Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[ "This looks good to me! I think it's a good idea to add `--source_prefix` for the 5 t5 checkpoints in the examples. We should specify though that it's only for these 5 checkpoints. Let's discuss summarization in the issue.", "As discussed in https://github.com/huggingface/transformers/issues/10733#issuecomment-800123545 updated the summarization example to use cnn_dailymail - now the prefix works for t5! Thank you, @patrickvonplaten ", "Sorry for disturbing you, but has the support for MBart been removed? This model needs source_lang and target_lang arguments, but the scripts don't accept them now.", "They are sill [here](https://github.com/huggingface/transformers/blob/fd1d9f1ab89805fb2a8e773edbc27531b449ddea/examples/seq2seq/run_translation.py#L96).", "Thank you very much! I can copy that part into summarization file.", "Sorry to disturb you again. Even after the copy of the MBart part code, the error occurred.\r\n```py\r\nTraceback (most recent call last):\r\n File \"examples/seq2seq/run_summarization.py\", line 609, in <module>\r\n main()\r\n File \"examples/seq2seq/run_summarization.py\", line 443, in main\r\n train_dataset = train_dataset.map(\r\n File \"/home/zchelllo/anaconda3/envs/ex/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 1407, in map\r\n update_data = does_function_return_dict(test_inputs, test_indices)\r\n File \"/home/zchelllo/anaconda3/envs/ex/lib/python3.8/site-packages/datasets/arrow_dataset.py\", line 1378, in does_function_return_dict\r\n function(*fn_args, indices, **fn_kwargs) if with_indices else function(*fn_args, **fn_kwargs)\r\n File \"examples/seq2seq/run_summarization.py\", line 424, in preprocess_function\r\n with tokenizer.as_target_tokenizer():\r\n File \"/home/zchelllo/anaconda3/envs/ex/lib/python3.8/contextlib.py\", line 113, in __enter__\r\n return next(self.gen)\r\n File \"/home/zchelllo/anaconda3/envs/ex/lib/python3.8/site-packages/transformers/models/mbart/tokenization_mbart_fast.py\", line 214, in as_target_tokenizer\r\n self.set_tgt_lang_special_tokens(self.tgt_lang)\r\n File \"/home/zchelllo/anaconda3/envs/ex/lib/python3.8/site-packages/transformers/models/mbart/tokenization_mbart_fast.py\", line 240, in set_tgt_lang_special_tokens\r\n suffix_tokens_str = self.convert_ids_to_tokens(self.suffix_tokens)\r\n File \"/home/zchelllo/anaconda3/envs/ex/lib/python3.8/site-packages/transformers/tokenization_utils_fast.py\", line 286, in convert_ids_to_tokens\r\n index = int(index)\r\nTypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'\r\n```\r\n\r\nThe command is \r\n```sh\r\nCUDA_VISIBLE_DEVICES=0 python examples/seq2seq/run_summarization.py --model_name_or_path facebook/mbart-large-cc25 \\\r\n --do_train --do_predict --train_file ../data/lang8_ja_train.csv --test_file ../data/lang8_ja_test.csv \\\r\n--output_dir ../pre_models/mbart_lang8_ja \\\r\n--per_device_train_batch_size=4 --per_device_eval_batch_size=4 \\\r\n--predict_with_generate --text_column errorful_sent --summary_column correct_sent \\\r\n--save_steps=2000 --save_total_limit=3 --overwrite_output_dir \\\r\n--source_lang ja_XX --target_lang ja_XX\r\n```\r\n\r\nAnd the possible reason is that even I input the source and target language, these arguments haven't been sent into prepare_seq2seq_batch function in line 195 tokenization_mbart_fast.py. The `self.src_lang` is `en_XX` and `self.tgt_lang` is `None`.", "Sorry, I found that I missed this part. After copying this part, it works well! Thank you very much!\r\n```py\r\n # For translation we set the codes of our source and target languages (only useful for mBART, the others will\r\n# ignore those attributes).\r\nif isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):\r\n if data_args.source_lang is not None:\r\n tokenizer.src_lang = data_args.source_lang\r\n if data_args.target_lang is not None:\r\n tokenizer.tgt_lang = data_args.target_lang\r\n```" ]

[]

[]

[ "`--max_train_samples 50` is a tiny sample, I am not surprised that the model doesn't learn anything here. Note that `t5-small` was pretrained on CNN/Dailymail and WMT, but **not** on XSum. So, it makes sense that one gets reasonable results when fine-tuning the model on just 50 translation samples of WMT16 because the model has already seen the whole training data in pretraining. However, the model has never seen XSum in pretraining, so fine-tuning on 50 samples will get us nowhere here I think. We could try to switch to `CNN/Dailymail`. I have fine-tuned the model on the whole corpus for CNN/Dailymail and have gotten good results. In the paper, it was reported that with `t5-small` a ROUGE-2 score of 19.56 can be achieved on CNN/Dailymail. So we should get something like 17 or 18 ROUGE-2 for full fine-tuning. \r\n\r\nAlso, IMO for such low ROUGE number we cannot really say that \"no prefix\" works better than \"with prefix\" because both cases don't work well at all. \r\n\r\nLet's just try it with CNN/Dailymail instead and see what we get. Maybe first with just very few samples & if this doesn't work then let's run one full fine-tuning.", "Definitely a jackpot on the example using a new dataset and too short of training: might be a good idea to add some of your notes to the README as well.\r\n\r\nNew stats, this time on `--dataset_name cnn_dailymail --dataset_config \"3.0.0\"`\r\n\r\n```\r\n# w/o --source_prefix \"summarize: \"\r\n\r\npython examples/seq2seq/run_summarization.py --model_name_or_path t5-small --do_train --do_eval --dataset_name cnn_dailymail --dataset_config \"3.0.0\" --output_dir /tmp/tst-summarization --per_device_train_batch_size=4 --per_device_eval_batch_size=4 --overwrite_output_dir --predict_with_generate --max_train_samples 50 --max_val_samples 50\r\n\r\n***** eval metrics *****\r\n epoch = 3.0\r\n eval_gen_len = 65.24\r\n eval_loss = 2.3276\r\n eval_rouge1 = 25.4707\r\n eval_rouge2 = 7.2334\r\n eval_rougeL = 18.3807\r\n eval_rougeLsum = 23.2505\r\n eval_runtime = 6.4841\r\n eval_samples = 50\r\n eval_samples_per_second = 7.711\r\n\r\n```\r\n\r\n```\r\n# w/ --source_prefix \"summarize: \"\r\n\r\npython examples/seq2seq/run_summarization.py --model_name_or_path t5-small --do_train --do_eval --dataset_name cnn_dailymail --dataset_config \"3.0.0\" --output_dir /tmp/tst-summarization --per_device_train_batch_size=4 --per_device_eval_batch_size=4 --overwrite_output_dir --predict_with_generate --max_train_samples 50 --max_val_samples 50 --source_prefix \"summarize: \"\r\n\r\n***** eval metrics *****\r\n epoch = 3.0\r\n eval_gen_len = 62.38\r\n eval_loss = 2.3243\r\n eval_rouge1 = 30.0675\r\n eval_rouge2 = 10.2052\r\n eval_rougeL = 22.154\r\n eval_rougeLsum = 27.2161\r\n eval_runtime = 6.0876\r\n eval_samples = 50\r\n eval_samples_per_second = 8.213\r\n```\r\n\r\nThis is much better. I will try a longer train sequence next.", "Pretrained with 5000 samples the score goes up nicely, this is 1/100th of the full dataset.\r\n\r\n```\r\n***** eval metrics *****\r\n epoch = 3.0\r\n eval_gen_len = 61.66\r\n eval_loss = 2.0773\r\n eval_rouge1 = 30.174\r\n eval_rouge2 = 12.0182\r\n eval_rougeL = 23.5012\r\n eval_rougeLsum = 27.3718\r\n eval_runtime = 5.9836\r\n eval_samples = 50\r\n eval_samples_per_second = 8.356\r\n```\r\n\r\nSo I updated https://github.com/huggingface/transformers/pull/10734 with the recommendation you made @patrickvonplaten.\r\n\r\nClosing this." ]

[]

[]

[ "Indeed! Would you mind making a PR with that change since you found the correct fix?", "Sure, I'll get that prepared!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "Yes, that sounds clearer. Would you mind making a PR with this?", "Will make time this week. ", "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." ]

[]

[]

[]

[]

[]

[ "cc @sgugger ", "I'm unsure what you think is not supported. Launching any of the example scripts with \r\n```\r\npython -m torch.distributed.launch --nproc_per_node=xxx \\\r\n --node_rank=$THIS_MACHINE_INDEX \\\r\n --master_addr=\"192.168.1.1\" \\\r\n --master_port=1234 \\\r\n run_xxx.py\r\n```\r\nis going to work.", "Hi @sgugger , thank you for your answer. My understanding is that \"--nproc_per_node\" is the number of gpus will be used for the launched process? Also, if I want to launch another training node, I assume I will just run the same command with \"node_rank=1\"?", "Yes, that is the number of GPUs.\r\n\r\nYou can refer to the [PyTorch documentation](https://pytorch.org/docs/stable/distributed.html#launch-utility) for all the arguments of the PyTorch launcher as all the example scripts are fully compatible with it. You will also need to pass `--nnodes=$NUMBER_OF_NODES` for completeness.", "Thanks, I am able to make it work now." ]

[]

[]

[ "For broken models you would just tag the model author(s) – we'll add a feature to the model hub to tag someone in a conversation thread, but in the meantime you can use the Forum to ping them", "That's a great idea to tag the model author - how would a user know the model author's corresponding forum username? I guess this is temporary so probably can be somehow figured out...", "We use SSO so one's username on Forum is guaranteed to be one's username on hf.co", "Ah, that's perfect then! Thank you!" ]

[]

[]

[]

[]

[]

[ "I fixed this model (it had `bart` as a `model_type` instead of `mbart` in `config.json`). \r\n\r\nBut the point of this issue is that perhaps we could add to a todo list to run a cronjob that will validate the models and tokenizers? \r\n\r\n@theo-m, is this something that would fit with the project you're currently working on, since you will have to run these things anyway for each model? Just asking - not piling it up on you.\r\n\r\nIn fact if I understand the idea correctly this will be a requirement for your things to work, right?", "We would not be aiming for 100% coverage, but yes, getting a sense of what's runnable on the hub would be awesome. \r\n\r\nMaybe a once-a-week CI job, as I expect the run to be terribly long/expensive? cc infra people @julien-c @n1t0 ", "I think we'll hook something into the git-push-event driven ML analytics system we've been talking about internally\r\n\r\nThat's a medium-term goal though", "It's a different thing though: on-push would ensure integrity on upload, which is indeed needed, but a recurrent job would enable us to detect regression in what the lib can support and give estimates on what is actually runnable.", "The lib can change after the upload was made and the model/tokenizer stop working. We have seen this before with older models.\r\n\r\nI think @theo-m you're saying the same thing. ", "@theo-m, btw the low hanging fruit would be to just validate that the listed in \"use in transformers\" instructions indeed work. i.e. we just load the model and tokenizer and do nothing with it if it works and do something with it if it doesn't. ", "Note that this is not necessarily a low hanging fruit (depending on your definition of a low hanging fruit 😂) given that:\r\n- we have 7,000+ models whose total weights represent multiple TBs of data\r\n- they change over time", "the lowest hanging fruit is loading all that can be associated to a pipeline and run a single example in the associated pipeline, the results of this are stronger than just loading\r\n\r\nand yes it sure is a big big job, but it's the best we can do in order to build a good understanding of what is runnable on the hub - _in fine_ for non hf we won't be able to do much, but we can't give guarantees to code we don't manage.", "I meant that just loading a model / tokenizer is cheaper/faster/requires almost 0 extra code to write - hence low-hanging fruit.\r\n\r\nI hear you that the hub is huge, a little bit at a time. It would have been the same code to validate 10 models or 7K models if there is no urgency to complete it fast, it just would take much much longer to complete.\r\n\r\n> * they change over time\r\n\r\nThat was exactly my point, they and the codebase too, so it's not enough to check it once, even if we track when it was changed and when it was validated last.\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." ]

[ { "id": 2934977194, "node_id": "MDU6TGFiZWwyOTM0OTc3MTk0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Flax", "name": "Flax", "color": "4862AD", "default": false, "description": "" } ]

[]

[]

[]

[]

[]

[]

[]

[ "HuggingFace has another library called [tokenizers](https://github.com/huggingface/tokenizers) especially for this.", "Currently, the training of Deberta Tokenizer is not supported directly by huggingface. Of course, you can create the required files by yourself from BPETokenizer training output, but you could also simply wait until #10703 is merged into the master branch and released. :-)", "How would be the process of creating the required files from the BPETokenizer training output? @cronoik I'd really appreciate a little bit of explanation, as I tried to do so and I failed.", "You can save me a lot of time by simply using the mentioned patch above. Just copy the DebertaTokenizer class to your runtime.", "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 would result in an inaccurate metric in most cases, as metric functions are seldom linear. I'm afraid you will have to run evaluation on smaller chunks or use your own evaluation loop.", "Yes of course I thought maybe to gather a reduced version of the predictions before actually computing the metric\r\n\r\ne.g. for sentence level accuracy, in my example, a bool array of shape `(200206, )` where the boolean value represents the accuracy of the output (i.e. `predictions == labels`). \r\n\r\nThe actual `compute_metrics` would only have to reduce this array to a single value (using `np.mean` in my example).", "You can do that using `Trainer` if your model returns that. `Trainer` is too generic to be able to guess that in this case it should gather a reduced version of the predictions (and how would it do it?). Otherwise writing the evaluation loop yourself is super easy (there is one example in [run_glue_no_trainer](https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_glue_no_trainer.py) for instance).", "Ok, thanks for the advice :)" ]

[]

[]

[ "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.", "Kindly help in resolving the issue.\r\nIt will help to build a seq2seq conversation model using pretrained bert model.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Maybe of interest to @patrickvonplaten and @patil-suraj ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "That example only runs with `BERT`, which is why it has been moved to a separate research project.", "I tried this script with albert and it worked, which script should I use to train a Roberta model from scratch with Whole word Masking??", "Is that intended:\r\n`--model_type deberta`\r\n?\r\n@alexvaca0 ", "Sorry, that was from the previous launch script, now it is roberta @cronoik ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "The fixup target fails because `get_modified_files.py` reports `tokenization_layoutlm_fast.py` as differing from the master branch.\r\nThen calling e.g. `black` to format that file fails because it doesn't exist anymore. How would I fix that?", "Hello! Thank you for your PR, and sorry for getting back to you so late. First of all, incredible work on doing the implementation and on overriding the tests that should be.\r\n\r\nI wonder if the approach here shouldn't make use of the recently introduced feature processors. @NielsRogge you have had extensive experience with feature processors and you were part of the initial conversation, what do you think would be the best approach here? \r\n\r\nIt's a bit different to handling images as it's simply handling the bbox, so I might be wrong here.\r\n\r\nAs a high level overview, I'm not keen on removing the fast tokenizer, and I'm wondering if we really need to accept two sequences or if LayoutLM is only made for single sequences - I haven't played with the model, so please tell me if I'm mistaken.\r\n\r\nAlso cc @sgugger and @patil-suraj who might have some insights.", "Yes I think you should look at the design used for `SpeechToText` or `Wav2Vec2`: there is a processor that combines a tokenizer and a feature extractor in those models. We should do the same here: leave the tokenizer unchanged and add a feature extractor to treat the bounding boxes separately, then merge the two in a `LayoutLMProcessor`.", "Thank you for the input, I wasn't aware of feature processors. It sounds like this could be a way nicer solution here, I agree.\r\n\r\n> I'm wondering if we really need to accept two sequences or if LayoutLM is only made for single sequences\r\n\r\nThe basic difference between LayoutLM and BERT is that the additional `bbox` input is [added to the embeddings](https://github.com/huggingface/transformers/blob/master/src/transformers/models/layoutlm/modeling_layoutlm.py#L104-L126).\r\n\r\nSo the two sequences are processed slightly different inside the model. However, there's a one to one relationship between their items. That's also why the processing of the `bbox` sequence depends on how the tokenizer splits the input. In case of a split into N sub-tokens the corresponding bounding box is repeated N times to retain the one to one relationship.\r\nWill this be possible with a feature processor? From a first glance I'm not too sure but I might be wrong. Can someone clarify?\r\n\r\n> As a high level overview, I'm not keen on removing the fast tokenizer\r\n\r\nI removed the fast tokenizer to first discuss if this is a suitable approach before investing more time. Eventually, I planned to add support for it as well.", "> However, there's a one to one relationship between their items. That's also why the processing of the bbox sequence depends on how the tokenizer splits the input. In case of a split into N sub-tokens the corresponding bounding box is repeated N times to retain the one to one relationship.\r\n\r\nIn the fast tokenizer, you can rely on the `word_ids` method of the `BatchEncoding` (the type of the return of the tokenizer) to get back the word associated to each token. For the slow tokenizer you may have to compute it. \r\n\r\nThe workflow I see is: the tokenizer returns a `BatchEnconding` with `input_ids`, `attention_mask` etc (like a usual tokenzier) and a field containing the mapping token to word, then the processor will extract that field form the batch encoding and pass it to the feature extractor responsible for the bounding boxes, so the proper repetition can happen. This way we still get a nice separation for the two modalities in two different objects.", "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 is not a priority for me right now and correctly marked as stale. I didn't forget about it, though and hope to be able to come back to it in a near future." ]

[]

[]

[]

[]

[]

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

[]

[]

[ "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\nYou can ping @patrickvonplaten on the forum as he's the best suited to help you out.\r\n\r\nThanks!", "I moved the question to the forum and @patrickvonplaten said that the feature is not supported for now, but will be soon. \r\nhttps://discuss.huggingface.co/t/language-model-for-wav2vec2-0-decoding/4434", "I'm now working on this topic full time. \r\n\r\nWe will most likely foster a closer collaboration between [pyctcdecode](https://github.com/kensho-technologies/pyctcdecode) and Transformers. [Here](https://github.com/patrickvonplaten/Wav2Vec2_PyCTCDecode) is a github repo that shows how to use `pyctcdecode` with Wav2Vec2 for LM supported decoding. It works quite well with KenLM." ]

[]

[]

[ "`pegasus-large` is a pre-trained checkpoint. It's not fine-tuned on downstream task. The finetuned checkpoint name will have the name of the datsets in them. \r\n\r\nAlso please the [forum](https://discuss.huggingface.co/) to ask such 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." ]

[]

[]

[]

[]

[]

[ "As mentioned in the [docs](https://huggingface.co/transformers/main_classes/pipelines.html), you can either provide a model identifier from the hub or a model that inherits from `PreTrainedModel` or `TFPreTrainedModel`. \r\n\r\nSo suppose you have fine-tuned a model and stored it into a variable called `model`, then you can initialize a corresponding pipeline by providing this model variable at initialization. Make sure that the model you're providing is suitable for the pipeline. \r\n\r\nSo suppose that you want to use the question answering pipeline, and you have a local `xxxForQuestionAnswering` model, then you can provide it as follows:\r\n\r\n```\r\nfrom transformers import pipeline\r\n\r\nmodel = ...\r\nnlp = pipeline (task='question-answering', model=model)\r\n```\r\n\r\n", "> docs\r\n\r\nThanks a lot." ]

[]

[]

[ "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.\r\n> \r\n> Please note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.\r\n\r\nThis issue causes incorrect parameter transfer between PyTorch and TensorFlow when the LSTM neural networks are used in code.", "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." ]

[]

[]

[ "Btw ,these 3 APIs show similar error:\r\n**API:**\r\n```\r\n trainer.log_metrics(\"train\", metrics)\r\n trainer.save_metrics(\"train\", metrics)\r\n trainer.save_state()\r\n```\r\n\r\n**error**\r\n```\r\nTraceback (most recent call last):\r\n File \"run_mlm.py\", line 476, in <module>\r\n main()\r\n File \"run_mlm.py\", line 452, in main\r\n trainer.save_state()\r\nAttributeError: 'Trainer' object has no attribute 'save_state'\r\n```", "See #10446 ", "Indeed. Please search the issues before opening one that is an exact duplicate of an existing one (see the link above for a resolution of your problem)." ]

[]

[]

[ "Thanks!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "As indicated in the [documentation](https://huggingface.co/transformers/main_classes/trainer.html#trainingarguments) `metric_for_best_model` _Must be the name of a metric returned by the evaluation with or without the prefix \"eval\\_\"_. You passed a Metric object to it instead of the name returned by your `compute_metric` function, which is what caused your error.", "Thank you so much, I'm sorry for that" ]

[]

[]

[ "We are organizing a \"fine-tuning XLSR-53\" event. Check this announcement: https://discuss.huggingface.co/t/open-to-the-community-xlsr-wav2vec2-fine-tuning-week-for-low-resource-languages/4467. Would be awesome if you want to participate :-)", "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." ]

[]

[]

[ "Looking at the [source code](https://github.com/huggingface/transformers/blob/4c32f9f26e6a84f0d9843fec8757e6ce640bb44e/src/transformers/pipelines/text_generation.py#L108) of the text-generation pipeline, it seems that the texts are indeed generated one by one, so it's not ideal for batch generation. \r\n\r\nIn order to genere contents in a batch, you'll have to use GPT-2 (or another generation model from the hub) directly, like so (this is based on PR #7552):\r\n\r\n```\r\nfrom transformers import GPT2Tokenizer, GPT2LMHeadModel\r\nimport torch\r\n\r\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\r\ntokenizer.padding_side = \"left\" \r\ntokenizer.pad_token = tokenizer.eos_token # to avoid an error\r\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\r\n\r\ndevice = 'cuda' if torch.cuda.is_available() else 'cpu'\r\n\r\ntexts = [\"this is a first prompt\", \"this is a second prompt\"]\r\nencoding = tokenizer(texts, padding=True, return_tensors='pt').to(device)\r\nwith torch.no_grad():\r\n generated_ids = model.generate(**encoding)\r\ngenerated_texts = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)\r\n```\r\n\r\nIn my case, this prints out: \r\n`['this is a first prompt for the user to enter the password.\\n\\nThe password is a string', \"this is a second prompt, but it's not a full-screen one.\\n\\nThe first\"]`", "> Looking at the [source code](https://github.com/huggingface/transformers/blob/4c32f9f26e6a84f0d9843fec8757e6ce640bb44e/src/transformers/pipelines/text_generation.py#L108) of the text-generation pipeline, it seems that the texts are indeed generated one by one, so it's not ideal for batch generation.\r\n> \r\n> In order to genere contents in a batch, you'll have to use GPT-2 (or another generation model from the hub) directly, like so (this is based on PR #7552):\r\n> \r\n> ```\r\n> from transformers import GPT2Tokenizer, GPT2LMHeadModel\r\n> import torch\r\n> \r\n> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')\r\n> tokenizer.padding_side = \"left\" \r\n> tokenizer.pad_token = tokenizer.eos_token # to avoid an error\r\n> model = GPT2LMHeadModel.from_pretrained('gpt2')\r\n> \r\n> device = 'cuda' if torch.cuda.is_available() else 'cpu'\r\n> \r\n> texts = [\"this is a first prompt\", \"this is a second prompt\"]\r\n> encoding = tokenizer(texts, return_tensors='pt').to(device)\r\n> with torch.no_grad():\r\n> generated_ids = model.generate(**encoding)\r\n> generated_texts = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)\r\n> ```\r\n> \r\n> In my case, this prints out:\r\n> `['this is a first prompt for the user to enter the password.\\n\\nThe password is a string', \"this is a second prompt, but it's not a full-screen one.\\n\\nThe first\"]`\r\n\r\nthanks. It seems that the standard workflow is to organize the components of `tokenizer`, `generate` and `batch_decode`\r\nin a cascade way. ", "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.", "@NielsRogge Is padding on the left still the way to go for batched generation? It seems odd to require a workaround for such a common feature" ]

[]

[]

[ "This is great @cronoik! I just tested it and reproduced identical results between the previous version and this one. Fantastic!\r\n\r\nCould you add an integration tests for the DeBERTa tokenizer to ensure the implementations don't diverge? You can just copy paste the [test for ALBERT tokenizers](https://github.com/huggingface/transformers/blob/master/tests/test_tokenization_albert.py#L106) and change the values to DeBERTa.\r\n\r\nIf you don't have time, let us know and we'll take care of it.", "@LysandreJik \r\nI will give it a try.", "Hi,\r\na new test is now added for the debertatokenizer. I stuck more to the robertatokenizer test since the debertatokenizer is basically a robertatokenizer.\r\n\r\nI can not interpret the failed check. I assume it is because the model hub is missing the new files? @LysandreJik could you please have a look?", "I am trying out this PR with the following:\r\n\r\n```\r\ntokenizer = DebertaTokenizer.from_pretrained(\"microsoft/deberta-base\")\r\ntarget_tokenized = tokenizer.tokenize(\"Some test text\")\r\n```\r\nbut I see the following error (`TypeError: expected str, bytes or os.PathLike object, not NoneType`):\r\n```\r\n def __init__(\r\n self,\r\n vocab_file,\r\n merges_file,\r\n errors=\"replace\",\r\n unk_token=\"<|endoftext|>\",\r\n bos_token=\"<|endoftext|>\",\r\n eos_token=\"<|endoftext|>\",\r\n add_prefix_space=False,\r\n **kwargs\r\n ):\r\n bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token\r\n eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token\r\n unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token\r\n super().__init__(\r\n errors=errors,\r\n unk_token=unk_token,\r\n bos_token=bos_token,\r\n eos_token=eos_token,\r\n add_prefix_space=add_prefix_space,\r\n **kwargs,\r\n )\r\n\r\n> with open(vocab_file, encoding=\"utf-8\") as vocab_handle:\r\nE TypeError: expected str, bytes or os.PathLike object, not NoneType\r\n\r\n../../../../jeswan_transformers/src/transformers/models/gpt2/tokenization_gpt2.py:179: TypeError\r\n```\r\n@LysandreJik potentially related to @cronoik's error above ", "@jeswan that will not work since the required files are not uploaded to the model hub yet. You can download them from the [link](https://drive.google.com/drive/folders/1gH5EMABR94iHO7SCb_AdNGCOOIdSloxh?usp=sharing) and load the tokenizer from local.", "@cronoik thanks for your effort. I just uploaded the files to the model repository and left one comment to the changes. ", "@BigBird01 Thank you for the review. I had only tested single sentences and completely ignored the sentence pairs before pushing.\r\n@LysandreJik Can you please help me with the test error? ", "@sgugger Thanks for the review. I have pushed/accepted your change requests.", "Thanks for your efforts @cronoik!" ]

[]

[]

[ "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!" ]

[]

[ { "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 @Ninja16180 \r\n\r\nCould you please post a short code snippet to reproduce the issue? Thanks.", "Hi Suraj,\r\n\r\nI was able to find out the issue; there was a variable name I wrongly passed into the decoder class and hence the error.\r\n\r\nCorrection made:\r\n`embedded = self.bert(sent2)[0] should be embedded = self.bert(input)[0] `\r\n\r\nThus closing this issue." ]

[]

[]

[ "Thanks for your interest! LUKE is not part of the master branch yet.\r\n\r\nActually, the current implementation of LUKE is here (at my `adding_luke_v2` branch): https://github.com/NielsRogge/transformers/tree/adding_luke_v2/src/transformers/models/luke\r\n\r\nNote that it is work-in-progress, but you can already use the base `EntityAwareAttentionModel` and the head models. It's mostly the tokenizer that needs some work.\r\n\r\ncc'ing the original author for visibility: @ikuyamada ", "Thanks, Niels!\r\n\r\nAs far as I'm concerned, this can be closed." ]

[]

[]

[ "Currently, there are some Torch/TF equivalent tests failing but it should be settled down once #10605 be merged.", "@jplu - I'm gonna run all slow tests today and will let you know if everything works or not.", "@jplu - I ran all (slow) non-GPU tests and it seems to me everything is passing :) ", "Ok, if all the tests for the involved models, including the slow ones, are passing, it is fine to merge for 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.", "still open ", "@jplu @patrickvonplaten After merging #10605 and rebasing the branch, all tests have passed now :)", "Great, thanks @stancld! \r\n\r\nThis looks good to merge for me.", "Pinging @Rocketknight1 for a final review here. Seq2Seq models like BART should also return cross attentions in TF (not just in PT)", "This is a big PR but LGTM! I haven't exhaustively checked everything but the changes seem correct and innocuous, so if it passes tests I'm happy to merge 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." ]

[]

[]

[]

[]

[]

[ "The model doesn't contains the tokenizer and preprocessing files.\r\nCheckout this notebook:\r\nhttps://huggingface.co/blog/fine-tune-xlsr-wav2vec2\r\n\r\nTo build your own vocab etc", "@flozi00 Thank you! \r\n\r\n> The model doesn't contains the tokenizer and preprocessing files.\r\n> Checkout this notebook:\r\n> https://huggingface.co/blog/fine-tune-xlsr-wav2vec2\r\n> \r\n> To build your own vocab etc\r\n\r\n", "Forget about `transformers.AutoProcessor`. This class is used to load a general `processor` model. By call the constructor of this class you have to submit `feature_extractor` and `tokenizer`, however `Wav2Vec2` just extract the features from raw speech data. Then, there is no `tokenizer` has been defined for it. To load the `processor` you can use `transformers.Wav2Vec2FeatureExtractor` as follow:\r\n\r\n```\r\nfrom transformers import Wav2Vec2FeatureExtractor\r\n\r\nprocessor = Wav2Vec2FeatureExtractor.from_pretrained('facebook/wav2vec2-large-xlsr-53')\r\n```", "> Forget about `transformers.AutoProcessor`. This class is used to load a general `processor` model. By call the constructor of this class you have to submit `feature_extractor` and `tokenizer`, however `Wav2Vec2` just extract the features from raw speech data. Then, there is no `tokenizer` has been defined for it. To load the `processor` you can use `transformers.Wav2Vec2FeatureExtractor` as follow:\r\n> \r\n> ```\r\n> from transformers import Wav2Vec2FeatureExtractor\r\n> \r\n> processor = Wav2Vec2FeatureExtractor.from_pretrained('facebook/wav2vec2-large-xlsr-53')\r\n> ```\r\n\r\nUsing this approach, I got segmentation fault on the same wav2vec2-large-xlsr-53 model.\r\nOutput:\r\n```bash\r\nSome weights of the model checkpoint at facebook/wav2vec2-large-xlsr-53 were not used when initializing Wav2Vec2Model: ['quantizer.weight_proj.bias', 'project_q.weight', 'project_q.bias', 'quantizer.codevectors', 'quantizer.weight_proj.weight', 'project_hid.bias', 'project_hid.weight']\r\n- This IS expected if you are initializing Wav2Vec2Model from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing Wav2Vec2Model from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nProcessing 24393.wav\r\nSegmentation fault (core dumped)\r\n```" ]

[]

[]

[]

[]

[]

[ "Great! Thanks a lot for correcting those :-)\r\n", "Hey @MikeG112, sorry could you run `make style` once to fix the code quality issue? Then we can merge :-)", "Hey @patrickvonplaten, absolutely, I added the changes made by `make style`. Thanks for the review and the wav2vec2 implementation :)" ]

[]

[ { "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 } ]

[ "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 Was this resolved? If so I'll close this issue", "HI @xhlulu \r\n\r\nI don't think if this is an issue with the model implementation, also the model is not really expected to do well on paraphrasing (which English-English), I've seen few other issues where models output text in the wrong language but it's the same with original model in `fairseq` as well. From my experience, multilingual models tend to do this in few cases.", "Thanks, that makes sense! Glad you clarified it 😊" ]

[ { "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" } ]

[]

[ "@LysandreJik I've been mostly following: https://huggingface.co/transformers/add_new_model.html so far.\r\nTrying to add the tokenizer. I think I am good for the slow one but not sure what to do for the fast one, in particular how to generate the tokenizer.json style files (e.g: https://huggingface.co/t5-small/resolve/main/tokenizer.json ). Do you have any pointers to that?\r\n\r\nI also see that the doc mentions that there is no fast version for sentencepiece, which this model uses. Is that the case given T5 seems to have one?\r\n\r\nEdit: Seem to have found a way to add a FastTokenizer version, doc still seems out of sync.", "For the TF code, I'm struggling a bit to initialize an output embedding layer in `TFRemBertLMPredictionHead()` that interacts well with `_resize_token_embeddings`. Welcoming any suggestions as what the right approach there is.", "@LysandreJik : They are still some things to iron out but I think this is ready for a first look.\r\n\r\nWhat is missing:\r\n- Model hub upload \r\n- Minor discrepancy between model and original tf implementation\r\n\r\n\r\nWhat I'd like some input on:\r\n- I'm having some issue on the `TFRemBertLMPredictionHead` implementation. I'd like to initialize a new projection from hidden_size to vocab_size (embeddings are decoupled) but I'm struggling to find how to make my implementation compatible with all the `get_bias`, `set_bias` details so that it's `resize_embeddings` friendly. Any chance you could help here? This is the culprit for tests failing AFAICT.\r\n- Model hub upload: should this be done on top level (how) or on the Google org model hub?\r\n- I'm finding a discrepancy between this implementation and the original tf one. Results are exactly equal up to the first hidden layer (so embeddings and upprojection). On the first layer it differs but by small amounts (~0.002), difference eventually increases up to 0.007. Any idea what are common culprits here? This is just the standard BERT model and differences are small so maybe numerical stability?\r\n", "> Model hub upload: should this be done on top level (how) or on the Google org model hub?\r\n\r\nToplevel models were for legacy (historical) integrations and we now namespace all models. If this work was conducted at Google yes google is the right namespace! Do you want us to add you to the `google` org?", "> A difference of *e-3 doesn't look too bad, but looking at the integration test you have provided, it seems that the difference is noticeable. Is it possible that a bias is missing, or something to do with attention masks?\r\n\r\nNot impossible but given the transformer section is simply Bert I doubt it. Also does seem like the results would change more.\r\n \r\n> If it proves impossible to get the two implementations closer to each other, then we'll rely on a fine-tuning tests: if we can obtain similar results on a same dataset with the two implementations, then we'll be good to go.\r\n\r\nI've tried to do that for a bit, unfortunately hard to fine-tune this model on a colab on XNLI (training gets interrupted too early on). Will try to see if I can get a better finetuning setup.\r\n\r\n\r\n", "> > Model hub upload: should this be done on top level (how) or on the Google org model hub?\r\n> \r\n> Toplevel models were for legacy (historical) integrations and we now namespace all models. If this work was conducted at Google yes google is the right namespace! Do you want us to add you to the `google` org?\r\n\r\nThat would be helpful, though I'm no longer affiliated with Google so not sure what the usual policy is there. If it is ok that will be easier than having to send the checkpoints to @hwchung so he uploads them.", "> That would be helpful, though I'm no longer affiliated with Google so not sure what the usual policy is there.\r\n\r\nUltimately the org admins should decide, but for now I think it's perfectly acceptable if you're a part of the org. I added you manually.", "@Iwontbecreative I opened a PR on your branch that should fix all the failing tests here: https://github.com/Iwontbecreative/transformers/pull/1\r\n\r\nI've separated each test suite (TF, style, docs) in three commits if you want to have a look at smaller portions at a time.", "Thanks Lysandre. Have not forgotten about this issue, just need one more approval from Google to open source the checkpoint so waiting for this.", "Sure @Iwontbecreative, let us know when all is good!", "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.", "Reopen: this is still in progress, just got the open sourcing done from google last Friday. Travelling for a few days and then can tackle this. See: https://github.com/google-research/google-research/tree/master/rembert", "That's great news @Iwontbecreative!", "(Finally) some updates on this now that the Google open sourcing is done.\r\n\r\n* Updated the PR for the lates transformers\r\n* Included @LysandreJik's change (manually instead of merging your PR since I messed up the order, sorry about that)\r\n* Uploaded the model to the model hub under `iwontbecreative/rembert`\r\n\r\nMain issue that still exists: the discrepancy between my implementation and the tf one. Welcoming ideas on this one. The code should now be easier for everyone to test now that the tf version is available and mine is uploaded to the model hub.\r\n\r\n@LysandreJik think this is ready for another look. Thanks for the patience here!\r\n", "Welcome back! We'll take a look shortly. Do you have an example of code to run to quickly test this version with the TF one?", "Sadly the tensorflow code to run the model is not available externally. \r\nI tried to replicate for some time this morning but it is hard because it relies on https://github.com/google-research/bert which has dependencies that are not available on pip anymore...\r\n\r\nThe main change to the modelling code is here:\r\nhttps://github.com/google-research/bert/blob/master/modeling.py#L813-L823\r\nneeding to be replaced with:\r\nhttps://github.com/google-research/albert/blob/master/modeling.py#L1083-L1088\r\non the tokenization front, it is mainly replacing `BertTokenizer` with `AlbertTokenizer`\r\n\r\nI do however have example inputs and outputs run by my coauthor:\r\n\r\n### Model outputs\r\nExample modelling outputs at several layers for different input_ids:\r\nhttps://pastebin.com/t9bPFmeM \r\nThis is the `[batch, length, :3]` section of the `[batch, length, hidden]` outputs.\r\n\r\n### Tokenization outputs\r\nhttps://pastebin.com/j6D6YE1e", "Fine-tuning was, here is what I was able to run, comnparing performance on XNLI:\r\n\r\nhttps://docs.google.com/spreadsheets/d/1gWWSLo7XxEZkXpX272tQoZBXTgs96IFvh-fwqVqihM0/edit#gid=0\r\n\r\nPerformance matches in English but does seem to be lower on other languages. We used more hyperparam tuning at Google but I do not think that explains the whole difference for those languages. I think there might be a subtle difference that is both causing the results to differ slightly and the worse fine-tuning outcomes. The model is still much better than random so most of it should be there.", "Performance does look pretty similar, and good enough for me to merge it.\r\n\r\nThere are a few `# Copied from` statements missing though as said in my previous message, in both the PyTorch and TensorFlow implementations. Do you mind adding them? If you're lacking time let me know and I'll do it myself.", "Hi @LysandreJik \r\n\r\n- Added copy statements\r\n- Merged with last master\r\n- Uploaded model to google org\r\n\r\nSeems like it is mostly ready, though tests fail at the `utils/check_copies.py` stage of `make quality`.\r\nI am actually not sure what the issue is in this instance, any chance you could help investigate/fix/merge after?", "Actually, managed to find the issue\r\n\r\n`utils/check_copies.py` crashes without a helpful error message if the \"Copied from\" statement is before the decorator. I was just overeager with my copied from statements.\r\n\r\nAlso renamed rembert-large to rembert since this is the only version we are open-sourcing at this time.\r\n\r\nEdit: Not sure why the model templates check is failing, but think this should be ready for merge with one last review. ", "Fantastic, thanks a lot @Iwontbecreative! I'll take a final look, fix the model templates issue and ping another reviewer.", "@patrickvonplaten Thanks for the helpful feedback. Incorporated most of it. See the comment on possible needed changes to the cookiecutter templates to address on of your comments in the future.\r\nFor the discrepancy in results, see my answer above.\r\n\r\n@sgugger Regarding older template: Yes, this PR ended up being delayed due to slow open-sourcing process at Google, so the templates were a bit out of date. Thanks for catching most of the mistakes.\r\n\r\n", "Hi @LysandreJik, any last remaining steps before this can be merged? Would like to get this in to avoid further rebases if possible. ", "I think this can be merged - thanks for your effort @Iwontbecreative, fantastic addition!", "I'm not entirely sure why there was 88 authors involved or 250 commits squashed into a single one - but I did verify only your changes were merged.\r\n\r\nCould you let me know how you handled the merge/rebasing of this branch so that I may understand what happened w.r.t the number of commits included?", "I think I just merged the master's changes into my branch to ensure it was up to date with upstream. Maybe I needed to rebase?", "Hi @Iwontbecreative thanks for adding the RemBERT model! Do you have a list of the 110 languages used in the pretraining of the model?", "Sure, here's the list:\r\n\r\n['af', 'am', 'ar', 'az', 'be', 'bg', 'bg-Latn', 'bn', 'bs', 'ca', 'ceb', 'co', 'cs', 'cy', 'da', 'de', 'el', 'el-Latn', 'en', 'eo', 'es', 'et', 'eu', 'fa', 'fi', 'fil', 'fr', 'fy', 'ga', 'gd', 'gl', 'gu', 'ha', 'haw', 'hi', 'hi-Latn', 'hmn', 'hr', 'ht', 'hu', 'hy', 'id', 'ig', 'is', 'it', 'iw', 'ja', 'ja-Latn', 'jv', 'ka', 'kk', 'km', 'kn', 'ko', 'ku', 'ky', 'la', 'lb', 'lo', 'lt', 'lv', 'mg', 'mi', 'mk', 'ml', 'mn', 'mr', 'ms', 'mt', 'my', 'ne', 'nl', 'no', 'ny', 'pa', 'pl', 'ps', 'pt', 'ro', 'ru', 'ru-Latn', 'sd', 'si', 'sk', 'sl', 'sm', 'sn', 'so', 'sq', 'sr', 'st', 'su', 'sv', 'sw', 'ta', 'te', 'tg', 'th', 'tr', 'uk', 'ur', 'uz', 'vi', 'xh', 'yi', 'yo', 'zh', 'zh-Hans', 'zh-Hant', 'zh-Latn', 'zu']\r\n\r\ncf https://github.com/google-research/google-research/tree/master/rembert\r\n" ]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "I think I might have come up with a bandaid solution for now. If the output dir exists and overwrite output dir flag is not set, load the configuration from the output dir and resume training from `pytorch_model.bin`. I'm going to give this a try, I think it's going to work. ", "Yes, as I expected, it worked. I modified the sample scripts as follows:\r\n\r\n```\r\n if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:\r\n # last_checkpoint = get_last_checkpoint(training_args.output_dir)\r\n # check there is a WEIGHTS_NAME model in the output directory and use that as the last_checkpoint\r\n if os.path.isfile(os.path.join(training_args.output_dir, WEIGHTS_NAME)):\r\n # use it as checkpoint\r\n last_checkpoint = training_args.output_dir\r\n```\r\n\r\nThe ramification of this change is that the checkpoint is going to be replaced upon training completion (which so happens is what I needed). \r\n\r\nI still think the checkpoint naming conventions should be reconciled (re: WEIGHTS_NAME and PREFIX_CHECKPOINT_DIR) so I'll leave this feature request open. ", "Nothing has changed in the way checkpoints are named since version 2 at least, e.g. the checkpoints are saved in `args.output_dir/checkpoin-xxx` where xxx is the number of training steps elapsed since the beginning of training.\r\n\r\nThe change you are suggesting would remove the ability for people to resume training from the last checkpoint saved which is not something we want. If you want to start your training from a specific saved model, you can pass along \r\n`--model_name_or_path path_to_folder_with_saved_model`.", "@sgugger Thanks for the reply. I think I wanted the behavior that you're talking about (not what I ended up doing, i.e., looking for a saved model directly in the output dir). \r\n\r\nBased on what you just said, I looked again at the code and there are two ways to save a checkpoint: `save_model` and `_save_checkpoint`. It so happens that the sample text classification scripts use `save_model` directly which does not create the `checkpoint-xxx` directory. Underneath, `_save_checkpoint` calls `save_model` but with the dir `checkpoint-xxx` which was the behavior that I wanted. \r\n\r\nBTW, what's the \"right\" way of saving models/checkpoints? the `_` in `_save_checkpoint` makes me believe it's supposed to be a utility function and there is another API function. \r\n\r\nSo I guess what's happening is the sample script calls `save_model` when the training ends (saving the `pytorch_model.bin` directly in the output dir) and that confused me. Bottom line: the current sample script for text classification can't resume training from the last checkpoint saved (because it doesn't \"save a checkpoint\" it \"saves a model\" https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_glue.py#L466).\r\n\r\nI completely agree with you that that is a behavior everybody wants, i.e., resuming training from last checkpoint. If you agree, I'll change this issue to read save checkpoint in sample text classification script.", "I think you are confusing:\r\n- saving checkpoints during training, done automatically by the `Trainer` every `save_steps` (unless you are using a different strategy)\r\nand\r\n- saving the final model, which is done at the end of training.", "Got it, thanks for clarifying the terminology! \r\n\r\nPS: It so happens that I needed a checkpoint to be saved at the end of training, now I understand how that's done. " ]

[]

[]

[ "Very nice ! " ]

[]

[]

[ "> This is not the proper way to proceed. Here you force the dtype to float32 which is not correct, because you might encounter conflicts with other values that are in float16.\r\n\r\nYes, you're right, I should not force the dtype to float32, maybe I should force it to float32 when using `mixed_precision`.\r\n\r\n> To get similar results, you have to increase the number of epochs. Usually, I get a similar model after multiplying the number of epochs by 2 or 2.5.\r\n\r\nBut in my sample: \r\nI always get 0.98 accuracy after 18 epochs without using the `mixed_precision`, \r\nbut sometimes the accuracy will stop at 0.7333 when using the `mixed_precision` policy, even I train it after 100 epochs.\r\n\r\nI think there must something wrong. I have updated my sample, have a look: https://colab.research.google.com/github/mymusise/gpt2-quickly/blob/main/examples/mixed_precision_test.ipynb", "> Yes, you're right, I should not force the dtype to float32, maybe I should force it to float32 when using mixed_precision.\r\n\r\nNo! This would be even worse. Set the layer norm and the embeddings directly to `float32` is a better temporary fix.\r\n\r\n> I always get 0.98 accuracy after 18 epochs without using the mixed_precision, but sometimes the accuracy will stop at 0.7333 when using the mixed_precision policy, even I train it after 100 epochs.\r\n\r\nHaving a lower accuracy in mixed precision is normal, but not that much. How much do you get by setting the layer norm and the embeddings directly to `float32` in mixed precision?\r\n\r\nI cannot run any test for now as I don't have access to a computer so I cannot really use your Colab.", "> No! This would be even worse. Set the layer norm and the embeddings directly to `float32` is a better temporary fix.\r\n\r\nMy apologies, I didn't get it, do you mean the first commit is better? :joy:\r\n\r\n> Having a lower accuracy in mixed precision is normal, but not that much. How much do you get by setting the layer norm and the embeddings directly to `float32` in mixed precision?\r\n> \r\n> I cannot run any test for now as I don't have access to a computer so I cannot really use your Colab.\r\n\r\nSorry to disturb your rest time. I can get 0.99 accuracy after setting the layer norm and the embeddings directly to `float32` in mixed precision.\r\nHere's a partial screenshot:\r\n\r\n", "> My apologies, I didn't get it, do you mean the first commit is better?\r\n\r\nYes.\r\n\r\n> Sorry to disturb your rest time. I can get 0.99 accuracy after setting the layer norm and the embeddings directly to float32 in mixed precision.\r\n\r\nThanks for the screenshot. Can you please revert your last commit then.", "> Thanks for the screenshot. Can you please revert your last commit then.\r\n\r\nSure!", "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.", "@Rocketknight1\r\nHello, what should I do to help with 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." ]

[]

[]

[ "No problem. I can't see what test failed in CircleCI as it wants to bind to my Github account and access private things in my org. Is there any way to see the failures without letting it into my account?" ]

[]

[]

[]

[]

[ { "login": "theo-m", "id": 17948980, "node_id": "MDQ6VXNlcjE3OTQ4OTgw", "avatar_url": "https://avatars.githubusercontent.com/u/17948980?v=4", "gravatar_id": "", "url": "https://api.github.com/users/theo-m", "html_url": "https://github.com/theo-m", "followers_url": "https://api.github.com/users/theo-m/followers", "following_url": "https://api.github.com/users/theo-m/following{/other_user}", "gists_url": "https://api.github.com/users/theo-m/gists{/gist_id}", "starred_url": "https://api.github.com/users/theo-m/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/theo-m/subscriptions", "organizations_url": "https://api.github.com/users/theo-m/orgs", "repos_url": "https://api.github.com/users/theo-m/repos", "events_url": "https://api.github.com/users/theo-m/events{/privacy}", "received_events_url": "https://api.github.com/users/theo-m/received_events", "type": "User", "site_admin": false } ]

[ "I'm talking with @n1t0 soon to get a better sense of where it might make sense as well yes! Will update then", "went looking for similar case with this regex `(\\[|\\()\"do_lower_case`, found only one more: 7e42461\r\n\r\nI think this is it, will move onto #10121 once this is merged", "There does seem to be an issue remaining though, as the tests in the suite currently fail with:\r\n```\r\n=========================== short test summary info ============================\r\nFAILED tests/test_tokenization_auto.py::AutoTokenizerTest::test_do_lower_case\r\n=== 1 failed, 5143 passed, 2103 skipped, 2207 warnings in 266.40s (0:04:26) ====\r\n```" ]

[]

[]

[]

[]

[]

[ "Here's an answer to your questions:\r\n\r\n> 1. Now that I have this model, how can I use it to answer to a \"custom\" question on a specific context? I am assuming that I will need to preprocess the question+context in a way similar to the preprocessing of training dataset, but how can I exactly do that?\r\n\r\nYes, I've created a small Colab notebook to illustrate inference for question answering models: https://colab.research.google.com/drive/1F-4rWIDythF4B8hS6SdNx9x3h3ffg2zw?usp=sharing\r\n\r\nBtw, this is also explained in the [docs](https://huggingface.co/transformers/task_summary.html#extractive-question-answering).\r\n\r\n> 2\\. As final goal I will have to fine-tune an italian language model. I'm assuming that this depends on the value of `model_checkpoint`, so here I would have to select an italian pre-trained model (for example [dbmdz/bert-base-italian-cased](https://huggingface.co/dbmdz/bert-base-italian-cased)), is that correct? And if I want to use a multilanguage model, how do I specify the language I want to use?\r\n\r\nIf you want to fine-tune on an Italian dataset, then it's indeed advised to start from a pre-trained Italian model. If you want to use a multilanguage model, you don't need to specify any language you want to use, because it's a general-purpose model. Just make sure that Italian is one of the languages on which the multi-lingual model was pre-trained.\r\n\r\n> 3\\. (this may be a very dumb question) Can I fine-tune an already fine-tuned model (for example [this model](https://huggingface.co/mrm8488/bert-italian-finedtuned-squadv1-it-alfa))? Would it make sense?\r\n\r\nYes you can, and it's maybe the best thing to do, because as this model is already fine-tuned on Italian questions, then it will already have reasonable performance out-of-the-box. You can just improve it a little more by fine-tuning on your specific dataset.\r\n\r\nBtw, it's advised to ask such questions on the [forum](https://discuss.huggingface.co/) rather than here, as the authors of HuggingFace like to keep Github issues for bugs and feature requests.\r\n\r\nCheers!", "Many thanks @NielsRogge , I discovered the existence of the forum 20 minutes after posting this, my bad :(\r\n\r\nI will use the forum now to ask some more things.\r\n\r\nThansk a lot!" ]

[]

[]

[ "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,\r\n\r\nwould it be possible to ask this question on HuggingFace's [forum](https://discuss.huggingface.co/)? Some people like Sylvain (who created the tutorial you mention) are very active there, and are happy to help you. The authors like to keep Github issues for bugs caused by the Transformers library or feature requests.\r\n\r\nThanks!", "I moved this question to: https://discuss.huggingface.co/t/token-classification-how-to-tokenize-and-align-labels-with-overflow-and-stride/4353" ]

[]

[]

[ "@stas00 I hear you regarding the installation of dependencies. We are installing every dependency except the main one (PT/TF) on the GPU machine, so I would say most errors are caught.\r\nFurthermore, we are installing these exact dependencies on the CircleCI runs, so I believe we are testing that already on every commit.\r\n\r\nIf there is a scenario I am missing, please let me know and I will do my best to adjust.\r\n\r\nAll other comments have been adressed in the last three commits.", "Ah, ok, for some reason I thought that if we are using a docker image then we can skip wasting time and resources on installing the same things million times a day and just run the tests right away, and in that case only an occasional test that installs from scratch will be needed. But perhaps this further speed up can be done in some future iteration.\r\n\r\nNothing more from my side.\r\n\r\nThank you, @LysandreJik " ]

[]

[]

[ "This is a very weird bug actually and can be reproduced the easiest ass follows:\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nfrom transformers import TFMarianMTModel, MarianMTModel, MarianTokenizer\r\n\r\ntokenizer = MarianTokenizer.from_pretrained(\"Helsinki-NLP/opus-mt-mt-en\")\r\n\r\nmodel_pt = MarianMTModel.from_pretrained(\"Helsinki-NLP/opus-mt-mt-en\")\r\nmodel_tf = TFMarianMTModel.from_pretrained(\"Helsinki-NLP/opus-mt-mt-en\", from_pt=True)\r\n\r\nmodel_tf.save_pretrained(\"./\")\r\nmodel_tf = TFMarianMTModel.from_pretrained(\"./\")\r\n\r\ninput_str = \"My name is Wolfgang and I live in Berlin\"\r\ninput_str = \"Billi messu b'mod ġentili, Ġesù fejjaq raġel li kien milqut bil - marda kerha tal - ġdiem.\"\r\n\r\n\r\noutput_tokens_pt = model_pt.generate(tokenizer(input_str, return_tensors=\"pt\").input_ids)\r\noutput_tokens_tf = model_tf.generate(tokenizer(input_str, return_tensors=\"tf\").input_ids)\r\n\r\nprint(\"Pt:\", tokenizer.batch_decode(output_tokens_pt))\r\nprint(\"Tf:\", tokenizer.batch_decode(output_tokens_tf))\r\n```\r\n\r\nfails while commenting out the save and load lines works:\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nfrom transformers import TFMarianMTModel, MarianMTModel, MarianTokenizer\r\n\r\ntokenizer = MarianTokenizer.from_pretrained(\"Helsinki-NLP/opus-mt-mt-en\")\r\n\r\nmodel_pt = MarianMTModel.from_pretrained(\"Helsinki-NLP/opus-mt-mt-en\")\r\nmodel_tf = TFMarianMTModel.from_pretrained(\"Helsinki-NLP/opus-mt-mt-en\", from_pt=True)\r\n\r\ninput_str = \"My name is Wolfgang and I live in Berlin\"\r\ninput_str = \"Billi messu b'mod ġentili, Ġesù fejjaq raġel li kien milqut bil - marda kerha tal - ġdiem.\"\r\n\r\n\r\noutput_tokens_pt = model_pt.generate(tokenizer(input_str, return_tensors=\"pt\").input_ids)\r\noutput_tokens_tf = model_tf.generate(tokenizer(input_str, return_tensors=\"tf\").input_ids)\r\n\r\nprint(\"Pt:\", tokenizer.batch_decode(output_tokens_pt))\r\nprint(\"Tf:\", tokenizer.batch_decode(output_tokens_tf))\r\n```\r\n\r\nMost other TFMarian models work correctly. This is pretty weird though and will need more time for investigation (cc @patil-suraj for info)", "Thanks for your investigation!", "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.", "Unstale", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[ "I don't think you should use `torch.cuda.empty_cache()`, as explained on [PyTorch's forum](https://discuss.pytorch.org/t/about-torch-cuda-empty-cache/34232/2), \"This function should not be used by the end-user except in very edge cases.\". \r\n\r\nAlso, you can set `truncation=True`, because currently you're only padding examples, but not truncating examples that are too long. \r\n\r\nBtw, it's better to ask training related questions which are not bugs caused by the Transformers library on the [forum](https://discuss.huggingface.co/) rather than here.", "@NielsRogge \r\nThanks for your help! It works!\r\nI use `torch.cuda.empty_cache()` because I have no idea but try it.\r\nAnd I will go to the website you said then.\r\nThank you again!" ]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "I agree with your assessment and your proposed fix, so by all means, please suggest a PR! Thanks!" ]

[]

[]

[ "@lhoestq If this looks good to you, I can help create a PR. Thanks!", "Hi !\r\nThis looks like a great solution, feel free to open a PR ! :)", "Great, thanks! Just made a PR here: https://github.com/huggingface/transformers/pull/10780" ]

[]

[]

[ "Hey @fostiropoulos,\r\n\r\nSorry to answer this late - could you provide a code example to reproduce the error? ", "@patrickvonplaten It would take me time to come up with a full working minimal example in colab. \r\n\r\nHowever you can try a model that you supply only the input_embeds and leave input_ids,position_ids to `None` during test time and when the sequence requires padding. \r\n\r\nThe positional encoding will use a ids (`torch.arange`) from index 0 (`start_idx_pos_encodings=0`) to `padded_sequence_length` for the padded ids. It should have been the start position of the end of the input embedding e.g. (`start_idx_pos_encodings=seq_len`)\r\n\r\nIt shouldn't affect the final results because the padded tokens are discarded at the end, but it is not the expected behavior. It would cause errors if padding happens outside of test time or the function is used elsewhere. ", "@patrickvonplaten does padding affect anything during sampling? e.g. the pad token being 0 vs 100 or any other random int. The casual attention / mask should only make it so that each token depends only on previously seen tokens (beyond the pad). Am I correct? ", "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": 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!" } ]

[]

[ "I am unsure what you want me to fix. Yes evaluation in FP16 is not as precise as evaluation in FP32 and results in NaNs more easily, in particular in the loss, that's precisely the reason I was reluctant to add the --fp16_full_eval option.", "well, first of all I'm reporting this behavior.\r\n\r\nSecondly, it looks like DeepSpeed Zero3 for now requires that we continue using the fp16 model for inference. We are discussing other possibilities but they aren't there yet. This is because of all the hooks that get installed into the model during training. But if we were to remove all those hooks suddenly the model won't fit into the gpu memory - if it was spread out over multiple gpus in the first place. \r\n\r\nWhich is why I added ` --fp16_full_eval` in first place. To enable fitting the model onto the gpu memory during inference if it fit during training (with fp32 master weights being offloaded to cpu). e.g. fitting t5-11b (45GB at fp32) into 40GB gpu. Can train with DeepSpeed zero2, so should be able to do inference too.\r\n\r\nFinally, according to DeepSpeed devs a model trained in mixed precision shouldn't need to be put back into fp32 and should provide similar results only slightly imprecise. \r\n\r\nwe will need a way to get the fp32 model out of DeepSpeed training https://github.com/microsoft/DeepSpeed/issues/800", "I understand all of that, and maybe a fine-tuned model in mixed precision will have better results that don't get any Nan losses on the evaluation dataset, but it only takes one sample of the whole evaluation dataset with a slightly bigger loss than usual to get one Nan that will then drive the whole evaluation loss to NaN. And with label smoothing enabled it only takes one log_probs over all the possibilities to get to that.\r\n\r\nNot sure it really matters since you still have proper metrics (the predictions are not driven to Nan, just the loss). Maybe we could try a flag to deactivate label smoothing when running evaluation, but then the loss wouldn't be comparable to the training loss, so not sure if it would really be useful either.\r\n", "Thank you for elucidating the situation and suggesting workarounds, @sgugger \r\n\r\nWould it be of a useful value if we approximated the nans with something that would lead to a non-nan loss? since NaN can come from a variety of combinations of inf/0 operations, do we know which one is it? and then perhaps pick a corresponding substitution that might lead to a sufficiently good estimate?\r\n\r\n", "Yes, we could maybe track when it arises and where it comes from exactly as a first step. Won't have time to dive into this in the near future, but if someone wants to tackle that issue to that effect and report here, that would be awesome!", "I'm interested in taking a stab on this!", "Awesome! Thank you, @vladdy\r\n\r\nJust be aware that the following PR should be merged shortly: https://github.com/huggingface/transformers/pull/10611\r\nand so the script in the reproduction line will most likely be renamed to `run_translation.py` and will slightly change the cl args.\r\n", "@vladdy, while you research this - it'd be great to understand the cause of NaNs - so if you discover which operation leads to it please do share. Thank you!", "Before falling asleep this idea came to me, tried it this morning and it worked:\r\n\r\n```\r\n--- a/src/transformers/trainer_pt_utils.py\r\n+++ b/src/transformers/trainer_pt_utils.py\r\n@@ -390,7 +390,9 @@ class LabelSmoother:\r\n\r\n def __call__(self, model_output, labels):\r\n logits = model_output[\"logits\"] if isinstance(model_output, dict) else model_output[0]\r\n+ #logits = logits.to(dtype=torch.float32)\r\n log_probs = -torch.nn.functional.log_softmax(logits, dim=-1)\r\n+ log_probs = log_probs.to(dtype=torch.float32)\r\n if labels.dim() == log_probs.dim() - 1:\r\n labels = labels.unsqueeze(-1)\r\n```\r\n\r\nBasically, flip `logits` or `log_probs` back to fp32 and the problem goes away.\r\n\r\nSo the problem here has nothing to do with full fp16 inference, but with how label smoothing is calculated.\r\n\r\nThe issue comes from:\r\n```\r\n smoothed_loss = log_probs.sum(dim=-1, keepdim=True)\r\n```\r\nin fp32, the return values are huge:\r\n```\r\n [ 863637.3750],\r\n [ 864242.0000],\r\n [ 865449.0000],\r\n [ 866092.9375],\r\n [ 867702.4375],\r\n```\r\nand in fp16 these turn `inf`.\r\n\r\nSo either:\r\n1. we do what I proposed on top of this comment, which will double the size of the `log_probs` tensor (4 times if we apply it to `logits`, rather than `log_probs`) . This will of course depend on the size of the dictionary and `target_max_len` - so say:\r\n```\r\nbs * max_len * dict_size * 2 more bytes\r\n32 * 128 * 32000 * 2 = 262MB - huge!\r\n```\r\n2. or we change the calculation to scale down huge numbers back to a numerical range where `sum` over fp16 numbers doesn't overflow. Note that `smoothed_loss` does another `sum` towards the end which would definitely blow things up.\r\n3. same as (1) but switch the calculations to `.cpu()` - a bit slower but no extra gpu memory will be required.\r\n\r\nOf course number 2 is a better solution since it doesn't require much more memory to solve this problem and will require a change in algorithm to avoid going into huge numbers.\r\n\r\nThe other question is: do we deal with the label smoother separately or do we have other parts which may be affected in which case we should change the logits back to fp32 when prediction has completed. But as explained above this will come at a large gpu memory cost.", "Yeah, I came to a similar conclusion regarding the cause and when I wanted to post an update on it, I saw @stas00's response above.\r\n\r\nIf I'm not wrong, Apex tried to solve [this compatibility issue between fp16 and losses](https://github.com/NVIDIA/apex/tree/a109f856840ebb3ff5578e0bddfc4cffd4b96ed0/apex/fp16_utils), but I'm not sure how much of that could be reused in addition to already stated options. @stas00, please let me know if you want to continue driving this and I'll try to find some other issue for my contribution.", "@vladdy, by all means please continue, I was just sharing what I have discovered and calculated that this won't be an efficient solution memory requirement-wise. And I now have a better understanding of where NaN came from.\r\n\r\nAs you're suggesting the most efficient generic solution would be around loss scaling. We are doing it already during the training, so this is just some of the same for label smoothing.\r\n\r\nBut we definitely don't want to depend on apex for this. \r\n\r\nI haven't looked closely but I think the idea is to scale the `log_probs` into a much smaller size, while ensuring that the scaled numbers and the sum of 30-60k elements remain within the dynamic range of fp16 (plus there is one more sum of sums at the end!). If we were to implement it in a non-vectorized way it'd be the simplest to create an fp32 variable and add the fp16 bits to it, so it won't overflow. It won't take any extra memory, but that won't be efficient speed wise.\r\n\r\nAnd of course perhaps you can think of other solutions. Anything that doesn't require an extra GPU memory to perform label smoothing is goodness.\r\n\r\nPerhaps pytorch has some ready-made solutions too...", "Any progress on this, @vladdy? I have one idea that may work, but would love to hear what you have discovered in your research.", "@stas00, I have not found anything better than switching to fp32 for that operation. The rest of the approaches appear to be more complicated or not as generic as I think we want them to be. What idea did you have in mind?", "Thank you for sharing the results of your research, @vladdy. So the need is the same, it's all about how to do it efficiently and not defeat the purpose of keeping things at fp16.\r\n\r\nMy discovery was switching to fp32 only for the aggregate and have it done by pytorch on the hardware level:\r\n```\r\ndiff --git a/src/transformers/trainer_pt_utils.py b/src/transformers/trainer_pt_utils.py\r\nindex ae8e24949..c2071733c 100644\r\n--- a/src/transformers/trainer_pt_utils.py\r\n+++ b/src/transformers/trainer_pt_utils.py\r\n@@ -399,7 +399,8 @@ class LabelSmoother:\r\n # will ignore them in any case.\r\n labels.clamp_min_(0)\r\n nll_loss = log_probs.gather(dim=-1, index=labels)\r\n- smoothed_loss = log_probs.sum(dim=-1, keepdim=True)\r\n+ # works for fp16 input tensor too, by internally upcasting it to fp32\r\n+ smoothed_loss = log_probs.sum(dim=-1, keepdim=True, dtype=torch.float32)\r\n\r\n nll_loss.masked_fill_(padding_mask, 0.0)\r\n smoothed_loss.masked_fill_(padding_mask, 0.0)\r\n```\r\nI was shocked that it took ~0 extra memory over pure fp16 solution (that is not even extra peak memory!) - that is it was doing the conversion to fp32 on the hardware level. At least that was so on my hardware - it might be not so on an older gpu that doesn't support fp16 natively. \r\n\r\nThis is pretty amazing that it can do that!\r\n\r\nI came to this idea while researching bfloat16 where its aggregate operations too require fp32 aggregates - so I thought why not try the same for our case and it seems to work. Similarly sometimes they use fp64 for fp32 inputs if they are too big, but I don't think we run into those things here.\r\n\r\nWhat do you think? \r\n", "I think, this simple solution makes sense to be applied as it is also generic enough to cover all the cases. I doubt it is possible to find a better approach within short time and it does not appear it is necessary to spend more time on this (at least, for now). Feel free to do the PR as you offered it!", "Dear @stas00 \r\nThank you very much for filing this issue. I am training mt5-small model and with deepspeed without label smoothing, I am getting NaNs, so far could not managed to fix it. I greatly appreciate your suggestions on this. if you think this can be appropriate I will open up a separate issue for mt5 model getting NaNs with deepspeed, and if not I follow this issue. Thank you very much", "@vladdy, thank you for doing your research and validating my suggestion.\r\n\r\n@dorost1234, the PR is here: https://github.com/huggingface/transformers/pull/10815 or you can just change it manually for your test - it's just one line.\r\nPlease let me know if it fixes your problem, if it's about eval_loss. If it doesn't, or it's about something else - then yes please open a separate issue. Based on your comments elsewhere the issue about mt5 and NaNs already exists, but not with deepspeed so definitely open one. Perhaps the Deepspeed team has some insights about this situation.\r\n\r\n", "Thank you so much @stas00 for the great response, I applied your PR and with deepspeed now mt5-small for me is not getting nan anymore, this is an incredible job you are doing, thanks a lot, I still getting nans with mt5-small with fp16, even after your PR, for this I made a separate issue here https://github.com/huggingface/transformers/issues/10819 I did not tag you since with deepspeed with your applied magic PR it is not getting nans so far, while still If you have time to give me an advice I would be really grateful. ", "@stas00 I tested mt5-small with run_translation.py model without this PR this also works fine without nans, if one does not use smoothing, with this PR this becomes much slower for me with deepspeed. is there a way to keep the speed as great as before?", "Please post the exact command lines that you're referring to.\r\n\r\nAs I wrote in https://github.com/huggingface/transformers/pull/10815 I definitely can see a 25% slowdown when enabling --fp16_full_eval and opened an issue about it https://github.com/huggingface/transformers/issues/10816 \r\n\r\nI don't see any speed difference with https://github.com/huggingface/transformers/pull/10815 w/o deepspeed, so once you show me what command line use then I can test.\r\n\r\nedit: Oh, I see you posted them in https://github.com/huggingface/transformers/issues/10819 - all is good then - I can test now." ]

[]

[]

[]

[]

[]

[ "No, `Optional[bool]` are dealt with later on at [this line](https://github.com/huggingface/transformers/blob/master/src/transformers/hf_argparser.py#L101) and need to stay wrapped until then.\r\n\r\nWithout more information on your error it's hard to help find the right fix.", "Hi @sugger, thank you for the quick response and for pointing that out. Sorry I did not catch it.\r\n\r\nI also see that you put a [tentative fix](https://github.com/huggingface/transformers/commit/fa1a8d102f273ee8118546f7b84133ab58032ac5) for this issue, but also just wanted to check to make sure that `Optional[bool]` values are caught by the expected if statement.\r\n\r\nLooking more into the issue, I'm actually seeing that `disable_tqdm` has its typing changed from `Optional[bool]` to `Union[bool, None]` during the `dataclasses.fields` parsing. So it is not getting caught by the line you reference and then the error occurs on the next if statement.\r\n\r\nI'm running Python 3.8 on ubuntu 18.04, maybe dataclasses parses optionals differently in other versions.\r\n\r\nPrintout of `disable_tqdm` fields value:\r\n```\r\n>>> [field for field in dataclasses.fields(TrainingArguments) if \"tqdm\" in field.name][0]\r\nField(name='disable_tqdm',type=typing.Union[bool, NoneType],default=None,default_factory=<dataclasses._MISSING_TYPE object at 0x7f2e6e7ff1f0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({'help': 'Whether or not to disable the tqdm progress bars.'}),_field_type=_FIELD)\r\n```\r\n\r\ndataclasses parsing of generic optional bool type:\r\n```\r\n>>> @dataclass\r\n... class OptionalBool:\r\n... value: Optional[bool]\r\n...\r\n>>> dataclasses.fields(OptionalBool)\r\n(Field(name='value',type=typing.Union[bool, NoneType],default=<dataclasses._MISSING_TYPE object at 0x7f2e6e7ff1f0>,default_factory=<dataclasses._MISSING_TYPE object at 0x7f2e6e7ff1f0>,init=True,repr=True,hash=None,compare=True,metadata=mappingproxy({}),_field_type=_FIELD),)\r\n```\r\n\r\nTo address this, would it be fine to change the `is Optional[bool]` check [here](https://github.com/huggingface/transformers/blob/master/src/transformers/hf_argparser.py#L101) to `==`?\r\n\r\n```\r\n>>> Optional[bool] is Union[bool, None]\r\nFalse\r\n>>> Optional[bool] == Union[bool, None]\r\nTrue\r\n```\r\n\r\nThanks again for looking into this so quickly.", "I think this change is acceptable, thanks! Trying to check what the failure in the test is and if it's spurious." ]

[]

[]

[]

[]

[]

[]