Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Pinging @vasudevgupta7 and @patrickvonplaten ", "There are several issues, we will have to keep in mind before we decide to implement bigbird block sparse attention in decoder.\r\n\r\n1) Initial number of tokens in decoder side must be > `5 x block_size + 2 x num_random_blocks x block_size` (for bigbird block sparse attention to work). Typically `block_size=64`, `num_random_blocks=3`, then we need at least 708 tokens initially.\r\n2) I am assuming your task involves at least 1024 tokens in decoder side. Since, else it is recommended to use `original_full` attention by authors.\r\n3) Also, if model is given 1024 tokens (block size = 64 let's say), then for predicting 1025th token, model will first `<pad>` to 1024 + 64 tokens, since this attention will work only when sequence length is multiple of block size. & similarly for predicting every single token, we will have to pad again & again. This way inference can become expensive, I believe.\r\n4) Also, if you want to use Encoder-Decoder model, we will have to write block sparse attention in cross-attention completely. Else cross-attention layer will be very expensive if both encoder & decoder are given very long sequences.\r\n\r\nYeah training decoder in the bigbird block sparse attention fashion is possible if we just change masks in this [line](https://github.com/huggingface/transformers/blob/3d339ee6595b9e42925559ae21a0f6e77f032873/src/transformers/models/big_bird/modeling_big_bird.py#L2119). But again, there are several issues:\r\n\r\n5) Currently, we have 1st & last block as global. So, if we mask for autoregressive, we will have only 1st block as global during training.\r\n6) Because of masking all the right tokens, we will reduce number of random tokens which were chosen from the complete sequence earlier.\r\n7) Sliding tokens will also reduce by half. \r\n\r\nSo, this way number of tokens which each query can attend, will be reduced by large amount. 5,6,7 can be resolved if we decide to implement modify a lot in current block sparse attention (instead of just changing masks). But 1,2,3,4 are still a problem.\r\n\r\nThere may be several other issues. But these are my initial thoughts on this. Let me know if I am unclear or wrong somewhere.", "Thanks for the response! I was on vacation when you posted and am coming back to this now.\r\n\r\n>Initial number of tokens in decoder side must be > 5 x block_size + 2 x num_random_blocks x block_size (for bigbird block sparse attention to work). Typically block_size=64, num_random_blocks=3, then we need at least 708 tokens initially.\r\nI am assuming your task involves at least 1024 tokens in decoder side. Since, else it is recommended to use original_full attention by authors.\r\nAlso, if model is given 1024 tokens (block size = 64 let's say), then for predicting 1025th token, model will first <pad> to 1024 + 64 tokens, since this attention will work only when sequence length is multiple of block size. & similarly for predicting every single token, we will have to pad again & again. This way inference can become expensive, I believe.\r\n\r\nYes, we are using a large input size so the motivation to implement causal is there regardless.\r\n\r\nI see how there is a minimum threshold input size below which BigBird is obsolete. That said, even in that case, hyper-parameters could be tweaked such as decreasing block size and number of random blocks. Not to mention, this minimum threshold applies to other efficient attention algos as well such as performer.\r\n\r\n>Also, if you want to use Encoder-Decoder model, we will have to write block sparse attention in cross-attention completely. Else cross-attention layer will be very expensive if both encoder & decoder are given very long sequences.\r\n\r\nAre you saying that BigBird is not possible in cross attention? Could use some more color on this.\r\n\r\n>Currently, we have 1st & last block as global. So, if we mask for autoregressive, we will have only 1st block as global during training.\r\nBecause of masking all the right tokens, we will reduce number of random tokens which were chosen from the complete sequence earlier.\r\nSliding tokens will also reduce by half.\r\n\r\nThe attention mask is applied after computing the sparse attention matrix, so doesn't the full global attention piece still pull weight? \r\n", "> Thanks for the response! I was on vacation when you posted and am coming back to this now.\r\n> \r\n> > Initial number of tokens in decoder side must be > 5 x block_size + 2 x num_random_blocks x block_size (for bigbird block sparse attention to work). Typically block_size=64, num_random_blocks=3, then we need at least 708 tokens initially.\r\n> > I am assuming your task involves at least 1024 tokens in decoder side. Since, else it is recommended to use original_full attention by authors.\r\n> > Also, if model is given 1024 tokens (block size = 64 let's say), then for predicting 1025th token, model will first to 1024 + 64 tokens, since this attention will work only when sequence length is multiple of block size. & similarly for predicting every single token, we will have to pad again & again. This way inference can become expensive, I believe.\r\n> \r\n> Yes, we are using a large input size so the motivation to implement causal is there regardless.\r\n> \r\n> I see how there is a minimum threshold input size below which BigBird is obsolete. That said, even in that case, hyper-parameters could be tweaked such as decreasing block size and number of random blocks. Not to mention, this minimum threshold applies to other efficient attention algos as well such as performer.\r\n> \r\n> > Also, if you want to use Encoder-Decoder model, we will have to write block sparse attention in cross-attention completely. Else cross-attention layer will be very expensive if both encoder & decoder are given very long sequences.\r\n> \r\n> Are you saying that BigBird is not possible in cross attention? Could use some more color on this.\r\n\r\nIt might be possible to implement as cross-attention if we think something but major problem will that:\r\nFor bigbird block sparse attention to work, this must hold `query sequence length // block size == key sequence length // block size`. Now this can be managed during training but during inference, I am not sure if we can.\r\n\r\n> \r\n> > Currently, we have 1st & last block as global. So, if we mask for autoregressive, we will have only 1st block as global during training.\r\n> > Because of masking all the right tokens, we will reduce number of random tokens which were chosen from the complete sequence earlier.\r\n> > Sliding tokens will also reduce by half.\r\n> \r\n> The attention mask is applied after computing the sparse attention matrix, so doesn't the full global attention piece still pull weight?\r\n\r\nDuring training, we will have to mask the last global block right; so they won't contribute to the context layer & effectively global tokens will reduce by half.\r\n\r\n\r\npinging @patrickvonplaten for putting some light on this issue.\r\n", "Thanks for the issue & the detailed answer @vasudevgupta7. To be honest, I just don't think it's worth yet to do any kind of sparse attention on neither the cross_attention layers nor the decoder attention layers because the output is usually quite small (think summarization, question_answering). For tasks like translation, it's often better to split per sentence anyways so here it also doesn't make too much sense. => Overall IMO, this is low priority", "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." ]

[]

[]

[ "Pinging @stas00 since I see in your stack trace you are working on jean-zay and I know Stas has some experience with it.", "@keloemma, please have a look at the special offline mode https://huggingface.co/transformers/installation.html#offline-mode - if your gpu instance is firewalled (which is the case on JZ) it explains how to solve this problem.", "@sgugger, I'm thinking the installation doc is not quite the right place for the offline mode feature. Somehow I'd never think to look there. Could you think of a better placement?\r\n\r\nAlso perhaps we should start a troubleshooting doc like we had @ fastai with pointers to solutions based on symptoms? So this could be the first entry.", "I wouldn't remove the doc for offline feature from the installation page, but we can certainly duplicate it elsewhere. Why not in the `main_classes/model` apge since it has the from_pretrained method?\r\n\r\nWe can also start a troubleshooting document, that's also helpful.", "Thank you for this feedback, @sgugger - I will do both.\r\n\r\nhttps://github.com/huggingface/transformers/pull/11236", "@stas00 Hello and thank you for your response. I have just a question, \r\n\r\n#Setting environment variable TRANSFORMERS_OFFLINE=1 will tell 🤗 Transformers to use local files only and will not try to look things up\"\r\n=> if I understand , this means that I have to save locally FlauBert model size on a specific directory in the local and then provided the path to that directory in my original script and then I set the variable as mentionned in the doc; \r\n\r\nso I should have something like this : \r\n\r\nHF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 \\\r\nsrun python test.py --model_name_or_path t5-small ( path is already written in the script \"test.py)\r\n\r\n", "If I understand correctly the instructions aren't too clear, right?\r\n\r\nSo you run in your login shell that is not firewalled:\r\n```\r\ncd /path/to/transformers\r\npython examples/seq2seq/run_translation.py --model_name_or_path t5-small --dataset_name wmt16 --dataset_config ro-en ...\r\n```\r\n\r\nthen for example let's do an interactive session with `srun`, e.g.:\r\n```\r\nsrun --pty --ntasks=1 --cpus-per-task=10 --gres=gpu:2 --hint=nomultithread --time=60 bash\r\n```\r\nwhich opens a new interactive bash shell on the gpu instance, and then you repeat exactly the same command, but this time with 2 env vars:\r\n```\r\ncd /path/to/transformers\r\nHF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 \\\r\npython examples/seq2seq/run_translation.py --model_name_or_path t5-small --dataset_name wmt16 --dataset_config ro-en ...\r\n```\r\nOf course, you will need to figure out to load the right `modules`, etc., to get the desired environment. But I trust you have already done so.\r\n\r\nIt should work exactly the same with the non-interactive shell if you use `sbatch`.\r\n\r\nBottom line: do the same thing as you normally do, but the key here is to first to launch the script in your login shell (no gpus!) so that it would download the model and the dataset files. You don't need to let the script complete - it will take forever since you're running on CPU, but just the first batch is enough since by that time all files would be downloaded.\r\n\r\nPlease let me know if this helped and if it did how can I improve the documentation to make it easier to understand.\r\n\r\n> HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1\r\n> srun python test.py --model_name_or_path t5-small ( path is already written in the script \"test.py)\r\n\r\nI suppose you could try that as well, I haven't tried, I would start a bash shell via `srun` and then start the script from there. Since this is what you'd do if you were to use SLURM normally via `sbatch`.\r\n", "@stas00 Thank you for your clear explanation, I will try all that and tell you back what works", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "You need to pass along `--version_2_with_negative` when using this script with a dataset that has samples with no answers (like squad v2).", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Ah, I don't match the `.. code-block:: xxx`, thanks for pointing that out!", "OK, the next problem is the one I was concerned about in the first place (which is why i didn't offer a perl one liner). Your fixing code isn't shifting the whole block, but only the outside lines. so it now results in broken python code - bad indentation. e.g. look at any code blocks in the diff that start with `class`, or `{`\r\n\r\nSo when you shift whitespace - you have to do it for the whole code block.", "That's too hard to fix and will result in other problems (what about bash commands that have different indents with the \\?). So I will remove the automatic fix and just put a hard error the user will have to manually fix.", "> That's too hard to fix and will result in other problems (what about bash commands that have different indents with the ?). \r\n\r\nBash has no indentation issues. Only python does AFAIK. Perhaps you meant multiline python fed into a bash shell?\r\n\r\nI think it's about finding the difference in the required whitespace, and applying this exact change to the whole block should make it work. Since everything will be shifted by the same number of characters, which is what the manual fix would do anyway.\r\n\r\n> So I will remove the automatic fix and just put a hard error the user will have to manually fix.\r\n\r\nSure, that would work just fine.\r\n\r\nThank you.", "Ok, I have something that can do the whole block, but since this PR already badly treated some parts, I need to go back from a fresh master, so closing this one." ]

[]

[]

[ "Hi @forest1988, thanks a lot for your PR! In parallel I've merged another PR that enforces proper indentation for those examples so now yours is conflicting. Could you rebase on master and solve the conflicts, or open a new PR from a fresh master, whichever is easier?\r\nThanks", "Hi @sgugger,\r\nThank you for telling me that there is another PR merged and this PR has conflicts with it.\r\nI've just rebased on master and solved the conflicts.\r\n\r\nIt seems there is a code quality problem, so I'll soon fix it.", "`check_code_quality` shows the error message as below.\r\nIt seems something wrong happens during installing packages.\r\n\r\nI'm sorry, but can you try to run circleci again?\r\n\r\n```\r\n#!/bin/bash -eo pipefail\r\npip install .[all,quality]\r\nDefaulting to user installation because normal site-packages is not writeable\r\nProcessing /home/circleci/transformers\r\n Installing build dependencies ... -\b \b\\\b \b|\b \b/\b \bdone\r\n Getting requirements to build wheel ... -\b \bdone\r\n Preparing wheel metadata ... -\b \b\\\b \bdone\r\nRequirement already satisfied: filelock in /usr/local/lib/python3.6/site-packages (from transformers==4.6.0.dev0) (3.0.12)\r\nCollecting dataclasses\r\n Using cached dataclasses-0.8-py3-none-any.whl (19 kB)\r\nRequirement already satisfied: importlib-metadata in /usr/local/lib/python3.6/site-packages (from transformers==4.6.0.dev0) (1.7.0)\r\nCollecting tqdm>=4.27\r\n Using cached tqdm-4.60.0-py2.py3-none-any.whl (75 kB)\r\nRequirement already satisfied: requests in /usr/local/lib/python3.6/site-packages (from transformers==4.6.0.dev0) (2.25.1)\r\nCollecting tokenizers<0.11,>=0.10.1\r\n Using cached tokenizers-0.10.2-cp36-cp36m-manylinux2010_x86_64.whl (3.3 MB)\r\nRequirement already satisfied: packaging in /usr/local/lib/python3.6/site-packages (from transformers==4.6.0.dev0) (20.9)\r\nCollecting regex!=2019.12.17\r\n Using cached regex-2021.4.4-cp36-cp36m-manylinux2014_x86_64.whl (722 kB)\r\nCollecting numpy>=1.17\r\n Using cached numpy-1.19.5-cp36-cp36m-manylinux2010_x86_64.whl (14.8 MB)\r\nCollecting sacremoses\r\n Using cached sacremoses-0.0.44-py3-none-any.whl\r\nRequirement already satisfied: isort>=5.5.4 in /home/circleci/.local/lib/python3.6/site-packages (from transformers==4.6.0.dev0) (5.8.0)\r\nCollecting black>=20.8b1\r\n Using cached black-20.8b1-py3-none-any.whl\r\nCollecting flake8>=3.8.3\r\n Using cached flake8-3.9.0-py2.py3-none-any.whl (73 kB)\r\nCollecting jaxlib>=0.1.59\r\n Using cached jaxlib-0.1.65-cp36-none-manylinux2010_x86_64.whl (44.7 MB)\r\nCollecting soundfile\r\n Using cached SoundFile-0.10.3.post1-py2.py3-none-any.whl (21 kB)\r\nCollecting tensorflow>=2.3\r\n Using cached tensorflow-2.4.1-cp36-cp36m-manylinux2010_x86_64.whl (394.3 MB)\r\nCollecting torchaudio\r\n Using cached torchaudio-0.8.1-cp36-cp36m-manylinux1_x86_64.whl (1.9 MB)\r\nCollecting Pillow\r\n Using cached Pillow-8.2.0-cp36-cp36m-manylinux1_x86_64.whl (3.0 MB)\r\nCollecting keras2onnx\r\n Using cached keras2onnx-1.7.0-py3-none-any.whl (96 kB)\r\nCollecting jax>=0.2.8\r\n Using cached jax-0.2.12-py3-none-any.whl\r\nCollecting sentencepiece==0.1.91\r\n Using cached sentencepiece-0.1.91-cp36-cp36m-manylinux1_x86_64.whl (1.1 MB)\r\nCollecting protobuf\r\n Using cached protobuf-3.15.8-cp36-cp36m-manylinux1_x86_64.whl (1.0 MB)\r\nCollecting flax>=0.3.2\r\n Using cached flax-0.3.3-py3-none-any.whl (179 kB)\r\nCollecting torch>=1.0\r\n\r\nReceived \"killed\" signal\r\n```", "Thanks, I applied all suggestions!\r\n\r\nI'm sorry, I misunderstood the meaning of the following two that appear in https://huggingface.co/transformers/_sources/quicktour.rst.txt, and assumed that double # were required for comments in the Transformers documentation.\r\n`## PYTORCH CODE` ` ## TENSORFLOW CODE`\r\n\r\nIf I'm not mistaken in my current understanding, these are special codes to switch between PyTorch and TensorFlow versions, right?\r\n", "Yes those are special markers, which is why they have the double #, for regular comments, we use just one #" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "@LysandreJik @sgugger all comments are addressed, CI is green! Incredible work by the original author. ", "Thanks, addressed your comments.\r\n\r\nAdded LUKE to the README, and added 3 community notebooks. ", "Thanks again for all your work on this!" ]

[]

[]

[ "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.", "Should be addressed.", "Taking a look at the `pytorch_model.bin` saved on the `microsoft/DialoGPT-small` repository, one can see it's made up of float16 weights. When loading the model in the `GPT2Model` and saving it, the weights are saved in float32, resulting in the large increase.\r\n\r\nIf you want to keep the model in half precision, add the following line after initializing your model:\r\n```py\r\nmodel.half()\r\n```" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thanks a lot for your help!", "@sgugger @LysandreJik \r\nThank you for your quick reviewing this PR!\r\n\r\n@sgugger \r\nThank you for telling me how to format the code block!\r\n\r\n> Thanks a lot for your PR! I made a couple of suggestions. Mostly, the example should follow the syntax\r\n> \r\n> ```\r\n> >>> code_line_1\r\n> >>> code_line_2\r\n> result\r\n> ```\r\n\r\nI've split the big code block into two, and then make all lines prefixed with `>>>`.\r\nNow I think I can understand the format.\r\nHowever, I wonder why BertTweet and BertGeneration, which I referred to before opening this PR, has code blocks without using `>>>` in them.\r\nAre there any specific reasons? (Could it be because the output is not specifically described?)\r\nOr, may I correct them using `>>>`?\r\n", "We should always have the `>>>` as it allows us to use `doctest` which will test the example (it's been deactivated for a while but we will bring it back to life soon). So if you want to add those to some examples where it's missing, go ahead!\r\n\r\nThe only instance where we would not want those `>>>` is if we don't want the example to be tested.", "Thanks for the detailed explanation about the prefix!\r\nNow, I would like to add `>>>` to examples without the prefix, as far as I can find (except for which you don't want to be tested)." ]

[ { "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 Stas! I'm not sure what exactly introduced this memory/speed regression, so I'm going to investigate it and won't merge this PR before that.", "> Thank you for doing this refactoring, @patil-suraj!\r\n> \r\n> It's a bit hard to review since all the code is moved around, so no easy diff to follow - so while I skimmed through it - I trust your expertise and the tests on the correctness.\r\n> \r\n> With regards to memory/performance regression - (thank you for running this important check!) could it be that it was introduced in the initial Bart refactor? i.e. perhaps running the same check on Bart pre and post PR that did the main refactoring (when all the Barts were split up)? And if so then the issue is bigger and needs to be looked in that PR that introduced it.\r\n\r\nI remember that I checked that the [Bart refactor](https://github.com/huggingface/transformers/pull/8900) didn't show any regression both on the forward pass and `generate()`. I might however have overlooked something. Would definitely be a good idea to verify this first with the exact same testing params (batch_size=4, ...)!", "@patil-suraj, what needs to be done to complete this? \r\n\r\nLast we talked there was a performance regression and the suggestion was to test Bart's performance pre and post its original refactoring.", "@patil-suraj, FYI, recently I made a few fixes to the model to make it work with Deepspeed:\r\nhttps://github.com/huggingface/transformers/pull/12477/files#diff-564f6d9b78eec17b410c924f868840770a9ad9649032bcf3754827317b9eaba3\r\n\r\nAre we still planning to merge this PR? As we said earlier if there is a regression it'll be on the whole Bart family, so perhaps it might be easier to just merge this? Otherwise a lot of time gets waste getting back to it again and again and not getting anywhere.\r\n\r\nThank you." ]

[]

[]

[ "Hi,\r\n\r\nYou can use Stackoverflow for that: https://stats.stackexchange.com/questions/19048/what-is-the-difference-between-test-set-and-validation-set\r\n\r\nWe like to keep Github issues for feature requests/bug reports. There's also the [forum](https://discuss.huggingface.co/) where you can ask training-related questions.\r\n\r\nThanks!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This seems to be an issue with `simpletransformers` so please post it there since we won't get time to look into other code bases to fix such issues.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hi @ashleylew \r\n\r\nThe `run_translation.py` uses the `Seq2SeqTrainer` which does not pass the `num_return_sequences` argument to `generate`, this is because if multiple sequences are returned then its not clear what sequence should be used to compute the metrics. \r\n\r\nyou could generate test set predictions by using the `generate` method and passing the `num_return_sequences` argument. But if you want to do this using `Seq2SeqTrainer` then you would need to modify it.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[ { "id": 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 } ]

[ "Would that information not be better on the respective model cards? I think that's more on that side that it should go with the rest of the training setup.", "1. It'd be difficult to enforce that it's being filled out\r\n2. It'd be difficult to tell the user at run time that model A pre-trained in fp32, is attempted to be run in fp16 or bf16\r\n\r\nBut if we can't do it on config level at the very least it could be a required card entry (but I don't think anything is required in the cards).", "How would you enforce it being filled out via the config? You would get the default for pretty much all models too: if a user is too lazy to fill the model card they will also be too lazy to fill the config.\r\n\r\nI don't understand what you mean in 2, could you elaborate? Why would it be bad to fine-tune a model trained in FP32 in FP16 or bfloat16?", "> How would you enforce it being filled out via the config? You would get the default for pretty much all models too: if a user is too lazy to fill the model card they will also be too lazy to fill the config.\r\n\r\nHmm, right. I was thinking about the tests, which can enforce the field existing in the config object but have no way to enforce the values.\r\n\r\n> I don't understand what you mean in 2, could you elaborate? Why would it be bad to fine-tune a model trained in FP32 in FP16 or bfloat16?\r\n\r\nIt won't work out of the box and will require finetuning, which may not succeed if running into infs/nans. I suppose I was thinking more about inference, which won't work w/o finetuning first. If the model was pre-trained in mixed precision it can be used in fp16 inference, but this won't be the case if it was pretrained in fp32.", "I think this feature would be welcome indeed and would save us a lot of trouble as we've seen in the past. Regarding whether we want to have this in the model card or in the configuration, I guess it really depends on whether we want to enforce that with errors or warnings.\r\n\r\nI think the main point of having that field is to actually warn the user when they're doing inference with a model trained with a different precision, and to that extent, having it in the configuration makes more sense.\r\n\r\nI also think the configuration is here to detail how a checkpoint is configured: how the architecture fits the weights (hidden size, layers) and how it should be used (model type, architecture). I think it would make sense to have this as a configuration field, as not knowing that can result in an unusable checkpoint in other environments. I think that's different from other training-related arguments, such as `gradient_checkpointing`, which don't really make sense once ported to a different environment.", "Excellent! So that makes the 2 of us who think it would be most strategically placed in the model config.\r\n\r\nSo let's look at specifics. I can think of the following:\r\n\r\n1. at conversion point - it'd be a responsibility of the porter to fill it out - but could also look at the imported state_dict first - perhaps the weights are already in non-fp32 (some models are saved in `.half()` so in this situation it could be derived automatically)\r\n\r\n2. at `save_pretrained` - this is the difficult one. what do we set here? As `save_pretrained` has no way to determine how the model was trained. So we will need to require the precision to be passed explicitly then? The Trainer can be adapted since it knows the precision, but for non-Trainer users will have to specify it explicitly.\r\n\r\n3. rewriting history. what do we do about the thousands of models already on the hub? do a massive script that will push `config.trained_precision = unknown` and then over time start correcting this? at least for the main/popular models and problematic ones - m?t5/pegasus/gpt-neo\r\n\r\nany others cases that I missed?\r\n\r\nwhat would be a good not too long keyword for this one? would `config.trained_precision` be not too long and clear enough?", "I think the name is good. I would leave it to a default of `\"unknown\"` for all existing models, so that we don't have to add it anywhere (especially when we don't have the info). I would personally not try to guess it too much and only set that information when we have it from the people who trained the model.\r\n\r\nFor 2, I don't think we should try to guess it either when people are not using the `Trainer` and just focus on the trainer. We just need to add a `model.config.trained_precision = xxx` from the args and the env at the beginning of training, then the `save_pretrained` method, which also saves the config, will properly handle that.\r\n\r\nFor 3, I would only populate the popular models, for which we have the info.", "> I think the name is good. I would leave it to a default of `\"unknown\"` for all existing models, so that we don't have to add it anywhere (especially when we don't have the info). I would personally not try to guess it too much and only set that information when we have it from the people who trained the model.\r\n\r\nBut we could require this new key for when new models are added. That's why I thought that if we were to massively rewrite the hub's config with `trained_precision = unknown` then we could start enforcing this new field.\r\n\r\n> For 2, I don't think we should try to guess it either when people are not using the `Trainer` and just focus on the trainer. We just need to add a `model.config.trained_precision = xxx` from the args and the env at the beginning of training, then the `save_pretrained` method, which also saves the config, will properly handle that.\r\n\r\nYes! \r\n\r\nThe only trouble here is that someone taking a model in fp32, training it for 10 steps in mixed precision doesn't quite qualify it for fp16.\r\n\r\n----------------\r\n\r\nI think the problem is that we can't make `save_pretrained` require this new field (for outside Trainer) since it'd be a breaking change.\r\n\r\nAnd also the main event where this field needs to be set is when the model is ported from another system (since that's where the current problems all originate from). So how could we at the very least enforce this in conversion scripts?\r\n", "Made a [wiki post](https://discuss.huggingface.co/t/compiling-data-on-how-models-were-trained-fp16-fp32-bf16/5671) - hoping to gather more info via the community input, so that we can have enough data to do some initial seeding of this new field.\r\n", "This PR is related - adding `config.torch_dtype` field: https://github.com/huggingface/transformers/pull/12316\r\n\r\nI guess I can tackle this one next in line.\r\n", "New development:\r\n\r\n8-bit quantized models have arrived: https://github.com/huggingface/transformers/issues/14839 - need to make sure we don't load those in fp32!\r\n", "> > How would you enforce it being filled out via the config? You would get the default for pretty much all models too: if a user is too lazy to fill the model card they will also be too lazy to fill the config.\r\n> \r\n> Hmm, right. I was thinking about the tests, which can enforce the field existing in the config object but have no way to enforce the values.\r\n> \r\n> > I don't understand what you mean in 2, could you elaborate? Why would it be bad to fine-tune a model trained in FP32 in FP16 or bfloat16?\r\n> \r\n> It won't work out of the box and will require finetuning, which may not succeed if running into infs/nans. I suppose I was thinking more about inference, which won't work w/o finetuning first. If the model was pre-trained in mixed precision it can be used in fp16 inference, but this won't be the case if it was pretrained in fp32.\r\n\r\nHi, thanks for this nice analysis. Do you mean the weights of a model pre-trained in mixed precision (computation precision=bf16, **data/param=fp32**) can be loaded in bf16 during inference? Would that change the model weights and hurt inference performance? @stas00 stas00", "It should work just fine, @LeoXinhaoLee - at least for the situation you are describing.\r\n\r\nThe problem usually emerges when a model is trained in bf16 but used in fp16.", "> It should work just fine, @LeoXinhaoLee - at least for the situation you are describing.\r\n> \r\n> The problem usually emerges when a model is trained in bf16 but used in fp16.\r\n\r\nThanks for your kind reply. Is this because fp32 and bf16 have the same range, so it will only lose something small after the decimal point when converting fp32 weights to bf16 for inference, but converting to fp16 may lead to overflow and significantly change weights? @stas00 ", "You will find this thread enlightening https://github.com/huggingface/transformers/pull/10956\r\n\r\nHopefully it'll address your question, @LeoXinhaoLee " ]

[]

[]

[ "I am unsure on what you want us to do: all the example scripts have been tested with evaluation and work in a distributed setup. Unless you share your script, there is little we can do to fix an issue we have not encountered.\r\n\r\nAlso make sure you have the latest version of Transformers installed as this bug might have been fixed already.", "@sgugger Hello and thank you for the reply! Understood it's not clear how to help here. Unfortunately the error persists in version `4.5.0`. For a minimal example, will you need data or is code good enough? It's a very nonstandard dataset, composed of DNA strings ( and only 4 coding tokens plus a pad), but the only nonstandard way I am interacting with `transformers` is by feeding the custom datasets to a `Trainer`. \r\n\r\ntraining code:\r\n```python\r\n#!/usr/bin/env python\r\n# -*- coding: utf-8 -*-\r\n\r\nimport os\r\nimport sys\r\nimport logging\r\nfrom typing import Optional\r\n\r\nimport shutil\r\nimport argparse\r\nimport itertools\r\nfrom pathlib import Path\r\nfrom functools import partial\r\n\r\nimport torch\r\nfrom tokenizers import ByteLevelBPETokenizer\r\nfrom tokenizers.pre_tokenizers import Whitespace\r\n\r\nimport transformers\r\n\r\nfrom min_data import DNADataset, read_list, chunk_fasta\r\n\r\ndef arguments():\r\n parser = argparse.ArgumentParser(description=\"Train GPT-2 model on DNA data.\")\r\n parser.add_argument(\"--partitions\", type=Path, nargs=2,\r\n help=\"Train, validation partition files\")\r\n\r\n parser.add_argument(\"--session\", type=Path,\r\n help=\"Training session directory; models are saved here along with other important metadata\")\r\n\r\n parser.add_argument(\"--log-to\", type=Path,\r\n help=\"Tensorboard logging root directory; logs will write to log_dir/session_name\",\r\n default=\"tensorboard_logs\",\r\n dest='log_dir')\r\n\r\n parser.add_argument(\"--tokenizer\", \r\n type=Path,\r\n help=\"Specify pre-trained tokenizer if it exists\",\r\n required=True)\r\n\r\n # architecture specs\r\n parser.add_argument(\"--n-layer\", type=int, default=4, help=\"# of layers\")\r\n parser.add_argument(\"--n-embed\", type=int, default=16, help=\"embedding dim.\")\r\n parser.add_argument(\"--n-inner\", type=int, default=1024, help=\"hidden dim.\")\r\n parser.add_argument(\"--chunk-size\", type=int, default=2000, help=\"max base pair width\", dest=\"chunksize\")\r\n parser.add_argument(\"--lr\", default=1e-4, type=float)\r\n\r\n # training/logging specs\r\n parser.add_argument(\"--train-epochs\", type=int, default=1)\r\n parser.add_argument(\"--save-steps\", type=int, default=250)\r\n parser.add_argument(\"--save-up-to\", type=int, default=5)\r\n parser.add_argument(\"--batch-size\", type=int, default=8)\r\n parser.add_argument(\"--lens\", nargs=2, type=int, default=(10_000, 2000))\r\n parser.add_argument(\"--progress-bar\", action='store_true', default=False, dest='tqdm')\r\n parser.add_argument(\"--local_rank\", type=int, default=-1)\r\n return parser\r\n\r\ndef get_training_args(args: argparse.Namespace):\r\n output_dir = args.session / \"outputs\"\r\n save_steps = args.save_steps\r\n save_limit = args.save_up_to\r\n batch_size = args.batch_size\r\n log_dir = args.log_dir / args.session.name\r\n lr = args.lr\r\n \r\n max_steps = args.train_epochs * args.lens[0]\r\n return transformers.TrainingArguments(\r\n output_dir=str(output_dir),\r\n overwrite_output_dir=True,\r\n ddp_find_unused_parameters=False,\r\n per_device_train_batch_size=batch_size,\r\n per_device_eval_batch_size=batch_size,\r\n evaluation_strategy=\"steps\",\r\n learning_rate=lr,\r\n local_rank=args.local_rank,\r\n disable_tqdm=not args.tqdm,\r\n max_steps=max_steps,\r\n eval_steps=save_steps,\r\n save_steps=save_steps,\r\n prediction_loss_only=True,\r\n #logging_dir=str(log_dir),\r\n save_total_limit=save_limit\r\n )\r\n\r\ndef construct_dataset_from(fasta_list: Path, tokenizer, **kwargs):\r\n dataset = DNADataset( read_list(fasta_list), tokenizer, **kwargs )\r\n return dataset\r\n\r\nclass DNATrainer(transformers.Trainer): # overwritten because error occur wrt sampler for IterableDataset, still seems necessary in 4.5.0\r\n def _get_eval_sampler(self, eval_dataset: Dataset) -> Optional[torch.utils.data.sampler.Sampler]:\r\n if isinstance(eval_dataset, torch.utils.data.IterableDataset):\r\n return None\r\n\r\nif __name__ == '__main__':\r\n args: argparse.Namespace = arguments()\r\n \r\n # get tokenizer, gpt2 config, training arguments \r\n tokenizer: tokenizers.ByteLevelBPETokenizer = LOAD_TOKENIZER(args) # standard loading of tokenizer\r\n config: transformers.GPT2Config = GET_GPT2_CONFIG(args) # generates a GPT2Config \r\n training_args: transformers.TrainingArguments = get_training_args(args)\r\n\r\n model = transformers.GPT2LMHeadModel(config=config)\r\n model.resize_token_embeddings(len(tokenizer))\r\n model.train()\r\n\r\n kwargs = {'pad_token': '<pad>', 'chunksize': args.chunksize}\r\n train_part, val_part = args.partitions # list of files for each dataset to stream from\r\n \r\n train_data = construct_dataset_from(train_part, tokenizer, asserted_len=args.lens[0], **kwargs)\r\n val_data = construct_dataset_from(val_part, tokenizer, asserted_len=args.lens[1], **kwargs)\r\n\r\n collator = transformers.DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)\r\n trainer = DNATrainer(model=model,\r\n args=training_args,\r\n data_collator=collator,\r\n train_dataset=train_data,\r\n eval_dataset=val_data,\r\n )\r\n\r\n trainer.train()\r\n```", "Oh, but you're overriding the sampler part of the `Trainer` code. There is no way distributed evaluation can work then, as it relies on this.", "Aha! Thank you! I'm sure this is the right track, but now I am back to an error in how the `Trainer` chooses a sampler and constructs the `DataLoader`: (this is `transformers` version `4.5.0`)\r\n\r\n```\r\nValueError: DataLoader with IterableDataset: expected unspecified sampler option, but got sampler=<transformers.trainer_pt_utils.SequentialDistributedSampler object at 0x1554e8f1e2b0>\r\nself._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/mnt/home/dberenberg/projects/metagenomics/huggingface_meta/lib/python3.7/site-packages/transformers/trainer.py\", line 1265, in _maybe_log_save_evaluate\r\n eval_dataloader = self.get_eval_dataloader(eval_dataset)\r\n File \"/mnt/home/dberenberg/projects/metagenomics/huggingface_meta/lib/python3.7/site-packages/transformers/trainer.py\", line 612, in get_eval_dataloader\r\n pin_memory=self.args.dataloader_pin_memory,\r\n File \"/mnt/home/dberenberg/projects/metagenomics/huggingface_meta/lib/python3.7/site-packages/torch/utils/data/dataloader.py\", line 231, in __init__\r\nmetrics = self.evaluate()\r\n File \"/mnt/home/dberenberg/projects/metagenomics/huggingface_meta/lib/python3.7/site-packages/transformers/trainer.py\", line 1754, in evaluate\r\n \"sampler option, but got sampler={}\".format(sampler))\r\nValueError: DataLoader with IterableDataset: expected unspecified sampler option, but got sampler=<transformers.trainer_pt_utils.SequentialDistributedSampler object at 0x1554e8039dd8>\r\neval_dataloader = self.get_eval_dataloader(eval_dataset)\r\n File \"/mnt/home/dberenberg/projects/metagenomics/huggingface_meta/lib/python3.7/site-packages/transformers/trainer.py\", line 612, in get_eval_dataloader\r\n pin_memory=self.args.dataloader_pin_memory,\r\n File \"/mnt/home/dberenberg/projects/metagenomics/huggingface_meta/lib/python3.7/site-packages/torch/utils/data/dataloader.py\", line 231, in __init__\r\n \"sampler option, but got sampler={}\".format(sampler))\r\nValueError: DataLoader with IterableDataset: expected unspecified sampler option, but got sampler=<transformers.trainer_pt_utils.SequentialDistributedSampler object at 0x1554f2898c88>\r\n```", "Yes you need to forego the inheritance to `IterableDataset` as PyTorch does not let you take a sampler for those, so you will nee to implement a `__getitem__` instead of the `__iter__` for the evaluation.", "Ok, that makes sense. So just to conclude, `transformers.Trainer` won't work in distributed setting with an `torch.utils.data.IterableDataset`, in principal due to the fact that `IterableDataset`s are not amenable to that use case, since it isn't clear how describe a distributed sampling procedure for them. Is that correct? Thanks in advance", "That's correct, especially for distributed evaluation." ]

[]

[]

[ "Would you consider this solution instead?\r\n\r\n```\r\n--- a/src/transformers/models/bert_generation/configuration_bert_generation.py\r\n+++ b/src/transformers/models/bert_generation/configuration_bert_generation.py\r\n@@ -78,7 +78,7 @@ class BertGenerationConfig(PretrainedConfig):\r\n >>> # Accessing the model configuration\r\n >>> configuration = model.config\r\n \"\"\"\r\n- model_type = \"bert-generation\"\r\n+ model_type = \"bert\"\r\n\r\n def __init__(\r\n self,\r\n```\r\n\r\nSo it's adjusted to say that it's actually a `bert`-type of model that it works with. And if some model can work with 5 types or any, it could list those 5 types or have `any`.\r\n\r\nperhaps we need to split this field - one of them should say the type of model it works with and keep this one that is just a unique identifier of the sub-type of the model.\r\n\r\nSo e.g. `model.config.works_with = \"(type1|...|typeN|any)\"`\r\n", "Then the all the calls to `from_pretrained` with models that have a `model_type` of `bert-generation` will fail (and there are several checkpoints on the hub with that model type). I also don't think this is the only instance where it's possible to load some model's weights in another model, it's just the first one that is reported.", "see my edits - perhaps we need a specific config entry that lists all the model types the model can work with?\r\n\r\nSo that field doesn't control too many functions?", "But why would we actively prevent a user to load some weights in another model if that doesn't cause any error? Of course they would not work as is, but perhaps it could be a strategy of transfer learning.", "Load the model weights with the model class it was pre-trained with, then do whatever you want with those weights - copy them into the new model, etc. Nothing stops the user here from using those weights. i.e. unless I'm missing something here we aren't preventing anything. \r\n\r\nI just fail to see how loading weights in a model class that is totally different can be of any direct use, even for transfer learning. If you can see such ways do you have an example?\r\n\r\nThat's said if you strongly feel that the enforcement of the match is not logical, then I'm totally fine with the proposed change.", "I don't have strong opinions but the change made was breaking for existing use-cases. I have no way to know which other use cases have been broken by it too, so leaving the warning makes the most sense to me to avoid having to do a new patch release in ten days if a user comes with another case of `XxxModel.from_pretrained(yyy_model)` not working anymore.\r\n\r\nLet's see what @LysandreJik thinks!\r\n\r\n", "I agree wrt breaking changes. How far are we from v5.0? We could postpone the enforcement until then and use your proposed change until then.\r\n\r\nBut functionality-wise do you agree that the model type match enforcement would be useful and that it doesn't prevent the user from using the weights from a mismatched model?", "The issue with raising an error is that I'm nearly 100% sure we're forgetting some use-cases. Bert Generation is one example, but I really wouldn't be surprised that there exist other niche use-cases and that we're involuntarily blocking those. Printing a warning instead seems safer, and while it's not as visible as an error, a user that doesn't obtain whatever performance they're looking for will look at the warnings and should still understand where the issue is coming from.\r\n\r\n> How far are we from v5.0? \r\n\r\nIt isn't on the horizon yet. The breaking changes we've wanted to make until now are mostly cosmetic, so there's nothing pushing us to release a breaking release as of now.\r\n\r\nLGTM, thanks for taking care of this @sgugger.", "That's all said - we ideally should start stirring users towards loading models with the exact classes that created them, and once loaded do whatever is wanted (copy weights, etc.). What is happening now in the edge cases is a misuse of not having a strict verification - it kind of works, so \"why not\" seems to be the way. If this is done, e.g. by changing the documentation, this issue will just disappear and we can reinstate the assert-check.\r\n\r\nI was just thinking about this whole issue of warnings and how they don't quite work. A warning sign on the road is not surrounded by 20 other signs - it stands alone and acts as a warning - loud and clear. A warning in the logs is like a vendor in the bazaar shouting how good his wares are - nobody can hear unless you're right in front of that vendor.\r\n\r\nJust 2 days ago my [PR](https://github.com/huggingface/transformers/pull/11168) trying to help with invalid warning, ended up introducing a bug which I didn't see because it got covered up by yet another warning. The first warning was from incomplete design. And the second warning was covering a real bug.\r\n\r\nWarnings should be a last resort and usually indicate that some aspect of the software design isn't fully thought out. IMHO, of course. " ]

[]

[]

[]

[]

[]

[ "Well someone would need to go other all the examples and add it as an argument in those calls to `load_dataset`.", "Would you still keep it in \r\n```python\r\n@dataclass\r\nclass ModelArguments:\r\n```\r\nI can adjust the examples in the near future. I just wanted to align on we can adjust this. ", "We can switch it but to be honest it's more an internal class to the script than something of real significance so we don't really care." ]

[]

[]

[ "Which version of Transformers are you using?\r\n\r\nIf you're using one of the latest versions of Transformers, `modeling_camembert.py` will be located at `transformers.models.camembert.modeling_camembert.py`. ", "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." ]

[]

[]

[ "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!" ]

[]

[]

[ "Hi! I'm the new Tensorflow maintainer at Hugging Face. Your PR looks good, and you're right that unique weight names is a better strategy than relying on `name_scope`. Now that you've raised the issue, doing a check for that across the whole codebase is definitely on my to-do list.\r\n\r\nThe main issue for us is backward compatibility and ensuring that cross-loading weights from PyTorch checkpoints still works as expected after the change. Can you leave this with me for a few days until I get a chance to review that properly? Hopefully there are no issues, but I don't want to be the guy who breaks the whole codebase in his first week, lol.", "> Hi! I'm the new Tensorflow maintainer at Hugging Face. Your PR looks good, and you're right that unique weight names is a better strategy than relying on `name_scope`. Now that you've raised the issue, doing a check for that across the whole codebase is definitely on my to-do list.\r\n> \r\n> The main issue for us is backward compatibility and ensuring that cross-loading weights from PyTorch checkpoints still works as expected after the change. Can you leave this with me for a few days until I get a chance to review that properly? Hopefully there are no issues, but I don't want to be the guy who breaks the whole codebase in his first week, lol.\r\n\r\nHi @Rocketknight1 that sounds great! Thx for taking a look at this :) Just to clarify a bit: It is my current understanding that tf.name_scope() makes absolutely no difference when it comes to variables in tf 2.x and is not comparable to the old tf.compat.v1.variable_scope. There is, to the best of my knowledge, no point in adding these name scopes for variables in tf 2.x.\r\nThey might still be useful for grouping certain ops in the graph under logical names, but TensorFlow 2.x generally relies on the object hierarchy of tf.Module subclass objects rather than global variable names / name spaces. See: https://www.tensorflow.org/guide/migrate#2_use_python_objects_to_track_variables_and_losses", "I checked the test logs and we have several failing tests involving loading model weights, so it seems like there might be backward compatibility issues with this change, even though you're totally right about the `name_scope()` issues. So unfortunately, I probably can't merge this PR as is.\r\n\r\nI'd like to resolve the underlying problem, though - if you want to try to figure out the compatibility issues yourself you can, or if not (which would be completely understandable, lol) I'll try to take a look when I get a chance.", "@Rocketknight1 Hey, I think I have pretty much fixed the issues. But all of this arcane template usage is giving me a headache, any ideas where/how I might be make the final test pass?\r\n(Model templates runner / run_tests_templates)\r\n\r\nThe error message form the test suggests running \"make fix-copies\" but that does not seem to do anything in the current state.", "Hey! Don't worry too much about the template issues, we can fix those up for you before we merge it.\r\n\r\nThis is something that will affect a few teams, though - we're currently in the process of making sure everyone knows about it and they don't think it'll catastrophically break anything, but we might have to make some changes, which will probably be Monday because it's 7pm at the French office on a Friday right now!\r\n\r\nThanks again for the work you put into this and for identifying the problem, though - I'll try to keep you updated as we figure out if we can use your solution, or which tweaks we'll have to make to make it fit in with our other projects.", "> Hey! Don't worry too much about the template issues, we can fix those up for you before we merge it.\r\n> \r\n> This is something that will affect a few teams, though - we're currently in the process of making sure everyone knows about it and they don't think it'll catastrophically break anything, but we might have to make some changes, which will probably be Monday because it's 7pm at the French office on a Friday right now!\r\n> \r\n> Thanks again for the work you put into this and for identifying the problem, though - I'll try to keep you updated as we figure out if we can use your solution, or which tweaks we'll have to make to make it fit in with our other projects.\r\n\r\nHave a great weekend :) ", "So this seems to be taking quite a while. Is there anything I can do to help and/or expedite this process? Thx in advance. ", "I'm sorry about the delay! I've checked with everyone and we think it's okay, but there's an issue with ensuring this code stays in sync with the other BERT-based models. It's going slowly because I only started a couple of weeks ago, so I'm very paranoid about breaking things, and I'm double-checking things as I go.", "> I'm sorry about the delay! I've checked with everyone and we think it's okay, but there's an issue with ensuring this code stays in sync with the other BERT-based models. It's going slowly because I only started a couple of weeks ago, so I'm very paranoid about breaking things, and I'm double-checking things as I go.\r\n\r\nNo worries :) You should be moving fast and breaking things. That's what you have tests for ;) \r\nAlso why on earth are you guys doing all of this templating in the first place? It seems like a total maintenance nightmare and a textbook example of what not to do? You could cleanup and remove like 80% of your code for your MLMs with some standard object oriented programming? Or is there something I'm not seeing here?", "It's a good question! The underlying idea is that we want code to be self-contained and easy to separate from the rest of the library, so that users can work on the model they care about in isolation without needing to understand our whole hierarchy of abstractions and imports. \r\n\r\nIt's also helpful because we care a lot about supporting a variety of models that were trained outside of Hugging Face, which often involves reproducing their particular quirks rather than just importing the same single function in every case.", "Any updates? I would like to contribute in any way I can :) Other people in my organisation are starting to use HugginFace transformers at the same scale as me and will likely face the same issues as I did. \r\nIs there a branch I can follow?", "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.", "Was this ever fixed?", "No, but we're seeing other issues being caused by the same underlying problem, such as #12245 . I'm very aware of it, but finding a way to fix it without breaking changes to backward compatibility is difficult! It might be something that'll have to wait until a major release when we can break a lot of things at once." ]

[]

[]

[ "The error seems to come from your dataset, and you did not share the code you used to create and process it, so there is little we can do to help.", "Thanks for the quick reply. Here is the code I used to prepare the data:\r\n\r\n```\r\ntrain_data = datasets.load_dataset(\"cnn_dailymail\", \"3.0.0\", split=\"train\")\r\nval_data = datasets.load_dataset(\"cnn_dailymail\", \"3.0.0\", split=\"validation[:10%]\")\r\ntest_data = datasets.load_dataset(\"cnn_dailymail\", \"3.0.0\", split=\"test[:5%]\")\r\n\r\nencoder_max_length = 512\r\ndecoder_max_length = 128\r\nbatch_size = 4 # 16\r\n\r\ndef process_data_to_model_inputs(batch):\r\n inputs = tokenizer(batch[\"article\"], padding=\"max_length\", truncation=True, max_length=encoder_max_length)\r\n outputs = tokenizer(batch[\"highlights\"], padding=\"max_length\", truncation=True, max_length=decoder_max_length)\r\n\r\n batch[\"input_ids\"] = inputs.input_ids\r\n batch[\"attention_mask\"] = inputs.attention_mask\r\n batch[\"decoder_input_ids\"] = outputs.input_ids\r\n batch[\"decoder_attention_mask\"] = outputs.attention_mask\r\n batch[\"labels\"] = outputs.input_ids.copy()\r\n batch[\"labels\"] = [[-100 if token == tokenizer.pad_token_id else token for token in labels] for labels in batch[\"labels\"]]\r\n\r\n return batch\r\n\r\ntrain_data = train_data.shuffle()\r\n\r\ntrain_data = train_data.map(\r\n process_data_to_model_inputs, \r\n batched=True, \r\n batch_size=batch_size, \r\n remove_columns=[\"article\", \"highlights\"] # \"id\"\r\n)\r\n\r\ntrain_data.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\",\r\n \"attention_mask\",\r\n \"decoder_input_ids\",\r\n \"decoder_attention_mask\",\r\n \"labels\"]\r\n)\r\n\r\nval_data = val_data.shuffle()\r\n\r\nval_data = val_data.map(\r\n process_data_to_model_inputs, \r\n batched=True, \r\n remove_columns=[\"article\", \"highlights\"] # id\r\n)\r\n\r\nval_data.set_format(\r\n type=\"torch\",\r\n columns=[\"input_ids\",\r\n \"attention_mask\",\r\n \"decoder_input_ids\",\r\n \"decoder_attention_mask\",\r\n \"labels\"]\r\n)\r\n```\r\n\r\nThen I loaded the pre-trained models and set parameters as in the original notebook. If more information is needed, please let me know.", "The evaluation runs without any problem on my side, with the code you provided and the rest from the notebook you mentioned. Are you sure you have the latest version of the `Datasets` library installed?\r\nOtherwise, could you share a colab or a full script reproducing the error?", "Thank you! An update to the latest version of `Datasets` solved my problem." ]

[]

[]

[ "I am unable to reproduce: the notebook with your added code works smoothly on my side.", "Thank you @sgugger.\r\nI want ask you sorry, I can't figure out on what's going on my side.\r\nNow I have cloned again the notebook and the example works.\r\nIn the next days I want test it again and I will tell you more about it.\r\n\r\nThank you again for your help\r\nKind Regards,\r\nAndrea", "Hi @sgugger,\r\nI worked on the notebook and I found the problem.\r\nI have not yet had the opportunity to test it with the original squad dataset but this happens to me both on colab and on my machine.\r\nI warn you it seems an absurd and paradoxical situation, moreover I in no way manage the device.\r\nI can provide you with a video while running the notebook.\r\nAs you can see from the screenshot I am forced to keep two versions of the training args, one original from the notebook and one customized by me.\r\n\r\nIf I perform these operations I get the error\r\n1) I instantiate my training args\r\n2) I instantiate the Trainer\r\n3) I run trainer.fit\r\nI get the error `Input, output and indices must be on the current device`\r\n\r\nTo solve I have to:\r\nInstantiate the original training args of the notebook, instantiate the trainer, perform the fit to check that it has started and then do it all over again with the training args I customized.\r\n\r\n\r\n\r\n\r\nKind regards,\r\nAndrea", "Hi @sgugger,\r\nI can confirm, the same bug happens in the original notebook with this TrainingArguments (I have tested with squad v2), the temporary fix is to start the train with the original one, stop it and then run with the customized args. ", "It looks like a bug in colab (from the screenshots I assume that is what you are using for training?) since I didn't get any error on my side by executing this as a notebook.", "Hi @sgugger \r\nDo you have tested the notebook replacing the trainer args with the following?\r\n```python3\r\nargs = TrainingArguments(\r\n f\"my-experiment\",\r\n evaluation_strategy = \"epoch\",\r\n learning_rate=2e-5,\r\n per_device_train_batch_size=batch_size,\r\n per_device_eval_batch_size=250,\r\n num_train_epochs=2,\r\n weight_decay=0.01,\r\n fp16=True,\r\n gradient_accumulation_steps=2,\r\n eval_accumulation_steps=2,\r\n fp16_opt_level='O2',\r\n fp16_full_eval=True,\r\n save_strategy='epoch',\r\n metric_for_best_model='eval_loss',\r\n logging_strategy='epoch'\r\n)\r\n```\r\nBecause I encountered the same issue on my machine.\r\nCan you kindly test with it? Please\r\nTo test it: remove the old trainer args use the attached one and run the trainer.fit\r\n\r\nKind regards,\r\nAndrea", "Ah you're right, I must have made a mistake. This comes from the option `fp16_full_eval=True`.\r\n\r\n@stas00 I'm not sure what the best place is for fixing this but if someone uses `fp16_full_eval=True` with training, the model is never sent to the proper device and training fails.", "But there is no `do_train` in the args at https://github.com/huggingface/transformers/issues/11200#issuecomment-822566973\r\n\r\nThe logic is very explicit to not place on the device only for non-train when`fp16_full_eval=True` is used:\r\n\r\n```\r\n if (\r\n self.is_model_parallel\r\n or (args.deepspeed and args.do_train)\r\n or (args.fp16_full_eval and not args.do_train)\r\n or (self.sharded_ddp in [ShardedDDPOption.ZERO_DP_2, ShardedDDPOption.ZERO_DP_3])\r\n ):\r\n self.place_model_on_device = False\r\n```\r\n\r\nYou need to add `do_train=True` to your `TrainingArguments`, otherwise it defaults to eval only because you have `evaluation_strategy` set.\r\n", "Hi @stas00 & @sgugger,\r\n\r\n> You need to add `do_train=True` to your `TrainingArguments`, otherwise it defaults to eval only because you have `evaluation_strategy` set.\r\n\r\nOk so `do_train=True` is also compatible with `fp16_full_eval=True`? \r\nMy objective is to train the model and pick the best one at the lowest point of eval loss.\r\n\r\nRegarding the notebook, can I use the same Trainer object for fit and predict? Because these Booleans are never set in the notebook. I mean when I am doing trainer.predict() is obvious for the trainer to set model.eval() and torch.no_grad()?\r\n\r\nThank you both,\r\nAndrea", "> Ok so do_train=True is also compatible with fp16_full_eval=True?\r\n\r\nWhy did you think it shouldn't be compatible?\r\n\r\nThe only reason there is a special case for non-training is to avoid placing the full model on device before it was `half()`'ed - as it might not fit in its full size, but might fit in `half()`.\r\n\r\n> Regarding the notebook, can I use the same Trainer object for fit and predict? Because these Booleans are never set in the notebook. I mean when I am doing trainer.predict() is obvious for the trainer to set model.eval() and torch.no_grad()?\r\n\r\nOf course. It was designed for you to pass all the init args at once and then you can call all its functions.\r\n\r\n", "@stas00 Ok clear, I have just checked and the trainer works perfectly.\r\nWhat do you think to place a warning to alert the user when call trainer.fit having the trainer.do_train = False?\r\n\r\nBecause it's clear in the point of view of performance as you said but the documentation don't bring out this things for this reason I have open then issue.\r\n\r\nKind regards,\r\nAndrea", "Oh, I see. Until recently `do_train` was sort of optional when using user's custom code and what you're saying we need to then require `do_train=True` if `trainer.train()` is called. But we started relying on `do_train` for more than just knowing to call `train()` from scripts. This makes sense to me.\r\n\r\n@sgugger, do you agree if we add this?\r\n```\r\n def train(...):\r\n[...]\r\n if not self.args.do_train:\r\n raise ValueError(\"To use `train` please make sure you set `do_train=True` when instantiating the Trainer object\")\r\n```\r\n", "I would rather avoid adding this, as users have been used to not have to set that argument to True when not using example scripts. Can we just add the proper line in `train` to put the model on the device if it was not done already?\r\n\r\n(Sorry I didn't catch you were using `do_train` in the PR you added that test, I should have caught it and commented there.)", "We will probably have to rethink the design then, since it's not a simple \"put on device if it wasn't already\" - there are multiple cases when it shouldn't happen. For now added a hardcoded workaround: https://github.com/huggingface/transformers/pull/11322\r\n" ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "You will need to update to the last version of Transformers (I'm seeing 4.0.1 in your report), we fixed this issue so the evaluation loop uses the `compute_loss` function too.", "> You will need to update to the last version of Transformers (I'm seeing 4.0.1 in your report), we fixed this issue so the evaluation loop uses the `compute_loss` function too.\r\n\r\nThanks sgugger!\r\n\r\nI also found this out after taking a dive into your code base. I overwrote the prediction_step function in my case, since i dont know if the rest of my code supports transformers 4.5.0.\r\n\r\nYou can close the issue now! :)" ]

[]

[]

[ "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.", "Great job @lexhuismans !\r\n\r\nError is unrelated -> merging" ]

[]

[]

[ "Hey @rajvi-k, I believe there is just the styling issue left to fix before we can merge this. Just run `make style` on your branch!" ]

[ { "id": 1897896961, "node_id": "MDU6TGFiZWwxODk3ODk2OTYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Migration", "name": "Migration", "color": "e99695", "default": false, "description": "" } ]

[]

[ "#### Update: \r\nI changed the following line of code from \r\n`outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)` to \r\n `outputs = (context_layer,)` \r\nand my code seems to run fine with the latest transformers - https://github.com/gsrivas4/Oscar_latest/blob/latest_transformer/oscar/modeling/modeling_bert.py#L74-L75. However, I am still not sure if this change can break something in code logically.", "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." ]

[]

[]

[ "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!" ]

[]

[]

[]

[]

[]

[ "Actually that's not an issue, this warning shouldn't be here. I'll open a PR to remove it shortly.", "If you try generating text with it, you should get sensible results!", "Great to hear, thanks." ]

[]

[]

[ "I get `rate` printed out on my setup! Do you mind sharing your `pip list`?", "This was mentioned in #4857 and #5359 ....fiixed it after reinstalling. ", "For me, i didn't have sentencepiece installed. I needed to import torch before transformers to fix this" ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false }, { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "Hi, I think VQ only works for Pretraining, it doesn't look like Transformers currently support Pretrain", "Vector quantization is only required for pretraining which is currently not supported. It should be added soon: https://github.com/huggingface/transformers/issues/10873.", "Thanks! Just to clarify, was the [base model](https://huggingface.co/facebook/wav2vec2-base) pretrained with quantization, and it's just that the port to HF doesn't include the quantization module?", "The port didn't include the quantization module - we should re-port the model :-) ", "It was trained with quantization if I remember correctly", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[ "This 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." ]

[]

[]

[ "Which model are you using? What's the command you're using to launch the training?", "Hi\r\nI will close this issue and open up a proper reporting as I see the issue is arising from loading a checkpointin trainer class " ]

[]

[]

[]

[]

[]

[ "I met the same issues.", "This was fixed by #11207 and was released in patch 4.5.1. Please install the latest version and let us know if it works for you!", "It works in patch 4.5.1, thanks.\r\nI still wonder what is the difference between the following two methods to instantiate the BERT encoder-decoder model.\r\nProbably I should ask in another thread.\r\n\r\nmodel_name='bert-base-multilingual-cased'\r\nencoder = BertGenerationEncoder.from_pretrained(model_name, bos_token_id=bos_token_id, eos_token_id=eos_token_id)\r\ndecoder = BertGenerationDecoder.from_pretrained(model_name, add_cross_attention=True, is_decoder=True, bos_token_id=bos_token_id, eos_token_id=eos_token_id)\r\nbert2bert = EncoderDecoderModel(encoder=encoder, decoder=decoder)\r\n\r\nv.s.\r\n\r\nmodel_name='bert-base-multilingual-cased'\r\nbert2bert = EncoderDecoderModel.from_encoder_decoder_pretrained(model_name, model_name)\r\nbert2bert.config.decoder.decoder_start_token_id = bos_token_id\r\nbert2bert.config.encoder.bos_token_id = bos_token_id\r\nbert2bert.config.encoder.eos_token_id = eos_token_id\r\nbert2bert.config.encoder.pad_token_id = pad_token_id\r\n", "Hi @ken-arf \r\n\r\nBoth methods are doing the same thing.\r\n\r\nThe difference is that in the second method you don't need to initialize the encoder and decoder, you could just pass the name of the model two the `from_encoder_decoder_pretrained` method and takes care of initializing the encoder, decoder and adding cross_attention in the decoder etc.", "Hi Suraj\r\n\r\nThank you for your answer, I understand. I used both methods interchangeably, so that sounds good to me. " ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[ { "id": 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" } ]

[]

[ "Has anyone working on this yet?", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "If no one else beats me to it, I may tackle this mid-summer. No promises. ", "@kcarnold any update? \r\n\r\nThere is another recent paper on constrained decoding with complex constraints (fairseq only has positive constraints): [NeuroLogic Decoding: (Un)supervised Neural Text Generation with Predicate Logic Constraints\r\n](https://arxiv.org/abs/2010.12884)" ]

[]

[]

[ "Ah, I think this is an issue indeed, the device statement shouldn't be inside that `if` statement. Do you want to open a PR to fix this?", "Sure. ", "Fixed by #11292 " ]

[]

[]

[]

[]

[]

[ "from transformers import LongformerModel, LongformerTokenizer, RobertaTokenizer, AutoTokenizer\r\n \r\npretrain_model_path = 'schen/longformer-chinese-base-4096'\r\ntokenizer = LongformerTokenizer.from_pretrained(pretrain_model_path)\r\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "You could try the another one [run_language_modeling.py](https://github.com/Adapter-Hub/adapter-transformers/blob/master/examples/contrib/legacy/run_language_modeling.py) with `line-by-line` and add a `batch_size` along with `batched=True` to eliminate the memory error:\r\n\r\n```\r\n tokenized_datasets = datasets.map(\r\n ...\r\n batched=True,\r\n batch_size=200,\r\n ...\r\n )\r\n```", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "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.", "Works fine when using the model.generate function" ]

[]

[]

[ "I was able to get a fix working for ProphetNetForConditionalGeneration", "Hi @ManavR123 \r\n\r\n> I figured out the main cause of the error: apex converts BaseModelOutput objects into dictionaries, but a lot of the code functionality relies on receiving the former. I don't know if there is a way to avoid this.\r\n\r\nYou could pass `return_dict=False` to `forward` if you don't want the mode to return the output as model output classes, when `return_dict` is `False`, `tuple` is returned", "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." ]

[]

[]

[]

[]

[]

[ "@Rocketknight1 can you help in this", "@sgugger can you help in this", "Hi, we have a new, simpler and more robust text classification in TensorFlow script contributed by @Rocketknight1 [here](https://github.com/huggingface/transformers/blob/master/examples/tensorflow/text-classification/run_text_classification.py), could you check it out?", "Hi @LysandreJik and @Rocketknight1 \r\n\r\nI used it run_text_classification.py in jupyter\r\n%run run_text_classification.py \\\r\n--model_name_or_path roberta-base \\\r\n--output_dir classificationoutput \\\r\n--train_file PreparedData.csv \\\r\n--validation_file PreparedData.csv \\\r\n--do_train\r\n\r\n\r\nPreparedData.csv looks like below\r\nsentence,label\r\nsent1,l1\r\nsent2,l1\r\nsent3,l2\r\nsent3,l2\r\n\r\nI got following error\r\n\r\n---------------------------------------------------------------------------\r\nTypeError Traceback (most recent call last)\r\n~\\run_text_classification.py in <module>\r\n 532 \r\n 533 if __name__ == \"__main__\":\r\n--> 534 main()\r\n\r\n~\\run_text_classification.py in main()\r\n 492 \r\n 493 callbacks = [SavePretrainedCallback(output_dir=training_args.output_dir)]\r\n--> 494 model.fit(\r\n 495 training_dataset, validation_data=eval_dataset, epochs=training_args.num_train_epochs, callbacks=callbacks\r\n 496 )\r\n\r\nc:\\python38\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_batch_size, validation_freq, max_queue_size, workers, use_multiprocessing)\r\n 1086 self._maybe_load_initial_epoch_from_ckpt(initial_epoch))\r\n 1087 logs = None\r\n-> 1088 for epoch, iterator in data_handler.enumerate_epochs():\r\n 1089 self.reset_metrics()\r\n 1090 callbacks.on_epoch_begin(epoch)\r\n\r\nc:\\python38\\lib\\site-packages\\tensorflow\\python\\keras\\engine\\data_adapter.py in enumerate_epochs(self)\r\n 1132 with self._truncate_execution_to_epoch():\r\n 1133 data_iterator = iter(self._dataset)\r\n-> 1134 for epoch in range(self._initial_epoch, self._epochs):\r\n 1135 if self._insufficient_data: # Set by `catch_stop_iteration`.\r\n 1136 break\r\n\r\nTypeError: 'float' object cannot be interpreted as an integer\r\n\r\n​", "Good catch, thank you! This was totally my fault, and has now been fixed in #11379 . If you pull the latest version of the library, training should work.", "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." ]

[]

[]

[ "@patil-suraj , can you please help in suggesting how to finetune m2m100 on more than one-pair.I am able to finetune for one lang pair using below script:\r\n\r\nCUDA_VISIBLE_DEVICES=0,1,2,3,6 python -m torch.distributed.run --nproc_per_node=5 run_translation.py --model_name_or_path=m2m100_418M_new_token --do_train --do_eval --source_lang ja --target_lang en --fp16=True --evaluation_strategy epoch --output_dir bigfrall --per_device_train_batch_size=48 --per_device_eval_batch_size=48 --overwrite_output_dir --forced_bos_token \"en\" --train_file orig_manga/orig/train_exp_frame_50k.json --validation_file orig_manga/orig/valid_exp_frame_50k.json --tokenizer_name tokenizer_new_token --num_train_epochs 50 --save_total_limit=5 --save_strategy=epoch --load_best_model_at_end=True --predict_with_generate\r\n\r\nBut, now I want to finetune it on ja-en and ja-zh pairs. How to pass these both languages?", "Hi @nikhiljaiswal !\r\nIt would be nice if you ask this question on the [forum ](https://discuss.huggingface.co/). PR comments won't be a good place to discuss this. Thanks!" ]