out, the model has non-convergence problem. I try to load the pretrained checkpoint and finetuned checkpoint to carry out further fine-tuning, all of their loss is optimized to 7.x and keeps that loss. With the finetuned checkpoint plus further fine-tuning, the results are still reasonable but a bit worse. I suspect the most part of the model is frozen and only a small part is trainable but I failed to find this bug. I also tried overfitting experiments and the model still can not converge. I will try 1) old Transformers version and 2) fairseq model to compare the intermediate hidden states with the latest Transformers prophetnet model to localize the bug this weekend.", "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.", "Has the problem of ProphetNet non-convergence been solved? I want to fine tune it based on its checkpoint.", "I think the code to compute the loss may be wrong.\r\n\r\nThis is the code to compute the loss in [`ProphetNetForConditionalGeneration`](https://github.com/huggingface/transformers/blob/master/src/transformers/models/prophetnet/modeling_prophetnet.py#L1968):\r\n```python\r\n predicting_streams = outputs[1].view(batch_size, self.config.ngram, sequence_length, -1)\r\n predict_logits = self.lm_head(predicting_streams)\r\n\r\n ...\r\n\r\n loss = None\r\n if labels is not None:\r\n loss = self._compute_loss(predict_logits, labels)\r\n```\r\nThe shape of `predicting_streams` is `(batch_size, ngram, sequence_length, hidden_size)`.\r\nThe shape of `predict_logits` is `(batch_size, ngram, sequence_length, vocab_size)`.\r\nThe shape of `labels` is `(batch_size, sequence_length)`.\r\n\r\nThen pass `predict_logits` and `labels` to `_compute_loss`, the code of [`_compute_loss`](https://github.com/huggingface/transformers/blob/master/src/transformers/models/prophetnet/modeling_prophetnet.py#L2001) is: \r\n```python\r\n def _compute_loss(self, logits, labels, ignore_index=-100):\r\n expend_targets = labels.new_zeros(self.config.ngram, labels.size(0), labels.size(1)).fill_(ignore_index)\r\n\r\n for i in range(self.config.ngram):\r\n if i > 0 and self.disable_ngram_loss:\r\n break\r\n expend_targets[i, :, :] = labels\r\n\r\n lprobs = nn.functional.log_softmax(\r\n logits.view(-1, logits.size(-1)),\r\n dim=-1,\r\n dtype=torch.float32,\r\n )\r\n\r\n loss = nn.functional.nll_loss(lprobs, expend_targets.view(-1), reduction=\"mean\")\r\n\r\n ...\r\n\r\n return loss\r\n```\r\nThe shape of `expend_targets` is `(ngram, batch_size, sequence_length)`, the shape of `expend_targets.view(-1)` is `(ngram * batch_size * sequence_length)`, .\r\nThe shape of `lprobs` is `(batch_size * ngram * sequence_length, vocab_size)`.\r\nThen computing the `nll_loss` of `lprobs` and `expend_targets` leads to the mismatch.", "@patrickvonplaten ", "This is the code of the prophetnet [hub](https://github.com/microsoft/ProphetNet/blob/master/ProphetNet_En/prophetnet/ngram_criterions.py#L36).\r\nYou can see [line 62](https://github.com/microsoft/ProphetNet/blob/master/ProphetNet_En/prophetnet/ngram_criterions.py#L62) that the shape of `logits` is `(ngram * batch_size * sequence_length, vocab_size)`.", "Hey @StevenTang1998,\r\n\r\nThanks a lot for taking a closer look here! Would you be interested in opening a PR to fix it?" ]

[ { "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!" } ]

[]

[ "Thank you very much, @oborchers for opening a new ticket and re-testing with other models and verifying that this problem is project-wide.\r\n\r\nI hope @mfuntowicz gets a chance to have a look at it, or tag someone else who understands this sub-domain.", "Hi @mfuntowicz @stas00 , is this a known issue with GPT2 as well? Please let me know if there is a workaround.\r\n\r\nI was considering to convert ```gpt2``` or ```gpt2-medium``` to ONNX using the notebook provided [here](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb).\r\n\r\nOn executing the line of code below:\r\n```convert(framework=\"pt\", model=\"gpt2-medium\", output=Path(\"onnx/gpt2-medium.onnx\"), opset=11)```\r\n\r\nI get this error:\r\n```~/miniconda3/envs/onnx/lib/python3.9/site-packages/torch/onnx/utils.py in _validate_dynamic_axes(dynamic_axes, model, input_names, output_names)\r\n 1115 for i, x in enumerate(value):\r\n 1116 if not isinstance(x, int):\r\n-> 1117 raise ValueError(\"The type of axis index is expected to be an integer\")\r\n 1118 if x in value_dict:\r\n 1119 warnings.warn('Duplicate dynamic axis index {} was provided for input {}.'\r\n\r\nValueError: The type of axis index is expected to be an integer```", "I recently stumbled upon this issue myself. Specifically case 2. The same error appears for `facebook/bart-large`, `facebook/bart-large-cnn`, `IlyaGusev/mbart_ru_sum_gazeta`. The main issue here is that for some outputs the tokenizer/model gives not a tensor, but rather a **tuple of tensors**, which is then converted into a list of shape dicts. \r\n\r\n`torch.onnx._validate_dynamic_axes` (line 1193 in the latest release) expects a dict (and does nothing) or a list of ints for `dynamic_axes` (and mocks up some axes names), however (for the reason above) it gets a __list of dicts ([map int -> string])__\r\n\r\n```python3\r\n for key, value in dynamic_axes.items():\r\n if key not in valid_names:\r\n warnings.warn(\"Provided key {} for dynamic axes is not a valid input/output name\".format(key))\r\n if isinstance(value, list):\r\n warnings.warn('No names were found for specified dynamic axes of provided input.'\r\n 'Automatically generated names will be applied to each dynamic axes of input {}'.format(key))\r\n\r\n value_dict = {}\r\n for i, x in enumerate(value):\r\n if not isinstance(x, int):\r\n raise ValueError(\"The type of axis index is expected to be an integer\")\r\n if x in value_dict:\r\n warnings.warn('Duplicate dynamic axis index {} was provided for input {}.'\r\n .format(x, key))\r\n else:\r\n value_dict[x] = str(key) + '_dynamic_axes_' + str(i + 1)\r\n dynamic_axes[key] = value_dict\r\n```\r\n\r\nI will keep digging into that, but the core question here is why Bart and related models return tuple of tensors (for outputs 1 to 12; outputs 0 and 13 are fine)? Although, I'm not an expert in either transformers, pytorch or onnx, so I might be missing something.\r\n\r\nOn a slight tangent here, is there a specific reason why `summarization` pipeline is not in the supported pipeline types for this script? ", "any update?", "any update?", "any update? ", "We're currently working on a rework of the ONNX implementation within Transformers, which is available here: https://github.com/huggingface/transformers/pull/11786\r\n\r\nInstead of offering a script to enable conversions for all models (which was not kept up to date with recent model releases), we're opting for a case-by-case approach, while offering the tools to convert models manually in a straightforward and simple manner; by creating `OnnxConfig` configuration objects to specify the input and output types of each model.\r\n\r\nPlease take a look at the PR and give us your feedback.", "@LysandreJik: Thank you very much! I think this is an excellent way to go. Having converted a dozen models myself, we internally went for something similar, albeit not nearly as streamlined / sophisticated.\r\n\r\n```\r\[email protected](auto_attribs=True)\r\nclass TransformersONNXConfig(BaseConfig):\r\n \"\"\"Provides the basic configuration for all models.\"\"\"\r\n\r\n base_model: str\r\n trans_cfg: PretrainedConfig\r\n\r\n input_names: List[str]\r\n output_names: List[str]\r\n dynamic_axes: Dict\r\n model_args: Set[torch.tensor]\r\n tokenizer: PreTrainedTokenizerFast\r\n extra_args: Dict\r\n```\r\n\r\nand\r\n\r\n```\r\ndef create_and_export_onnx_model(self):\r\n \"\"\"Creates a new model if the current model does not exist and exports it.\"\"\"\r\n torch.onnx.export(\r\n self.create_torch_model(),\r\n self.cfg.model_args,\r\n f=self.onnx_posix_pth,\r\n input_names=self.cfg.input_names,\r\n output_names=self.cfg.output_names,\r\n dynamic_axes=self.cfg.dynamic_axes,\r\n do_constant_folding=True,\r\n use_external_data_format=False,\r\n enable_onnx_checker=True,\r\n opset_version=12,\r\n )\r\n```\r\n\r\nWhere the most important part is `self.create_torch_model`, as we regularly modify the basic torch model with custom layers down the line. Is support for such a feature planned? If not, is it considerable? As it would substantially easy conversion of custom models, such as the [sbert](https://www.sbert.net) ones.\r\n\r\nFurthermore, would it make sense to make `OnnxConfig` a part of the `PreTrainedModel` config to enable support from the get-go?\r\n\r\nAnd finally, I assume this leaves us with the export, so that for seq2seq models we need still need to re-write the `.generate` function? Or is it possible to add support for an ONNX model from your side (probably difficult, as it's a part of the pre-trained model already, which would require double loading the model)? ", "Thanks @oborchers for your comments and use-cases.\r\n\r\nI will let @LysandreJik speak about a potential integration of the `OnnxConfig` within the `PreTrainedModel` config, my initial plan was to have 100% backward compatibility, this explain why I put this somewhere else _(currently)_.\r\n\r\nRegarding `generate`, this is something that might require some investigations but I'm seeing good opportunities to have something within the ONNX graph with [the recent knobs released by Microsoft](https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/python/tools/transformers/notebooks/Inference_GPT2-OneStepSearch_OnnxRuntime_CPU.ipynb) folks on the ONNXRuntime project _(cc @tianleiwu for visibility on this)_.\r\n\r\nStill, for this initial rework of the ONNX exporting capabilities we focused on \"model only\", with the ability to extend to full pipelines in the future. Generation is definitively one of the hardest task to get within the graph, but also one where I can see the biggest benefits.", "@mfuntowicz: Thank you for your feedback! Yes, I understand the point for the compatibility to the fullest. After all, it's not that difficult to get to the config if done once or twice. \r\n\r\nRegarding the `.generate` function. Thanks for the link! Will look into this more! Yes, absolutely!!", "> Hi @mfuntowicz @stas00 , is this a known issue with GPT2 as well? Please let me know if there is a workaround.\r\n> \r\n> I was considering to convert `gpt2` or `gpt2-medium` to ONNX using the notebook provided [here](https://github.com/huggingface/transformers/blob/master/notebooks/04-onnx-export.ipynb).\r\n> \r\n> On executing the line of code below: `convert(framework=\"pt\", model=\"gpt2-medium\", output=Path(\"onnx/gpt2-medium.onnx\"), opset=11)`\r\n> \r\n> I get this error:\r\n> \r\n> ```python\r\n> 1115 for i, x in enumerate(value):\r\n> 1116 if not isinstance(x, int):\r\n> -> 1117 raise ValueError(\"The type of axis index is expected to be an integer\")\r\n> 1118 if x in value_dict:\r\n> 1119 warnings.warn('Duplicate dynamic axis index {} was provided for input {}.'\r\n> \r\n> ValueError: The type of axis index is expected to be an integer```\r\n> ```\r\n\r\nHello @mriganktiwari \r\nAny update on this? I am still facing the issue with GPT-2. I have used the same code as yours. Please guide, thanks!", "can i work on this?\r\nplease assign this to me\r\n", "Hey! This is no longer something handled at the `transformers` level, will close it! Sorry for the inconvenience \r\n\r\nThe way to handle it now is through optimum! See this documentation page for more information: [ONNX exporter](https://huggingface.co/docs/optimum/exporters/onnx/overview)" ]

[]

[]

[ "Edit: I am not sure why this param is always set to `True` except for gradient checkpointing. We can certainly make a training argument control its value to avoid hard-coding it. At least to experiment and benchmark whether it's best at True/False." ]

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

[]

[ "Mmm, which scripts use `n_obs`? I don't remember seeing this one the official maintained examples.\r\n\r\n`--max_steps` is different from `n_train`/`n_val`/`n_test`: `--max_steps` runs training for `max_steps`, using the *full training set*. `--n_train` restrains the training set to its first `n_train` samples. The first has its place inside `Trainer` for obvious reason, the second is part of the processing of the training (or eval/test) dataset so I don't think this has its place in `Trainer`.\r\n\r\nAs for a consistent way to do this in all examples, it doesn't really matter in non seq2seq scripts as their evaluation runs quite fast. I imagine those arguments were introduces in the seq2seq script originally because its evaluation is super long. We can add them with a need-to basis on other datasets, but I haven't felt the need to do this.", "> Mmm, which scripts use `n_obs`? I don't remember seeing this one the official maintained examples.\r\n\r\nall `seq2seq/run_*py`\r\n\r\n> `--max_steps` is different from `n_train`/`n_val`/`n_test`: `--max_steps` runs training for `max_steps`, using the _full training set_. `--n_train` restrains the training set to its first `n_train` samples. The first has its place inside `Trainer` for obvious reason, the second is part of the processing of the training (or eval/test) dataset so I don't think this has its place in `Trainer`.\r\n\r\nright, so this confusion leads to an incorrect benchmark. that's what I thought last night but it was too late to see.\r\nhttps://github.com/huggingface/transformers/issues/9371#issuecomment-767323420\r\n\r\nWe need a way to be able to truncate the dataset to an identical size and then compare say 1-gpu vs 2-gpu benchmark on the same total number of input objects.\r\n\r\nSo how do we currently do that with other scripts that aren't `finetune_trainer.py`?\r\n\r\n> As for a consistent way to do this in all examples, it doesn't really matter in non seq2seq scripts as their evaluation runs quite fast. I imagine those arguments were introduces in the seq2seq script originally because its evaluation is super long. We can add them with a need-to basis on other datasets, but I haven't felt the need to do this.\r\n\r\nfast? try `run_clm.py` on gpt2/wiki - it's multiple hours\r\ne.g. see: https://github.com/huggingface/transformers/issues/9371#issuecomment-759074475", "> all seq2seq/run_*py\r\n\r\nThose are not official maintained examples except for the new `run_seq2seq`. No one has really touched them since Sam left and they are in need for cleanup ;-)\r\n\r\n> fast? try run_clm.py on gpt2/wiki - it's multiple hours e.g. see: #9371 (comment)\r\n\r\nYou are pointing to a comment that does not contain any evaluation. So I stand by what I say. Evaluation on wikitext-2 runs in a couple of seconds.\r\n\r\n> We need a way to be able to truncate the dataset to an identical size and then compare say 1-gpu vs 2-gpu benchmark on the same total number of input objects.\r\n\r\nLike I said, if it's needed it can be added.\r\n\r\n> So how do we currently do that with other scripts that aren't finetune_trainer.py?\r\n\r\nBy opening a PR adding this ;-)", "Thank you for clarifying which is which, @sgugger \r\n\r\nOK, so what should we call a new flag in HF Trainer that would be an equivalent of --n_train? or use the same?\r\n\r\nDo you suggest it should be train-specific?", "I think it should be in the scripts, not the Trainer, as it's part of the preprocessing. I don't think it should be train-specific, we can do eval/test like in the finetune_trainer script.", "but then we have to change all the scripts. Why not have an option to truncate the dataset at trainer level and solve it at once for all scripts?", "Because it doesn't have much to do with the Trainer itself IMO. It's like putting all the arguments of all the scripts about tokenization in the Trainer, it doesn't really make sense as the Trainer is supposed to take the lead after the data preprocessing.\r\n\r\nLet's see if @LysandreJik and @patrickvonplaten think differently maybe?", "This makes sense, then perhaps having a Trainer-subclass that all scripts can tap into?\r\n\r\nAlso may I suggest that `--max_steps` is an ambiguous argument as it tells the user nothing about whether this is per gpu or per the whole thing?", "The documentation says number of training steps. I don't see how the number GPU intervenes here as a training step is the full combination of forward, backward (perhaps multiple times if gradient accumulation is activated) and optimizer step.\r\n\r\nOne training step can have a different number of training samples depending on the number of GPUs, but also depending on the batch size, gradient accumulation steps etc. This information is logged at the beginning of training (`logger.info(f\" Total train batch size (w. parallel, distributed & accumulation) = {total_train_batch_size}\")` in Trainer.train)", "Right, so what you're saying is that `--max_steps` is just the wrong tool for the truncating job and we need an explicit `--use-that-many-total-train-records`.\r\n\r\nHonestly, I have been staring at all these different trainer options for a long time now and I still get confused at which is which, and which are impacted by number of gpus and which aren't. Every time this happens I have to go through the source code to see how it's used and then I get it. To me some of these arg names are hard to make sense of in the multi-gpu vs single gpu env.\r\n\r\n* `--per_device_train_batch_size` is loud and clear.\r\n* `--max_steps` is not.\r\n\r\nI propose we use `total` and `per_device` prefix for any cl arg that behaves differently depending on the number of gpus.", "The problem is that this then is a breaking change. I'm not necessarily super fond of the name `max_steps` myself but I'm not sure it's worth going through the trouble of a deprecation cycle for this one.", "Do you think it's actually used a lot? \r\n\r\nI agree with avoiding break changes, but since we are trying to make the API intuitive, such changes in the long run will benefit a much larger community than the annoyance it'd cause to those who use it right now.\r\n\r\nI think the main issue we have here is that all these proposals to renames happen dynamically. But instead I think it'd make sense for a group of us to sit down, review all the cl args and do a single adjustment. Surely, this won't guarantee that in the future we won't find we missed something, but it's definitely better than doing it a little bit at a time, which is much more annoying.\r\n\r\nIn some previous projects for such things we also had a back-compat mode, which ones enabled supported a whole bunch of old ways until the user was ready to make the shift to the new code. Surely a rename of a cl arg could be easily supported by such feature. So here, instead of a deprecation cycle per item the approach is to keep anything old around but only if it's loaded from a helper module. So that the main code remains clean of deprecated things. This was in a different programming environment where it was developer, so I will have to think how to do the same here.", "Note that this is not just a CI arg rename, since `TrainingArguments` is also a public class users may very well directly use in their code (you need to instantiate one each time you use a `Trainer`). We can certainly have a discussion around the arguments and decide which one we want to rename, though it should be in a separate issue. We're starting to derail this one ;-)\r\n\r\nAnd from the issues, I'd say that half the users use `num_train_epochs` and half use `max_steps` to control the length of their training, so it is used a lot.", "Thank you for flagging that we are diverging from the topic at hand, @sgugger \r\nAs you suggested I opened a new one: https://github.com/huggingface/transformers/issues/9821\r\n\r\nAnd thank you for confirming that these are used a lot.", "> Because it doesn't have much to do with the Trainer itself IMO. It's like putting all the arguments of all the scripts about tokenization in the Trainer, it doesn't really make sense as the Trainer is supposed to take the lead after the data preprocessing.\r\n> \r\n> Let's see if @LysandreJik and @patrickvonplaten think differently maybe?\r\n\r\nSo for the benefit of reviewers, and to bring us back to the focus of this Issue. I proposed to have a cl arg that will truncate the dataset (train, others?) (total!) across all example scripts.\r\n\r\n@sgugger, correctly suggested that perhaps this shouldn't belong to Trainer, and then I suggested that perhaps there should be a sub-class that does such nice little tweaks consistently across all example scripts, rather than manually replicating the same code and which often leads to scripts diverging.\r\n\r\nPlus, @sgugger points out that `examples/seq2seq/run*.py` haven't yet been converted to the new way.", "I always thought that `max_steps` defines the total number of weight update steps (which is then not really influenced by other parameters such as number of GPUs or `gradient_accumalation_steps` or whatever). To me it defines: \"How often do I want to update my weights?\" or am I wrong here?. Think the name is clear and does not need to be changed, the documentation could be updated with a sentence that makes clear that `max_steps` = number of weight updates. Also, I use this arg quite often when training and think it's important to keep.\r\n\r\nI agree with @sgugger here that I think a `--max_num_train_samples` arg (or whatever the name) should not go into the trainer, but should be added to all examples scripts. It's actually incredibly easy to do this with `datasets`: \r\n\r\n```python\r\nds = load_dataset(\"crime_and_punish\", split=\"train\")\r\nds = ds.select(range(arg.max_num_train_samples))\r\n```\r\n\r\nI'm totally fine with having this as another cl arg for the scripts, but don't think it's the responsibility of the `trainer`.", "I agree with Sylvain and Patrick about `max_steps`.\r\n\r\nAnd for controlling the number of examples, this should go in scripts than `Trainer`, as we do all the pre-processing in the scripts. We could add two arguments to `DataTrainingArguments` in every script.\r\n`--max_train_samples` = number of training examples\r\n`--max_val_samples` = number of validation examples\r\n\r\nThese args are already there in the new `run_seq2seq.py` script.\r\n", "Thank you for your input, guys. Your suggestions work for me.\r\n\r\n> We could add two arguments to DataTrainingArguments in every script.\r\n> --max_train_samples = number of training examples\r\n> --max_val_samples = number of validation examples\r\n> \r\n> These args are already there in the new run_seq2seq.py script.\r\n\r\nbut not in other `run_*.py` scripts.\r\n\r\nand then we have `test` too - at least in `finetune_trainer.py`\r\n\r\nI proposed to have a Trainer subclass that implements this for all scripts vs repeating the same cl arg definition and code in every script a new (and forgetting to sync some) - could you please address that? \r\n\r\n---------------------------\r\n\r\nThe other slight confusion across some scripts is `val` vs `eval` - it's inconsistent - some reports say `val` others `eval` - train/val/test are splits and are orthogonal to train/evaluate/predict - and while they are the same for train, the rest are just confusing, since you can have predict for val split and evaluate for test split. Should we discuss this in a separate issue?", "> I proposed to have a Trainer subclass that implements this for all scripts vs repeating the same cl arg definition and code in every script a new (and forgetting to sync some) - could you please address that?\r\n\r\nI don't think this is a good idea personally. The goal of the scripts is to provide examples for our users. Having examples that don't use the main object of the library is counterproductive. It's one other instance where we have to bear the burden of duplicate code to make the user experience easier IMO.\r\n\r\n> The other slight confusion across some scripts is val vs eval - it's inconsistent - some reports say val others eval - train/val/test are splits and are orthogonal to train/evaluate/predict - and while they are the same for train, the rest are just confusing, since you can have predict for val split and evaluate for test split. Should we discuss this in a separate issue?\r\n\r\nI think this is mostly `finetune_trainer` (and maybe `run_seq2seq2` since I may have copied some names) not using the same terminology as the other scripts in this instance. So those two scripts should get aligned with the rest on this matter. Again, let's keep the examples simple (I feel like I'm repeating this all day long but *they are just examples* we cannot have scripts that will solve every use case and trying to do so make them un-understandable for our users) and match train/eval/test with what is done (training/evaluation/predict).", "> > I proposed to have a Trainer subclass that implements this for all scripts vs repeating the same cl arg definition and code in every script a new (and forgetting to sync some) - could you please address that?\r\n> \r\n> I don't think this is a good idea personally. The goal of the scripts is to provide examples for our users. Having examples that don't use the main object of the library is counterproductive. It's one other instance where we have to bear the burden of duplicate code to make the user experience easier IMO.\r\n\r\nYou're correct. I didn't think of that. \r\n\r\nSo we have a conflict here between example scripts and them being used for more than that.\r\n\r\nI, for one, need a solid set of scripts to do:\r\n\r\n1. integration validation\r\n2. benchmarking\r\n\r\nIn the absence of these I have been heavily relying on the example scripts. And this is probably where the conflict is.\r\n\r\nSo I keep on bringing this up - should we have a set of scripts that are not examples, but real production work horses and we treat them as such? Perhaps they can have much less functionality but do it consistently across different domains and simple?\r\n\r\nPerhaps, instead of `run_(foo|bar|tar).py` it's one script that can tap into any of these domains and then it can have a simple identical cl args. And all we change is model names and most other args are almost the same.\r\n\r\n> > The other slight confusion across some scripts is val vs eval - it's inconsistent - some reports say val others eval - train/val/test are splits and are orthogonal to train/evaluate/predict - and while they are the same for train, the rest are just confusing, since you can have predict for val split and evaluate for test split. Should we discuss this in a separate issue?\r\n> \r\n> I think this is mostly `finetune_trainer` (and maybe `run_seq2seq2` since I may have copied some names) not using the same terminology as the other scripts in this instance. So those two scripts should get aligned with the rest on this matter. Again, let's keep the examples simple (I feel like I'm repeating this all day long but _they are just examples_ we cannot have scripts that will solve every use case and trying to do so make them un-understandable for our users) and match train/eval/test with what is done (training/evaluation/predict).\r\n\r\nYou're absolutely correct, please see my response in the comment above.\r\n", "> So I keep on bringing this up - should we have a set of scripts that are not examples, but real production work horses and we treat them as such? Perhaps they can have much less functionality but do it consistently across different domains and simple?\r\n\r\nIf the basic examples do not suffice, then yes, definitely.", "But we are walking in circles. If these are examples and they are treated as examples, these aren't tools to be relied upon. I hope you can see the irony...\r\n\r\nI need a solid tool that will not change its API, start doing all the benchmarks in it so that we could go back to benchmarks from 6 months or a year ago and be able to run those and re-check.\r\n", "I'm not sure why you say we are walking in circles. I just dais yes to having benchmark-specific scripts if the examples do not have all the functionality you need.", "I see what you mean. But you asked a tricky question - can I figure out how to the use the example scripts to meet my needs - mostly yes - but then every time I ask for something that ensures consistency, you say - but the audience is wrong - it should be for users. And I say, yes, of course, you're right. And we end up nowhere. Do you see where the circle is?\r\n\r\nIdeally there should be just one benchmarking tool that can handle any model (or at least the majority of them) and support the different tasks and it probably won't need all the possible flags the various scripts have. If that makes sense.\r\n\r\nI was using `finetune_trainer.py` for many things, but then a user asks to validate/benchmark/integrate a model not supported by that script, so I go into that subdomain in examples and things aren't the same there. And I know we are trying to make the example scripts consistent, but the example of this Issue I know for a fact that when one manually copies the same feature across scripts they are bound to become inconsistent. At least that's the experience with transformers so far.\r\n\r\nComplaining and frustration expression aside - perhaps we could start with one best script that you think is a good model and then making it non-examples and to start transforming it to support a multitude of tasks/models/features? Would that be a good way to move forward?\r\n", "The issue is derailing a bit as I think adding the `max_train_samples` etc to all scripts has been validated (and is useful to quickly test the example is running on the user data).\r\n\r\nIf you want to look at a benchkmarking script, I think a good starting point is `run_glue` for fine-tuning on text classification, `run_mlm` for language modeling. Those are more for BERT-like models than seq2seq models however. `finetune_trainer` is aimed at being deprecated and once `run_seq2seq` has all its features, it can be the one good script to be based on for all things seq2seq.", "> The issue is derailing a bit as I think adding the `max_train_samples` etc to all scripts has been validated (and is useful to quickly test the example is running on the user data).\r\n\r\nExcellent!\r\n\r\n> If you want to look at a benchkmarking script, I think a good starting point is `run_glue` for fine-tuning on text classification, `run_mlm` for language modeling. Those are more for BERT-like models than seq2seq models however. `finetune_trainer` is aimed at being deprecated and once `run_seq2seq` has all its features, it can be the one good script to be based on for all things seq2seq.\r\n\r\nI feel I'm not managing to successfully communicate the need here. I will let it go for now.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "This is getting resolved by https://github.com/huggingface/transformers/pull/10551\r\n", "> I always thought that `max_steps` defines the total number of weight update steps (which is then not really influenced by other parameters such as number of GPUs or `gradient_accumalation_steps` or whatever). To me it defines: \"How often do I want to update my weights?\" or am I wrong here?. Think the name is clear and does not need to be changed, the documentation could be updated with a sentence that makes clear that `max_steps` = number of weight updates. Also, I use this arg quite often when training and think it's important to keep.\r\n> \r\n> I agree with @sgugger here that I think a `--max_num_train_samples` arg (or whatever the name) should not go into the trainer, but should be added to all examples scripts. It's actually incredibly easy to do this with `datasets`:\r\n> \r\n> ```python\r\n> ds = load_dataset(\"crime_and_punish\", split=\"train\")\r\n> ds = ds.select(range(arg.max_num_train_samples))\r\n> ```\r\n> \r\n> I'm totally fine with having this as another cl arg for the scripts, but don't think it's the responsibility of the `trainer`.\r\n\r\nhi,I want to use the crime_and_punish dataset to do evaluation on model reformer,which task code should I use?", "@LeopoldACC, it looks like you posted your question in a very unrelated discussion. Please try https://discuss.huggingface.co/. Thank you." ]

[]

[]

[ "This fix works, but only for this particular enum. I'm wondering if it shouldn't be better to just change [this line](https://github.com/huggingface/transformers/blob/master/src/transformers/hf_argparser.py#L83) in `HFArgumentParser`? For instance\r\n```bash\r\n$ python examples/text-classification/run_glue.py -h | grep evaluation_strategy\r\n```\r\n\r\nreturns\r\n```\r\n--evaluation_strategy {EvaluationStrategy.NO,EvaluationStrategy.STEPS,EvaluationStrategy.EPOCH}\r\n```\r\n\r\nwhich probably does not work as well. (Or if does, we want to display \"no\"/\"steps\"/\"epoch\" here.)", "That's an excellent suggestion, @sgugger - thank you! Please have a look at this variation.", "Weird, somehow it broke `finetune_trainer.py`:\r\n\r\n```\r\nexamples/seq2seq/test_finetune_trainer.py:84: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\nexamples/seq2seq/test_finetune_trainer.py:76: in finetune_trainer_quick\r\n output_dir = self.run_trainer(1, \"12\", MBART_TINY, 1, distributed, deepspeed, extra_args_str)\r\nexamples/seq2seq/test_finetune_trainer.py:210: in run_trainer\r\n main()\r\nexamples/seq2seq/finetune_trainer.py:160: in main\r\n model_args, data_args, training_args = parser.parse_args_into_dataclasses()\r\nsrc/transformers/hf_argparser.py:150: in parse_args_into_dataclasses\r\n namespace, remaining_args = self.parse_known_args(args=args)\r\n/usr/local/lib/python3.6/argparse.py:1773: in parse_known_args\r\n self.error(str(err))\r\n/usr/local/lib/python3.6/argparse.py:2393: in error\r\n self.exit(2, _('%(prog)s: error: %(message)s\\n') % args)\r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\n\r\nself = HfArgumentParser(prog='finetune_trainer.py', usage=None, description=None, formatter_class=<class 'argparse.HelpFormatter'>, conflict_handler='error', add_help=True)\r\nstatus = 2\r\nmessage = \"finetune_trainer.py: error: argument --evaluation_strategy: invalid choice: <EvaluationStrategy.STEPS: 'steps'> (choose from 'no', 'steps', 'epoch')\\n\"\r\n\r\n```\r\n\r\nThe test is just passing `--evaluation_strategy steps`\r\n", "Ok, tested locally and doing this does not work indeed (e.g. you can't launch any script with `--evaluation_strategy steps` or `--lr_scheduler_type linear`). To have it work, we have to change the type of the enum to `str`, but then the actual values of the dataclass are strings.\r\n\r\nSo there is no easy solution. I'm fine with leaving as is or also to remove the choices and expand the help to show the actual possible values, but it looks like it won't work as is.", "Thank you for validating it.\r\n\r\nHow about moving my initial PR's `get_arg_names` to `ExplicitEnum` so any sub-class has access to it - perhaps a different name? And then we change just `meta[\"choices\"]` in the arg parser to get these values?", "No the initial PR doesn't work either (this is not caught by the tests since the test do not use `--lr_scheduler_type` in any of the example scripts). The field ends up being a `str` if you try on your side (and not a `SchedulerType` despite the cast in the post_init so then all tests comparing `self.args.lr_scheduler_type` to `SchedulerType.XXX` will fail.", "Ah, my bad! I was testing on the `parse` method of `HfArgumentParser`, not `parse_into_dataclasses`. There is a way to make this work :-)", "I'm all ears.\r\n\r\n", "The easy fix is to force `kwargs[\"type\"] = type(kwargs[\"choices\"][0])` for the Enum subclasses, after your line `kwargs[\"choices\"] = [x.value for x in field.type]`, and let the dataclass set them back to their proper enum types in the postinit (as is done right now).\r\n\r\nI even have a function that will automagically do the casting back after the init, which is the following:\r\n```\r\n for dtype in self.dataclass_types:\r\n keys = {f.name for f in dataclasses.fields(dtype) if f.init}\r\n keys_to_enum_types = {f.name: f.type for f in dataclasses.fields(dtype) if isinstance(f.type, type) and issubclass(f.type, Enum)}\r\n inputs = {k: v for k, v in vars(namespace).items() if k in keys}\r\n for k in keys:\r\n if k in keys_to_enum_types:\r\n inputs[k] = keys_to_enum_types[k](inputs[k])\r\n delattr(namespace, k)\r\n obj = dtype(**inputs)\r\n outputs.append(obj)\r\n```\r\nin the `parse_into_dataclasses` method. The same would need to be done in the other special parse methods for consistency.", "Please feel free to take over, rather than me being the middle person as you know what you're doing and I will learn from your work when it's done. Thank you!", "Fabulous! \r\n\r\nThis is not a user facing code, correct? There are 3 pretty big identical chunks of code which can be refactored then.", "Not sure we actually need them since the dataclasses need to re-cast those args to the enum type anyway for when someone is using `TrainingArguments` not in CLI. I was hoping to remove the lines\r\n```\r\nself.evaluation_strategy = EvaluationStrategy(self.evaluation_strategy)\r\nself.lr_scheduler_type = SchedulerType(self.lr_scheduler_type)\r\n```\r\nby using those four lines, but they are still necessary. So we'll probably just remove those three blocks of identical code (let's see what @LysandreJik thinks!)" ]

[]

[]

[ "Yeah I forgot to add this in a comment -> Talked to @n1t0 about it and he says those are unwanted breaking changes, so we will pin to the next patch release he is going to make to tokenizers. Leaving the PR open in the meantime!", "Added a few more things to this PR:\r\n- Use last tokenizers RC release\r\n- Update the conversion from slow to fast tokenizers as described in https://github.com/huggingface/transformers/issues/9637\r\n- Added a script to verify the conversion from slow to fast tokenizers looks good\r\n- Fix some links to the hub", "Regarding the masks, should that be applied to all SentencePiece-based tokenizers? Should it be added to XLNet/ALBERT/T5 as well?", "Can't approve since this is my PR originally, but this looks good to me. We just need to make sure all special masks tokens are taken into account.", "Most of the tests are unrelated (a rebase on `master` will make them pass), but these two aren't:\r\n\r\n```\r\nFAILED tests/test_tokenization_pegasus.py::PegasusTokenizationTest::test_mask_tokens_rust_pegasus\r\nFAILED tests/test_tokenization_mbart.py::MBartTokenizationTest::test_embeded_special_tokens\r\n=== 2 failed, 3352 passed, 3280 skipped, 2512 warnings in 765.06s (0:12:45) ====\r\n```" ]

[]

[]

[ "Very cool integration.🚀 🚀 🔥 🔥 \r\nI'm doing some tests over the day for single-GPU and multi-GPU and let you know if I find something strange" ]

[]

[]

[ "You can see from the message that it's skipping the optimizer states. We're only saving the model, so it makes sense that the optimizer states are discarded :)", "Thanks you for your reply ! \r\n\r\nI will close the issue then." ]

[]

[]

[ "Thanks for addressing the comments, this is good to go IMO.\r\nPro-tip, if you edit your description to replace `Issue #9752` by `Fixes #9752`, the issue will be automatically closed when we merge this PR." ]

[]

[]

[ "Ah, I just noticed `Seq2SeqTrainingArguments` sets `local_rank=-1` by default" ]

[]

[]

[]

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

[]

[ "If anyone else comes here, until this is fixed you can find the version in `transformers/dependency_versions_check.py` - e.g. change `\"tokenizers\": \"tokenizers==0.9.4\"` to `\"tokenizers\": \"tokenizers>=0.9.4\"`", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "One more thing: this behavior still persists after I fine tuned the model on my own dataset", "Hi @tomdzh \r\n\r\nfirst of all, unlike the original T5, mT5 is not pre-trained on any supervised downstream task (like summarization, translation etc), so generation work without fine-tuning it.\r\n\r\nAlso, it would be hard to answer why it's happening in fine-tuned model without looking at any code.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "I think the following comment of my own can be left as a PR for the future.\r\nThis change seems to require a modification of the `convert` code, and I think it may be necessary to separate it from this broken link issue.\r\n\r\nI'm sorry for saying such a thing, even though the following is what I wrote.\r\n\r\n> I think there are some outdated explanations.\r\n> As I discussed in issue #9657, I think it should be better to explain `from_pretrained()` instead of `torch.save()`.\r\n> Hence, I think the explanation below should be updated.\r\n> \r\n> ```\r\n> You can then disregard the TensorFlow\r\n> checkpoint (the three files starting with ``bert_model.ckpt``\\ ) but be sure to keep the configuration file (\\\r\n> ``bert_config.json``\\ ) and the vocabulary file (\\ ``vocab.txt``\\ ) as these are needed for the PyTorch model too.\r\n> ```\r\n> \r\n\r\nI'll remove the WIP for now, but if you think this matter should be worked in this PR, please let me know." ]

[]

[]

[]

[]

[]

[ "@lhoestq Thanks for the suggestions! All the CI checks pass now." ]

[]

[]

[ "I fully rework the keywords addition part to keep only those that seemed the most meaningful." ]

[]

[]

[]

[]

[]

[ "I've read these similar issue\r\n\r\n[#5654](https://github.com/huggingface/transformers/issues/5656)\r\n[#7500](https://github.com/huggingface/transformers/issues/7500)\r\n\r\nbut couldn't get the information needed.\r\n\r\n@sshleifer @patil-suraj Any idea why this is happening? (Sorry for tagging everyone, I've been facing this issue for a while, looking for answers 🙂)", "Hi @sumanthd17 ,\r\nI had this issue with standard T5 in which it was ignoring the `min_length` flag. I fixed this simply by upping the `num_beams` to 4.\r\n\r\nCould you test this and see if it's quick fix for you?", "@FL33TW00D Thanks. Yeah the `min_length` flag is being ignored. \r\n\r\nThanks for the quick fix. I think we can close this issue for now, But it might be a good idea to know why the min_length is being ignored.", "@sumanthd17 \r\nThis blog post from Patrick explains why min_length may not be satisfied by the beam_search:\r\nhttps://huggingface.co/blog/how-to-generate\r\n\r\nAlthough I do agree perhaps it should throw a warning message when it is unable to satisfy the min_length flag with the current generation parameters." ]

[]

[]

[ "As explained in the documentation: \"`run_glue.py`: This script can fine-tune the following models: BERT, XLM, XLNet and RoBERTa.\"\r\n\r\n=> GPT-2 is a Transformer decoder, which can learn to generate text in an autoregressive way. It is not aimed at GLUE tasks, which are sequence classification tasks. ", "Hi! Actually we've recently added `GPT2ForSequenceClassification` to enable support for sequence classification tasks (like GLUE). The support was added to enable some models such as EDIT: linked wrong model. Updated: [DialogRPT](https://huggingface.co/microsoft/DialogRPT-updown)!\r\n\r\nHowever, as you have seen @nlp-student, the GPT-2 model isn't trainable out of the box with batch size > 1, as it has no padding token defined. Furthermore, MNLI is a three-way classification so you would need to set the number of labels appropriately.\r\n\r\nI invite you to run the following script to create a model that you can then use with the `run_glue.py` script, initialized from the GPT-2 weights:\r\n```py\r\nfrom transformers import GPT2ForSequenceClassification, GPT2Tokenizer\r\n\r\nmodel = GPT2ForSequenceClassification.from_pretrained(\"gpt2\", num_labels=3)\r\ntokenizer = GPT2Tokenizer.from_pretrained(\"gpt2\")\r\n\r\n# Define a padding token\r\ntokenizer.pad_token = tokenizer.eos_token\r\nmodel.config.pad_token_id = tokenizer.pad_token_id\r\n\r\n# Save the model and tokenizer in a directory\r\nmodel.save_pretrained(\"directory\")\r\ntokenizer.save_pretrained(\"directory\")\r\n```\r\n\r\nThen you can launch the `run_glue.py` script by specifying that checkpoint.\r\n\r\nHowever, as @NielsRogge pointed out, GPT-2 will not obtain as good results as a bi-directional encoders such as BERT.", "Thank you so much for your responses. I will try that out.\r\n\r\nIs GPT-2 generally expected to perform worse than BERT on sequence classification ? \r\n\r\nI've seen a lot of examples of BERT finetuned for classification, but not really for GPT-2. Is it not common (or that useful) to finetune GPT-2 for classification?\r\n\r\nThank you for any help or direction!", "> Is GPT-2 generally expected to perform worse than BERT on sequence classification ?\r\n\r\nNormally, yes. The reason for this is that GPT-2 is not really designed for such a task, whereas BERT is, as it has a special [CLS] token for classification tasks. GPT-2 is designed to process text autoregressively (i.e. left to right) in order to generate new text, whereas BERT is designed to process all tokens at once (hence creating a bidirectional representation of all input tokens), which is useful for tasks like sequence classification or extractive question answering for example. \r\n\r\nThat said, you can still use GPT-2 to perform sequence classification. You can simply let GPT-2 process a sentence word by word from left to right, and then train it to predict the class of the sentence by placing a linear layer on top of the hidden representation of the final token of the sentence, which is done by looking at the [code](https://github.com/huggingface/transformers/blob/285c6262a84490270d2f1a1c06ee9ccfc1b60e8f/src/transformers/models/gpt2/modeling_gpt2.py#L1233) of `GPT2ForSequenceClassification`.\r\n\r\n So maybe an interesting thing to do is compare `BERTForSequenceClassification` and `GPT2ForSequenceClassification` on the same dataset, and see which one performs best. ", "Many thanks for your detailed response! It was very helpful for my understanding.", "You're welcome, I'm closing this issue! Feel free to reopen if you have other issues down the road.", "> Hi! Actually we've recently added `GPT2ForSequenceClassification` to enable support for sequence classification tasks (like GLUE). The support was added to enable some models such as EDIT: linked wrong model. Updated: [DialogRPT](https://huggingface.co/microsoft/DialogRPT-updown)!\r\n> \r\n> However, as you have seen @nlp-student, the GPT-2 model isn't trainable out of the box with batch size > 1, as it has no padding token defined. Furthermore, MNLI is a three-way classification so you would need to set the number of labels appropriately.\r\n> \r\n> I invite you to run the following script to create a model that you can then use with the `run_glue.py` script, initialized from the GPT-2 weights:\r\n> \r\n> ```python\r\n> from transformers import GPT2ForSequenceClassification, GPT2Tokenizer\r\n> \r\n> model = GPT2ForSequenceClassification.from_pretrained(\"gpt2\", num_labels=3)\r\n> tokenizer = GPT2Tokenizer.from_pretrained(\"gpt2\")\r\n> \r\n> # Define a padding token\r\n> tokenizer.pad_token = tokenizer.eos_token\r\n> model.config.pad_token_id = tokenizer.pad_token_id\r\n> \r\n> # Save the model and tokenizer in a directory\r\n> model.save_pretrained(\"directory\")\r\n> tokenizer.save_pretrained(\"directory\")\r\n> ```\r\n> \r\n> Then you can launch the `run_glue.py` script by specifying that checkpoint.\r\n> \r\n> However, as @NielsRogge pointed out, GPT-2 will not obtain as good results as a bi-directional encoders such as BERT.\r\n\r\nHi, met the same error as the original post although but I've set the per_device_train_batch_size=1. \r\nI was running run_swag.py with GPT2 (I added a class GPT2ForMultipleChoice referring to BertForMultipleChoice), could you offer some help with this problem?　Thanks a lot.", "I follow the suggestion but got this error:\r\n\r\nRuntimeError: Error(s) in loading state\\_dict for GPT2ForSequenceClassification: size mismatch for score.weight: copying a param with shape torch.Size(\\[100, 768\\]) from checkpoint, the shape in current model is torch.Size(\\[2, 768\\]). You may consider adding \\`ignore\\_mismatched\\_sizes=True\\` in the model \\`from\\_pretrained\\` method.\r\n\r\nI then tried to flag `ignore_mismatched_sizes` as True like below but still, get the same error. \r\n\r\n```python\r\nfrom transformers import GPT2ForSequenceClassification, GPT2Tokenizer\r\n\r\nmodel = GPT2ForSequenceClassification.from_pretrained(\"gpt2\", num_labels=100, ignore_mismatched_sizes=True)\r\ntokenizer = GPT2Tokenizer.from_pretrained(\"gpt2\")\r\n\r\n# Define a padding token\r\ntokenizer.pad_token = tokenizer.eos_token\r\nmodel.config.pad_token_id = tokenizer.pad_token_id\r\n\r\n# Save the model and tokenizer in a directory\r\nmodel.save_pretrained(\"directory\")\r\ntokenizer.save_pretrained(\"directory\")\r\n```" ]

[ { "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" } ]

[]

[ "Hello, Thanks for you work on that @oborchers ! I also saw the notebook on SentenceTransformers and it helped a lot ! \r\n\r\nAny status about this feature ? I also need to run models in onnx but most of them need to call a `.generate` function which is for now not supported... (I could replicate all the generate code in nodejs but i'm sure there is a nicer solution)\r\nIs there any fix, status update or hack ? \r\n\r\nThanks a lot in advance,\r\nHave a great day. ", "Hi @lerezell! Welcome! Glad the notebook helped.\r\n\r\nNot from my side, unfortunately. I had postponed the issue from our side, because we had more pressing stuff to work on. But lately, the need for this feature starts to become larger as well at our company. Does your solution of porting the `.generate` function work by placing it on top of the ONNX version?\r\n\r\n**Edit:**\r\nJust out of curiosity I went through the `.generate` code and it should be possible to place the existing `.generate` code on top of an `CausalLMOutput` model, very similar as done in the [notebook](https://github.com/oborchers/sentence-transformers/blob/master/examples/onnx_inference/onnx_inference.ipynb). This requires an extension of the forward method. \r\n\r\nIn an initial implementation, it should be perfectly sufficient to port just the `sample` section and see if it works. However, this does not necessarily apply to beam_search, which I haven't figured out how it works. And the raw implementation shouldn't be too complex, because one might strip away a set of the convenience functions/arguments.\r\n\r\nDownsides of this are, that, there needs to be some way of defining the arguments of `.generate` at runtime for inference. For example, the `min_length` and `max_length` and `eos_token_id` parameter should be included in the `forward` method arguments, because otherwise they would be static and defined via configuration at runtime. This may be sensible for some applications, but requires re-exporting the model every-time those change, which isn't really a nice way of doing this. Or at least if I didn't miss something completely\r\n\r\nBest regards and have a nice eastern", "Hi @oborchers, \r\n\r\nI still haven't implemented \"my\" solution as I wanted to know if there was any other solution than writing all the logic again. \r\nI would rather not and exporting the logic in the forward (and then in the onnx model) seems to be the best solution. \r\n\r\nFor the `x_length` arguments, that a downside, passing them as optional in the forward method could do ? \r\n\r\nI need to focus on other things right now but I definitely keep an eye open for that ! \r\n\r\nHave a great day \r\n\r\n", "Hi, any update on how to export full pipelines to onnx? \r\n\r\nFor now, we're still obliged to keep a custom/hugging face lib code to handle the \"post output embeddings\" logic.... \r\n\r\nThanks in advance, \r\nHave a great day", "Hi @Ierezell!\r\n\r\nSorry for not coming back on the issue. To be honest, for our use case there are quite a few problems we've encountered in exporting full pipelines to ONNX:\r\n\r\n- How to best deal with caching (`past_key_values`)\r\n- Less than optimal performance when used with some generative models (https://github.com/microsoft/onnxruntime/issues/7238)\r\n- The problem of batching requests on inference servers which is very difficult due to the dynamic dimensions of `past_key_values`\r\n- Similar gains in inference time by using custom kernels (e.g. deepspeed inference) + regular pytorch\r\n\r\nThis blog post from Microsoft may help though:\r\n- https://cloudblogs.microsoft.com/opensource/2021/06/30/journey-to-optimize-large-scale-transformer-model-inference-with-onnx-runtime/\r\n- https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/python/tools/transformers/notebooks/Inference_GPT2-OneStepSearch_OnnxRuntime_CPU.ipynb", "Hi @oborchers, thanks a lot for the feedback! \r\n\r\nOnnx is nice to be able to change stack for me (javascript etc...) but in the light of what you're saying it will be better to keep my GPU inference server. \r\n\r\nThanks a lot, \r\nHave a great day ! \r\n\r\n", "Hi,\r\n\r\nIs there an alternative to onnx that you'd recommend? The able to keep and manipulate past_key_values is the most crucial part that I cannot find for many inference optimizations.\r\n\r\nThank you!" ]

[]

[]

[ "If it's a bug users relied on, then it's a feature, not a bug. \r\n\r\nAnyway I'll merge this.", "Good point, but I don't think that's the case here. Do you think users rely on the previous behavior? If that's the case, then we should revert this PR until we change of major version." ]

[]

[]

[ "Hey @sadakmed, sorry I might have been a bit unclear here. Could you instead of opening new PRs add your added code directly to your PR here: https://github.com/huggingface/transformers/pull/9128? The `#Copy from` statements force us to have all the code in the same PR.\r\n\r\nAlso we need a decoder-only test for this." ]

[]

[]

[ "Hey @sadakmed, sorry I might have been a bit unclear here. Could you instead of opening new PRs add your added code directly to your PR here: https://github.com/huggingface/transformers/pull/9128? The `#Copy from` statements force us to have all the code in the same PR.\r\n\r\nAlso we need a decoder-only test for this." ]

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

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "model_name = \"distilbert-base-uncased-finetuned-sst-2-english\"\r\n\r\nmodel = AutoModelForSequenceClassification.from_pretrained(model_name)\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\n\r\n\r\ntrain_encoding = tokenizer(X_train, truncation=True, padding=True, max_length=512, return_tensors=\"pt\")\r\n\r\n\r\n# Training the model with PyTorch\r\n\r\n\r\nwith torch.no_grad():\r\n\r\n outputs = model(**train_encoding)\r\n \r\n # Normalize logits and spans \r\n start = F.softmax(outputs.start_logits, dim=1)\r\n end = F.softmax(outputs.end_logits, dim=1)\r\n\r\n\r\n # Getting the start and end index of the tensor to retrieve the answer\r\n start_index = torch.argmax((start), dim=1)\r\n end_index = torch.argmax((end), dim=1)\r\n\r\n\r\n\r\n# Computing score\r\n\r\nouter = np.matmul(np.expand_dims(start, -1), np.expand_dims(end, 1))\r\nmax_answer_len = 512\r\ncandidates = np.tril(np.triu(outer), max_answer_len - 1)\r\nidx_sorts = [np.argmax(candidates[i].flatten()) for i in range(len(candidates))]\r\nscores = [candidates[[i], start_indices[i], end_indices[i]][0] for i in range(len(candidates))]\r\n\r\n\r\n# Extracting answer\r\ntokenizer.decode(train_encoding[\"input_ids\"][start_index:end_index], skip_special_tokens=True)\r\n\r\n\r\n\r\n \r\n" ]

[]

[]

[ "Hi @Sai-Ashish \r\n\r\nPlease use the forum https://discuss.huggingface.co/ for asking such questions. Use issues to report bugs, feature requests, etc. Thanks!\r\n\r\nClosing it", "have you came up with solution?" ]

[]

[]

[ "Thanks for pointing it out, the files are now renames to `distil_marian_enro_teacher.sh`, `distil_marian_no_teacher.sh` and are available here \r\nhttps://github.com/huggingface/transformers/blob/master/examples/research_projects/seq2seq-distillation/distil_marian_enro_teacher.sh\r\n\r\nhttps://github.com/huggingface/transformers/blob/master/examples/research_projects/seq2seq-distillation/distil_marian_no_teacher.sh\r\n\r\nFeel free to open a PR to fix the links if you want to contribute :) \r\n\r\nAlso, we now recommend using the new `Seq2SeqTrainer` for fine-tuning seq2seq models, https://github.com/huggingface/transformers/tree/master/examples/seq2seq", "Thanks a lot. Also, there a way to train the MarianMT model on my dataset ?", "Sure, the script let's you pass your own dataset as well, have a look at the readme.", "Thanks a lot for the help." ]

[]

[]

[ "Could you post a short code snippet so we can reproduce ?", "> Could you post a short code snippet so we can reproduce ?\r\n\r\nThis is my code.\r\n\r\n```\r\nclass GPReviewDataset(Dataset):\r\n def __init__(self, reviews, targets, tokenizer, max_len):\r\n self.reviews = reviews\r\n self.targets = targets\r\n self.tokenizer = tokenizer\r\n self.max_len = max_len\r\n \r\n def __len__(self):\r\n return len(self.reviews)\r\n \r\n def __getitem__(self, item):\r\n review = str(self.reviews[item])\r\n target = self.targets[item]\r\n \r\n encoding = self.tokenizer.encode_plus(text=review, max_length=self.max_len,\r\n add_special_tokens=True, padding='max_length', \r\n return_attention_mask=True, \r\n return_token_type_ids=False, return_tensors='pt')\r\n \r\n return {'review': review,\r\n 'input_ids': encoding['input_ids'].flatten(), \r\n 'attention_mask': encoding['attention_mask'].flatten(),\r\n 'targets': torch.tensor(target, dtype=torch.long)}\r\n\r\n\r\nerror_list = []\r\n\r\nfor i in range(len(free_df)):\r\n if len(GPReviewDataset(free_df['content'].to_numpy(), free_df['score'].to_numpy(), tokenizer, 160).__getitem__(i)['attention_mask']) != 160:\r\n error_list.append((i, len(GPReviewDataset(free_df['content'].to_numpy(), free_df['score'].to_numpy(), tokenizer, 160).__getitem__(i)['input_ids'])))\r\n \r\nerror_list\r\n```\r\n\r\nand the results\r\n```\r\n[(95, 184),\r\n (948, 218),\r\n (1025, 162),\r\n (3679, 204),\r\n (3680, 164),\r\n (4150, 220),\r\n (6139, 185),\r\n (7139, 165),\r\n (7201, 166),\r\n (7237, 256),\r\n (7381, 181),\r\n (7599, 254),\r\n (7600, 204),\r\n (7679, 170),\r\n (8111, 202),\r\n (8378, 193),\r\n (8773, 583),\r\n (9041, 583),\r\n (9321, 161),\r\n (10466, 279)]\r\n```", "You should set your truncation parameter as well, otherwise it won't truncate the texts that are too long. See the docs about the tokenizer methods [here (look for `truncation`)](https://huggingface.co/transformers/internal/tokenization_utils.html?highlight=truncation#transformers.tokenization_utils_base.PreTrainedTokenizerBase.__call__)", "> You should set your truncation parameter as well, otherwise it won't truncate the texts that are too long. See the docs about the tokenizer methods [here (look for `truncation`)](https://huggingface.co/transformers/internal/tokenization_utils.html?highlight=truncation#transformers.tokenization_utils_base.PreTrainedTokenizerBase.__call__)\r\n\r\nThank you! That worked." ]

[]

[]

[ "I was having a problem getting this to actually resume training on my system, and I had to make three small changes to the new code in trainer_utils.py:\r\n\r\n1. `checkpoints = [path for path in content if _re_checkpoint.search(path) is not None and os.path.isdir(path)]` was returning empty. I changed `os.path.isdir(path)` to `os.path.isdir(os.path.join(folder, path))` and now it returns a list of the checkpoint folders as expected.\r\n2. Similarly, the `get_last_checkpoint` function was returning the basename of the checkpoint folder, not the full path, which seems to be expected based on the updates to the example scripts. I changed the last line of the function to `return os.path.join(folder, max(checkpoints, key=lambda x: int(_re_checkpoint.search(x).groups()[0])))`\r\n3. After I made those update, it was resuming from the oldest checkpoint, not the newest. I noticed the checkpoint regex was only capturing the final digit in the directory name. I changed it to `_re_checkpoint = re.compile(r\"^\" + PREFIX_CHECKPOINT_DIR + r\"\\-(\\d+)$\")` with the `+` inside the capture group, and now `get_last_checkpoint` is giving me the newest checkpoint as expected.\r\n\r\nI'm just a novice, so I'm not sure if those tweaks would break anything other systems. Does `os.listdir()` return full paths instead of basenames under some OS/python combinations?\r\n\r\nBut with these changes I'm able to resume an aborted training from within the same folder. ", "Oh, I went a bit too fast and all the issues you list are completely valid! You should make a PR with all your changes since you were the one to find the problems and fix them :-)", "Got sucked back into work for my day job, but I'll try getting to that soonish.\r\n\r\nUnrelated, do you guys have an office in DUMBO? I think we're in the same building (at least until we all started working from home)", "DUMBO offices are at 20 Jay street, you should come and say hi for a coffee when the pandemic is over if you're around :-) ", "Yeah, I run a small animation studio up on the 10th floor. Turns out the fancy GPUs I have for animation are also good for messing around with machine learning :) And so my pandemic side project became teaching my computer to write poetry. \r\n\r\nSomeday when it's safe again I'll definitely stop by.", "Haha small world @jncasey! I'm missing 20 Jay right now. What is your company's name? ", "We're called Mixtape Club. We moved into the building from Manhattan last January, and were only there for about 10 weeks before everyone started working from home. Now my business partner is the only one who goes in, since he can't bring our whole sound studio back to his apartment. " ]

[]

[]

[ "Pinging @jplu ", "Hello!!\r\n\r\nYes this is a known issue for the Seq2Seq models, that is already fixed in master.", "Hi @jplu \r\n\r\nI'll close the issue as the fix will be in the next release for sure. \r\n\r\nThanks again. " ]

[]

[]

[ "Hey @sadakmed, sorry I might have been a bit unclear here. Could you instead of opening new PRs add your added code directly to your PR here: https://github.com/huggingface/transformers/pull/9128? The `#Copy from` statements force us to have all the code in the same PR.\r\n\r\nAlso we need a test for this.", "Hey @sadakmed, sorry I might have been a bit unclear here. Could you instead of opening new PRs add your added code directly to your PR here: https://github.com/huggingface/transformers/pull/9128? The `#Copy from` statements force us to have all the code in the same PR.\r\n\r\nAlso we need a decoder-only test for this." ]

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

[]

[ "I'm pretty sure it's just a bad naming of that variable. It should be question_encoder_pooler_output.\r\n\r\nIndeed this variable is computed as the first output of a `DPRQuestionEncoder`. The DPR encoders don't return the hidden states (they basically skip it) to return the pooler output (aka DPR embeddings) of shape `(batch_size, vector_size)` instead.\r\n\r\nIf you want to contribute, feel free to open a PR to rename this variable :) ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "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" } ]

[]

[ "Hi @LysandreJik is this feature request being worked on by someone? If not, I would love to take it up. Thanks!", "Hello! \r\n\r\nThere is existing support for this thanks to the `problem_type` configuration attribute for some Sequence classification models, see here: https://huggingface.co/transformers/main_classes/configuration.html#transformers.PretrainedConfig\r\n\r\n(search for `problem_type`)\r\n\r\nIt could be better documented though by having an example for each model that supports this/more complete documentation on the model themselves. Would you like to try your hand at it?", "Hi,\r\n\r\nIt seems like the documentation for this feature has already been merged so I think this issue should be closed. \r\n", "Indeed! Thanks @astern21 " ]

[]

[]

[ "Hello!\r\n\r\nThis is the expected behavior, if you want any reduction on the loss, you have to do it yourself on your side, not inside the respective compute_loss function.", "Hi, thanks for the explanation. \r\nI realized it was probably by design, it's just odd that it differs so much in behavior from the Pytorch version. Is there any plan to bring those more inline in this regards? Probably a breaking change, I don't have a clear overview of how much would break, even internally within the Transformers library.", "Nothing planed to align this with Python and we won't. The reason is because when training with a distribute strategy, TensorFlow doesn't allow a reduction other than `None` or `Sum`. Knowing that we have our own custom trainer and we cannot apply the change you would like as it will make it fails for such cases.", "Makes sense, didn't think about the incompatibility of the AUTO reduction with the distribute strategies and the custom trainer.\r\n \r\nI'll try to make a small patch over the weekend with an update to the documentation in the docstrings, as it's currently not in line with the actual (and intended) output.", "That would be awesome! Thanks!" ]

[]

[]

[ "I found that TFBart models use ```padding token``` for masking ```decoder_input_ids``` instead of using ```-100 token```, which is different from T5 models. So this is not a bug, but a little confusing because some of the bart code and documents talk about ```-100 token```.\r\n\r\nAlso, losses from ```torch``` and ```tensorflow``` are different with the same dataset (text shown above).\r\nI also directly convert ```pytorch_model.bin``` to ```tf_model.h5``` instead of using uploaded model, but they shows different losses.", "Hi @kiyoungkim1 \r\n\r\nIn both T5 and BART `decoder_input_values` can never contain -100, for both models the padding is done using pad token.\r\nIn labels, however, we usually replace pad tokens with -100 so as to not include them while computing the loss.\r\n\r\nBut what you pointed out is correct, TFBart uses pad token as the ignore index \r\nhttps://github.com/huggingface/transformers/blob/9152f16023b59d262b51573714b40325c8e49370/src/transformers/models/bart/modeling_tf_bart.py#L1340-L1348\r\n\r\nWhere as TFT5 ignores -100\r\nhttps://github.com/huggingface/transformers/blob/9152f16023b59d262b51573714b40325c8e49370/src/transformers/modeling_tf_utils.py#L146-L152\r\n\r\nnot an expert on TF side, so @jplu , @patrickvonplaten will know more", "What says @patil-suraj is correct, in T5 we expect the pad token id to always be -100 while in TF BART it can be any digit assigned to `config.pad_token_id`.", "Great catch @kiyoungkim1!\r\n\r\nIt's not very consistent what we are doing here...TFBart should have never ignored the `pad_token_id` as a default setting, but -100 as all other models do.\r\n\r\nTo fix the problem, I think we should add a couple of lines that check if -100 are in the labels and if yes replaces them with the `pad_token_id` to have consistency with PyTorch's Bart. It would be a pretty big breaking change to just replace `pad_token_id` with -100 so I think the first option is the better one. @kiyoungkim1 if you feel like opening a PR to correct this behavior we would be more than happy :-) ", "We also plan to turn all the loss computation, not anymore as a method but as a layer, so it will be much easier to use, to configure and TensorFlow workflow compliant." ]

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

[]

[ "Have you taken a look at the [Keras docs](https://www.tensorflow.org/api_docs/python/tf/keras/callbacks/ModelCheckpoint)?", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "@LysandreJik Yes i looked the following doc and implemented it.\r\n\r\n```\r\nconfig = AutoConfig.from_pretrained(\r\n PRETRAINED_MODEL_VERSION, num_labels=len(classes), label2id=label2id, id2label=id2label,\r\n finetuning_task=\"text-classification\")\r\nmodel = TFAutoModelForSequenceClassification.from_pretrained(PRETRAINED_MODEL_VERSION, config=config)\r\n\r\nloss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)\r\nmetric = tf.keras.metrics.Accuracy()\r\noptimizer = tf.keras.optimizers.Adam(learning_rate=2e-6, epsilon=1e-08)\r\nmodel.compile(loss=loss, optimizer=optimizer, metrics=[metric])\r\nprint(model.summary())\r\n\r\nes_callback = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=3, verbose=1)\r\nchkpt_calllback = tf.keras.callbacks.ModelCheckpoint(MODEL_DIRECTORY+\"{epoch:02d}\",\r\n monitor='val_loss', verbose=2,\r\n save_best_only=True, save_weights_only=False)\r\n\r\nmodel.fit([input_ids_train, attention_mask_train], train_labels,\r\n validation_data=([input_ids_valid, attention_mask_valid], valid_labels),\r\n epochs=1, batch_size=32, callbacks=[es_callback, chkpt_calllback])\r\n```\r\nbut doing so saves the model checkpoint only and I'm not sure how to load these weights again into the **model**. As it only accepts **.h5** format. Facing difficulty in loading these weights\r\n```\r\ncheckpoint\r\n02.index\r\n02.data-00000-of-00001\r\n```" ]

[]

[]

[ "Hey @sadakmed, sorry I might have been a bit unclear here. Could you instead of opening new PRs add your added code directly to your PR here: https://github.com/huggingface/transformers/pull/9128? The `#Copy from` statements force us to have all the code in the same PR.\r\n\r\nAlso we need a decoder-only test for this." ]

[]

[]

[]

[ { "id": 2627272588, "node_id": "MDU6TGFiZWwyNjI3MjcyNTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Model%20Parallel", "name": "Model Parallel", "color": "8B66A5", "default": false, "description": "Model Parallelilsm Implementations" }, { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" }, { "id": 2682576896, "node_id": "MDU6TGFiZWwyNjgyNTc2ODk2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Pipeline%20Parallel", "name": "Pipeline Parallel", "color": "1F75CB", "default": false, "description": "" }, { "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": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[]

[ { "id": 2682576896, "node_id": "MDU6TGFiZWwyNjgyNTc2ODk2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Pipeline%20Parallel", "name": "Pipeline Parallel", "color": "1F75CB", "default": false, "description": "" }, { "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" } ]

[]

[ "Thanks @stas00 , I am getting what looks like a torch error when I run this (I'm not sure if the \"Failed to look up the IP address for the hostname\" error is related -- I'm not able to find much on this except for an issue from a few days ago that mentions this: https://github.com/pytorch/pytorch/issues/50700 ):\r\n\r\n```\r\ngit clone https://www.github.com/huggingface/transformers.git\r\ncd transformers/\r\ngh pr checkout 9765\r\n\r\nconda create -y -n py38-pt18 python=3.8\r\nconda activate py38-pt18\r\npip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cu110/torch_nightly.html -U\r\npip install -e .[dev]\r\npip install -r examples/_tests_requirements.txt\r\n\r\ncd examples/seq2seq/\r\nln -s ~/github/transformers/examples/seq2seq/wmt_en_ro wmt_en_ro\r\n\r\nexport BS=160 MODEL=t5-base; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py --model_name_or_path $MODEL --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 50 --n_train 1000 --n_val 1000 --pipeline \"chunks=4 device_map=0:0-3,1:3-12\" --dataloader_num_workers 4 \r\n```\r\n\r\n\r\nOutput:\r\n```\r\nexport BS=160 MODEL=t5-base; rm -r output_dir; PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py --model_name_or_path $MODEL --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_train --evaluation_strategy=steps --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size $BS --per_device_train_batch_size $BS --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 50 --n_train 1000 --n_val 1000 --pipeline \"chunks=4 device_map=0:0-3,1:3-12\" --dataloader_num_workers 4 \r\n\r\n01/25/2021 11:39:51 - WARNING - __main__ - Process rank: -1, device: cuda:0, n_gpu: 4, distributed training: False, 16-bits training: False\r\n01/25/2021 11:39:51 - INFO - __main__ - Training/evaluation parameters Seq2SeqTrainingArguments(output_dir='output_dir', overwrite_output_dir=True, do_train=True, do_eval=True, do_predict=False, evaluation_strategy=<EvaluationStrategy.STEPS: 'steps'>, prediction_loss_only=False, per_device_train_batch_size=160, per_device_eval_batch_size=160, per_gpu_train_batch_size=None, per_gpu_eval_batch_size=None, gradient_accumulation_steps=1, eval_accumulation_steps=None, learning_rate=3e-05, weight_decay=0.0, adam_beta1=0.9, adam_beta2=0.999, adam_epsilon=1e-06, max_grad_norm=1.0, num_train_epochs=1.0, max_steps=-1, lr_scheduler_type=<SchedulerType.LINEAR: 'linear'>, warmup_steps=50, logging_dir='runs/Jan25_11-39-51_seahorse', logging_first_step=True, logging_steps=1000, save_steps=500, save_total_limit=None, no_cuda=False, seed=42, fp16=False, fp16_opt_level='O1', fp16_backend='auto', local_rank=-1, tpu_num_cores=None, tpu_metrics_debug=False, debug=False, pipeline='chunks=4 device_map=0:0-3,1:3-12', dataloader_drop_last=False, eval_steps=25000, dataloader_num_workers=4, past_index=-1, run_name='output_dir', disable_tqdm=False, remove_unused_columns=True, label_names=None, load_best_model_at_end=False, metric_for_best_model=None, greater_is_better=None, ignore_data_skip=False, sharded_ddp=False, deepspeed=None, label_smoothing_factor=0.1, adafactor=False, group_by_length=False, report_to=['tensorboard'], sortish_sampler=True, predict_with_generate=False)\r\n[INFO|configuration_utils.py:445] 2021-01-25 11:39:51,546 >> loading configuration file https://huggingface.co/t5-base/resolve/main/config.json from cache at /home/pajansen/.cache/huggingface/transformers/91e9fe874e06c44883b535d6c950b8b89d6eaa3298d8e7fb3b2c78039e9f8b7b.66b9637a52aa11e9285cdd6e668cc0df14b3bcf0b6674cf3ba5353c542649637\r\n[INFO|configuration_utils.py:481] 2021-01-25 11:39:51,547 >> Model config T5Config {\r\n \"architectures\": [\r\n \"T5WithLMHeadModel\"\r\n ],\r\n \"d_ff\": 3072,\r\n \"d_kv\": 64,\r\n \"d_model\": 768,\r\n \"decoder_start_token_id\": 0,\r\n \"dropout_rate\": 0.1,\r\n \"eos_token_id\": 1,\r\n \"feed_forward_proj\": \"relu\",\r\n \"initializer_factor\": 1.0,\r\n \"is_encoder_decoder\": true,\r\n \"layer_norm_epsilon\": 1e-06,\r\n \"model_type\": \"t5\",\r\n \"n_positions\": 512,\r\n \"num_decoder_layers\": 12,\r\n \"num_heads\": 12,\r\n \"num_layers\": 12,\r\n \"output_past\": true,\r\n \"pad_token_id\": 0,\r\n \"relative_attention_num_buckets\": 32,\r\n \"task_specific_params\": {\r\n \"summarization\": {\r\n \"early_stopping\": true,\r\n \"length_penalty\": 2.0,\r\n \"max_length\": 200,\r\n \"min_length\": 30,\r\n \"no_repeat_ngram_size\": 3,\r\n \"num_beams\": 4,\r\n \"prefix\": \"summarize: \"\r\n },\r\n \"translation_en_to_de\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to German: \"\r\n },\r\n \"translation_en_to_fr\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to French: \"\r\n },\r\n \"translation_en_to_ro\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to Romanian: \"\r\n }\r\n },\r\n \"transformers_version\": \"4.3.0.dev0\",\r\n \"use_cache\": true,\r\n \"vocab_size\": 32128\r\n}\r\n\r\n[INFO|configuration_utils.py:445] 2021-01-25 11:39:51,740 >> loading configuration file https://huggingface.co/t5-base/resolve/main/config.json from cache at /home/pajansen/.cache/huggingface/transformers/91e9fe874e06c44883b535d6c950b8b89d6eaa3298d8e7fb3b2c78039e9f8b7b.66b9637a52aa11e9285cdd6e668cc0df14b3bcf0b6674cf3ba5353c542649637\r\n[INFO|configuration_utils.py:481] 2021-01-25 11:39:51,741 >> Model config T5Config {\r\n \"architectures\": [\r\n \"T5WithLMHeadModel\"\r\n ],\r\n \"d_ff\": 3072,\r\n \"d_kv\": 64,\r\n \"d_model\": 768,\r\n \"decoder_start_token_id\": 0,\r\n \"dropout_rate\": 0.1,\r\n \"eos_token_id\": 1,\r\n \"feed_forward_proj\": \"relu\",\r\n \"initializer_factor\": 1.0,\r\n \"is_encoder_decoder\": true,\r\n \"layer_norm_epsilon\": 1e-06,\r\n \"model_type\": \"t5\",\r\n \"n_positions\": 512,\r\n \"num_decoder_layers\": 12,\r\n \"num_heads\": 12,\r\n \"num_layers\": 12,\r\n \"output_past\": true,\r\n \"pad_token_id\": 0,\r\n \"relative_attention_num_buckets\": 32,\r\n \"task_specific_params\": {\r\n \"summarization\": {\r\n \"early_stopping\": true,\r\n \"length_penalty\": 2.0,\r\n \"max_length\": 200,\r\n \"min_length\": 30,\r\n \"no_repeat_ngram_size\": 3,\r\n \"num_beams\": 4,\r\n \"prefix\": \"summarize: \"\r\n },\r\n \"translation_en_to_de\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to German: \"\r\n },\r\n \"translation_en_to_fr\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to French: \"\r\n },\r\n \"translation_en_to_ro\": {\r\n \"early_stopping\": true,\r\n \"max_length\": 300,\r\n \"num_beams\": 4,\r\n \"prefix\": \"translate English to Romanian: \"\r\n }\r\n },\r\n \"transformers_version\": \"4.3.0.dev0\",\r\n \"use_cache\": true,\r\n \"vocab_size\": 32128\r\n}\r\n\r\n[INFO|tokenization_utils_base.py:1766] 2021-01-25 11:39:52,522 >> loading file https://huggingface.co/t5-base/resolve/main/spiece.model from cache at /home/pajansen/.cache/huggingface/transformers/684a47ca6257e4ca71f0037771464c5b323e945fbc58697d2fad8a7dd1a2f8ba.3b69006860e7b5d0a63ffdddc01ddcd6b7c318a6f4fd793596552c741734c62d\r\n[INFO|tokenization_utils_base.py:1766] 2021-01-25 11:39:52,523 >> loading file https://huggingface.co/t5-base/resolve/main/tokenizer.json from cache at /home/pajansen/.cache/huggingface/transformers/90de37880b5ff5ac7ab70ff0bd369f207e9b74133fa153c163d14c5bb0116207.8627f1bd5d270a9fd2e5a51c8bec3223896587cc3cfe13edeabb0992ab43c529\r\n[INFO|modeling_utils.py:1027] 2021-01-25 11:39:52,809 >> loading weights file https://huggingface.co/t5-base/resolve/main/pytorch_model.bin from cache at /home/pajansen/.cache/huggingface/transformers/ab4e948915b067f5cb6e5105f6f85044fd717b133f43240db67899a8fc7b29a2.26934c75adf19ceac3c268b721ba353356b7609c45f5627550326f275a2163b4\r\n[INFO|modeling_utils.py:1143] 2021-01-25 11:39:58,232 >> All model checkpoint weights were used when initializing T5ForConditionalGeneration.\r\n\r\n[INFO|modeling_utils.py:1151] 2021-01-25 11:39:58,233 >> All the weights of T5ForConditionalGeneration were initialized from the model checkpoint at t5-base.\r\nIf your task is similar to the task the model of the checkpoint was trained on, you can already use T5ForConditionalGeneration for predictions without further training.\r\n01/25/2021 11:39:58 - INFO - utils - setting model.config to task specific params for translation_en_to_ro:\r\n {'early_stopping': True, 'max_length': 300, 'num_beams': 4, 'prefix': 'translate English to Romanian: '}\r\n01/25/2021 11:39:58 - INFO - utils - note: command line args may override some of these\r\n[INFO|modeling_t5.py:1536] 2021-01-25 11:39:58,479 >> enabling pipeline with chunks=4\r\n[INFO|modeling_t5.py:1545] 2021-01-25 11:39:58,479 >> using user-provided device_map\r\n[INFO|modeling_t5.py:1563] 2021-01-25 11:39:58,479 >> using pipeline partitioning: {0: [0, 1, 2], 1: [3, 4, 5, 6, 7, 8, 9, 10, 11]}\r\n[W ProcessGroupGloo.cpp:532] Warning: Unable to resolve hostname to a (local) address. Using the loopback address as fallback. Manually set the network interface to bind to with GLOO_SOCKET_IFNAME. (function operator())\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n[W tensorpipe_agent.cpp:63] Failed to look up the IP address for the hostname (EADDRNOTAVAIL: address not available), defaulting to 127.0.0.1\r\n01/25/2021 11:40:01 - INFO - __main__ - *** Train ***\r\n[INFO|trainer.py:807] 2021-01-25 11:40:01,659 >> ***** Running training *****\r\n[INFO|trainer.py:808] 2021-01-25 11:40:01,659 >> Num examples = 1000\r\n[INFO|trainer.py:809] 2021-01-25 11:40:01,659 >> Num Epochs = 1\r\n[INFO|trainer.py:810] 2021-01-25 11:40:01,659 >> Instantaneous batch size per device = 160\r\n[INFO|trainer.py:811] 2021-01-25 11:40:01,659 >> Total train batch size (w. parallel, distributed & accumulation) = 160\r\n[INFO|trainer.py:812] 2021-01-25 11:40:01,659 >> Gradient Accumulation steps = 1\r\n[INFO|trainer.py:813] 2021-01-25 11:40:01,659 >> Total optimization steps = 7\r\n2021-01-25 11:40:01.766436: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0\r\n 0%| | 0/7 [00:00<?, ?it/s]Traceback (most recent call last):\r\n File \"./finetune_trainer.py\", line 373, in <module>\r\n main()\r\n File \"./finetune_trainer.py\", line 303, in main\r\n train_result = trainer.train(\r\n File \"/home/pajansen/stass-test1/transformers/src/transformers/trainer.py\", line 904, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/home/pajansen/stass-test1/transformers/src/transformers/trainer.py\", line 1271, in training_step\r\n loss = self.compute_loss(model, inputs)\r\n File \"/home/pajansen/stass-test1/transformers/src/transformers/trainer.py\", line 1301, in compute_loss\r\n outputs = model(**inputs)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/stass-test1/transformers/src/transformers/models/t5/modeling_t5.py\", line 1704, in forward\r\n encoder_outputs = self.encoder(\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/stass-test1/transformers/src/transformers/models/t5/modeling_t5.py\", line 1088, in forward\r\n outputs = block_pipe(inputs)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/pipe.py\", line 362, in forward\r\n self.pipeline.run(batches)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/pipeline.py\", line 117, in run\r\n self.compute(batches, schedule, skip_trackers)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/pipeline.py\", line 257, in compute\r\n raise exc_info[0].with_traceback(exc_info[1], exc_info[2])\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/worker.py\", line 79, in worker\r\n batch = task.compute()\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/worker.py\", line 60, in compute\r\n return self._compute()\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/pipeline.py\", line 222, in compute\r\n return batch.call(partition)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/distributed/pipeline/sync/microbatch.py\", line 70, in call\r\n return Batch(function(self.value))\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/container.py\", line 119, in forward\r\n input = module(input)\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/pajansen/stass-test1/transformers/src/transformers/models/t5/modeling_t5.py\", line 838, in forward\r\n layer_outputs = self.layer_module(hidden_states,\r\n File \"/home/pajansen/anaconda3/envs/py38-pt18/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\nTypeError: forward() got an unexpected keyword argument 'head_mask'\r\n```\r\n\r\n\r\n", "It looks more like a warning as it recovers with a fallback, make sure you have:\r\n\r\n```\r\n$ cat /etc/hosts\r\n127.0.0.1 localhost\r\n```\r\n\r\nIt looks like I forgot to commit the last change. My apologies. Could you please update and try again?", "Thanks -- appears to be working -- on t5-3B it spreads it evenly across the 4 A100s (13.0-13.3GB each with a batch size of 1). For t5-11B there's an out of memory error -- I suppose (naively) if 11b is ~3.7x larger than 3B then it would require ~49gb per card without some form of offloading?\r\n", "Thank you for confirming that you were able to use it with t5-3b on your 4 gpus. \r\n\r\nWere you able to get a decent gpu utilization across the board? Or were they all under 25%?\r\n\r\n--------------\r\n\r\nPlease make sure you read my notes on the balancing in OP and experiment with the device map so that all gpus get a balanced GPU memory usage. gpu0 is already busy with many things, so I'd try a spread of 2/4/4/4 parts or perhaps 1/3/3/3 in your definition of:\r\n\r\n`--pipeline \"chunks=4 device_map=0:0-3,1:3-12\"`\r\n\r\nin this example we have 1/3 parts balance between gpu 0 and 1. i.e. 3 times more layers for gpu 1.\r\n\r\nOf course, it needs to be adjusted to 4 gpus and I don't remember how many encoder blocks t5-11b has, but as I mentioned if you look at the logs you will find a ready map there, just re-adjust it to balance things better. Please let me know if I communicated clearly what I'm trying to say - we want all 4 gpus to have about the same memory usage - then we maximize the chance to fit t5-11b on those 4 gpus.\r\n\r\n--------------\r\n\r\nNext we need to try to bolt DeepSpeed on it. So we will try to use 2 gpus for pipeline and 2 gpus for ZeRO-DP and perhaps some ZeRO-Offload too. I should get access to 4 gpus soon and I will start working on figuring that step out. I will post back once I have something practical to share.", "Thanks -- the autobalancing (just \"chunks=4\") actually seemed to give nearly entirely even results on -3B (the ~13.0-3GB each), so I tried that with 11B instead of manually supplying the device map (since it seemed a bit uneven when I tested on -base) -- but I'll tinker on 11B and report back. ", "> the autobalancing \r\n\r\nFYI, currently the automatic device map just tries to split `n_layers/n_gpus` per gpu, and not taking into an account gpu0's extra load. Once everything else is working we will come up with much better heuristics based on actual gpu capacity and each layer's real memory demands.\r\n\r\n", "What's interesting is that I'm not generally observing GPU0 to have a higher load. Here's an example with unifiedqa-t5-3b (essentially just a further pre-trained t5-3b, not relevant here), chunks=4, autobalancing (with a different visualization tool). They all tend to show about the same RAM usage over time. The graph also shows the utilization (also generally under 30% most of the time): \r\n\r\n\r\n\r\nBTW -- I tinkered with different manual device_map settings for t5-11b, but it always quickly gave out of memory errors.", "Oh, what tool is that? I want it too!\r\n\r\nIt looks like different GPUs behave differently, it will take some experimentation to make sense of it all.\r\n\r\nBut clearly you're also not seeing much benefit from the pipeline over the native MP. Same as I. Either my workaround to make it work slow everything down or there is another problem elsewhere. As I mentioned I'd like to redesign my implementation in hope \r\nto reduce the unnecessary logic and data-copying.\r\n\r\n> BTW -- I tinkered with different manual device_map settings for t5-11b, but it always quickly gave out of memory errors.\r\n\r\nThank you for the experimentation. I'm still waiting to get access to a 4-gpu setup and when it happens will immediately start experimenting with bolting DeepSpeed on it and then will get back to you.", "Thanks -- this handy cool visualization tool is nvtop -- I just found it to plot the relative changes rather than stare at nvidia-smi and hope to keep it all in my brain. It's available with apt ( sudo apt-get install nvtop ). \r\n\r\nHappy to offer my rig for some testing if you need a 4 GPU setup sooner. :) ", "Oh, yes, I had it and forgot about its usefulness. Thank you!\r\n\r\nI typically use \r\n```\r\nalias wn='watch -n 1 nvidia-smi'\r\n```\r\nbut this is a way better.\r\n\r\n> Happy to offer my rig for some testing if you need a 4 GPU setup sooner. :)\r\n\r\nIf don't find access by tomorrow I will gladly accept your generous offer, @PeterAJansen. Thank you!", "hmm, how do you get a split screen per card in nvtop? for some reason my version reports both cards as one card. I don't see any command line options to configure that.", "hmmm, it actually worked out-of-the-box for me (but looks very different depending on the dimensions of the terminal). Does it show only 1 GPU (with memory for both?), or two separate GPUs? ", "It reports 2 gpus but shows the report only for gpu 0. could be a bug. I just saw that for you it showed all 4 gpus.\r\n\r\n", "What happens if you make the window really tall/wide? It changes the display for me if I resize the terminal -- if I make it really tiny, it looks something like yours: \r\n\r\n\r\n", "Sorry, I forgot to mentioned I tried this already to no avail. I give it a huge console.\r\n\r\nI even tried various terminals - same.\r\n\r\nI think it may have to do with my 2nd card being rtx-3090 - and it doesn't work with cuda < 11.1 - most likely nvtop was built against cuda-10, so while it replicates the nvidia-smi stats, it can't access nvml for that card and thus doesn't show the graph.\r\n\r\nYup, installed nvtop on a machine with 2 normal gpus and it shows them both in the same-size terminal. So it just can't handle rtx-30* unless it's rebuilt from source against cuda-11.1+\r\n\r\nBut even then when it works it gives no way to separate the 2 gpu other than colors and 4 lines often around the same magnitude for different things are impossible to make sense of. This is an odd design. ", ":-/ That's unfortunate (though I suppose the cost of using bleeding-edge hardware). The A100s are supported with CUDA 11.0, so they must just squeak in on the current version available with apt. \r\n\r\n(And, the usability is a little unusual, but with ASCII graphics there are strong limits... :) )", "pytorch w/ cuda-11.2 nightly should be available any day now. cuda-11.2 has been out for a month now.\r\n\r\n>(And, the usability is a little unusual, but with ASCII graphics there are strong limits... :) )\r\n\r\nThis is a good point. But at least one could use a double line or asterisks or something to differentiate 4 different things. Perhaps some people can track 4 similar colors and remember which is which. Not me. I guess the source code is there, if I really need to I could probably hack it to do be more user-friendly.", "Update: this overload of the term MP to mean totally different things is a big problem.\r\n\r\nI was sure I could easily combine non-DeepSpeed pipeline with Deepspeed after reading \r\nhttps://www.deepspeed.ai/features/#support-for-custom-model-parallelism\r\nExcept, I have just now realized that it's not PP but the super-confusing-mean-different-things-in-different-contexts abbreviation MP, which in this particular context means horizontal MP and not vertical MP/PP. And there are no instructions on how to integrate non-DeepSpeed PP. So I have been trying to fix the wrong thing. https://github.com/microsoft/DeepSpeed/issues/710\r\n\r\nSo this particular branch takes us nowhere closer to integration of PP with DeepSpeed.\r\n\r\nBack to the drawing board.", "too long. closing.", "We will test this branch soon.", "There are probably some things that can be salvaged from this PR, but the main utility of it is to see the difficulties I run into. And of course, this is not a good solution not only because the code is absolutely nuts, but because it's very inefficient.\r\n\r\nAs I mentioned in the other thread, pytorch now has a better API, so some of the encoding/decoding of non-tensor inputs/outputs I did won't be needed anymore as it now supports non-tensor inputs/output." ]