Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

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

[]

[]

[ "Still not letting me assign you @sgugger :(" ]

[]

[]

[ "Can't pin you for review @sgugger, so tagging you!" ]

[]

[]

[ "Just rename your `.bin` to `.pt`", "Hi @julien-c \r\n\r\nSorry to come back to this but I am having trouble with this. As the .bin file also has a much larger accompanying ``optimizer`` file that I assume holds the weights.\r\n\r\nI am trying to deploy a fine tuned model to google cloud. And even when using a custom prediction routine to load the entire folder for distil GPT2 the folder size exceeds the limit of ``500MB``.\r\n\r\nIs there a way to export the fine tuned model to be used alone. Either export to a standalone pytorch model or tensorflow model.\r\n\r\nI searched, but could not find any documentation on this. Would appreciate any help on this, or even direct me towards relevant documentation that can help me.\r\n\r\nI'm using ``transformers==2.8.0``\r\n\r\nThank you", "The optimizer file does not contain the weights of your model, but the state of the optimizer during your training.\r\n\r\nIf you do not plan on continuing training, then you can safely discard that file. You can have more information about [optimizers here](https://pytorch.org/docs/stable/optim.html)." ]

[]

[ { "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 } ]

[]

[]

[]

[ "👍 " ]

[]

[]

[ "I suppose I need to stick to the common bullet format, as it couldn't handle this `\\d\\) ` style of bullets. Leading to this rewrite:\r\n```\r\n-1) Optimizer State Partitioning (stage 1)\r\n-2) Add Gradient Partitioning (stage 2)\r\n+1) Optimizer State Partitioning (stage 1) 2) Add Gradient Partitioning (stage 2)\r\n```\r\nThis is not a problem - will fix the style.", "We also need a new line injector for bulleted lists in .rst checker pretty please.\r\n\r\n In .rst I had:\r\n\r\n```\r\nMiscellaneous notes:\r\n- DeepSpeed works with the PyTorch Trainer but not TF Trainer.\r\n- While DeepSpeed has a pip installable PyPI package, \r\n```\r\nthe style wrapper broke the bullets and made them into one paragraph/line.\r\n```\r\nMiscellaneous notes: - DeepSpeed works with the PyTorch Trainer but not TF Trainer. - While DeepSpeed has a pip installable PyPI package, \r\n```\r\n\r\nsame problem as with docstring - it's missing a new line again. Could we do the same fix for .rst to inject a new line before bullets if an unwary writer forgot to add one? \r\n\r\nThank you!\r\n\r\n", "Mmm, the patch should be applied to the rst files too (can't link to the diff but it's line 384 of the last file in the diff shown by GitHub).", "I re-based just in case, and no, it still doesn't insert the line. Here is the exact para:\r\n```\r\nMiscellaneous notes:\r\n* DeepSpeed works with the PyTorch Trainer but not TF Trainer.\r\n* While DeepSpeed has a pip installable PyPI package, it is highly recommended that it be `installed from source\r\n <https://github.com/microsoft/deepspeed#installation>`__ to best match your hardware and also to enable features like\r\n 1-bit Adam, which aren't available in the pypi distribution.\r\n```", "Indeed, I made some stupid mistake, #9488 should fix this." ]

[]

[]

[ "Applied the change", "Hi @patrickvonplaten. Sorry for the late reply.\r\n\r\nActually I tested the model with `BartForConditionalGeneration` and everything was working well. On the other hand, after the modification I am getting the following error: \r\n```\r\nUnrecognized configuration class for this kind of AutoModel: AutoModelForMaskedLM. Model type should be one of LayoutLMConfig, DistilBertConfig, AlbertConfig, BartConfig, CamembertConfig, XLMRobertaConfig, LongformerConfig, RobertaConfig, SqueezeBertConfig, BertConfig, MobileBertConfig, FlaubertConfig, XLMConfig, ElectraConfig, ReformerConfig, FunnelConfig, MPNetConfig, TapasConfig.\r\n```\r\n\r\nMaybe we only need to change `model_type` (even if I am not sure why) and not the architecture, because [mBART](https://huggingface.co/facebook/mbart-large-cc25/blob/main/config.json) itself is using `BartForConditionalGeneration`.\r\n\r\nWe still have the problem of the tokenizer when using AutoTokenizer:\r\n```\r\nTokenizer class BarthezTokenizer does not exist or is not currently imported.\r\n```\r\n\r\nIs it possible to force the api to import and use `BarthezTokenizer` instead of `AutoTokenizer`?", "Hey @moussaKam,\r\n\r\nThanks for your answer! Yeah the `AutoTokenizer` is still a problem and actually showcases a deeper problem we're having for the `AutoTokenziers` in the lib. We'll need a new design, something like proposed here: https://github.com/huggingface/transformers/pull/9305 to fix this issue. It's on my Todo list.\r\n\r\n", "Regarding the error with `AutoTokenizer` I cannot reproduce it :-/ Could you maybe provide code snippet showcasing the problem?", "Hi @patrickvonplaten,\r\n\r\nHere's a snippet:\r\n```python\r\ntext_sentence = \"Paris est la capitale de la <mask>\"\r\nimport torch\r\n\r\nfrom transformers import (\r\n AutoTokenizer,\r\n BartForConditionalGeneration\r\n)\r\n\r\nbarthez_tokenizer = AutoTokenizer.from_pretrained(\"moussaKam/barthez\")\r\nbarthez_model = BartForConditionalGeneration.from_pretrained(\"moussaKam/barthez\")\r\n\r\ninput_ids = torch.tensor(\r\n [barthez_tokenizer.encode(text_sentence, add_special_tokens=True)]\r\n)\r\nmask_idx = torch.where(input_ids == barthez_tokenizer.mask_token_id)[1].tolist()[0]\r\n\r\npredict = barthez_model.forward(input_ids)[0]\r\n\r\nbarthez_tokenizer.decode(predict[:, mask_idx, :].topk(5).indices[0])\r\n```\r\n```\r\n----> 9 barthez_tokenizer = AutoTokenizer.from_pretrained(\"moussaKam/barthez\")\r\n 10 barthez_model = BartForConditionalGeneration.from_pretrained(\"moussaKam/barthez\")\r\n 11 \r\n\r\n~/anaconda3/envs/transformers/lib/python3.8/site-packages/transformers-4.1.1-py3.8.egg/transformers/models/auto/tokenization_auto.py in from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs)\r\n 357 \r\n 358 if tokenizer_class is None:\r\n--> 359 raise ValueError(\r\n 360 \"Tokenizer class {} does not exist or is not currently imported.\".format(tokenizer_class_candidate)\r\n 361 )\r\n\r\nValueError: Tokenizer class BarthezTokenizer does not exist or is not currently imported.\r\n```\r\nThe expected output (if we use BarthezTokenizer instead of AutoTokenizer):\r\n```\r\n'France culture francophonie gastronomie mode'\r\n```", "Ok, @LysandreJik found a nice fix for the tokenizer. Regarding the model, I think from now on we should use `MBart` for Barthez since after the new release Bart is not compatible with Barthez anymore", "However, there seems to be an issue remaining with the `BarthezTokenizer`, as the code shared by @moussaKam outputs the following in v4.1.0:\r\n```\r\nFrance culture francophonie gastronomie mode\r\n```\r\nbut outputs the following on `master`:\r\n```\r\nompeolin corporelleenfin1-1\r\n```\r\n\r\nIt also mentions the following:\r\n```\r\nSome weights of the model checkpoint at moussaKam/barthez were not used when initializing BartForConditionalGeneration: ['encoder.layer_norm.weight', 'encoder.layer_norm.bias', 'decoder.layer_norm.weight', 'decoder.layer_norm.bias']\r\n```", "My bad, changing from `BartForConditionalGeneration` to `MBartForConditionalGeneration` fixes the issue.", "Yeah, Barthez is the only model that is not longer compatible with Bart looking forward - we have to stick to MBart. But the model architecture corresponds 1-to-1 to MBart, so I think it's fine. Hope it's ok for you @moussaKam ", "It's OK @patrickvonplaten if BARThez works well with `AutoModel`. Currently the shared code outputs (on the master): \r\n\r\n'France culture francophonie gastronomie mode' if we use `MBartForConditionalGeneration`\r\n'édappraiav comme' if we use `AutoModel`\r\n'ompeolin corporelleenfin1-1' if we use `BartForConditionalGeneration`", "Ah yeah, so instead of `AutoModel`, you'll have to use `AutoModelForSeq2SeqLM`.\r\nAnd it should not work anymore on master with `BartForConditionalGeneration`, but only with `MBartForConditionalGeneration`. Is the output of `MBartForConditionalGeneration` correct/reasonable in your opinion? \r\n\r\n=> so the model classes to use in the future are `AutoModelForSeq2SeqLM` (as before) and `MBartForConditionalGeneration` (this worked before as well), but now `BartForConditionalGeneration` should not work anymore.\r\n\r\nIf you could verify that this is actually the case on master now, that would be super nice", "yes the output is reasonable with `MBartForConditionalGeneration` and `AutoModelForSeq2SeqLM`.\r\n\r\nHowever we still have one last (I hope) problem when using `pipeline`. \r\nThe following code returns an error:\r\n```python\r\nfrom transformers import pipeline\r\n\r\npbase = pipeline(task=\"fill-mask\", model=\"moussaKam/barthez\")\r\nsrc_text = [\"Paris est la capitale de la <mask>\"]\r\nresults = [x[\"token_str\"] for x in pbase(src_text)]\r\n```\r\n```\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n<ipython-input-12-d7b2e5a78b7c> in <module>\r\n 1 from transformers import pipeline\r\n 2 \r\n----> 3 pbase = pipeline(task=\"fill-mask\", model=\"moussaKam/barthez\")\r\n 4 src_text = [\"Paris est la capitale de la <mask>\"]\r\n 5 results = [x[\"token_str\"] for x in pbase(src_text)]\r\n\r\n/datadisks/datadisk1/transformers/src/transformers/pipelines/__init__.py in pipeline(task, model, config, tokenizer, framework, revision, use_fast, **kwargs)\r\n 403 )\r\n 404 \r\n--> 405 model = model_class.from_pretrained(model, config=config, revision=revision, **model_kwargs)\r\n 406 if task == \"translation\" and model.config.task_specific_params:\r\n 407 for key in model.config.task_specific_params:\r\n\r\n/datadisks/datadisk1/transformers/src/transformers/models/auto/modeling_auto.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)\r\n 1123 pretrained_model_name_or_path, *model_args, config=config, **kwargs\r\n 1124 )\r\n-> 1125 raise ValueError(\r\n 1126 \"Unrecognized configuration class {} for this kind of AutoModel: {}.\\n\"\r\n 1127 \"Model type should be one of {}.\".format(\r\n\r\nValueError: Unrecognized configuration class <class 'transformers.models.mbart.configuration_mbart.MBartConfig'> for this kind of AutoModel: AutoModelForMaskedLM.\r\nModel type should be one of LayoutLMConfig, DistilBertConfig, AlbertConfig, BartConfig, CamembertConfig, XLMRobertaConfig, LongformerConfig, RobertaConfig, SqueezeBertConfig, BertConfig, MobileBertConfig, FlaubertConfig, XLMConfig, ElectraConfig, ReformerConfig, FunnelConfig, MPNetConfig, TapasConfig.\r\n```\r\n\r\nWe got the same error when using the the inference [api](https://huggingface.co/moussaKam/barthez?text=Paris+est+la+%3Cmask%3E+de+la+France.).", "Ah yeah, that's something unrelated to the Bart Split PR I think. Do you mind opening a new issue where you can copy paste your code example from above? Feel free to tag me on 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." ]

[]

[]

[]

[]

[]

[ "Hi @ayushjain1144 \r\n\r\nThat's an interesting question, would be better if you ask it on the [forum](https://discuss.huggingface.co/) " ]

[]

[]

[ "Humm looks like the quick test I have added to make sure that a saved model can be properly created is a bit too long for at least one of the models. @LysandreJik any idea how I can figure out for which model?", "Ok, the test `test_saved_model_creation` is skipped if it needs more than 30sec to be executed for a model. For now these models are skipped:\r\nBART\r\nBlenderBot\r\nFunnel\r\nLongformer\r\nLxmert\r\nMarian\r\nMBart\r\nMobilebert\r\nPegasus\r\nT5\r\n\r\nLet's see if the test becomes faster once I will optimise these model like I did for BERT. LGTM!", "Cool, maybe @sgugger can take a look as well :-) ", "Oops forgot to rebase and then the changes for the LED model is missing, and also the changes in the Seq2Seq template. Please wait my next push before merging.", "I should have addressed all the comments. The saved model creation tests are silent for the Seq2Seq models until I find a proper fix.", "Great should we merge @jplu @LysandreJik @? - it's blocking the TF-Bart Split PR a bit. ", "For me it is good to merge if there are no other comments :)", "Cool merging then", "I had some more comments actually! With the short names, most of the outputs of the serving methods fit on one line now. black does not put things back on the same line once it has split on several, so it's not fixed by the quality scripts.\r\n\r\nI also think it would make future maintenance easier to add the # Copied from comments for dupe code.", "They are all on one line (when it is possible, which means not too many characters to fit in).\r\n\r\nI will open a PR to take care of adding the `#copied from` comments once I finish to fix the S2S models.", "See comment above, and this is just one example, most of those now fit in one line with your last changes." ]

[]

[]

[ "I like this PR in general! \r\n\r\nJust wondering about two things:\r\n\r\n1) Do we need this `get_config` function?\r\n2) Not a huge fan of the `Add()` keras layer...does this really improve performance much?", "Good point @LysandreJik! Basically here most of the models share the similar embedding computation that stay inside their respective file. What has been exported is just the specific computation, which means that `WordEmbeddings`, `PositionalEmbeddings` and `TokenTypeEmbeddings` are always the same doesn't matter who is using it.\r\n\r\nThe same logic that is currently applied to `TFSharedEmbeddings`.", "> Just reviewed the general approach on one model for now and I have some questions before going further. If I understand correctly, the computation of the three different types of embeddings is split in three different ways to maximize the speedup but I wonder if it's documented from TF or just some tests on one particular setup. Before adding the extra complexity, I would like to be sure it brings a speedup on almost all possible environments (CPU, GPU, multi-GPU, TPU) without any loss in memory footprint (one-hot encoding the token type ids seems harmless, but we never know).\r\n\r\nI basically took example on the official implementation of Transformer encoder available in the Google Repo https://github.com/tensorflow/models/tree/master/official/nlp/keras_nlp . After having done several experiments (only on CPU and GPU though), I end up to extract from this an optimal version for each embedding.\r\n\r\n> As for putting those in modeling utils versus the model file, I agree with Lysandre that this breaks our philosophy of putting everything in each model file. I emitted the same reserves for TFSharedEmbeddings when it was introduced.\r\n\r\nI don't mind to copy/paste the same layers in all the concerned files if it is the recommended way. @sgugger @LysandreJik Will you be more confident if I create a version for each model and add the comment `# copied from ....` everytime it is a strong copy/paste?\r\n\r\n> I don't understand how it can be used above (line 420) in a tf.matmul if it's a layer and not a weight.\r\n\r\nNow the `get_input_embeddings` returns a `WordEmbeddings` layer that has a `word_embeddings` attribute. If you look at the Bert model for example, the layer `TFBertLMPredictionHead` takes a `WordEmbeddings` layer as `input_embeddings` and use the `WordEmbeddings.word_embeddings` attribute into the `tf.matmul`.", "> Now the `get_input_embeddings` returns a `WordEmbeddings` layer that has a `word_embeddings` attribute. If you look at the Bert model for example, the layer `TFBertLMPredictionHead` takes a `WordEmbeddings` layer as `input_embeddings` and use the `WordEmbeddings.word_embeddings` attribute into the `tf.matmul`.\r\n\r\nSo this part confuses me. Why name `word_embeddings` the weights inside the `WordEmbeddings`? It causes so much headache when reading the code afterward as we keep seeing some `word_embeddings` attributes which might either be an embedding layer or a weight.\r\n\r\nAlso, how does the new organization not screw up pretrained weights? From what I understand, the old `world_embeddings` in the `BertEmbeddings` layer used to be a weight and now it's a layer with an added `world_embeddings` attribute?", "> So this part confuses me. Why name word_embeddings the weights inside the WordEmbeddings? It causes so much headache when reading the code afterward as we keep seeing some word_embeddings attributes which might either be an embedding layer or a weight.\r\n\r\nI agree it is confusing, if you prefer it can be called `weight` such as in `TFSharedEmbeddings` I think it would be a more suitable name. This renaming will make easier the kind of checking (from the incoming PR on ebd resizing)\r\n```python\r\n def _get_word_embedding_weight(self, embedding_layer):\r\n if hasattr(embedding_layer, \"word_embeddings\"):\r\n return embedding_layer.word_embeddings\r\n elif hasattr(embedding_layer, \"weight\"):\r\n return embedding_layer.weight\r\n elif hasattr(embedding_layer, \"decoder\"):\r\n return embedding_layer.decoder\r\n else:\r\n # Here we build the word embeddings weights if not exists.\r\n # And then we retry to get the attribute once built.\r\n self(self.dummy_inputs)\r\n if hasattr(embedding_layer, \"word_embeddings\"):\r\n return embedding_layer.word_embeddings\r\n elif hasattr(embedding_layer, \"weight\"):\r\n return embedding_layer.weight\r\n elif hasattr(embedding_layer, \"decoder\"):\r\n return embedding_layer.decoder\r\n else:\r\n return None\r\n```\r\nNo more `word_embeddings` or `weight`, only `weight`. What do you think?\r\n\r\n> Also, how does the new organization not screw up pretrained weights? From what I understand, the old world_embeddings in the BertEmbeddings layer used to be a weight and now it's a layer with an added world_embeddings attribute?\r\n\r\nThis is because before we where using a [name score](https://www.tensorflow.org/api_docs/python/tf/name_scope) and not anymore in this PR. Let's say that defining a name scope or creating a layer represents the same thing. In both cases the weight is named `'tf_bert_model/bert/embeddings/word_embeddings/weight:0'` until now the `word_embeddings` part of the naming was because the embeddings was created in the context of `tf.name_scope(\"word_embeddings\"):` , in this PR it has the same name but because of the name of the new `WordEmbeddings` layer.", "Yes, having only \"weight\" makes more sense to me, and it would make the code easier to read. Thanks for explaining why the name of the weight doesn't change for loading!", "I found another advantage of these new embedding computation. It allows our models to be compiled in XLA_GPU and XLA_TPU which was not the case before. Small proof test on a machine with a GPU:\r\n```python\r\nfrom transformers import TFBertModel\r\nimport tensorflow as tf\r\n\r\nmodel = TFBertModel.from_pretrained(\"bert-base-cased\")\r\n\r\[email protected](experimental_compile=True)\r\ndef run():\r\n return model(model.dummy_inputs)\r\n\r\noutputs = run()\r\n```\r\nOn master fails with:\r\n```\r\ntensorflow.python.framework.errors_impl.InvalidArgumentError: Trying to access resource _AnonymousVar4 located in device /job:localhost/replica:0/task:0/device:CPU:0 from device /job:localhost/replica:0/task:0/device:GPU:0 [Op:__inference_run_4637]\r\n```\r\n\r\nOn this PR works as expected. The reason is because the `tf.keras.layers.Embeddings` layers are initialized when the model is instanciated instead of being initialized at build time.", "Now, each model has its own `WordEmbedding`, `TokenTypeEmbeddings` and `PositionEmbedding` layer in the model file decorated with the comment `#Copied from...` and the `words_embeddings` weights have been renamed into `weight` to make it more understandable and aligned with the name in `TFSharedEmbeddings`.", "> LGTM in general. One thing I'm not 100% sure about is whether we really need to add keras layers like tf.keras.layers.Add() if we start doing this for the embeddings now, I'm wondering if we should do the same for all residual connections in the self-attention blocks\r\n\r\nIn the absolute, yes we should. In an ideal world, everytime TF proposes a function/layer for doing something we should use it, as it is part of the optimization process. I know and I understand that it might seems confusing and starts to diverge with what PT looks like." ]

[]

[]

[ "Hey @jethrokuan, we've merged a big BART PR yesterday just as a heads up, I think this might solve this problem for Bart -> could you check again?", "@patrickvonplaten I think it does solve this problem: code looks good and runs fine, model results not great, but possibly a mistake of mine. Thanks!" ]

[]

[ { "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 } ]

[ "This class is not necessary anymore: it was the same as `DataCollatorForLanguageModeling` while keeping the `nsp_labels` but `DataCollatorForLanguageModeling` will keep any extra things (like `nsp_labels`) you pass to it. So you can just replace it with `DataCollatorForLanguageModeling`.", "Thank you for your quick reply. \r\nDo you mean you just have to use TextDatasetForNextSentencePrediction before DataCollatorForLanguageModeling to conduct NSP?", "Yes.", "That makes sense.\r\nThank you very much for your help!" ]

[]

[]

[ "Hi there, those kinds of questions should be asked on the [forums](https://discuss.huggingface.co/) as we keep issues for bugs and feature requests only, so closing this.\r\n\r\nFor a quick answer (move the discussion to the forums for a longer discussion!), it's normal that each GPU takes a slightly different time to train as all CUDA operations are asynchronous and the program is launched twice in parallel to be executed by both. It's also normal to see two different losses as the loss is not gathered across devices during training, only the gradients.", "Thank you for the answer. I will move this to forums." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thanks!" ]

[]

[]

[ "@sgugger \r\nThank you for checking and merging this PR!" ]

[]

[]

[ "I retried to install `transformers [dev]` with Python 3.9.1. \r\nThe latest `git tag` in the cloned repository is `v4.2.1`.\r\n\r\nI assumed that the same error would occur, but in this time it failed in installing `tensorflow`. \r\n\r\n``` sh\r\n****@**** $ conda create -n transformers-py39-dev\r\nCollecting package metadata (current_repodata.json): done\r\nSolving environment: done\r\n\r\n## Package Plan ##\r\n\r\n environment location: ****/.pyenv/versions/anaconda3-2020.07/envs/transformers-py39-dev\r\n\r\n\r\n\r\nProceed ([y]/n)? y\r\n\r\nPreparing transaction: done\r\nVerifying transaction: done\r\nExecuting transaction: done\r\n#\r\n# To activate this environment, use\r\n#\r\n# $ conda activate transformers-py39-dev\r\n#\r\n# To deactivate an active environment, use\r\n#\r\n# $ conda deactivate\r\n\r\n****@**** $ conda activate transformers-py39-dev\r\n(transformers-py39-dev) ****@**** $ conda install pip\r\nCollecting package metadata (current_repodata.json): done\r\nSolving environment: done\r\n\r\n## Package Plan ##\r\n\r\n environment location: ****/.pyenv/versions/anaconda3-2020.07/envs/transformers-py39-dev\r\n\r\n added / updated specs:\r\n - pip\r\n\r\n\r\nThe following packages will be downloaded:\r\n\r\n package | build\r\n ---------------------------|-----------------\r\n setuptools-51.1.2 | py39h06a4308_4 743 KB\r\n ------------------------------------------------------------\r\n Total: 743 KB\r\n\r\nThe following NEW packages will be INSTALLED:\r\n\r\n _libgcc_mutex pkgs/main/linux-64::_libgcc_mutex-0.1-main\r\n ca-certificates pkgs/main/linux-64::ca-certificates-2020.12.8-h06a4308_0\r\n certifi pkgs/main/linux-64::certifi-2020.12.5-py39h06a4308_0\r\n ld_impl_linux-64 pkgs/main/linux-64::ld_impl_linux-64-2.33.1-h53a641e_7\r\n libedit pkgs/main/linux-64::libedit-3.1.20191231-h14c3975_1\r\n libffi pkgs/main/linux-64::libffi-3.3-he6710b0_2\r\n libgcc-ng pkgs/main/linux-64::libgcc-ng-9.1.0-hdf63c60_0\r\n libstdcxx-ng pkgs/main/linux-64::libstdcxx-ng-9.1.0-hdf63c60_0\r\n ncurses pkgs/main/linux-64::ncurses-6.2-he6710b0_1\r\n openssl pkgs/main/linux-64::openssl-1.1.1i-h27cfd23_0\r\n pip pkgs/main/linux-64::pip-20.3.3-py39h06a4308_0\r\n python pkgs/main/linux-64::python-3.9.1-hdb3f193_2\r\n readline pkgs/main/linux-64::readline-8.0-h7b6447c_0\r\n setuptools pkgs/main/linux-64::setuptools-51.1.2-py39h06a4308_4\r\n sqlite pkgs/main/linux-64::sqlite-3.33.0-h62c20be_0\r\n tk pkgs/main/linux-64::tk-8.6.10-hbc83047_0\r\n tzdata pkgs/main/noarch::tzdata-2020d-h14c3975_0\r\n wheel pkgs/main/noarch::wheel-0.36.2-pyhd3eb1b0_0\r\n xz pkgs/main/linux-64::xz-5.2.5-h7b6447c_0\r\n zlib pkgs/main/linux-64::zlib-1.2.11-h7b6447c_3\r\n```\r\n\r\n``` sh\r\n(transformers-py39-dev) ****@**** $ pwd\r\n****/workspace/Clone/transformers\r\n(transformers-py39-dev) ****@**** $ pip install -e \".[dev]\"\r\nObtaining file:///****/workspace/Clone/transformers\r\n Installing build dependencies ... done\r\n Getting requirements to build wheel ... done\r\n Preparing wheel metadata ... done\r\nERROR: Could not find a version that satisfies the requirement tensorflow>=2.3; extra == \"dev\" (from transformers[dev])\r\nERROR: No matching distribution found for tensorflow>=2.3; extra == \"dev\"\r\n```\r\n\r\nIs it possible that you have decided not to support Python 3.9+ at this time because of the compatibility with the libraries `transformers` depends on?\r\n\r\nI apologize if there is any misunderstanding.\r\n", "You can use transformers without TensorFlow or FLAX installed, there is nothing in the code of transformers that is incompatible with Python 3.9. It looks like you want TensorFlow support for Python 3.9, which you should ask on the TensorFlow GitHub.", "@sgugger \r\nThank you for your comment.\r\nExcuse me for making you confused. It seems that there was a lack of information in my explanation.\r\n\r\nIn this case, my aim is not to use transformers with TensorFlow or FLAX.\r\nWhat I'd like to do is install `transformers [dev]` to open PRs in the future, so I'm a bit confused about whether I can install it with Python 3.9+.\r\nI'm not familiar with installing a `[dev]` version software, so I opened this issue to ask if we can install the [dev] version transformers with Python 3.9+ and open PRs using it.\r\n\r\nI can use Python <= 3.8, so this question is not urgent.\r\nI apologize for making you confused.", "You will be able to open PRs without installing `transformers [dev]`, it just mean you won't be able to run all the tests locally. \r\n`pip install transformers [torch, sentencepiece, tokenizers, testing, quality, ja, docs, sklearn, modelcreation]` might work to install all the depencies except TensorFlow and Flax (I just took all what is in dev and removed TensorFlow and Flax to create this command) but no guarantee.\r\n\r\nIf you're not an advanced user, I would recommend sticking with Python 3.6 to 3.8 while waiting for TensorFlow and Flax to support Python 3.9, as installing things with it might have some challenges :-)", "Hi @sgugger,\r\n\r\nThank you for telling me how to install it!\r\n\r\nWhen I tried to open a PR before, the auto-formatting of the code didn't work properly (I think it was when I tried to open a PR in `datasets`, not in `transformer`), and I assumed that I had to use `[dev]` versions when I want to open a PR. \r\nNow I think that the matter was caused by my have not installed the proper version of `testing`, `quality`, and `docs` then.\r\n\r\nI would like to become an advanced user eventually, but not now, so I would like to use Python 3.6 to 3.8 for now.\r\n\r\nThanks again!\r\n", "Yep getting the same error, \r\nOn a fresh 3.9 Python conda:\r\n`ERROR: No matching distribution found for jaxlib==0.1.55; extra == \"dev\"`", "To fix it, I moved back to Python 3.8.8, then `pip install -e \".[dev]\"` worked fine" ]

[]

[]

[ "Thanks!" ]

[]

[]

[ "The styling script does indeed break your list in this instance, which is kind of a bug that is a feature instead. Let me explain. If you use sphinx to convert this docstring to HTML, here is the result that it will produce:\r\n\r\nSo the styler is really only showing you in advance there is going to be a problem with your list when putting everything in the same paragraph.\r\n\r\nTo avoid the breaking (and properly rendering your list in the docs), you have to add a new empty line:\r\n```\r\n \"\"\"\r\n number of training steps is either \r\n\r\n 1. args.max_steps if --max_steps > 1 \r\n 2. else derive from dataset if we can get its size\r\n \"\"\"\r\n```\r\n\r\nI know it's a bit annoying for docstrings that are just there as internal documentation and not really designed to be shown in the main documentation, but the script can't guess which docstrings to check and which not...\r\n\r\n> I also am not sure why there is a need to merge lines when the writer meant them to be shorter.\r\n\r\nAgain, this will be shown as one paragraph in the actual documentation. If you want to keep lines separated, they need to have an extra new line in-between.", "Thank you for explaining that it is the sphinx that is lacking.\r\n\r\nCould the autoformatter detect such situations and fix that so that it remains a list by inserting a new line, rather than unwrapping the whole thing?\r\n\r\nIf such a parser would be complicated we could make it easier by having a stricter format. Usually, in English a proper list is preceded by a colon as in:\r\n```\r\nHere is what you do:\r\n1. ....\r\n2. ....\r\n```\r\nSo `r':\\s*[\\r\\n]+\\s+(\\d+\\.|[\\-\\*] )'` would match 3 types of lists." ]

[]

[]

[ "+1 on the `model_revision` part\r\n\r\nOn the `use_auth_token`, I was thinking we could try to implement auto-use of the token _iff_ the model is not public (i.e. send the token for non-existent models and private models, as the server doesn't make a difference for those two cases when unauthorized – you get a 404 in both cases – said differently if you don't have access to a model you shouldn't know whether it's an existing private model)\r\n\r\nThis will require some implementation changes in file_utils though so it might take a bit of time. \r\n\r\nIf you think it's helpful to expose this PR's manual option first, I'm ok with that.", "@LysandreJik Yep totally right! I won't personally get around to adding the feature in file_utils/huggingface_hub in the next 2-3 weeks though, so maybe worth it to merge it like this in the meantime:)", "I think it's important to provide the option right now to let the user play with their private models for those scripts. We can have that flag become `None` later on and default to the right thing when the implementation in `file_utils` permits it then remove it entirely a bit later.", "sounds good", "Sounds good!" ]

[]

[]

[ "Hello! Can you try with master instead of the old `2.0.0` release? In order to know if the problem is still here or not.", "By \"with master\", do you mean installed from source?\r\n\r\ngit clone https://github.com/huggingface/transformers.git\r\ncd transformers\r\npip install -e .", "After installing from source transformers 4.2.0.dev0,\r\n\r\nI have this error:\r\n\r\n```\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n<ipython-input-2-1503c944af5c> in <module>\r\n----> 1 from transformers import AutoTokenizer, TFAutoModel\r\n\r\n~/transformers/src/transformers/__init__.py in <module>\r\n 38 \r\n 39 # Data\r\n---> 40 from .data import (\r\n 41 DataProcessor,\r\n 42 InputExample,\r\n\r\n~/transformers/src/transformers/data/__init__.py in <module>\r\n 18 \r\n 19 from .metrics import glue_compute_metrics, xnli_compute_metrics\r\n---> 20 from .processors import (\r\n 21 DataProcessor,\r\n 22 InputExample,\r\n\r\n~/transformers/src/transformers/data/processors/__init__.py in <module>\r\n 18 \r\n 19 from .glue import glue_convert_examples_to_features, glue_output_modes, glue_processors, glue_tasks_num_labels\r\n---> 20 from .squad import SquadExample, SquadFeatures, SquadV1Processor, SquadV2Processor, squad_convert_examples_to_features\r\n 21 from .utils import DataProcessor, InputExample, InputFeatures, SingleSentenceClassificationProcessor\r\n 22 from .xnli import xnli_output_modes, xnli_processors, xnli_tasks_num_labels\r\n\r\n~/transformers/src/transformers/data/processors/squad.py in <module>\r\n 22 \r\n 23 from ...file_utils import is_tf_available, is_torch_available\r\n---> 24 from ...models.bert.tokenization_bert import whitespace_tokenize\r\n 25 from ...tokenization_utils_base import BatchEncoding, PreTrainedTokenizerBase, TruncationStrategy\r\n 26 from ...utils import logging\r\n\r\n~/transformers/src/transformers/models/bert/__init__.py in <module>\r\n 43 \r\n 44 if is_tf_available():\r\n---> 45 from .modeling_tf_bert import (\r\n 46 TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,\r\n 47 TFBertEmbeddings,\r\n\r\n~/transformers/src/transformers/models/bert/modeling_tf_bert.py in <module>\r\n 21 import tensorflow as tf\r\n 22 \r\n---> 23 from ...activations_tf import get_tf_activation\r\n 24 from ...file_utils import (\r\n 25 MULTIPLE_CHOICE_DUMMY_INPUTS,\r\n\r\n~/transformers/src/transformers/activations_tf.py in <module>\r\n 66 \"gelu\": tf.keras.layers.Activation(gelu),\r\n 67 \"relu\": tf.keras.activations.relu,\r\n---> 68 \"swish\": tf.keras.activations.swish,\r\n 69 \"silu\": tf.keras.activations.swish,\r\n 70 \"gelu_new\": tf.keras.layers.Activation(gelu_new),\r\n\r\nAttributeError: module 'tensorflow_core.python.keras.api._v2.keras.activations' has no attribute 'swish'\r\n```\r\n\r\nIt's due to the line of code:\r\nfrom transformers import AutoTokenizer, TFAutoModel", "The next release of transformers (from source) now requires TF >= 2.3", "It seems to work now, but I have a lot of warnings. Should I be worried about any of them?\r\n\r\n```\r\nWARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7fb1e4e82280>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nWARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7fb1e4e82280>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nWARNING: AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7fb1e4e82280>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\nWARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.\r\n```", "They look ok for me!", "I'm fine-tuning the model on my dataset and the accuracy is 0.02 after one epoch and it didn't really change during training. Also, it takes 7 hours per epoch. I'm wondering if the low accuracy might be due to the things mentioned in the warnings.\r\n\r\n`WARNING:tensorflow:Gradients do not exist for variables ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0'] when minimizing the loss.`\r\nIf there are no gradients, it cannot learn. Do you know if it's really correct that all those layers have no gradient? I would expect the layers to have gradients.\r\n\r\n```\r\nWARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7fa069898210>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nWARNING: AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7fa069898210>> and will run it as-is.\r\n```\r\nThis warning seems to say that it is a bug worth mentioning to the TensorFlow team. Could it be the cause of the bad training time?\r\n\r\n`The parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).`\r\nDo those parameters must be set? I have no set them or tried to modify them, I'm simply using the .fit method that all tensorflow 2 models have.\r\n", "Also, do we need to use training=True somewhere? It's mentioned to use it when using the call of the model, but I'm calling .fit() rather than using the model's call, so I don't have this option as far as I know.\r\n\r\nI'm asking because the training doesn't seem to be working and I'm wondering if it could be the problem.", "Unfortunately, no issues for me with `TFXLNetForSequenceClassification`, just test over the MRPC dataset and got around 0.93 accuracy on training and 0.83 accuracy on validation. The version I used is the master branch from source.\r\n\r\nThe issue might come from the way you are training the model. Are you using the same one that you shared in your first post?", "yes, but some parameters are different. For example, I needed a batchsize of 1 to fit in memory. Even 2 crashes. Also, I'm using callbacks. But I don't think it can make that the model doesn't learn.\r\n\r\n # Save the model after each epoch.\r\n ModelCheckpoint_callback = tf.keras.callbacks.ModelCheckpoint(\r\n filepath = self.params['save_model_weight_filepath']+'_{epoch:02d}.hdf5', \r\n monitor='val_loss', verbose=0, save_best_only=False,\r\n save_weights_only=False, mode='auto', save_freq='epoch'\r\n )\r\n \r\n # Stop when val loss stops decreasing.\r\n EarlyStopping_callback = tf.keras.callbacks.EarlyStopping(\r\n monitor='val_loss', min_delta=self.params['min_delta'], patience=self.params['patience'], verbose=0,\r\n mode='auto', baseline=None, restore_best_weights=True\r\n\r\n history = self.clf.fit(x=padded_inputs, y=y,\r\n batch_size=1,\r\n epochs=40,\r\n verbose=1,\r\n validation_split=0.2,\r\n max_queue_size=10,\r\n workers=-1,\r\n use_multiprocessing=True,\r\n callbacks=[ModelCheckpoint_callback, EarlyStopping_callback])\r\n\r\nAfter epoch 2, the accuracies and losses are loss: 6.7701 - accuracy: 0.0197 - val_loss: 11.5031 - val_accuracy: 0.0025\r\n\r\nEpoch 3 is still in process with loss: 6.7662 - accuracy: 0.0204\r\n\r\nIt doesn't seem to learn at all.\r\n\r\nAlso, I have this warning \r\n`WARNING:tensorflow:Callbacks method `on_test_batch_end` is slow compared to the batch time (batch time: 0.0089s vs `on_test_batch_end` time: 0.6593s). Check your callbacks.`\r\nbut none of my callbacks are used on batch_end, they are used on epoch ends, so infrequently and shouldn't affect the time too much.", "Ok, from what I see in your script, the reason why your model don't learn anything is because the labels are not seen by the model which is normal with the way you set your dataset. The models in the lib have to be feed in a specific way, the data have to be a `Tuple(x, y)` where `x` can be either a list or a dict with tf.Tensor or np.ndarray, same for `y`. And then feed your model with:\r\n\r\n```python\r\nhistory = model.fit(\r\n train_dataset,\r\n epochs=3,\r\n)\r\n```\r\n\r\nYou can take example on how to do this in our examples or on our datasets website https://huggingface.co/docs/datasets/torch_tensorflow.html to know how to format your dataset.", "I tried that (using a tuple with x and y, my x and y were already numpy arrays) and I got an error.\r\n\r\n```\r\n~/ticket-analysis-releasev3/ticket-analysis/src/model/xlnet/xlnet.py in fit(self, df)\r\n 181 workers=self.params['workers'],\r\n 182 use_multiprocessing=self.params['use_multiprocessing'],\r\n--> 183 callbacks=[ModelCheckpoint_callback, EarlyStopping_callback])\r\n 184 self.history_df = pd.DataFrame({'epochs':history.epoch, 'loss': history.history['loss'], \r\n 185 'validation_loss': history.history['val_loss'], 'accuracy': history.history['accuracy'],\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)\r\n 106 def _method_wrapper(self, *args, **kwargs):\r\n 107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access\r\n--> 108 return method(self, *args, **kwargs)\r\n 109 \r\n 110 # Running inside `run_distribute_coordinator` already.\r\n\r\n~/.local/lib/python3.7/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 1096 batch_size=batch_size):\r\n 1097 callbacks.on_train_batch_begin(step)\r\n-> 1098 tmp_logs = train_function(iterator)\r\n 1099 if data_handler.should_sync:\r\n 1100 context.async_wait()\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)\r\n 778 else:\r\n 779 compiler = \"nonXla\"\r\n--> 780 result = self._call(*args, **kwds)\r\n 781 \r\n 782 new_tracing_count = self._get_tracing_count()\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _call(self, *args, **kwds)\r\n 821 # This is the first call of __call__, so we have to initialize.\r\n 822 initializers = []\r\n--> 823 self._initialize(args, kwds, add_initializers_to=initializers)\r\n 824 finally:\r\n 825 # At this point we know that the initialization is complete (or less\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _initialize(self, args, kwds, add_initializers_to)\r\n 695 self._concrete_stateful_fn = (\r\n 696 self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access\r\n--> 697 *args, **kwds))\r\n 698 \r\n 699 def invalid_creator_scope(*unused_args, **unused_kwds):\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)\r\n 2853 args, kwargs = None, None\r\n 2854 with self._lock:\r\n-> 2855 graph_function, _, _ = self._maybe_define_function(args, kwargs)\r\n 2856 return graph_function\r\n 2857 \r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)\r\n 3211 \r\n 3212 self._function_cache.missed.add(call_context_key)\r\n-> 3213 graph_function = self._create_graph_function(args, kwargs)\r\n 3214 self._function_cache.primary[cache_key] = graph_function\r\n 3215 return graph_function, args, kwargs\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)\r\n 3073 arg_names=arg_names,\r\n 3074 override_flat_arg_shapes=override_flat_arg_shapes,\r\n-> 3075 capture_by_value=self._capture_by_value),\r\n 3076 self._function_attributes,\r\n 3077 function_spec=self.function_spec,\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)\r\n 984 _, original_func = tf_decorator.unwrap(python_func)\r\n 985 \r\n--> 986 func_outputs = python_func(*func_args, **func_kwargs)\r\n 987 \r\n 988 # invariant: `func_outputs` contains only Tensors, CompositeTensors,\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)\r\n 598 # __wrapped__ allows AutoGraph to swap in a converted function. We give\r\n 599 # the function a weak reference to itself to avoid a reference cycle.\r\n--> 600 return weak_wrapped_fn().__wrapped__(*args, **kwds)\r\n 601 weak_wrapped_fn = weakref.ref(wrapped_fn)\r\n 602 \r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)\r\n 971 except Exception as e: # pylint:disable=broad-except\r\n 972 if hasattr(e, \"ag_error_metadata\"):\r\n--> 973 raise e.ag_error_metadata.to_exception(e)\r\n 974 else:\r\n 975 raise\r\n\r\nValueError: in user code:\r\n\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:806 train_function *\r\n return step_function(self, iterator)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:796 step_function **\r\n outputs = model.distribute_strategy.run(run_step, args=(data,))\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/distribute/distribute_lib.py:1211 run\r\n return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/distribute/distribute_lib.py:2585 call_for_each_replica\r\n return self._call_for_each_replica(fn, args, kwargs)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/distribute/distribute_lib.py:2945 _call_for_each_replica\r\n return fn(*args, **kwargs)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:789 run_step **\r\n outputs = model.train_step(data)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:757 train_step\r\n self.trainable_variables)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:2737 _minimize\r\n trainable_variables))\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:562 _aggregate_gradients\r\n filtered_grads_and_vars = _filter_grads(grads_and_vars)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:1271 _filter_grads\r\n ([v.name for _, v in grads_and_vars],))\r\n\r\n ValueError: No gradients provided for any variable: ['tfxl_net_for_sequence_classification/transformer/mask_emb:0', 'tfxl_net_for_sequence_classification/transformer/word_embedding/weight:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._0/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._1/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._2/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._3/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._4/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._5/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._6/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._7/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._8/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._9/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._10/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/q:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/k:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/v:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/o:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_r_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_s_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/r_w_bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/seg_embed:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/rel_attn/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/ff/layer_norm/gamma:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/ff/layer_norm/beta:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/ff/layer_1/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/ff/layer_1/bias:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/ff/layer_2/kernel:0', 'tfxl_net_for_sequence_classification/transformer/layer_._11/ff/layer_2/bias:0', 'tfxl_net_for_sequence_classification/sequence_summary/summary/kernel:0', 'tfxl_net_for_sequence_classification/sequence_summary/summary/bias:0', 'tfxl_net_for_sequence_classification/logits_proj/kernel:0', 'tfxl_net_for_sequence_classification/logits_proj/bias:0'].\r\n```", "I also tried with dataset, but I got this error\r\n\r\n```\r\nSome layers from the model checkpoint at xlnet-base-cased were not used when initializing TFXLNetForSequenceClassification: ['lm_loss']\r\n- This IS expected if you are initializing TFXLNetForSequenceClassification 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 TFXLNetForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome layers of TFXLNetForSequenceClassification were not initialized from the model checkpoint at xlnet-base-cased and are newly initialized: ['sequence_summary', 'logits_proj']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\nEpoch 1/40\r\nWARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f21748f1210>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nWARNING: AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f21748f1210>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n<ipython-input-6-db016531efec> in <module>\r\n 1 # Try to use tensorflow dataset\r\n 2 st = time.time()\r\n----> 3 model.fit(df)\r\n 4 print('time', time.time()-st)\r\n\r\n~/ticket-analysis-releasev3/ticket-analysis/src/model/xlnet/xlnet.py in fit(self, df)\r\n 187 workers=self.params['workers'],\r\n 188 use_multiprocessing=self.params['use_multiprocessing'],\r\n--> 189 callbacks=[ModelCheckpoint_callback, EarlyStopping_callback])\r\n 190 self.history_df = pd.DataFrame({'epochs':history.epoch, 'loss': history.history['loss'], \r\n 191 'validation_loss': history.history['val_loss'], 'accuracy': history.history['accuracy'],\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py in _method_wrapper(self, *args, **kwargs)\r\n 106 def _method_wrapper(self, *args, **kwargs):\r\n 107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access\r\n--> 108 return method(self, *args, **kwargs)\r\n 109 \r\n 110 # Running inside `run_distribute_coordinator` already.\r\n\r\n~/.local/lib/python3.7/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 1096 batch_size=batch_size):\r\n 1097 callbacks.on_train_batch_begin(step)\r\n-> 1098 tmp_logs = train_function(iterator)\r\n 1099 if data_handler.should_sync:\r\n 1100 context.async_wait()\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in __call__(self, *args, **kwds)\r\n 778 else:\r\n 779 compiler = \"nonXla\"\r\n--> 780 result = self._call(*args, **kwds)\r\n 781 \r\n 782 new_tracing_count = self._get_tracing_count()\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _call(self, *args, **kwds)\r\n 821 # This is the first call of __call__, so we have to initialize.\r\n 822 initializers = []\r\n--> 823 self._initialize(args, kwds, add_initializers_to=initializers)\r\n 824 finally:\r\n 825 # At this point we know that the initialization is complete (or less\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in _initialize(self, args, kwds, add_initializers_to)\r\n 695 self._concrete_stateful_fn = (\r\n 696 self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access\r\n--> 697 *args, **kwds))\r\n 698 \r\n 699 def invalid_creator_scope(*unused_args, **unused_kwds):\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)\r\n 2853 args, kwargs = None, None\r\n 2854 with self._lock:\r\n-> 2855 graph_function, _, _ = self._maybe_define_function(args, kwargs)\r\n 2856 return graph_function\r\n 2857 \r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _maybe_define_function(self, args, kwargs)\r\n 3211 \r\n 3212 self._function_cache.missed.add(call_context_key)\r\n-> 3213 graph_function = self._create_graph_function(args, kwargs)\r\n 3214 self._function_cache.primary[cache_key] = graph_function\r\n 3215 return graph_function, args, kwargs\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)\r\n 3073 arg_names=arg_names,\r\n 3074 override_flat_arg_shapes=override_flat_arg_shapes,\r\n-> 3075 capture_by_value=self._capture_by_value),\r\n 3076 self._function_attributes,\r\n 3077 function_spec=self.function_spec,\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)\r\n 984 _, original_func = tf_decorator.unwrap(python_func)\r\n 985 \r\n--> 986 func_outputs = python_func(*func_args, **func_kwargs)\r\n 987 \r\n 988 # invariant: `func_outputs` contains only Tensors, CompositeTensors,\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/eager/def_function.py in wrapped_fn(*args, **kwds)\r\n 598 # __wrapped__ allows AutoGraph to swap in a converted function. We give\r\n 599 # the function a weak reference to itself to avoid a reference cycle.\r\n--> 600 return weak_wrapped_fn().__wrapped__(*args, **kwds)\r\n 601 weak_wrapped_fn = weakref.ref(wrapped_fn)\r\n 602 \r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)\r\n 971 except Exception as e: # pylint:disable=broad-except\r\n 972 if hasattr(e, \"ag_error_metadata\"):\r\n--> 973 raise e.ag_error_metadata.to_exception(e)\r\n 974 else:\r\n 975 raise\r\n\r\nValueError: in user code:\r\n\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py:806 train_function *\r\n return step_function(self, iterator)\r\n /home/jovyan/transformers/src/transformers/models/xlnet/modeling_tf_xlnet.py:1452 call *\r\n transformer_outputs = self.transformer(\r\n /home/jovyan/transformers/src/transformers/models/xlnet/modeling_tf_xlnet.py:625 call *\r\n inputs[\"input_ids\"] = tf.transpose(inputs[\"input_ids\"], perm=(1, 0))\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/util/dispatch.py:201 wrapper **\r\n return target(*args, **kwargs)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/ops/array_ops.py:2107 transpose_v2\r\n return transpose(a=a, perm=perm, name=name, conjugate=conjugate)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/util/dispatch.py:201 wrapper\r\n return target(*args, **kwargs)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/ops/array_ops.py:2188 transpose\r\n return transpose_fn(a, perm, name=name)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/ops/gen_array_ops.py:11535 transpose\r\n \"Transpose\", x=x, perm=perm, name=name)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/framework/op_def_library.py:744 _apply_op_helper\r\n attrs=attr_protos, op_def=op_def)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/framework/func_graph.py:593 _create_op_internal\r\n compute_device)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py:3485 _create_op_internal\r\n op_def=op_def)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py:1975 __init__\r\n control_input_ops, op_def)\r\n /home/jovyan/.local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py:1815 _create_c_op\r\n raise ValueError(str(e))\r\n\r\n ValueError: Dimension must be 3 but is 2 for '{{node tfxl_net_for_sequence_classification/transformer/transpose}} = Transpose[T=DT_INT32, Tperm=DT_INT32](IteratorGetNext, tfxl_net_for_sequence_classification/transformer/transpose/perm)' with input shapes: [1,11929,2000], [2].\r\n```\r\n\r\nFor dataset, I used this code\r\n```\r\n train_dataset = (padded_inputs, y)\r\n train_dataset, val_dataset = train_test_split(train_dataset, test_size=0.02)\r\n train_dataset = tf.data.Dataset.from_tensors(train_dataset)\r\n val_dataset = tf.data.Dataset.from_tensors(val_dataset)\r\n \r\n \r\n # Fit model\r\n self.clf = TFAutoModelForSequenceClassification.from_pretrained(\"xlnet-base-cased\", num_labels=self.n_label)\r\n loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\r\n self.clf.compile(optimizer='adam',loss=loss,metrics=['accuracy'])\r\n history = self.clf.fit(\r\n train_dataset, #x=padded_inputs, y=y,\r\n validation_data = val_dataset,\r\n batch_size=self.params['batchsize'],\r\n epochs=self.params['epochs'],\r\n verbose=1,\r\n #validation_split=self.params['validation_split'],\r\n max_queue_size=self.params['max_queue_size'],\r\n workers=self.params['workers'],\r\n use_multiprocessing=self.params['use_multiprocessing'],\r\n callbacks=[ModelCheckpoint_callback, EarlyStopping_callback])\r\n```", "This is still not ok. Here an example for MRPC, from which your can take inspiration from:\r\n```\r\nimport tensorflow as tf\r\nfrom datasets import load_dataset\r\nfrom transformers import AutoTokenizer\r\ndataset = load_dataset('glue', 'mrpc', split='train')\r\ntokenizer = AutoTokenizer.from_pretrained('bert-base-cased')\r\ndataset = dataset.map(lambda e: tokenizer(e['sentence1'], truncation=True, padding='max_length'), batched=True)\r\n\r\ndataset.set_format(type='numpy', columns=['input_ids', 'token_type_ids', 'attention_mask', 'label'])\r\nfeatures = {x: dataset[x] for x in ['input_ids', 'token_type_ids', 'attention_mask']}\r\ntfdataset = tf.data.Dataset.from_tensor_slices((features, dataset[\"label\"])).batch(1)\r\n```\r\n\r\nAnd your data must have the shape:\r\n```\r\n<BatchDataset shapes: ({input_ids: (None, 512), token_type_ids: (None, 512), attention_mask: (None, 512)}, (None,)), types: ({input_ids: tf.int64, token_type_ids: tf.int64, attention_mask: tf.int64}, tf.int32)>\r\n```\r\nHere it is a tuple, where the first element is a dict that has tensors (built from numpy arrays), and the label is a label id.", "What kind of format or object is the dataset obtained from\r\ndataset = load_dataset('glue', 'mrpc', split='train')\r\n?\r\nI'm not loading a public dataset but using my own so I can't take this part from the code. Do you know how I can generate it from a input numpy array and a label numpy array?\r\n\r\nAlso, I don't think xlnet uses 'token_type_ids', 'attention_mask', should I use\r\ndataset.set_format(type='numpy', columns=['input_ids', 'label'])\r\nfeatures = {x: dataset[x] for x in ['input_ids']}\r\n", "The example I gave is just to show you how should looks like your dataset. And yes XLNet can take both attention_mask and token_type_ids arguments. The steps are simple:\r\n1. Tokenize your dataset\r\n2. Create a tf.data.Dataset and format it to make it looking like I showed you: `({\"input_ids\": [[ex1],[ex2],...], \"attention_mask\":[[ex1],[ex2],...], \"token_type_ids\":[ex1],[ex2],...]}, [label_id_ex1, label_id_ex_2,...])`\r\n3. Run your training with `model.fit(training_dataset, epochs=3)`\r\n\r\nAnd that's it.\r\n\r\n", "Okay, I'm using\r\n\r\n x_train, x_val, y_train, y_val = train_test_split(x, y, test_size=0.02)\r\n tokenized_inputs = [xlnet_tokenizer.encode(text) for text in x_train.values.tolist()]\r\n max_length = max(1,min(np.array([len(inp) for inp in tokenized_inputs]).max(), self.params['MAX_LENGTH']))\r\n padded_inputs = (tf.keras.preprocessing.sequence.pad_sequences(tokenized_inputs, maxlen=max_length, \r\n value=0,\r\n padding='post', truncating='post',dtype='int32'))\r\n train_dataset = tf.data.Dataset.from_tensor_slices(({\"input_ids\":padded_inputs}, y_train)).batch(1)\r\n val_tokenized_inputs = [xlnet_tokenizer.encode(text) for text in x_val.values.tolist()]\r\n val_padded_inputs = (tf.keras.preprocessing.sequence.pad_sequences(val_tokenized_inputs, maxlen=max_length, \r\n value=0,\r\n padding='post', truncating='post',dtype='int32'))\r\n val_dataset = tf.data.Dataset.from_tensor_slices(({\"input_ids\":x_val}, y_val)).batch(1)\r\n print('dataset',train_dataset,val_dataset)\r\n history = self.clf.fit(\r\n train_dataset,\r\n validation_data = val_dataset,\r\n batch_size=self.params['batchsize'],\r\n epochs=self.params['epochs'],\r\n verbose=1,\r\n max_queue_size=self.params['max_queue_size'],\r\n workers=self.params['workers'],\r\n use_multiprocessing=self.params['use_multiprocessing'],\r\n callbacks=[ModelCheckpoint_callback, EarlyStopping_callback])\r\n\r\nThe format is <BatchDataset shapes: ({input_ids: (None, 2000)}, (None,)), types: ({input_ids: tf.int32}, tf.int16)>\r\n\r\nShould it work?\r\n\r\nAlso, for the prediction, is it correct to use\r\npred_dataset = tf.data.Dataset.from_tensor_slices(({\"input_ids\":x_train})).batch(1)\r\nsince the model should not use the y at prediction time.\r\n\r\nI'm getting this error right now\r\n```\r\n---------------------------------------------------------------------------\r\nInternalError Traceback (most recent call last)\r\n<ipython-input-6-db016531efec> in <module>\r\n 1 # Try to use tensorflow dataset\r\n 2 st = time.time()\r\n----> 3 model.fit(df)\r\n 4 print('time', time.time()-st)\r\n\r\n~/ticket-analysis-releasev3/ticket-analysis/src/model/xlnet/xlnet.py in fit(self, df)\r\n 168 \r\n 169 # Fit model\r\n--> 170 self.clf = TFAutoModelForSequenceClassification.from_pretrained(\"xlnet-base-cased\", num_labels=self.n_label)\r\n 171 loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\r\n 172 self.clf.compile(optimizer='adam',loss=loss,metrics=['accuracy'])\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/modeling_tf_auto.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)\r\n 384 return TFBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)\r\n 385 elif 'xlnet' in pretrained_model_name_or_path:\r\n--> 386 return TFXLNetForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)\r\n 387 elif 'xlm' in pretrained_model_name_or_path:\r\n 388 return TFXLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/modeling_tf_utils.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)\r\n 266 \r\n 267 inputs = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])\r\n--> 268 ret = model(inputs, training=False) # build the network with dummy inputs\r\n 269 \r\n 270 assert os.path.isfile(resolved_archive_file), \"Error retrieving file {}\".format(resolved_archive_file)\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, *args, **kwargs)\r\n 983 \r\n 984 with ops.enable_auto_cast_variables(self._compute_dtype_object):\r\n--> 985 outputs = call_fn(inputs, *args, **kwargs)\r\n 986 \r\n 987 if self._activity_regularizer:\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/modeling_tf_xlnet.py in call(self, inputs, **kwargs)\r\n 911 \r\n 912 def call(self, inputs, **kwargs):\r\n--> 913 transformer_outputs = self.transformer(inputs, **kwargs)\r\n 914 output = transformer_outputs[0]\r\n 915 \r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py in __call__(self, *args, **kwargs)\r\n 983 \r\n 984 with ops.enable_auto_cast_variables(self._compute_dtype_object):\r\n--> 985 outputs = call_fn(inputs, *args, **kwargs)\r\n 986 \r\n 987 if self._activity_regularizer:\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/modeling_tf_xlnet.py in call(self, inputs, attention_mask, mems, perm_mask, target_mapping, token_type_ids, input_mask, head_mask, training)\r\n 607 \r\n 608 ##### Positional encoding\r\n--> 609 pos_emb = self.relative_positional_encoding(qlen, klen, bsz=bsz, dtype=dtype_float)\r\n 610 pos_emb = self.dropout(pos_emb, training=training)\r\n 611 \r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/modeling_tf_xlnet.py in relative_positional_encoding(self, qlen, klen, bsz, dtype)\r\n 490 if self.clamp_len > 0:\r\n 491 fwd_pos_seq = tf.clip_by_value(fwd_pos_seq, -clamp_len, clamp_len)\r\n--> 492 pos_emb = self.positional_embedding(fwd_pos_seq, inv_freq, bsz)\r\n 493 \r\n 494 return pos_emb\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/modeling_tf_xlnet.py in positional_embedding(pos_seq, inv_freq, bsz)\r\n 437 @staticmethod\r\n 438 def positional_embedding(pos_seq, inv_freq, bsz=None):\r\n--> 439 sinusoid_inp = tf.einsum('i,d->id', pos_seq, inv_freq)\r\n 440 pos_emb = tf.concat([tf.sin(sinusoid_inp), tf.cos(sinusoid_inp)], axis=-1)\r\n 441 pos_emb = pos_emb[:, None, :]\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/util/dispatch.py in wrapper(*args, **kwargs)\r\n 199 \"\"\"Call target, and fall back on dispatchers if there is a TypeError.\"\"\"\r\n 200 try:\r\n--> 201 return target(*args, **kwargs)\r\n 202 except (TypeError, ValueError):\r\n 203 # Note: convert_to_eager_tensor currently raises a ValueError, not a\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/ops/special_math_ops.py in einsum(equation, *inputs, **kwargs)\r\n 682 - number of inputs or their shapes are inconsistent with `equation`.\r\n 683 \"\"\"\r\n--> 684 return _einsum_v2(equation, *inputs, **kwargs)\r\n 685 \r\n 686 \r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/ops/special_math_ops.py in _einsum_v2(equation, *inputs, **kwargs)\r\n 1111 if ellipsis_label:\r\n 1112 resolved_equation = resolved_equation.replace(ellipsis_label, '...')\r\n-> 1113 return gen_linalg_ops.einsum(inputs, resolved_equation)\r\n 1114 \r\n 1115 # Send fully specified shapes to opt_einsum, since it cannot handle unknown\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/ops/gen_linalg_ops.py in einsum(inputs, equation, name)\r\n 1086 return _result\r\n 1087 except _core._NotOkStatusException as e:\r\n-> 1088 _ops.raise_from_not_ok_status(e, name)\r\n 1089 except _core._FallbackException:\r\n 1090 pass\r\n\r\n~/.local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in raise_from_not_ok_status(e, name)\r\n 6841 message = e.message + (\" name: \" + name if name is not None else \"\")\r\n 6842 # pylint: disable=protected-access\r\n-> 6843 six.raise_from(core._status_to_exception(e.code, message), None)\r\n 6844 # pylint: enable=protected-access\r\n 6845 \r\n\r\n/opt/conda/lib/python3.7/site-packages/six.py in raise_from(value, from_value)\r\n\r\nInternalError: Blas xGEMM launch failed : a.shape=[1,1,10], b.shape=[1,1,384], m=10, n=384, k=1 [Op:Einsum]\r\n```", "This error means that your GPU doesn't have enough RAM to run an einsum operation. But yes, your dataset looks better. Still, you are not properly using the tokenizer, use a proper way to use it:\r\n```\r\nfrom transformers import XLNetTokenizer\r\ntokenizer = XLNetTokenizer.from_pretrained(\"xlnet-base-cased\")\r\ntokenizer(\"hello\")\r\n``` \r\nTo get tokenized data that looks like:\r\n```\r\n{'input_ids': [24717, 4, 3], 'token_type_ids': [0, 0, 2], 'attention_mask': [1, 1, 1]}\r\n```", "I'm getting\r\n\r\n```\r\n---------------------------------------------------------------------------\r\nTypeError Traceback (most recent call last)\r\n<ipython-input-1-4dd755fb81bd> in <module>\r\n 1 from transformers import XLNetTokenizer\r\n 2 tokenizer = XLNetTokenizer.from_pretrained(\"xlnet-base-cased\")\r\n----> 3 tokenizer(\"hello\")\r\n\r\nTypeError: 'XLNetTokenizer' object is not callable\r\n```", "Which version of transformers are you using?", "It went back to '2.0.0'. I don't know why but I'm trying to et 4.2.0 again.", "Please stick to the 4.2.0 release :)", "Do we need int64 or is int32 enough? I'm memory limited so anything that allow me to use less memory would help.", "int32 is enough.", "It is running again, but still not training.\r\n\r\nI'm using\r\n```\r\n x_train, x_val, y_train, y_val = train_test_split(x, y, test_size=self.params['validation_split'])\r\n # train\r\n tokenized_inputs = xlnet_tokenizer(x_train.values.tolist(), padding=True, max_length=self.params['MAX_LENGTH'], truncation=True)\r\n numpy_inputs = {x:np.array(tokenized_inputs[x]) for x in tokenized_inputs.keys()}\r\n train_dataset = tf.data.Dataset.from_tensor_slices((numpy_inputs, y_train)).batch(1)\r\n # val\r\n tokenized_inputs = xlnet_tokenizer(x_val.values.tolist(), padding=True, max_length=self.params['MAX_LENGTH'], truncation=True)\r\n numpy_inputs = {x:np.array(tokenized_inputs[x]) for x in tokenized_inputs.keys()}\r\n val_dataset = tf.data.Dataset.from_tensor_slices((numpy_inputs, y_val)).batch(1)\r\n print(train_dataset,val_dataset)\r\n history = self.clf.fit(\r\n train_dataset,\r\n validation_data = val_dataset,\r\n batch_size=self.params['batchsize'],\r\n epochs=self.params['epochs'],\r\n verbose=1,\r\n max_queue_size=self.params['max_queue_size'],\r\n workers=self.params['workers'],\r\n use_multiprocessing=self.params['use_multiprocessing'],\r\n callbacks=[ModelCheckpoint_callback, EarlyStopping_callback])\r\n```\r\n\r\nThe shapes are\r\n`<BatchDataset shapes: ({input_ids: (None, 500), token_type_ids: (None, 500), attention_mask: (None, 500)}, (None,)), types: ({input_ids: tf.int64, token_type_ids: tf.int64, attention_mask: tf.int64}, tf.int16)> <BatchDataset shapes: ({input_ids: (None, 500), token_type_ids: (None, 500), attention_mask: (None, 500)}, (None,)), types: ({input_ids: tf.int64, token_type_ids: tf.int64, attention_mask: tf.int64}, tf.int16)>`\r\n\r\nThe results are:\r\nepochs | loss | validation_loss | accuracy | validation_accuracy\r\n-- | -- | -- | -- | --\r\n0 | 6.980901 | 7.149147 | 0.015823 | 0.047779\r\n1 | 7.054768 | 7.217787 | 0.015194 | 0.047779\r\n2 | 7.099029 | 7.474302 | 0.014880 | 0.047779\r\n3 | 7.145690 | 7.359528 | 0.015509 | 0.047779\r\n4 | 7.183013 | 7.395905 | 0.013937 | 0.005448\r\n5 | 7.210382 | 7.452353 | 0.016137 | 0.047779\r\n\r\n", "Ok, now your data looks correct. Can you try with just:\r\n```\r\nself.clf.fit(train_dataset, epochs=self.params['epochs'])\r\n```\r\n\r\nIf it still not working, try to use multiple other model such as Bert and see if it is still the cases. ", "It's not working with that, and it's not working with 'bert-base-uncased' either.\r\n\r\nThis is the output for 'bert-base-uncased'.\r\n```\r\nDownloading: 100%\r\n433/433 [00:02<00:00, 199B/s]\r\n\r\nDownloading: 100%\r\n232k/232k [00:00<00:00, 1.25MB/s]\r\n\r\nDownloading: 100%\r\n466k/466k [00:00<00:00, 1.32MB/s]\r\n\r\ndataset <BatchDataset shapes: ({input_ids: (None, 500), token_type_ids: (None, 500), attention_mask: (None, 500)}, (None,)), types: ({input_ids: tf.int32, token_type_ids: tf.int32, attention_mask: tf.int32}, tf.int16)> <BatchDataset shapes: ({input_ids: (None, 500), token_type_ids: (None, 500), attention_mask: (None, 500)}, (None,)), types: ({input_ids: tf.int32, token_type_ids: tf.int32, attention_mask: tf.int32}, tf.int16)>\r\nDownloading: 100%\r\n536M/536M [00:11<00:00, 45.4MB/s]\r\n\r\nAll model checkpoint layers were used when initializing TFBertForSequenceClassification.\r\n\r\nSome layers of TFBertForSequenceClassification were not initialized from the model checkpoint at bert-base-uncased and are newly initialized: ['classifier']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\nEpoch 1/40\r\nWARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f0c5c70d210>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nWARNING: AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f0c5c70d210>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\n9543/9543 [==============================] - 1784s 187ms/step - loss: 6.9231 - accuracy: 0.0181\r\nEpoch 2/40\r\n9543/9543 [==============================] - 1451s 152ms/step - loss: 7.0684 - accuracy: 0.0153\r\nEpoch 3/40\r\n9543/9543 [==============================] - 1157s 121ms/step - loss: 7.1112 - accuracy: 0.0171\r\nEpoch 4/40\r\n9543/9543 [==============================] - 1160s 122ms/step - loss: 7.1445 - accuracy: 0.0170\r\nEpoch 5/40\r\n9543/9543 [==============================] - 1160s 122ms/step - loss: 7.1869 - accuracy: 0.0159\r\n```\r\n\r\n", "Is there anything else that could cause the model not to learn?", "The problem seems not to come from the models, I would guess that they might come from the way you build your data before the creation of the tf.data.Dataset or from the data themselves (sometime we cannot learn anything from the data, it happend). But I cannot be sure of anything without being able to reproduce the same behavior on my side sorry.", "That's strange because I can get 32% accuracy on validation data using a baseline model that finds the closest text and predicts its label. So it should be possible to learn from the data." ]

[]

[]

[ "Maybe @patrickvonplaten or @lhoestq can chime in here.", "It may very well be that this feature is not implemented. @ola13, @lhoestq do you have more insight here maybe?", "The gradient does propagate to the question encoder. What configuration did you use ?", "We used the default configuration and 'compressed' DPR. We are indeed seeing that the gradient is propagating to the question encoder; before training starts, a retriever retrieves related articles, but after training, the retriever universally retrieves the same few documents regardless of the query. That shows that the gradients are propagating, but not well.", "Following up on this. Any thoughts? We are curious whether there is an issue with the implementation of DPR training in HF", "Hi @JamesDeAntonis! Retrieval collapse is a problem we encountered in some setups and not necessarily caused by a bug in the retriever - basically what it means is that the passages retrieved at the beginning of the training are not useful enough so the models learns to ignore them.\r\n\r\nWe experienced collapse when training a RAG-Sequence model for FEVER, but we were successful with RAG-Token and RAG-Classifier. An option to move forward here could be:\r\n- try training a RAG-Token generative model (it'd be generating the labels)\r\n- share the classification code, maybe there's some issue there? Are you performing marginalization on top of the BART classification head logprobs?", "Thanks for the response! Good to hear that you were able to train successfully. When you trained, did you use the two-label or three-label dataset? (we are currently using the three-label) I'm curious whether the inconclusive samples are contributing to the collapse.\r\n\r\n* We are using the final hidden state of RAG-token as input into a classification head, and the model properly trains with the generator and classifier heads unfrozen (just the retriever is frozen in this case). This gets to 72% accuracy, same as the paper. I think this implies that the generator and classifier head are configured properly.\r\n* We are indeed marginalizing on top of the BART classification head logprobs", "Here is our classification head, mostly taken from HF:\r\n\r\n```{python}\r\nclass BartClassificationHead(Module):\r\n \"\"\"Head for sentence-level classification tasks.\"\"\"\r\n\r\n def __init__(\r\n self,\r\n input_dim: int,\r\n inner_dim: int,\r\n num_classes: int,\r\n pooler_dropout: float,\r\n **config_kwargs\r\n ):\r\n super().__init__(**config_kwargs)\r\n self.dense = Linear(input_dim, inner_dim)\r\n self.dropout = Dropout(p=pooler_dropout)\r\n self.out_proj = Linear(inner_dim, num_classes)\r\n\r\n def forward(self, hidden_states: torch.Tensor):\r\n hidden_states = self.dropout(hidden_states)\r\n hidden_states = self.dense(hidden_states)\r\n hidden_states = torch.tanh(hidden_states)\r\n hidden_states = self.dropout(hidden_states)\r\n hidden_states = self.out_proj(hidden_states)\r\n return hidden_states\r\n```", "..and here is the high-level model code:\r\n\r\n```{python}\r\n # input_ids shape = (batch_size, 512)\r\n outputs = super().forward(input_ids=input_ids, attention_mask=attention_mask, **rag_kwargs)\r\n\r\n ### the following code is inspired by BartForSequenceClassification forward method\r\n # best practice for bart classification is to use the last hidden state\r\n # hidden.shape=(batch_size * n_documents, 300, 1024)\r\n hidden = outputs.generator_dec_hidden_states[-1] # last hidden state;\r\n #print (hidden)\r\n\r\n # eos_mask.shape = (batch_size * n_documents, 300)\r\n eos_mask = outputs.context_input_ids.eq(self.rag.generator.config.eos_token_id)\r\n\r\n if len(torch.unique(eos_mask.sum(1))) > 1:\r\n raise ValueError(\"All examples must have the same number of <eos> tokens.\")\r\n\r\n # pass along the hidden state at the eos token\r\n # (batch_size * n_documents, 1024)\r\n sentence_representation = hidden[eos_mask, :].view(hidden.size(0), -1, hidden.size(-1))[:, -1, :]\r\n\r\n # (batch_size * n_documents, 1, 3)\r\n document_level_logits = self.classification_head(sentence_representation)\r\n\r\n # finally, marginalize across all the retrieved documents\r\n # (batch_size, 1, 3)\r\n logits = self.marginalize(document_level_logits, outputs.doc_scores)\r\n\r\n # (batch_size, 3)\r\n logits = logits.squeeze(1)\r\n```", "We were able to train RAG-Token and RAG-Classifier successfully both on 2-way and the 3-way variant of FEVER. One important thing to note though is that those were on our internal `fairseq` implementation.\r\n\r\n> try training a `RagToken` generative model (it'd be generating the labels)\r\n\r\nWhat I meant when suggesting to use `RagToken` would be to use it as-is, without a classification head - it might seem counterintuitive but the generative model is actually able to learn to generate the labels.\r\n\r\nAs for the classification implementation - what you're proposing is quite different from our implementation in `fairseq`. What happens currently in your implementation is that you marginalize twice - once inside the forward pass on `RagToken`, and then again after applying your classification head. What we do instead is the following: \r\n1) take the generator hidden states (not marginalized)\r\n2) apply BART-like classification head on top of that\r\n3) marginalize\r\n\r\nSo basically - you don't want to just add a `BartClassificationHead` on top of `RagToken` hidden states. You want to implement something similar to [`BartForSequenceClassification`](https://github.com/huggingface/transformers/blob/master/src/transformers/models/bart/modeling_bart.py#L1254-L1346) - a `RagForSequenceClassification` of sorts, doing what I outlined above - if you're interested in implementing that I think it'd be a great contribution to the repo, cc @patrickvonplaten :)\r\n\r\nLet me know if this makes sense!\r\n\r\n", "Thanks for the advice! I have a few questions\r\n\r\n(1) what makes you think we marginalize twice? `do_marginalize=False` in `RagToken` by default.\r\n(2) What's the difference between a `BartClassificationHead` on top of `Bart`, and `BartClassificationHead` on top of `RagToken`? Isn't the generator of `RagToken` simply `Bart` already, so the final hidden state of `RagToken` == final hidden state of `Bart`?\r\n(3) Did you use 'adam' as your optimizer", "And yes, I would be delighted to contribute it to the repo :)", "Hi there! I'm a colleague of @JamesDeAntonis and I just wanted to chime in and clarify that we are using RagTokenForGeneration only as a neat wrapper to have a `self.rag` variable if we ever needed it. During training we take `outputs.generator_dec_hidden_states[-1]` as posted in the code above. Then we proceed with the 3 steps you listed above, essentially ending up with a SequenceClassification head from the rag generator hidden state outputs. We don't interact with the generation aspect at all, as you correctly identified.", "> (1) what makes you think we marginalize twice? do_marginalize=False in RagToken by default.\r\n\r\nHey James and @suriyakode, I didn't realize that was the default configuration, in such case indeed your implementation should be equivalent to what I was suggesting.\r\n\r\nAnd yes, we did use Adam as optimizer.\r\n\r\nIn such case I don't see anything obvious unfortunately. What accuracy do you get with the collapse?", "Bummer :/\r\n\r\nWith frozen retriever, we achieved 72% accuracy, then unfroze the retriever and the accuracy fell to 68% after collapse\r\n\r\nTo be clear, you used `RagToken`, aka the current `RagTokenForGeneration` object that was implemented in HF? My original thought was that something could be wrong with the gradients or something in this specific implementation", "@JamesDeAntonis all of my FEVER experiments were done on `fairseq`, but I have been able to replicate RAG paper results training HF `RagToken` models on Natural Questions, which gives me some level of confidence in the implementation.", "Ok, thanks. What was your learning rate?", "It was 1e-05 for training the classifier.", "I also noticed an issue with the finetuning script. I ended up printing `doc_scores` while using the RAG finetuning scripts and saw that there was no gradient. Is there no gradient passed to the question encoder from the generator?", "The `doc_scores` is supposed to have gradients. And the gradients are propagated to the weights of the RAG question encoder.", "I cloned the transformers repo and pip installed transformers using the repo cloned to verify. I printed `doc_scores` in `RagModel` and got the following for what `doc_scores` was:\r\n\r\nI don't see gradients in the tensor.\r\n", "Is there a fix to the `doc_scores` gradient?", "It looks like there's no grad_fn on your `doc_scores`. Are the weights of the question encoder updated during finetuning ?", "Nope the weights aren't updated.", "Though should there even be a `grad_fn` when it's running \"Validation sanity check\" (as pictured above)?", "Good catch @dblakely indeed during validation it makes sense to not have any gradient", "Thanks for the clarification! I continued printing after and got a `grad_fn` in the tensor.", "@ola13 is there any code that we can see regarding the internal `fairseq` implementation of RAG and the training you did with it? I don't think there's any RAG in the public `fairseq` repo, but would be useful for me to be able to compare the two implementations", "@ola13 I have a couple of questions on this\r\n(1) how many docs did you retrieve?\r\n(2) regarding the `RagForFeverClassification` idea, what is meant by \"[we] first regenerate the claim\" in section C of the RAG paper's appendix? Does that mean that we should only provide tokens from the claim as `decoder_input_ids` in the generator? Does it mean we should run multiple passes of the generator? Curious what the correct interpretation is", "Hi @JamesDeAntonis - sorry just noticed your previous commend, the `fairseq` implementation is not available publicly. \r\n\r\nRegarding your latest questions:\r\n1) we used 5 docs at training time and tried evaluation with up to 50 docs\r\n2) This idea is adapted from BART - https://arxiv.org/pdf/1910.13461.pdf - section 3.3 or Figure 3.a - in case of RAG, we don't copy all contextualized input, just the claim tokens. Since BART model will copy all input to `decoder_input_ids` if you just leave it at `None` ([here](https://github.com/huggingface/transformers/blob/master/src/transformers/models/bart/modeling_bart.py#L1135-L1140)) you can try adding extra logic to only pass claim tokens to the BART decoder with RAG (that may mean explicitly setting `decoder_input_ids` like you mention). This does not require multiple runs of the generator." ]

[]

[]

[ "Dear @patrickvonplaten and the rest of HuggingFace group.\r\n\r\nI implemented the concept of `head_mask` from BERT into BART so that the internal of decoder-encoder-like models can be studied as well. However, as this is my very first attempt to contribute to such a large-scale open-source project, I have been a bit struggling to pass the tests. Would you be, please, able to guide me what everything needs to be done in this case in order to achieve a valid pull request?\r\n\r\nThank you very much for all your time in advance. I really do appreciate it.", "Hi @stancld - thanks a lot for pinging me! I'm happy to help you here :-) I think you're PR is a nice addition. Sadly, we did many changes to Bart recently (see https://github.com/huggingface/transformers/pull/9343) so that you'll probably have to rebase your PR to the current version of master. ", "After that I'm happy to get the tests passing together!", "Hi @patrickvonplaten, the model should be rebased according to the commit #9343 at this moment. :) I'll be more than happy to finish this PR with you. Thanks a lot in advance :) ", "@stancld, please do let me know if you're stuck and need help or if your PR is ready for review, just ping me here :-)", "Hi @patrickvonplaten, I would like to bring an update after the weekend off.\r\n\r\nFirst of all, I would like to apologise for a bit of messy PR, as I was initially struggling with on my local (I'll do better next time).\r\nRegarding this PR: To pass all the tests, `head_mask` and `decoder_head_mask` is now implemented for the following PyTorch BART-based models:\r\n\r\n- **BART**,\r\n- **MBart**,\r\n- **Blenderbot**,\r\n- **BlenderbotSmall**,\r\n- **Marian**,\r\n- **Pegasus**.\r\n\r\nBesides, I think some additional tests for head_mask for these models might be desired to implement, but I leave this decision up to you. In any case, please, let me know what it needs to do to complete this PR.\r\n", "@patrickvonplaten I think this PR is ready for review. I've currently resolved one conflict arose last night after a commit to `master` and now I've been tracking changes on my local and everything still seems to be working.", "Hey @stancld,\r\n\r\nThis is a super nice PR. It's very clean and that without any help - awesome!\r\n\r\nI think there are 3 things we should change/add:\r\n\r\n1) I think we should change the order of the forward args of all `...Model` and `...ForConditionalGeneration` as explained above. This a) means that there is no breaking change in the way Bart is used with torchscript and it's the better option IMO as well since the first 4 args should always be `input_ids, attention_mask, decoder_input_ids, decoder_attention_mask` for EncDec models\r\n\r\n2) Let's try to remove all \"hard-coded\" model names in the common tests. I've commented above. We don't really need to test torchscript with head_mask and for the signature it'd be better to change it according to 1)\r\n\r\n3) It would be awesome if you could a `if model.config.is_encoder_decoder` part to the `test_headmasking` test in `test_modeling_common.py` that tests headmasking correctly for Seq2Seq models. To enable this test for all Bart-like models you'll have to set `test_head_masking` to True in `BartModelTest` and others. One thing we'll have to adapt in the test is we should change the line:\r\n```\r\nattentions = outputs[-1]\r\n```\r\n\r\nto \r\n```python\r\nattentions = outputs.attetions\r\n```\r\n\r\nfor the `model.config.is_encoder_decoder is False` case and to \r\n\r\n```python\r\nencoder_attentions = outputs.encoder_attentions\r\ndecoder_attentions = outputs.decoder_attentions\r\n```\r\n\r\nfor the other case.\r\n\r\nI can also help you with 3) in case you're stuck.\r\nReally impressed by how clean the PR is! Think there is not much left to do. 1) and 2) are very easy changes and 3) will require a bit more time, but should be fine as well.", "Hey @patrickvonplaten, thanks a lot for your thorough feedback. I believe to come back later today with a new commit fixing the listed issues :)", "Hey @patrickvonplaten, this PR is again ready for review after making some changes according to your notes above. The one problem at this moment is that BART-like models do not satisfy one condition in `test_headmasking`:\r\n```\r\nself.assertNotEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0).\r\n```\r\n\r\nI am not sure whether the formula for masking attention heads (in BART-like models) is implemented correctly. Now, if `head_mask` in the test case is specified as\r\n```\r\nhead_mask = torch.ones(\r\n self.model_tester.num_hidden_layers,\r\n self.model_tester.num_attention_heads,\r\n device=torch_device,\r\n)\r\nhead_mask[0, 0] = 0\r\nhead_mask[-1, :-1] = 0\r\n```\r\nthen `outputs.encoder_attentions[1][..., :, :, :]` or `outputs.decoder_attentions[1][..., :, :, :]` equals tensor of `0.0` for all examples over all heads but the last one. This is not the case, however, for **non**-encoder-decoder models with `attentions[1][..., :, :, :]`. Do you have any idea where the problem can be?\r\n\r\nAnyway, I hope we will solve this issue and merge this PR. :) ", "I made some mistakes during updating my branch, which resulted in the problem with tracking files not edited actually by myself. I find this quite inconvenient and I have failed to repair this issue so far. Therefore, I've created a new (clean) branch, which might be found here https://github.com/stancld/transformers/tree/head_mask_for_bart_new.\r\n\r\nIf you, @patrickvonplaten, were okay with that, I would close this PR (after resolving those rather minor issues raised in our discussion above) and create a new one from the new branch referenced above to make everything nice and clean before an eventual merge. ", "@stancld absolutely! Feel free to close this PR and open a new one :-) This happens to me all the time as well ", "We can just link this closed PR to the new PR to have a reference to the discussion we had", "@patrickvonplaten - Great, you can find a newly open PR at #9569 :) " ]

[]

[]

[]

[ { "id": 1905493434, "node_id": "MDU6TGFiZWwxOTA1NDkzNDM0", "url": "https://api.github.com/repos/huggingface/transformers/labels/dependencies", "name": "dependencies", "color": "0366d6", "default": false, "description": "Pull requests that update a dependency file" } ]

[]

[ "Thanks dependabot!" ]

[]

[]

[ "Anyway, I doubt Lxmert is currently not supported for customized/ personal datasets (nlp+images, images are harder to prepare). See issues relevant to feature extraction in [Lxmert](https://github.com/airsplay/lxmert#faster-r-cnn-feature-extraction), for example, [issue#79](https://github.com/airsplay/lxmert/issues/79), [issue#86](https://github.com/airsplay/lxmert/issues/86)." ]

[]

[]

[]

[]

[]

[ "Perhaps this will help: https://huggingface.co/patrickvonplaten/longformer2roberta-cnn_dailymail-fp16", "`LongformerEncoderDecoder` will be added to the lib once #9278 is merged! It can be used for summarization or any other seq2seq task.", "@patil-suraj, @christianversloot I tried this model [longformer2roberta](https://huggingface.co/patrickvonplaten/longformer2roberta-cnn_dailymail-fp16). It is actually giving better summary than Pegasus (reddit-tifu). I will be waiting for LongformerEncoderDecoder to be added to the library.\r\n\r\nI just have one question - how many input tokens longformer2roberta model supports? I believe it's 2048. Could you please confirm,", "`longformer2roberta` should support 4096 tokens.\r\n\r\nAnd LED is now on master!", "@patil-suraj Awesome! I am trying LED but getting below error. Could you please take a look?\r\n\r\n```\r\nfrom transformers import LEDForConditionalGeneration, LEDTokenizer\r\nmodel_name = 'allenai/led-base-16384'\r\ntokenizer = LEDTokenizer.from_pretrained(model_name) \r\nmodel = LEDTokenizer.from_pretrained(model_name)\r\nbatch = tokenizer.prepare_seq2seq_batch(article, truncation=True, padding='longest', return_tensors=\"pt\").to(torch_device)\r\ntranslated = model.generate(**batch)\r\ntgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)\r\n```\r\n\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n<ipython-input-33-d4827d7b5770> in <module>()\r\n 5 model = LEDTokenizer.from_pretrained(model_name)\r\n 6 batch = tokenizer.prepare_seq2seq_batch(src_text, truncation=True, padding='longest', return_tensors=\"pt\").to(torch_device)\r\n----> 7 translated = model.generate(**batch)\r\n 8 tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)\r\n\r\nAttributeError: 'LEDTokenizer' object has no attribute 'generate'", "```python\r\nmodel = LEDTokenizer.from_pretrained(model_name)\r\n```\r\nhere you are assigning tokenizer as the model, it should be \r\n```python\r\nmodel = LEDForConditionalGeneration.from_pretrained(model_name)\r\n```", "@patil-suraj Thank you! My bad. I have corrected the typo. Below code seems to be working fine.\r\n\r\n```\r\nfrom transformers import LEDForConditionalGeneration, LEDTokenizer\r\nmodel_name = 'allenai/led-base-16384'\r\ntokenizer = LEDTokenizer.from_pretrained(model_name) \r\nmodel = LEDForConditionalGeneration.from_pretrained(model_name)\r\ninput_ids = tokenizer(src_text, return_tensors=\"pt\").input_ids\r\noutput_ids = model.generate(input_ids)\r\noutput = tokenizer.decode(output_ids[0], skip_special_tokens=True)\r\n```\r\n\r\nI am getting very short summary of just 20 tokens from the above code. So I was looking for the default values for below parameters for LED model. I could not find it in [config.json](https://huggingface.co/allenai/led-base-16384/blob/main/config.json) file. For the **longformer2roberta** model I found these values in [config.json](https://huggingface.co/patrickvonplaten/longformer2roberta-cnn_dailymail-fp16/blob/main/config.json). Could you please let me know where can I find these values for LED model.\r\n\r\n- num_beams\r\n- no_repeat_ngram_size,\r\n- early_stopping,\r\n- length_penalty,\r\n- min_length,\r\n- max_length", "@patil-suraj Any inputs/suggestions here?", "Hi, \r\nI have a question about the `LEDForConditionalGeneration` forward args. \r\nThe `decoder_input_ids` has a comment that `decoder_input_ids (torch.LongTensor of shape (batch_size, target_sequence_length), optional) – Provide for translation and summarization training. By default, the model will create this tensor by shifting the input_ids to the right, following the paper.`. \r\nForm the forward method in `LEDForConditionalGeneration`, i can see that when not assigning the `decoder_input_ids` in the forward method of `LEDForConditionalGeneration` object , the `decoder_input_ids` will be generated by [shifting the `labels` value one token to right in the forward method](https://github.com/huggingface/transformers/blob/17b6e0d474b797cdddf5225b0f51bf0e928091b9/src/transformers/models/led/modeling_led.py#L2337). \r\n\r\nSo my question is if i want to explictly pass the `decoder_input_ids` to the forward method, do i need to explictly shift it one token as the [code](https://github.com/huggingface/transformers/blob/17b6e0d474b797cdddf5225b0f51bf0e928091b9/src/transformers/models/led/modeling_led.py#L2337) shows before the forward pass?\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." ]

[]

[]

[ "Hi,\r\nI have the same problem different results using `model()` vs `trainer.predict()`.", "> Hi,\r\n> I have the same problem different results using `model()` vs `trainer.predict()`.\r\n\r\nThanks for replying. I assumed this feature was made for other usages so I ended up using `model()`. Anyway, I would like to seek an answer for sure.\r\n\r\nReading the new issue's template I guess @sgugger could help us here. I'm sorry to disturb you, could you please give some details on `Trainer.predict()` here?", "I solved it by returning to 4.0.1, here both methods return the same results. \r\n\r\nBut I still got a problem, before saving the model (so just at the end of the finetuning) with `TrainingArguments(..., load_best_model_at_end=True)` the `trainer.predict()` still differs from `model()`. But after reloading the model with `from_pretrained` with transformers==4.0.1 both methods are equal. So I guess the `trainer.predict()` does really load the best model at the end of the training.", "I'm unsure of what the problem is since the code you indicate is not reproducible (what is `model`, `load_pt_data()` etc.). On my side, using an installation form source on current master, here is what I get. First I instantiate a model and tokenizer and preprocess some data (with padding to be able to batch):\r\n```\r\nfrom transformers import AutoModelForSequenceClassification, AutoTokenizer, TrainingArguments, Trainer\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-cased\")\r\nmodel = AutoModelForSequenceClassification.from_pretrained(\"bert-base-cased\")\r\ntexts = [\"Hello there!\", \"This is another text\"]\r\ntokenized_texts = tokenizer(texts, padding=True)\r\n```\r\nThen I create a Dataset class to be able to feed my `tokenized_texts` to `Trainer`:\r\n```\r\nclass SimpleDataset:\r\n def __init__(self, tokenized_texts):\r\n self.tokenized_texts = tokenized_texts\r\n \r\n def __len__(self):\r\n return len(self.tokenized_texts[\"input_ids\"])\r\n \r\n def __getitem__(self, idx):\r\n return {k: v[idx] for k, v in self.tokenized_texts.items()}\r\n\r\ntest_dataset = SimpleDataset(tokenized_texts)\r\n```\r\nThen predicting through `Trainer` like this:\r\n```\r\ntrainer = Trainer(model=model)\r\npredictions = trainer.predict(test_dataset)\r\npredictions.predictions\r\n```\r\nreturns this:\r\n```\r\narray([[-0.68212456, 0.07081275],\r\n [-0.59134895, 0.16735002]], dtype=float32)\r\n```\r\nand predicting directly with the model:\r\n```\r\nimport torch\r\n\r\nmodel.eval()\r\npt_inputs = {k: torch.tensor(v).to(trainer.args.device) for k, v in tokenized_texts.items()}\r\nwith torch.no_grad():\r\n output = model(**pt_inputs)\r\noutput.logits.cpu().numpy()\r\n```\r\ngives me the exact same result.\r\n\r\nMake sure that you preprocess your inputs the same way in both instances, and when using the model directly, that it is in evaluation mode.", "> pt_inputs = {k: torch.tensor(v).to(trainer.args.device) for k, v in tokenized_texts.items()}\r\n> with torch.no_grad():\r\n> output = model(**pt_inputs)\r\n> output.logits.cpu().numpy()\r\n\r\nHi, thanks for your answers! I think the reason I'm having different results is I did not use `model.eval()` but I only had <1000 lines of test data to predict. \r\n\r\nThank you so much! :)", "> I'm unsure of what the problem is since the code you indicate is not reproducible (what is `model`, `load_pt_data()` etc.). On my side, using an installation form source on current master, here is what I get. First I instantiate a model and tokenizer and preprocess some data (with padding to be able to batch):\r\n> \r\n> ```\r\n> from transformers import AutoModelForSequenceClassification, AutoTokenizer, TrainingArguments, Trainer\r\n> \r\n> tokenizer = AutoTokenizer.from_pretrained(\"bert-base-cased\")\r\n> model = AutoModelForSequenceClassification.from_pretrained(\"bert-base-cased\")\r\n> texts = [\"Hello there!\", \"This is another text\"]\r\n> tokenized_texts = tokenizer(texts, padding=True)\r\n> ```\r\n> \r\n> Then I create a Dataset class to be able to feed my `tokenized_texts` to `Trainer`:\r\n> \r\n> ```\r\n> class SimpleDataset:\r\n> def __init__(self, tokenized_texts):\r\n> self.tokenized_texts = tokenized_texts\r\n> \r\n> def __len__(self):\r\n> return len(self.tokenized_texts[\"input_ids\"])\r\n> \r\n> def __getitem__(self, idx):\r\n> return {k: v[idx] for k, v in self.tokenized_texts.items()}\r\n> \r\n> test_dataset = SimpleDataset(tokenized_texts)\r\n> ```\r\n> \r\n> Then predicting through `Trainer` like this:\r\n> \r\n> ```\r\n> trainer = Trainer(model=model)\r\n> predictions = trainer.predict(test_dataset)\r\n> predictions.predictions\r\n> ```\r\n> \r\n> returns this:\r\n> \r\n> ```\r\n> array([[-0.68212456, 0.07081275],\r\n> [-0.59134895, 0.16735002]], dtype=float32)\r\n> ```\r\n> \r\n> and predicting directly with the model:\r\n> \r\n> ```\r\n> import torch\r\n> \r\n> model.eval()\r\n> pt_inputs = {k: torch.tensor(v).to(trainer.args.device) for k, v in tokenized_texts.items()}\r\n> with torch.no_grad():\r\n> output = model(**pt_inputs)\r\n> output.logits.cpu().numpy()\r\n> ```\r\n> \r\n> gives me the exact same result.\r\n> \r\n> Make sure that you preprocess your inputs the same way in both instances, and when using the model directly, that it is in evaluation mode.\r\n\r\nI have a more question that how can I load the model without using \"from_pretrained\" \r\n\r\nBecause I have some custom for the the model, nn.Model, it does not inherent from \"PreTrainedModel\", so I can't load it using \"from_pretrained\"\r\n" ]

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

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "Hi, did you fix this issue? I also got this problem when using BART-large." ]

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

[]

[ "Hi @Twsschx \r\n\r\nIn the newer version of transformers, PyTorch models have outputs that are instances of subclasses of `ModelOutput`,\r\nto access it as a tuple, you can use slicing, or to get a particular tensor, just provide the key to the output class.\r\n\r\ni.e \r\n```python3\r\nlast_hidden_states, pooling_output =transformer_model(input_ids_tensor, attention_mask_tensor, segment_ids_tensor)[:] # slice\r\n```\r\n\r\nor \r\n```python3\r\noutput =transformer_model(input_ids_tensor, attention_mask_tensor, segment_ids_tensor)\r\nlast_hidden_state = output[\"last_hidden_states\"]\r\npooling_output = output[\"pooler_output\"]\r\n```\r\n\r\nAnd if you want the models to output tuple like the previous versions, then pass `return_dict=False` to `forward`\r\n\r\n```python3\r\nlast_hidden_states, pooling_output = transformer_model(**inputs, return_dict=False)\r\n```\r\n\r\nYou can find more about output classes in this [doc](https://huggingface.co/transformers/main_classes/output.html).", "Thanks. It helps me a lot!", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[]

[]

[ { "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 } ]

[ "This is the correct fix indeed (though we can group this with the previous test with `elif data_args.task_name is None and not is_regression`)! Thanks for flagging this, do you want to open a PR with the fix you found?", "@sgugger \r\nThank you for checking this issue and giving the comment.\r\nI'd love to open a PR. \r\nI'm sorry but could you please wait for a while? I think I can open it by the end of the week.", "Thanks for the PR!" ]

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

[]

[ "Hey @farazk86,\r\n\r\nthanks for your issue! I've sadly never worked with tflite so not really sure how to best help you here. Maybe @jplu (or @LysandreJik) ? ", "> Hey @farazk86,\r\n> \r\n> thanks for your issue! I've sadly never worked with tflite so not really sure how to best help you here. Maybe @jplu (or @LysandreJik) ?\r\n\r\nThanks @patrickvonplaten but the ``print(model.inputs)`` line is before the tflite conversion. For starters, I wanted to convert the fine-tuned distilgpt2 from pytorch to tensorflow and then to tensorflow lite.\r\n", "Hello @farazk86 \n\nIt is normal that the inputs/outputs are not set when using from_pretrained because they are not explicitely given when the model is built. You have to create yourself a model by setting them.", "> Hello @farazk86\r\n> \r\n> It is normal that the inputs/outputs are not set when using from_pretrained because they are not explicitely given when the model is built. You have to create yourself a model by setting them.\r\n\r\nThanks for your reply.\r\n\r\nIs my method for converting the pretrained model to tflite wrong? I followed the code and explanation mentioned here: https://towardsdatascience.com/on-device-machine-learning-text-generation-on-android-6ad940c00911", "I would also like to add that when I convert the model to tflite, using the above code I get the following warnings\r\n\r\n```\r\nWARNING:absl:Found untraced functions such as wte_layer_call_fn, wte_layer_call_and_return_conditional_losses, wpe_layer_call_fn, wpe_layer_call_and_return_conditional_losses, dropout_layer_call_fn while saving (showing 5 of 380). These functions will not be directly callable after loading.\r\nWARNING:absl:Found untraced functions such as wte_layer_call_fn, wte_layer_call_and_return_conditional_losses, wpe_layer_call_fn, wpe_layer_call_and_return_conditional_losses, dropout_layer_call_fn while saving (showing 5 of 380). These functions will not be directly callable after loading.\r\n```\r\n does this help identify the issue?", "These warnings are not directly related to your issue, you can safely ignore them for now and the tutorial you linked is wrong for the TFLite creation part.\r\n\r\nUnfortunately, the current state of the TF models in transformers are not fully compliant with TFLite so, I suggest to do not push to far the conversion. It is in our plans to have a better compliancy, but we don't know when yet.\r\n\r\nYou use the following piece of code to create your TFLite model:\r\n```python\r\nfrom transformers import TFGPT2LMHeadModel\r\nimport tensorflow as tf\r\n\r\nbase_model = TFGPT2LMHeadModel.from_pretrained(\"gpt2\")\r\ninput_ids = tf.keras.layers.Input((128, ), batch_size=1, dtype=tf.int32, name='input_ids')\r\nattention_mask = tf.keras.layers.Input((128, ), batch_size=1, dtype=tf.int32, name='attention_mask')\r\ninputs = {\"input_ids\": input_ids, \"attention_mask\": attention_mask}\r\noutput = base_model(inputs)\r\nmodel = tf.keras.models.Model(inputs=inputs, outputs=output)\r\nconverter = tf.lite.TFLiteConverter.from_keras_model(model)\r\nconverter.experimental_new_converter = True\r\nconverter.optimizations = [tf.lite.Optimize.DEFAULT]\r\nconverter.inference_input_type = tf.float32\r\nconverter.target_spec.supported_ops = [tf.lite.OpsSet.SELECT_TF_OPS]\r\ntflite_quant_model = converter.convert()\r\n\r\nwith open(\"model.tflite\", \"wb\") as f:\r\n f.write(tflite_quant_model)\r\n```\r\n\r\nWith this piece of code you should be able to convert your model into a TFLite one. Note also that the current TF models are not compliant with float16 so you have to keep with float32.", "Thank you, but with the code above I am getting the following error: \r\n\r\n```\r\nRegular TensorFlow ops are not supported by this interpreter. Make sure you apply/link the Flex delegate before inference. \r\nNode number 0 (FlexIdentity) failed to prepare.\r\n```\r\n\r\nI even tried with ``tf-nightly`` as a google search of this error suggested that the nightly has flex delegate support. But still got the above error.\r\n\r\nThe android tflite interpreter I am using works fine with all the models presented here: https://github.com/huggingface/tflite-android-transformers/tree/master/gpt2#change-the-model . They are even quantized.\r\n\r\nWould it be possible for me to train the models using a previous version of transformers? Which version was used at the time of writing the article and for providing the above tflite models? Would it help to change to that version?\r\n\r\nP.S. I would like to add that I am currently using the April 21st 2020 git version: ``!git checkout b1ff0b2ae7d368b7db3a8a8472a29cc195d278d8`` as I needed the ``line-by-line`` parameter during training.\r\n\r\nThank you\r\n", "To make this work you have to use the current release of TF and Transformers, not below.", "> To make this work you have to use the current release of TF and Transformers, not below.\r\n\r\nhmm, that would make sense why your code did not work. The reason I was using the April 21st version is because I needed the ``line-by-line`` parameter during fine tuning. Does the current version and ``run_clm.py`` have ``line-by-line`` support?", "I don't think the new `run_clm.py` still supports `line-by-line`. Better to open a new issue to discuss of this.", "> To make this work you have to use the current release of TF and Transformers, not below.\r\n\r\nSo I am now using the current version of transformers and tf and fine-tuned a model using ``run_clm.py`` and used your above code to convert that model to tflite but still got the same error:\r\n\r\n```\r\nRegular TensorFlow ops are not supported by this interpreter. Make sure you apply/link the Flex delegate before inference. \r\nNode number 0 (FlexIdentity) failed to prepare.\r\n```\r\n\r\nWhen converting this model, I got a lot of messages in console, \r\n\r\n```\r\nTensorflow version: 2.4.0\r\nWARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f2df19a0660>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: <cyfunction Socket.send at 0x7f2e091e6e58> is not a module, class, method, function, traceback, frame, or code object\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nWARNING: AutoGraph could not transform <bound method Socket.send of <zmq.sugar.socket.Socket object at 0x7f2df19a0660>> and will run it as-is.\r\nPlease report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.\r\nCause: <cyfunction Socket.send at 0x7f2e091e6e58> is not a module, class, method, function, traceback, frame, or code object\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nWARNING:tensorflow:AutoGraph could not transform <function wrap at 0x7f2e06b7a8c8> and will run it as-is.\r\nCause: while/else statement not yet supported\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nWARNING: AutoGraph could not transform <function wrap at 0x7f2e06b7a8c8> and will run it as-is.\r\nCause: while/else statement not yet supported\r\nTo silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nWARNING:absl:Found untraced functions such as wte_layer_call_and_return_conditional_losses, wte_layer_call_fn, wpe_layer_call_and_return_conditional_losses, wpe_layer_call_fn, dropout_layer_call_and_return_conditional_losses while saving (showing 5 of 385). These functions will not be directly callable after loading.\r\nWARNING:absl:Found untraced functions such as wte_layer_call_and_return_conditional_losses, wte_layer_call_fn, wpe_layer_call_and_return_conditional_losses, wpe_layer_call_fn, dropout_layer_call_and_return_conditional_losses while saving (showing 5 of 385). These functions will not be directly callable after loading.\r\nINFO:tensorflow:Assets written to: /tmp/tmpus6vmwet/assets\r\nINFO:tensorflow:Assets written to: /tmp/tmpus6vmwet/assets\r\nThe parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nThe parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\nINFO:absl:Using new converter: If you encounter a problem please file a bug. You can opt-out by setting experimental_new_converter=False\r\n```\r\ncomparatively, a model fine tuned using the old v2.4 version of transformers did not generate any such messages:", "Those are just warnings and expected messages, this is ok. Can you try with usual `gpt2` models? If it works, the issue is certainly coming from your model. Otherwise, we will check deeper what is going wrong. Nevertheless, TFLite compliancy is not our priority for now, so if we have to fix something it will certainly be postponed to a bit later in the coming months.", "Hi, I'm getting the same error as above when using the default pretrained gpt2.", "Humm with Transformers from source and TF 2.4 I get no errors. What is your env?", "> Humm with Transformers from source and TF 2.4 I get no errors. What is your env?\r\n\r\nI'm trying to run on android using a flutter tflite interpreter. and yes, I was also considering that maybe the fault is not in the converted model but the interpreter. \r\n\r\nTo confirm this I wanted to use the python interpreter, if I can get an output here then that means that the converted models are fine and its an issue with the flutter interpreter.\r\n\r\nBut when using the python tflite interpreter, everytime I invoke the model tflite generated model using your above provided code, my colab runtime crashes. \r\n\r\nI'm using current version of transformers and TF2.4\r\n\r\nThe same interpreter works for the tflite models provided here and produces output in the expected shape : https://github.com/huggingface/tflite-android-transformers/tree/master/gpt2#change-the-model\r\n\r\nBelow is the code I am using for tflite interpreter\r\n\r\n```python\r\nfrom transformers import *\r\nimport tensorflow as tf\r\nimport numpy as np\r\n\r\n\r\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\r\n\r\n# start with their provided model, the -O will change downloaded file name to model.tflite\r\n\r\n# !wget -O model.tflite https://s3.amazonaws.com/models.huggingface.co/bert/distilgpt2-64.tflite\r\n\r\n# Encode random strings\r\nsentance = (\"\"\"You are John, a punk living in the futuristic city of Zail. You have a small xore blaster hidden in you jacket and a holoband on your wrist. You are John, a punk living in the futuristic city of Zail. You have a small xore blaster hidden in you jacket and a holoband on your wrist. You are John, a punk living in the futuristic city of Zail. You have a small xore blaster hidden in you jacket and a holoband on your wrist. You are John, a punk living in the futuristic city of Zail. You have a small xore blaster hidden in you jacket and a holoband on your wrist. You are John, a punk living in the futuristic city of Zail. You have a small xore blaster hidden in you jacket and a holoband on your wrist. You are John, a punk living in the futuristic city of Zail. You have a small xore blaster hidden in you jacket and a holoband on your wrist.\"\"\")\r\nreview_token = tokenizer.encode(sentance)\r\nprint(len(review_token))\r\nreview_token = np.array(review_token, dtype=np.int32)\r\nreview_token = review_token[:128]\r\nreview_token = np.expand_dims(review_token, axis=0) # unsqueeze to add the batch dimension\r\nprint(sentance)\r\nprint(review_token)\r\n\r\ntflite_interpreter = tf.lite.Interpreter(model_path='/content/model.tflite')\r\ntflite_interpreter.allocate_tensors()\r\n\r\ninput_details = tflite_interpreter.get_input_details()\r\noutput_details = tflite_interpreter.get_output_details()\r\n\r\nprint(\"== Input details ==\")\r\nprint(\"name:\", input_details[0]['name'])\r\nprint(\"shape:\", input_details[0]['shape'])\r\nprint(\"type:\", input_details[0]['dtype'])\r\n\r\nprint(\"\\n== Output details ==\")\r\nprint(\"name:\", output_details[0]['name'])\r\nprint(\"shape:\", output_details[0]['shape'])\r\nprint(\"type:\", output_details[0]['dtype'])\r\n\r\ntflite_interpreter.set_tensor(input_details[0]['index'], review_token)\r\n\r\ntflite_interpreter.invoke()\r\n\r\ntflite_model_predictions = tflite_interpreter.get_tensor(output_details[0]['index'])\r\nprint(\"Prediction results shape:\", tflite_model_predictions.shape)\r\n```\r\n", "Ok, thanks for sharing, I will check this once I can dedicate some time.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 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": "" } ]

[]

[ "Hey @forest1988,\r\n\r\nIn order to fully use `BartDecoder` separately from `BartModel` as a `BartForCausalLM` model, we're still waiting on this PR: #9128.\r\nAnd again, you're very correct in your assessment that the behavior between `BartDecoder` and `GPT2` is not fully aligned here. IMO, we should change GPT2's cache format from a single tensor of `[2, batch_size, ...,]` to `tuple([batch_size, ...])` => If you're keen feel free to open a PR for it! Actually, we could make this a \"Good second issue\" here. \r\n\r\nSo to answer your question, no you should not contact the 2 tensors in `self_attn_past_key_value` to a single tensor, but we should rather change the code in GPT2 slightly to also have a tuple of 2 tensors instead of one tensor. \r\nIn GPT2, we create a new tensor at each iteration when using `use_cache` here: https://github.com/huggingface/transformers/blob/d944966b19a4d6860bddc7cdc1ba928ca8a0da91/src/transformers/models/gpt2/modeling_gpt2.py#L235 => this is a bit unnecessary IMO. When the inputs are getting longer allocating new memory for `key` and `value` can actually lead to a small slow-down IMO. If instead we would just use a tuple => `present = (key, value)` we would not allocate new memory. \r\n\r\nSo 1) As soon as #9128 is merged you can use `BartForCausalLM` the same way as `GPT2` without having to change anything.\r\n2) Let's see if someone is interested in tackling this \"inconsistency\" issue in GPT2. This \"First good issue\" should replace this line: https://github.com/huggingface/transformers/blob/d944966b19a4d6860bddc7cdc1ba928ca8a0da91/src/transformers/models/gpt2/modeling_gpt2.py#L235\r\nwith \r\n```python\r\npresent = (key.transpose(-2, -1), value))\r\n```\r\n(I think it should actually be that simple)\r\n", "Hi @patrickvonplaten, \r\n\r\nThank you for answering this issue!\r\nI'm sorry I haven't checked the PR https://github.com/huggingface/transformers/pull/9128 before creating this issue. I'll check it!\r\n\r\nAnd, thanks for telling me your opinion about the need to change GPT2's cache format from a single tensor to a tuple of 2 tensors.\r\nI'd love to open a PR, but I'm afraid I don't have enough time now.\r\nI will work on it as soon as I find the time, but of course, if someone else who is interested in the same issue would like to work on it, I would appreciate that!\r\n\r\nI now understand that I should not contact the 2 tensors in `self_attn_past_key_value` to a single tensor, but rather the code in GPT2 should be changed.\r\n\r\nI'm looking forward to seeing the PR https://github.com/huggingface/transformers/pull/9128 is merged.\r\nAlso, I would like to think about how to avoid contacting the 2 tensors in `self_attn_past_key_value` for what I am currently working on.\r\n\r\nThank you so much!\r\n", "Hi,\r\n\r\nI've started fixing the issue but failed in some tests.\r\n\r\n```\r\n========================================================================================= short test summary info ==========================================================================================\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_sample_generate - AttributeError: 'tuple' object has no attribute 'index_select'\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_search_generate - AttributeError: 'tuple' object has no attribute 'index_select'\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_gpt2_gradient_checkpointing - TypeError: CheckpointFunctionBackward.forward: expected Tensor or tuple of Tensor (got tuple) for return value 1\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_group_beam_search_generate - AttributeError: 'tuple' object has no attribute 'index_select'\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_model_parallel_beam_search - AttributeError: 'tuple' object has no attribute 'index_select'\r\nFAILED tests/test_trainer_distributed.py::TestTrainerDistributed::test_trainer - OSError: [Errno 12] Cannot allocate memory\r\n================================================================== 6 failed, 4626 passed, 834 skipped, 642 warnings in 1766.16s (0:29:26) ==================================================================\r\n```\r\n\r\nI think this may be related to the `index_select` used in [generation_utils.py](\r\nhttps://github.com/huggingface/transformers/blob/143289dcf759a663c03317e30167e89ee6d86588/src/transformers/generation_utils.py).\r\n\r\nI will continue to look into this when I have more time.", "Hi @patrickvonplaten,\r\n\r\nNow that I have time, I'm thinking about how to fix this issue so that the testing part works well.\r\n\r\nI think where to modify in `generation_utils.py` is here:\r\n\r\nhttps://github.com/huggingface/transformers/blob/02e05fb0a532e572b56ba75dad6ba3db625bbdeb/src/transformers/generation_utils.py#L506-L517\r\n\r\n`past` is taken from `past_key_values` or other output variations in:\r\nhttps://github.com/huggingface/transformers/blob/02e05fb0a532e572b56ba75dad6ba3db625bbdeb/src/transformers/generation_utils.py#L477-L489\r\n\r\nThen, it is treated like this:\r\nhttps://github.com/huggingface/transformers/blob/02e05fb0a532e572b56ba75dad6ba3db625bbdeb/src/transformers/generation_utils.py#L1654-L1655\r\n\r\nCan I modify #L506-L517 so that `past` is replaced from `Tuple[torch.Tensor]` to `Tuple[Tuple[torch.Tensor]]`,\r\nor should I consider other output variations, `output.mem` and `outputs.past_buckets_states`?\r\n\r\nThank you.\r\n", "Hey @forest1988,\r\n\r\nThanks for you in-detail code snippet. I think the easiest solution would be to open a PR showcasing the required changes :-) I think you're right `Tuple[torch.Tensor]` should indeed be `Tuple[Tuple[torch.Tensor]]`.\r\n\r\nWould you be interested in opening a PR so that we can add the fixes there? ", "Hi @patrickvonplaten,\r\n\r\nThank you for your comment! \r\nAfter doing some additions to change `Tuple[torch.Tensor]` to `Tuple[Tuple[torch.Tensor]]`, I would like to open a PR and ask you all to add fixes.\r\nI'll open the PR in a few days!", "I'm sorry to keep you waiting. I have opened a PR #9596 for this issue.\r\nI marked the PR as WIP because it has not yet been resolved.\r\nI will continue to look into this issue myself in the future and any advice would be greatly appreciated." ]

[]

[]

[ "This should be good now, @sgugger - thanks a lot for all the suggestions!", "@LysandreJik GitHub Reviewers is down, so tagging you instead." ]

[]

[]

[]

[]

[]

[ "Hey @frankgandiao,\r\n\r\nNote that Encoder-Decoder models usually require both `input_ids` and `decoder_input_ids`. Bart is special in a sense that it can automatically create the `decoder_input_ids` from the `input_ids` if you **don't** provide the `decoder_input_ids`. However, the model is not able to automatically create the `decoder_inputs_embeds` from the `inputs_embeds` if you provide only the `inputs_embeds` => to solve your problem you should provide the `decoder_inputs_embeds` as well. What you could do is the following: \r\n\r\n```python\r\nfrom transformers.models.mbart.modeling_mbart.py import shift_tokens_right\r\n\r\ninput_ids = tokenizer(text, return_tensors='pt')['input_ids']\r\ndecoder_input_ids = shift_tokens_right(input_ids, tokenizer.pad_token_id)\r\n\r\ninputs_embeds = model.get_input_embeddings()(input_ids).squeeze()\r\ndecoder_inputs_embeds = model.get_input_embeddings()(decoder_input_ids).squeeze()\r\n\r\nmodel(inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds)\r\n```", "Hey @patrickvonplaten, \r\n\r\nThanks for your reply! That makes sense and the problem is resolved!\r\n" ]

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

[]

[ "If `output_attentions=True` memory consumption should increase significantly (for large `sequence_length`) since we now store all attentions of size (`batch_size`, `num_heas`, `sequence_length`, `sequence_length`). This is less significant in training since the stored activations for training consume most RAM anyways. Speed should not be really affected by this.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Merging since it's blocking #9343 ." ]

[]

[]

[]

[ { "id": 2627272588, "node_id": "MDU6TGFiZWwyNjI3MjcyNTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Model%20Parallel", "name": "Model Parallel", "color": "8B66A5", "default": false, "description": "Model Parallelilsm Implementations" }, { "id": 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" } ]

[]

[ "That looks great! Model parallelism would be very nice for Bart. We should coordinate here a bit with all the open PRs. I'm also more or less done with the big \"split-bart-into-separate-models\" PR: https://github.com/huggingface/transformers/pull/9343.\r\nThink the merge conflicts can become a bit painful here :D. \r\n\r\nI'd propose the following:\r\n-Merge: #9347, #9386 (they should be pretty trivial to merge)\r\n-Rebase and Merge the big Bart refactor (https://github.com/huggingface/transformers/pull/9343)\r\n-Discuss/Merge the \"new\" model parallel design: #9316 and #9323\r\n-Rebase and Discuss/Merge this PR", "Is this PR ready for review? There's a lot of comments that were probably here for debugging purposes. Let me know if you want a review or if we should come back to it after #9347 and #9386 have been merged.", "It's very ready, functionality/concept-wise. It's not ready 100% commentary, debug traces, etc. but that's very unimportant until the rest is sorted out, since there are multiple overlapping PRs happening.\r\n\r\nBecause of the holidays there is a lot of new code which is all inter-dependent and unreviewed and then there is a huge change to merge of #9343.\r\n\r\nSo I think it's the best to review it as it is - sort things out and then once everybody is happy with the logic, and #9343 I will most likely have to do a new PR anyway.\r\n\r\nBut I need your feedback that what I did is correct.\r\n\r\nThink of it as an ongoing code design and generalization PR.\r\n\r\nThanks. ", "@patrickvonplaten, your plan works for me. Please ping me when #9343 is in and you're happy with the outcome so that it'll be safe to add MP w/o creating conflicts. Thank you.\r\n\r\nBut as I commented above this blocking event doesn't have to interefere with this PR's review - we are just not going to merge it, but it should proceed as normal and I will take all the agreed upon changes to the new PR once the dust around Bart split settled down. \r\n", "@stas00 @LysandreJik @patrickvonplaten This PR introduces a `device_map` that is not backwards compatible with 4.1.0. We have to do that at some point (as @stas00 discovered), but let's not have three different versions. We really need to make sure that we have consensus on the final form of the `device_map` that will work for all models going forward or we will have to change it again when model parallelization is generalized and some of its functionality is placed in `PreTrainedModel`. Have you tested this on `gpt2`, @stas00 and is the code generalizable to models that don't have decoder architectures and can store their attention blocks in attributes like `self.h`?\r\n\r\nHas everyone read [this comment](https://github.com/huggingface/transformers/pull/9323#issuecomment-753518885)? Are we all on board for the plan to generalize model parallelism? Don't have to implement it now, but we need to make sure we've thought through any changes that affect user experience and backward compatibility.\r\n\r\nSorry, I'm in the middle of moving so not keeping close track of all the traffic and could easily have missed something. Also, this content is spread across several PRs, so sometimes I'm getting confused.", "@alexorona, I'm basically leaving all the old code in place, so that gpt2 works as is and t5 as is, so this PR only impacts Bart. And in any case it doesn't look like this PR will be merged since Bart went through a split https://github.com/huggingface/transformers/pull/9343, which isn't finalized yet and I will need to re-do it anyway. But it's no problem, since I know what to do. And see the end of this comment - the whole MP implementation might need to be redesigned altogether.\r\n\r\nSince there are so many moving parts, it's very difficult to manage things and definitely makes things difficult for reviewers. \r\n\r\nSo my intention was to merge each of the new things separately, while keeping the old code working and then to start integrating things in bit. The holidays made things pile up, but since the HF team is back I trust in the next few days we will form a plan.\r\n\r\nImportant notes:\r\n\r\n1. MP is new here and should be clearly marked as an experimental feature. Which means device maps are not fixed and can change at any moment. https://github.com/huggingface/transformers/pull/9412\r\n\r\n What we could commit to is having the default device map work - i.e users don't supply any device map and then it just works.\r\n\r\n That's why I propose we start with each model implementing its own device map format (while sharing bits with common code where possible) and then over time we will find a common format.\r\n\r\n If the HF team wants to allocate time then we need to sit down, look at all the models and decide on the format ahead of time. If I'm not mistaken it looks like at the moment it's just @alexorona and I that mostly understand what's going on, so it'd be great to have someone from HF to get on top of MP. I'd be happy to sit down with that person and explain what I learned in the last few weeks in person. It's not complicated.\r\n\r\n2. As it was just pointed out https://github.com/pytorch/pytorch/issues/49961#issuecomment-754342632 this implementation is highly inefficient since it doesn't take advantage of the idle gpus, so we might have to scratch a big part of it and re-implement it using PP or something similar. The current implementation just uses extra gpus to expand available memory, but doesn't take advantage of the extra hardware.\r\n\r\nUntil then we have deepspeed integration [almost ready](https://github.com/huggingface/transformers/pull/9211) and `sharded_ddp` should be available in the next few days, so users will have excellent ways to fit huge transformers models on limited hardware already. So let's not rush with MP here and think.\r\n\r\n ", "From what I understand, model parallelism as it's currently implemented is a naive implementation of what it's supposed to do: offer more memory so that bigger models may be trained using memory of several devices instead of a single device. It is indeed inefficient as devices as idle while others compute, so there's definitely a way of making it more efficient.\r\n\r\n@stas00, @alexorona, if you could walk us through what you have learned so that @patrickvonplaten, @sgugger and myself can understand the different options available, that would be great.\r\n\r\nSome thoughts to structure our approach towards MP:\r\n\r\n- You mention pipeline parallelism (PP) as a way to be more efficient than model parallelism (MP), as the idle devices can be used while other compute. This intuitively seems like an algorithm to set up during training, do you think we would have to modify the models themselves like what is currently done with model parallelism?\r\n- As noted by @sgugger and approved by @patrickvonplaten and myself, working on the MP API of the current models (GPT-2 and T5) is a better test-bed than trying to make it work for all models all at once. Most models are similar, and finding a generic approach (if possible!) should be feasible with just these two models for now.\r\n- You're right that we should not rush it, and take our time to understand what we can do best for both inference and training.", "@LysandreJik No, it's not a naïve implementation of model parallelism. In addition to **data parallelism** and **model parallelism**, there is **pipeline parallelism**, which is the next level of complexity along with **zero redundancy**. Model parallelism allows us to train bigger models on GPU. Pipeline parallelism + model parallelism would allow us to train these large models faster because the GPUs are not idle. I really think the next step is to make sure model parallelism is generalized and rely on a library -- probably deepspeed -- to implement pipeline parallelism and zero redundancy. deepspeed has something called **3D parallelism**, which I believe is a form of pipeline parallelism. @stas00 is that correct?\r\n\r\nFrom my understanding, deepspeed has three major enhancements:\r\n\r\n- 3D parallelism\r\n- zero-redundancy that reduces the GPU memory footprint of a given module\r\n- some support for clusters, but I'm hazy on the details\r\n\r\n**Practical feature implications:** We can currently train t5-11b -- I believe the largest model in the library -- in a practical and affordable amount of time on the newest cloud instances. There are three benefits to pursuing pipeline parallelism and zero redundancy libraries:\r\n\r\n- Users could train large models faster\r\n- Users could train large models on more modest hardware\r\n- We would be prepared for the eventual release of even larger models in the 20 billion and potentially up to 100 billion parameter range", "Some notes following up to the raised issues:\r\n\r\n- I need to study and experiment before I'm able to answer a lot of the questions you have been asking. For example one of the important questions @alexorona asks is whether the idling GPUs can be utilized to a high capacity by integrating other libraries like deepspeed. I will be able to answer that once I try that.\r\n\r\n- The \"naive\" part @LysandreJik referred to is that, say, you spread the model over 8 gpus - 7 gpus will be idling most of the time, so it'd a terribly expensive training as you would be paying per gpu and not per its utilization. So while the current solution works there must be a more efficient ways to do that. One much more complex solution suggested here: https://github.com/pytorch/pytorch/issues/49961#issuecomment-754306157 is with the RPC mechanism. Again, I don't have any experience with it, so I will eventually get to try it and comment back.\r\n\r\n- DeepSpeed's solution to huge model size is ZeRO - while it says it can support models implementing MP, it says it's not needed since we have a working solution (100B param model w/o needing MP) and my experiments showed that with sharded DDP on my weird hardware setup I can fit 3x more data, and with DeepSpeed 3-5x, and that's just with some default config.\r\n\r\n- We are on the same page wrt to making things working on a few models - t5, gpt2 and bart is ready too. Note that Bart is a better candidate than t5 because it can be asymmetrical wrt encoder/decoder-size - so it's slighly more complex (but not by much). We were discussing a specific issue of `device_map` design, which requires us to look at all models. But that's where it can stop.\r\n\r\nMy plan is to finish the DeepSpeed integration - almost there and then look into Pipelines next.\r\n\r\nOf course, nobody needs to wait for me, I'd be just as happy for others to experiment and teach me instead ;)\r\n\r\nI commented on the current design so that the HF team better understand what we have here:\r\nhttps://github.com/huggingface/transformers/issues/8771#issuecomment-755113545\r\nLet's keep the design discussion focused in one thread, otherwise we are all over multiple threads... doesn't matter which - just pick one... If you have questions or need for clarifications please don't hesitate to ask.\r\n", "I rebased on https://github.com/huggingface/transformers/pull/9343 so now it's no longer possible to develop anything on Bart - the check fails because it wants all copy-cats to be the same:\r\n```\r\npython utils/check_copies.py\r\nTraceback (most recent call last):\r\n File \"utils/check_copies.py\", line 305, in <module>\r\n check_copies(args.fix_and_overwrite)\r\n File \"utils/check_copies.py\", line 166, in check_copies\r\n raise Exception(\r\nException: Found the following copy inconsistencies:\r\n- src/transformers/models/pegasus/modeling_pegasus.py: copy does not match models.bart.modeling_bart.BartAttention at line 141\r\n- src/transformers/models/marian/modeling_marian.py: copy does not match models.bart.modeling_bart.BartAttention at line 140\r\n- src/transformers/models/marian/modeling_marian.py: copy does not match models.bart.modeling_bart.BartEncoderLayer at line 275\r\n- src/transformers/models/marian/modeling_marian.py: copy does not match models.bart.modeling_bart.BartDecoderLayer at line 331\r\n- src/transformers/models/blenderbot_small/modeling_blenderbot_small.py: copy does not match models.bart.modeling_bart.BartAttention at line 124\r\n- src/transformers/models/blenderbot_small/modeling_blenderbot_small.py: copy does not match models.bart.modeling_bart.BartEncoderLayer at line 259\r\n- src/transformers/models/blenderbot_small/modeling_blenderbot_small.py: copy does not match models.bart.modeling_bart.BartDecoderLayer at line 315\r\n- src/transformers/models/mbart/modeling_mbart.py: copy does not match models.bart.modeling_bart.BartAttention at line 133\r\n- src/transformers/models/blenderbot/modeling_blenderbot.py: copy does not match models.bart.modeling_bart.BartAttention at line 126\r\nRun `make fix-copies` or `python utils/check_copies.py --fix_and_overwrite` to fix them.\r\nmake: *** [Makefile:25: extra_quality_checks] Error 1\r\n```\r\n\r\nHow do I move forward with my work then? I suppose the only way to proceed is to drop Bart and use one of the derivatives? So Bart isn't going MP...\r\n\r\n@patrickvonplaten, @sgugger ", "That's also why we should pause the BART PR for MP and make sure the general API is solid enough. Any change in BART will impact all related models (that was true before the split, since the other models were subclasses) so the same PR will need to do BART/Pegasus/mBART/marian etc. And probably the ses2seq template. So better make sure we're happy with the design on a model independent from the others like GPT-2 or T5 first :-) ", "> I rebased on #9343 so now it's no longer possible to develop anything on Bart - the check fails because it wants all copy-cats to be the same:\r\n> \r\n> ```\r\n> python utils/check_copies.py\r\n> Traceback (most recent call last):\r\n> File \"utils/check_copies.py\", line 305, in <module>\r\n> check_copies(args.fix_and_overwrite)\r\n> File \"utils/check_copies.py\", line 166, in check_copies\r\n> raise Exception(\r\n> Exception: Found the following copy inconsistencies:\r\n> - src/transformers/models/pegasus/modeling_pegasus.py: copy does not match models.bart.modeling_bart.BartAttention at line 141\r\n> - src/transformers/models/marian/modeling_marian.py: copy does not match models.bart.modeling_bart.BartAttention at line 140\r\n> - src/transformers/models/marian/modeling_marian.py: copy does not match models.bart.modeling_bart.BartEncoderLayer at line 275\r\n> - src/transformers/models/marian/modeling_marian.py: copy does not match models.bart.modeling_bart.BartDecoderLayer at line 331\r\n> - src/transformers/models/blenderbot_small/modeling_blenderbot_small.py: copy does not match models.bart.modeling_bart.BartAttention at line 124\r\n> - src/transformers/models/blenderbot_small/modeling_blenderbot_small.py: copy does not match models.bart.modeling_bart.BartEncoderLayer at line 259\r\n> - src/transformers/models/blenderbot_small/modeling_blenderbot_small.py: copy does not match models.bart.modeling_bart.BartDecoderLayer at line 315\r\n> - src/transformers/models/mbart/modeling_mbart.py: copy does not match models.bart.modeling_bart.BartAttention at line 133\r\n> - src/transformers/models/blenderbot/modeling_blenderbot.py: copy does not match models.bart.modeling_bart.BartAttention at line 126\r\n> Run `make fix-copies` or `python utils/check_copies.py --fix_and_overwrite` to fix them.\r\n> make: *** [Makefile:25: extra_quality_checks] Error 1\r\n> ```\r\n> \r\n> How do I move forward with my work then? I suppose the only way to proceed is to drop Bart and use one of the derivatives? So Bart isn't going MP...\r\n> \r\n> @patrickvonplaten, @sgugger\r\n\r\nI agree with @sgugger that it would be better to just work on TF and GPT2 until we have a solid API for now...But in general the idea is to implement the feature in Bart and then run `make fix-copies` and all other models are updated automatically. In case you add a lot of code to Bart (outside of `BartAttention`) it can very well be that this code has to be manually copied inside the other models as well (happy to help then :-) )", "And big sorry for making this PR so much harder for you now! But that Bart split had to happen sooner or later", "> And big sorry for making this PR so much harder for you now! But that Bart split had to happen sooner or later\r\n\r\nSurprisingly, the rebasing was super-simple. So it wasn't a hurdle at all.", "1. Bart and t5 aren't exactly the same, so in order to generalize a variety of models is needed. \r\n2. And this PR is much further ahead than t5, albeit I can spend more time merging it back into t5.\r\n\r\nIf I switch to one of the original subclasses, say, MBart, and work with it instead - will the copy-checker complain just the same?", "> If I switch to one of the original subclasses, say, MBart, and work with it instead - will the copy-checker complain just the same?\r\n\r\nI'm afraid so, unless you remove all `# Copied from` comments, but that defeats the purpose.", "Understood. thank you!\r\n\r\nIt sounds like this change will make future development of the bart family somewhat painful. Since the developer will have to constantly sync multiple files with their new development and it won't help the reviewers since now there will be multiple duplicated diffs.\r\n\r\nIt'd be much more useful to run the check/sync periodically or at will, rather than enforcing them on each `make style`, IMO. I guess time will tell.", "Thinking more about the situation - the thing is - this PR works - I put a ton of work into it - users can start using MP with the Bart family yesterday, e.g. with `--model_parallel` flag in trainer - we don't have to expose the unstable device map and using the internal default device map is sufficient for most simple uses. And if we change to a different more efficient implementation down the road - it'd be totally transparent to the users. And if it's not trainer, they can just use `model.parallelize()` without the device map, or use the device map but know it may change down the road.\r\n\r\nI'd just need to enable `self.is_parallelizable` that was just added and clean up a bit.\r\n\r\nBut it's your call.", "> e.g. with --model_parallel flag in trainer\r\n\r\nThat's one of the thing to clean up: this flag is not necessary with the current API: we can detect if a model is parallelized and avoid a confusion with the name. I'm not saying we should throw this PR in the thrash, just that it should be paused until we have had time to do all clean up we want.", ">> e.g. with --model_parallel flag in trainer\r\n>\r\n> That's one of the thing to clean up: this flag is not necessary with the current API: we can detect if a model is parallelized and avoid a confusion with the name. \r\n\r\nDo tell more? Are you planning to launch MP just because a model supports it? It sounds that you are considering dropping the `--model_parallel` cl arg in trainer\r\n\r\nOr are we talking about different things?\r\n\r\n> I'm not saying we should throw this PR in the thrash, just that it should be paused until we have had time to do all clean up we want.\r\n\r\ntldr; \r\n1. **I'm fine with hitting the pause button as you suggested.**\r\n2. this is a fully functional implementation - so **you actually can send users to this PR branch if they want to use MP with Bart** (as the family has been cast out after I rebased on master, it will require quite some work to re-add it to other Bart-like models).\r\n\r\nthe full story:\r\n\r\nThe issue is simple. Is that things are complicated. This PR overlaps with https://github.com/huggingface/transformers/pull/9323 - both have multiple changes and improvements, and I have already documented and commented on each one of the changes in both PRs, well actually 3 PRs (this one too https://github.com/huggingface/transformers/pull/9316), so leaving such code spread out over several PRs is a recipe for a huge mess down the road. It all came to be as I was working over the holidays and wasn't getting feedback (No complaints, I'm just explaining how it came to be.). As a result of it I was working on new changes but with Bart so that I could see how to generalize better. Not knowing what you'd decide I tried to leave the existing code without any API changes, hence the separate independent PRs.\r\n\r\nThe bottom line is this. Regardless of whether the current implementation is efficient or not, it works. And any future more efficient implementation will use the same API on the user-side (or perhaps something more complicated) - at the moment its just one command to turn the feature on. \r\n\r\nSo you actually can send users to this PR branch if they want to use MP with Bart-only.\r\n\r\nSo the other approach I can take is to merge parts of this PR into t5-mp PR https://github.com/huggingface/transformers/pull/9323, but it'll be again a lot of work and nobody has even looked at any of those PRs...\r\n\r\nBut then we are talking about perhaps finding a more efficient solution, and perhaps deepspeed will render a lot of it pointless anyway... (Alex thinks not.) So why waste reviewers' time... makes sense not to.\r\n\r\nSo yes, let's freeze this up and I go back to work on deepspeed.\r\n\r\nI have convinced myself it's the right thing to do and you got to hear my inner talk.\r\n\r\nJust remember it's totally functional in case someone needs it.\r\n\r\nThank you for reading.\r\n\r\n", "As t5 MP is broken in the trainer, I needed to see if it was the same with my Bart MP port - but it works:\r\n\r\n```\r\nrm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py --model_name_or_path sshleifer/distilbart-xsum-6-6 --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --evaluation_strategy=steps --fp16 --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 4 --per_device_train_batch_size 4 --predict_with_generate --eval_steps 25000 --save_steps 25000 --sortish_sampler --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 1 --n_train 2 --n_val 2 --n_test 2 --do_predict --task summarization --data_dir xsum --model_parallel \r\n```\r\n\r\nSo with this PR you **can** use `--model_parallel` automatically with out trainer scripts with Bart models. ", "As I was trying to see if I can find a way to utilize the idling GPUs, I run these benchmarks - haven't found anything useful yet, but the interesting finding is that while we get a huge performance hit with evaluation and beam size > 1, actually the training time is faster than non-MP version, despite all the data copying \r\n\r\nThis PR beats master on training time almost by half 8.6sec vs 15.8 sec, but of course it has 2 gpus vs 1 gpus!!! But it beats even the DDP solution 10.6sec by 20%!\r\n\r\nSo perhaps there is something good in here we just need to understand why is it faster than DDP.\r\n\r\nUnfortunately I have an uneven GPUs setup, so it's hard to get very useful benchmarks. Perhaps someone with 2 identical GPUs could re-run these and report back.\r\n\r\nFor posterity here are the results I'm getting with 1x 8gb and 1x 24gb gpus:\r\n\r\n```\r\n# w/o MP w/o DDP\r\n\r\n\r\nrm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py --model_name_or_path sshleifer/distilbart-xsum-6-6 --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --evaluation_strategy=steps --fp16 --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 4 --per_device_train_batch_size 4 --predict_with_generate --eval_steps 25000 --save_steps 25000 --sortish_sampler --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 1 --n_train 200 --n_val 200 --task summarization --data_dir xsum\r\n\r\n2021-01-10 16:57:43 | INFO | __main__ | train_runtime = 15.8407\r\n2021-01-10 16:58:02 | INFO | __main__ | val_runtime = 19.0772\r\n\r\n# w/o MP w/ DDP\r\n\r\n\r\nrm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path sshleifer/distilbart-xsum-6-6 --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --evaluation_strategy=steps --fp16 --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 4 --per_device_train_batch_size 4 --predict_with_generate --eval_steps 25000 --save_steps 25000 --sortish_sampler --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 1 --n_train 200 --n_val 200 --task summarization --data_dir xsum\r\n\r\n2021-01-10 16:58:42 | INFO | __main__ | train_runtime = 10.6299\r\n2021-01-10 16:58:53 | INFO | __main__ | val_runtime = 11.4454\r\n\r\n# w/ MP w/o DDP\r\n\r\nrm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py --model_name_or_path sshleifer/distilbart-xsum-6-6 --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --evaluation_strategy=steps --fp16 --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 4 --per_device_train_batch_size 4 --predict_with_generate --eval_steps 25000 --save_steps 25000 --sortish_sampler --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 1 --n_train 200 --n_val 200 --model_parallel --task summarization --data_dir xsum\r\n\r\n2021-01-10 16:49:00 | INFO | __main__ | train_runtime = 8.6264\r\n2021-01-10 16:51:14 | INFO | __main__ | val_runtime = 134.0955\r\n\r\nruntime is very slow due to beam search (==4).\r\n\r\nsame w/ --eval_beams 1\r\n\r\n2021-01-10 16:56:10 | INFO | __main__ | train_runtime = 8.657\r\n2021-01-10 16:56:41 | INFO | __main__ | val_runtime = 31.4318\r\n\r\n\r\n# w/ MP w/ DDP\r\n\r\nrm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 python -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path sshleifer/distilbart-xsum-6-6 --output_dir output_dir --adam_eps 1e-06 --do_eval --do_train --evaluation_strategy=steps --fp16 --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_eval_batch_size 4 --per_device_train_batch_size 4 --predict_with_generate --eval_steps 25000 --save_steps 25000 --sortish_sampler --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 1 --n_train 200 --n_val 200 --model_parallel --task summarization --data_dir xsum\r\n\r\nthis doesn't work: can't mix this implementation of MP w/ DDP\r\n\r\nAssertionError: DistributedDataParallel device_ids and output_device arguments only work with single-device GPU modules, but got device_ids [0], output_device 0, and module parameters {device(type='cuda', index=0), device(type='cuda', index=1)}.\r\n``` ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "too long. closing.", "Hello, @stas00 is there any update on BART based model parallelism? also about model.parallelize() for BlenderBot? Thanks. ", "This line of work has been abandoned as it's highly inefficient. Please use DeeepSpeed which works with any model https://huggingface.co/docs/transformers/main/main_classes/deepspeed" ]

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

[]

[ "I think what's going on is that `config.add_final_bias_logits` is unused.\r\n\r\nIn modeling_bart.py line 148 we call\r\n\r\n```python\r\nself.register_buffer(\"final_logits_bias\", torch.zeros((1, self.model.shared.num_embeddings)))\r\n```\r\nregardless of the config, and then if it's in the state dict it will get loaded by `from_pretrained`.\r\n\r\nI do think that `final_bias_logits` is in the marian state dict, as this line would have `KeyError`'d during conversion otherwise: https://github.com/sshleifer/transformers_fork/blob/121ec9dced3d068352078e7c3523ecd66830e39e/src/transformers/models/marian/convert_marian_to_pytorch.py#L461-L461 \r\n\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hey @forest1988,\r\n\r\nThanks for your issue! You're 100% correct. The docs need to be updated here! The output is actually never a list, it should always be a `Tuple(Tuple(torch.FloatTensor))` - I'll make a PR afterward. \r\nAnd in Bart, `past_key_values` always consists of `selt_attn_present_key_value` and `cross_attn_present_key_value`.", "Hi @patrickvonplaten,\r\n\r\nThank you for your quick response to this issue!\r\nThe update of the docs and your answer to my question -- what `past_key_values` consists of -- are very helpful for me!\r\n", "Hi @patrickvonplaten,\r\n\r\nExcuse me for my frequent questions.\r\nI created a new issue https://github.com/huggingface/transformers/issues/9391, in which I ask your help about the `past_key_values` in Bart (Seq2SeqLM) and GPT-2 (CausalLM).\r\n\r\nI think it is not an error, but a feature request.\r\n\r\nIf you could check it out when you have time, it would be greatly appreciated." ]

[]

[]

[ "Thanks @Qbiwan!" ]

[]

[]

[ "Thank you for the pr. \r\nPlease also update the documentation here: https://huggingface.co/transformers/custom_datasets.html#qa-squad\r\nLine-> end_positions[-1] = encodings.char_to_token(i, answers[i]['answer_end'] + 1)\r\nto -> end_positions[-1] = tokenizer.model_max_length", "So what is the state of this issue? What version of processing script should we use?" ]

[]

[]

[ "I'm good with changing to `torch.nn.LayerNorm`. At @stas00 - do you know what the advantage of `apex.normalization.FusedLayerNorm` is supposed to be? Why did we add `apex.normalization.FusedLayerNorm` in the first place?", "Prior to about a year ago, `apex.normalization.FusedLayerNorm` was faster than `torch.nn.LayerNorm`, but then the former got ported to native `torch.nn.LayerNorm`, and now the native appears to be faster - at least the 2 cards I have experimented with. I checked with pt-1.4 .. pt-1.8.dev\r\n\r\nIf you have other than gtx-1070/rtx-3090 cards which I benchmarked with please run that benchmark and see if it stands true for other cards: https://github.com/pytorch/pytorch/issues/37713#issuecomment-753434842\r\nit only takes a few seconds if you have apex installed already. To install apex:\r\n\r\n```\r\ngit clone https://github.com/NVIDIA/apex\r\ncd apex\r\nrm -rf build\r\npip install --global-option=\"--cpp_ext\" --global-option=\"--cuda_ext\" .\r\n```\r\n\r\nThe benchmark measures and reports a total run time, so the smaller the numbers the faster it is.\r\n\r\nIf you do run the benchmarks please post your results at https://github.com/pytorch/pytorch/issues/37713 so that it can be seen whether it's safe to drop `apex.normalization.FusedLayerNorm` based on hard data and not anecdotal info.\r\n\r\nThank you." ]

[]

[]

[ "looks like this is what you're looking for: https://github.com/huggingface/transformers/blob/ae333d04b29a25be1a70eaccd6260c294c243c5b/src/transformers/file_utils.py#L842-L855", "Hey @ck37,\r\n\r\nThanks for your issue! Yes, this typo should be corrected -> it would be great if you could open a PR :-) " ]

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

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Hi!\r\n\r\nYou could try replacing the CrossEntropy loss with [this Dice Loss](https://github.com/fursovia/self-adj-dice), which may help you with the imbalance issues. In the paper linked in the repo they explain the design process. I have tried it with mixed results, although for my (skewed) dataset, weighting the CrossEntropy loss with the inverse frequency of each category has worked best.\r\n\r\nLet me know if it works for you 👍🏻 \r\n\r\nAs a last resort, you could try undersampling category 1 to match the second, and maybe combine this with a weighted loss as well.", "Thanks @viantirreau for your suggestions I actually tried to use the dice loss as well as class weights with crossentropy loss and the results I got from the crossentropyloss was actually better than what I am getting with the dice loss , however both of them fails to detect the categories 3-5 . I will try to do undersampling as the last resort however speaking of the class weights I have used the sklearn's compute_class_weight for getting my class_weights as follows:\r\n```\r\nfrom sklearn.utils.class_weight import compute_class_weight\r\n\r\n#compute the class weights\r\nclass_wts = compute_class_weight('balanced', np.unique(train_labels), train_labels)\r\n\r\n```\r\nCan you suggest any other workaround other than this strategy , I have came to know that neural networks tends to ignore class weights through an answer on one of the stackexchange sites . \r\n", "Hi!\r\n\r\nI think it shouldn't make any difference, using my method returns exactly the same as Sklearn's `compute_class_weight`, but normalized so as to add up to 1. \r\n\r\nUsing these class counts,\r\n\r\n> Category 1 10000\r\n> Category 2 2000\r\n> Category 3 400\r\n> Category 4 300 \r\n> Category 5 100\r\n\r\n I get the following weights for the respective categories `array([0.00608519, 0.03042596, 0.15212982, 0.20283976, 0.60851927])`. \r\nAnother nice (and extreme) experiment you could try is to over-emphasize the weights for the underrepresented classes, for example using something like `array([0.001, 0.001, 0.001, 0.001, 0.996])`, just as a sanity check to confirm the optimizer learns something about category 5.\r\nI would also start testing the model's predictions on the training data first (it should overfit), and only then try to measure its generalization abilities on a held out development set. Maybe your gradients are not backpropagating to the first layers, your learning rate is way too big or you need some warmup steps.\r\n\r\nLet me know if any of this works :)\r\nGood luck!", "Hi , @viantirreau Thanks for your suggestions. I did try it by reducing the class weights for majority classes and emphasizing the weights for minority class say category 5 , I found out that my neural network is still not able to learn anything about those classes , I have tried it with learning rates 1e-5,2e-5,5e-5 and warmup step of 1000 however no improvement is still being made on it . Any optimization strategies for the hyperparameters you can suggest?", "You're welcome! \r\nMmh interesting, what Transformers model are you using? Also, from what pretrained checkpoint are you initializing it? Are you sure there are no warnings like 'missing parameters, initializing from scratch'?\r\n\r\nI have faced some vanishing gradient problems in the past that manifest as an unexplainable \"preference\" for a class, so I'd make sure that your gradients are alright. I find [this](https://gist.github.com/viantirreau/ec591a428a5c0112bd8fa84f70968574) code snippet pretty useful to diagnose the gradient flow by plotting its values across each layer. If you use Weights&Biases as a logging tool, you can `watch` the model and create even nicer plots in their dashboard. A warmup strategy was crucial in my experience to eliminate the gradient problems.\r\n\r\nAlso, if you are manually adjusting the attention masks or some of the model inputs, make sure to not pass ignore_index in some/all of the inputs. Some prints will help in making sure that the model inputs are as expected.\r\nAnother idea I'd test is to completely eliminate your categories 1 and 2 from the training examples, and see if the same phenomenon happens to the most common class by then (should be category 3). Try this alone and see if including the inverse frequency weights in the loss helps in any way.\r\n\r\nGood luck!\r\nGood luck!", "Hi , @viantirreau sorry for the delay in response I haven't received any warnings as such . I am using the bert transformer with bert-base-multilingual-cased as the checkpoint , I was trying to first build a custom model from the final output layer of the BERT model in order to accomodate the class imbalance issue . I haven't tried the weights and biases yet will surely check it out. I will try your other suggestions and will let you know about it. Thanks for your suggestions.\r\n", "Hi , @viantirreau sorry for the delay in response .I finally figured out the reason behind the performance degradation it was because I was freezing the base layers and only fine tuning one extra layer which I added to the base model. Since, the model had the data imbalance issue already into it ,it was being biased towards the majority samples . It however performed much better when I unfreezed the base layers , however on the cost of additional gpu training time.", "Hi, @nikhil6041. I'm glad you figured it out. Thanks for reaching back with your experience and solution! 🙌🏻 ", "I am actually having the same problem you experienced. I am building a multi-label multi-class classification Bert/distilbert model and encountered the same issue with my 20 classes. Of course the data is imbalanced, and like you I thought I had locked down the base layers but I realized I hadn't and that model performed slight better with the imbalanced data than the locked down model. I could not figure out why other than knowing imbalanced data is a big deal. Unfortunately, the data set I have is extremely small so that is also probably playing a big role. @viantirreau and @nikhil6041, one method I have seen used is a weighted cost function like adacost. Has anyone had any success implementing this with Distilbert? I can provide more details or open a new ticket but this seemed very closely related.", "Hi @johnrodriguez190380 have you tried KL divergence loss function ? Try to use it once. Also there have been certain instances where the usage of weighted cost function doesn't help much. I don't remember the paper which pointed out this thing but I read it somewhere in a stackoverflow answer. If in case the dataset you are using is really small you can try some data augmentation techniques , you can also use this [](https://github.com/makcedward/nlpaug )repo it maybe helpful for u i guess. Let me know if anyone of this serves ur usecase.", "Hi nikhil, how are you? maybe can you share your code/colab/repo to see how you solve the issue? ", "Hi @nikhil6041 could you please share the script you wrote for changing the loss? I really appreciate it if you can share it with us!", "Hey @sasu08 and @un-lock-me sorry for the late response I used the sadice loss for this one however it didnt solve my problem fully however there are certain other ways you can possibly try for this thing try to use text augmentations (there are various ways for it like using synonyms ,back translation to name a few , there is one library named nlpaug which might come in handy for both of you , have a look at it. For the sadice loss part you can have a look at my repo [](https://github.com/nikhil6041/OLI-and-Meme-Classification) , here is the link to the nlpaug library [](https://github.com/makcedward/nlpaug). Hope it helps!!", "For the Sadie loss I could not find it in the repo could you please share the link here?", "@un-lock-me Here it is [sadice loss] https://github.com/fursovia/self-adj-dice ", "Hi @nikhil6041 can you please share your notebook? ", "Hi @pratikchhapolika you can find all my notebooks in this [repo](https://github.com/nikhil6041/OLI-and-Meme-Classification)", "@nikhil6041 thanks for the helpful repo. I was running the code on Google Colab and I used the provided dataset.\r\nHowever, I am getting this error in the Training Loop. \r\nTypeError: forward() got an unexpected keyword argument 'token_type_ids' \r\nWhat can be the issue? Thanks " ]

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

[]

[ "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "hi @vgoklani \r\n\r\n`datasets` is not required for tokenizers, so it's unlikely to get this error when just importing the tokenizer. Are you running any examples scripts? because those require `datasets` lib", "Hi @patil-suraj Happy New Year!\r\n\r\nHere is the stack trace:\r\n\r\n root@b5d80f9670ea:~/src# ipython\r\n Python 3.8.5 (default, Sep 4 2020, 07:30:14)\r\n Type 'copyright', 'credits' or 'license' for more information\r\n IPython 7.19.0 -- An enhanced Interactive Python. Type '?' for help.\r\n\r\n In [1]: from transformers.models.roberta.tokenization_roberta_fast import RobertaTokenizerFast\r\n 2021-01-01 12:03:22.340215: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0\r\n ---------------------------------------------------------------------------\r\n ImportWarning Traceback (most recent call last)\r\n <ipython-input-1-2758fed1e79c> in <module>\r\n ----> 1 from transformers.models.roberta.tokenization_roberta_fast import RobertaTokenizerFast\r\n\r\n /opt/conda/lib/python3.8/site-packages/transformers/__init__.py in <module>\r\n 32 absl.logging._warn_preinit_stderr = False\r\n 33\r\n ---> 34 from . import dependency_versions_check\r\n 35\r\n 36 # Configuration\r\n\r\n /opt/conda/lib/python3.8/site-packages/transformers/dependency_versions_check.py in <module>\r\n 32 if pkg == \"tokenizers\":\r\n 33 # must be loaded here, or else tqdm check may fail\r\n ---> 34 from .file_utils import is_tokenizers_available\r\n 35\r\n 36 if not is_tokenizers_available():\r\n\r\n /opt/conda/lib/python3.8/site-packages/transformers/file_utils.py in <module>\r\n 101\r\n 102 try:\r\n --> 103 import datasets # noqa: F401\r\n 104\r\n 105 # Check we're not importing a \"datasets\" directory somewhere\r\n\r\n /opt/conda/lib/python3.8/site-packages/datasets/__init__.py in <module>\r\n 51\r\n 52 if int(pyarrow.__version__.split(\".\")[1]) < 16 and int(pyarrow.__version__.split(\".\")[0]) == 0:\r\n ---> 53 raise ImportWarning(\r\n 54 \"To use `datasets`, the module `pyarrow>=0.16.0` is required, and the current version of `pyarrow` doesn't match this condition.\\n\"\r\n 55 \"If you are running this in a Google Colab, you should probably just restart the runtime to use the right version of `pyarrow`.\"\r\n\r\n ImportWarning: To use `datasets`, the module `pyarrow>=0.16.0` is required, and the current version of `pyarrow` doesn't match this condition.\r\n If you are running this in a Google Colab, you should probably just restart the runtime to use the right version of `pyarrow`.\r\n\r\n\r\n---\r\n\r\nAn older version of pyarrow was installed, but regardless, this happens immediately after the import. Upgrading pyarrow makes this warning disappear, but regardless, this shouldn't happen.\r\n", "cc @sgugger ", "This is because transformers imports all optional dependencies (like datasets) during its init. There will be some work to avoid doing that in the coming weeks.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 2604155188, "node_id": "MDU6TGFiZWwyNjA0MTU1MTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Benchmarks", "name": "Benchmarks", "color": "2DF372", "default": false, "description": "Issues related to Memory regressions in tests and scripts" }, { "id": 2690307185, "node_id": "MDU6TGFiZWwyNjkwMzA3MTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Performance", "name": "Performance", "color": "207F32", "default": false, "description": "" } ]

[]

[ "Not an answer to your issue/question, but have you tried running in distributed training (DDP), which is the recommended way of running over multiple GPUs: https://github.com/huggingface/transformers/tree/master/examples#distributed-training-and-mixed-precision\r\n\r\nWould be curious to see the same with/without NVLink experiment there.", "Hmm, I don't have much experience using torch.distributed. I tried just running the existing script with `python -m torch.distributed.launch --nproc_per_node 2 train.py`, but it runs out of GPU memory almost immediately, so I assume I'm doing something wrong.\r\n\r\nIf you have a link to some documentation that explains how to set up the training script so that it can be used with torch.distributed, I can give that a try.", "The command you posted \"should\" work.\r\n\r\n@sgugger might have links to better content when he's back, but the PyTorch tutorials are pretty good: https://pytorch.org/tutorials/beginner/dist_overview.html#data-parallel-training\r\n\r\nYour initial experiment is using `DataParallel` (not `DistributedDataParallel`) under the hood.", "OK, I got around to spending some more time with this today. I realized that the `run_language_modeling.py` script can do everything my script was doing, and it uses DDP by default (Note: looking at the most recent version on git, I see that `run_language_modeling.py` has been replaced by `run_clm.py`. However, after trying to upgrade transformers to that version, it seems to no longer use the GPU for reasons I don't have time to debug.).\r\n\r\nSo now I'm just using that, with:\r\n\r\n```\r\npython -m torch.distributed.launch --nproc_per_node 2 \\\r\n ~/git/transformers/examples/language-modeling/run_language_modeling.py \\\r\n --model_type gpt2 \\\r\n --config_name ./csrc_config \\\r\n --tokenizer_name ./csrc_tokenizer \\\r\n --fp16 --fp16_opt_level O3 \\\r\n --do_train --output_dir csrc_output \\\r\n --per_device_train_batch_size 4 \\\r\n --train_data_file plainsrc_all.txt --block_size 128\r\n```\r\n\r\nFor single GPU I drop the `torch.distributed.launch` and use `CUDA_VISIBLE_DEVICES=1`, to disable NVLINK I use `NCCL_P2P_DISABLE=1` as before. The `--block_size 128` argument is to match the default from my training script (without it I run out of GPU RAM).\r\n\r\nResults:\r\n\r\nModel | Block Size | GPUs | NVLINK | ETA | Perf\r\n------|------------|------|--------|-----|-----\r\nSmall | 512 | 2GPU | No | 17:08:12 | 4.75it/s\r\nSmall | 512 | 2GPU | Yes | 10:24:20 | 7.79it/s\r\nSmall | 512 | 1GPU | N/A | 18:37:17 | 8.74it/s\r\nMedium | 512 | 2GPU | No | 43:07:49 | 1.89it/s\r\nMedium | 512 | 2GPU | Yes | 26:19:09 | 3.09it/s\r\nMedium | 512 | 1GPU | N/A | 45:36:37 | 3.57it/s\r\nSmall | 128 | 2GPU | No | 48:12:05 | 6.75it/s\r\nSmall | 128 | 2GPU | Yes | 21:26:31 | 15.17it/s\r\nSmall | 128 | 1GPU | N/A | 30:54:41 | 21.06it/s\r\nMedium | 128 | 2GPU | No | 118:43:09 | 2.74it/s\r\nMedium | 128 | 2GPU | Yes | 51:55:58 | 6.27it/s\r\nMedium | 128 | 1GPU | N/A | 74:02:16 | 8.79it/s\r\nLarge | 128 | 2GPU | No | 239:19:44 | 1.36it/s\r\nLarge | 128 | 2GPU | Yes | 102:17:18 | 3.18it/s\r\nLarge | 128 | 1GPU | N/A | 143:34:42 | 4.54it/s\r\n\r\nSo the general observation is that for block size 512, two GPUs without NVLink are about the same performance as a single GPU. For block size 128, two GPUs without NVLink are typically quite a bit *slower* than a single GPU.\r\n\r\nIt doesn't seem like DistributedDataParallel helps with this issue, in other words.\r\n", "I think @sgugger has experience with multi-GPU, and works on the example scripts, pinging him!", "A friend was linking me to this issue. Thank you for your work on this benchmark! It is some interesting data. I still believe the poor performance could be a hardware issue though.\r\n\r\nAs far as I know, RTX 3090 GPUs have peer-to-peer access disable, or in other words, you cannot transfer memory from GPU to GPU on these GPUs. All data is first routed through the CPU, which is often slow because the CPU buffers are not pinned, meaning that memory transfers are _synchronous_. So in my eyes, slow performance without NVLink is a hardware issue in this case. It would be curious, though, if these numbers would be similar for peer-to-peer enabled GPUs. Do you have access to such a GPU?", "You're thinking of something like P2P over PCIe? You're right that NVIDIA has disabled that for the 3090s. The only other hardware I have access to is our HPC cluster, which has RTX8000s and V100s (non-NVLINKed); I believe both show similar slowdowns.\r\n\r\nOne thing I have been looking into is whether using something like DeepSpeed will help. I got their Megatron-LM example working and it does much better at scaling to two at least GPUs without NVLINK using the 1-bit Adam optimizer. I'm still waiting for my HPC job to get scheduled to confirm that it scales well there too. If that works then presumably something like what's being done for the t5-3b model here would help? https://github.com/huggingface/transformers/issues/8771", "If you confirm you have the same results for the RTX 8000 that would rule out any GPU issue. It could still be a hardware issue with PCIe lanes. There is a bandwidth test I believe among the NVIDIA samples that come with CUDA with which you can test the available bandwidth to/from GPUs. If this shows good numbers it should be purely an issue of software or network architecture.", "OK, I'll give this a try. Our HPC cluster is a bit busy so it may be a while before I can get a slot on the RTX 8000 nodes.", "I managed to get some time on a node with 4x V100s. For the Large model, it gets 3.83s/it with an ETA of 1248:01:43 (!).\r\n\r\nHere's the output of p2pBandwidthLatencyTest on the V100 system:\r\n\r\n```\r\n[bd52@gv02 p2pBandwidthLatencyTest]$ ./p2pBandwidthLatencyTest \r\n[P2P (Peer-to-Peer) GPU Bandwidth Latency Test]\r\nDevice: 0, Tesla V100-PCIE-32GB, pciBusID: 6, pciDeviceID: 0, pciDomainID:0\r\nDevice: 1, Tesla V100-PCIE-32GB, pciBusID: 2f, pciDeviceID: 0, pciDomainID:0\r\nDevice: 2, Tesla V100-PCIE-32GB, pciBusID: 86, pciDeviceID: 0, pciDomainID:0\r\nDevice: 3, Tesla V100-PCIE-32GB, pciBusID: d8, pciDeviceID: 0, pciDomainID:0\r\nDevice=0 CAN Access Peer Device=1\r\nDevice=0 CAN Access Peer Device=2\r\nDevice=0 CAN Access Peer Device=3\r\nDevice=1 CAN Access Peer Device=0\r\nDevice=1 CAN Access Peer Device=2\r\nDevice=1 CAN Access Peer Device=3\r\nDevice=2 CAN Access Peer Device=0\r\nDevice=2 CAN Access Peer Device=1\r\nDevice=2 CAN Access Peer Device=3\r\nDevice=3 CAN Access Peer Device=0\r\nDevice=3 CAN Access Peer Device=1\r\nDevice=3 CAN Access Peer Device=2\r\n\r\n***NOTE: In case a device doesn't have P2P access to other one, it falls back to normal memcopy procedure.\r\nSo you can see lesser Bandwidth (GB/s) and unstable Latency (us) in those cases.\r\n\r\nP2P Connectivity Matrix\r\n D\\D 0 1 2 3\r\n 0 1 1 1 1\r\n 1 1 1 1 1\r\n 2 1 1 1 1\r\n 3 1 1 1 1\r\nUnidirectional P2P=Disabled Bandwidth Matrix (GB/s)\r\n D\\D 0 1 2 3 \r\n 0 768.57 11.42 11.52 11.53 \r\n 1 11.39 770.46 11.50 11.53 \r\n 2 11.42 11.43 771.22 11.45 \r\n 3 11.42 11.43 11.44 769.70 \r\nUnidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)\r\n D\\D 0 1 2 3 \r\n 0 767.06 9.93 9.68 9.49 \r\n 1 9.93 769.33 9.33 9.50 \r\n 2 9.87 9.35 769.70 10.05 \r\n 3 9.66 9.68 9.92 770.08 \r\nBidirectional P2P=Disabled Bandwidth Matrix (GB/s)\r\n D\\D 0 1 2 3 \r\n 0 771.22 15.98 16.04 16.16 \r\n 1 16.00 773.51 16.11 16.07 \r\n 2 15.90 15.99 772.75 15.83 \r\n 3 16.05 16.01 15.85 772.55 \r\nBidirectional P2P=Enabled Bandwidth Matrix (GB/s)\r\n D\\D 0 1 2 3 \r\n 0 770.84 18.72 18.41 18.07 \r\n 1 18.52 772.94 18.82 18.30 \r\n 2 18.41 18.16 771.80 19.13 \r\n 3 18.40 17.99 18.94 771.22 \r\nP2P=Disabled Latency Matrix (us)\r\n GPU 0 1 2 3 \r\n 0 1.89 14.77 14.42 14.59 \r\n 1 14.52 1.91 15.50 15.50 \r\n 2 15.53 15.42 1.87 14.44 \r\n 3 14.76 14.71 14.51 1.82 \r\n\r\n CPU 0 1 2 3 \r\n 0 2.52 8.33 8.61 8.55 \r\n 1 8.20 2.49 8.50 8.49 \r\n 2 8.30 8.29 2.61 8.69 \r\n 3 8.41 8.36 8.74 2.56 \r\nP2P=Enabled Latency (P2P Writes) Matrix (us)\r\n GPU 0 1 2 3 \r\n 0 1.86 1.60 1.65 1.64 \r\n 1 1.59 1.91 1.64 1.65 \r\n 2 1.65 1.63 1.88 1.58 \r\n 3 1.65 1.64 1.59 1.82 \r\n\r\n CPU 0 1 2 3 \r\n 0 2.51 2.05 2.02 2.02 \r\n 1 2.14 2.54 2.04 2.02 \r\n 2 2.28 2.18 2.61 2.18 \r\n 3 2.32 2.19 2.24 2.73 \r\n\r\nNOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.\r\n```\r\n\r\nAnd for comparison, here's the dual 3090 w/NVLINK system:\r\n\r\n```\r\n[P2P (Peer-to-Peer) GPU Bandwidth Latency Test]\r\nDevice: 0, GeForce RTX 3090, pciBusID: 1, pciDeviceID: 0, pciDomainID:0\r\nDevice: 1, GeForce RTX 3090, pciBusID: 21, pciDeviceID: 0, pciDomainID:0\r\nDevice=0 CAN Access Peer Device=1\r\nDevice=1 CAN Access Peer Device=0\r\n\r\n***NOTE: In case a device doesn't have P2P access to other one, it falls back to normal memcopy procedure.\r\nSo you can see lesser Bandwidth (GB/s) and unstable Latency (us) in those cases.\r\n\r\nP2P Connectivity Matrix\r\n D\\D 0 1\r\n 0 1 1\r\n 1 1 1\r\nUnidirectional P2P=Disabled Bandwidth Matrix (GB/s)\r\n D\\D 0 1 \r\n 0 831.56 11.25 \r\n 1 11.33 831.12 \r\nUnidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)\r\n D\\D 0 1 \r\n 0 810.85 52.77 \r\n 1 52.85 832.89 \r\nBidirectional P2P=Disabled Bandwidth Matrix (GB/s)\r\n D\\D 0 1 \r\n 0 812.31 16.55 \r\n 1 16.75 838.03 \r\nBidirectional P2P=Enabled Bandwidth Matrix (GB/s)\r\n D\\D 0 1 \r\n 0 821.29 101.41 \r\n 1 101.80 835.34 \r\nP2P=Disabled Latency Matrix (us)\r\n GPU 0 1 \r\n 0 1.59 33.13 \r\n 1 20.55 1.48 \r\n\r\n CPU 0 1 \r\n 0 2.89 8.85 \r\n 1 8.81 2.85 \r\nP2P=Enabled Latency (P2P Writes) Matrix (us)\r\n GPU 0 1 \r\n 0 1.59 1.43 \r\n 1 1.40 1.47 \r\n\r\n CPU 0 1 \r\n 0 2.93 2.45 \r\n 1 2.39 2.90 \r\n```", "Thank you - these data are very valuable! It also shows that no hardware problem exists. It seems you could confirm poor performance on the V100 which makes it very likely that you can also reproduce performance issues with the RTX 8000. With that, it seems the only option is that it is an issue with the combination of parallelism and network architecture. ", "Great benchmarks! Thank you for sharing the data, @moyix \r\n\r\nDo you have the same benchmarks for V100s too - just one set is enough (1 vs 2).\r\n\r\nAlso, why are you running comparison benchmarks on such huge number of items? Running enough items so that runtime is around a few minutes should be plenty to see the difference. Or is it that you were aborting these early and just recording the projected ETA and it/s from tqdm? `e.g. --max_steps 1000`\r\n\r\nHere are some ideas that may address your issue\r\n\r\n1. If I understand things right 3090 won't work at full capacity until we get pytorch w/ cuda-11.2\r\nhttps://github.com/pytorch/pytorch/issues/50232\r\nI don't know the nuances yet, but could it be that the communication channel is limited with cuda-11.0?\r\n\r\n That's why I wanted to see the results from VT100\r\n\r\n2. In one place it was suggested to check how your GPUs are inter-connected with help of:\r\n ```\r\n nvidia-smi topo -m\r\n ```\r\n that's do this check with NVLink disconnected.\r\n\r\n3. Also are sure your both GPUs running on the same speed PCIx (e.g. 8x if it's a consumer MB)? It must be, but just checking. I suppose doing a single GPU test on the other GPU would show if it's somehow on a slow PCIx slot. But I'd just test to rule that out. Should you get a slower outcome doing the same test on the 2nd gpu would explain the situation.\r\n", "OK, so here is my benchmark with the same tool.\r\n\r\n**edit**: my initial benchmark had a bug in it as pointed out by @sgugger as one has to tweak `--max_steps` if changed to more gpus - I'm proposing to change that and have a way to have a fixed dataset truncation regardless of the number of gpus used. https://github.com/huggingface/transformers/issues/9801\r\n\r\nSo for 1 gpu, I had to double `--max_steps` to get the same number of items. The rest of this comment has been updated to reflect the corrected state:\r\n\r\nHardware 2x TITAN RTX 24GB each + NVlink\r\n\r\n|type| time secs |\r\n|----|-----|\r\n| 1: | 204 |\r\n| 2:DP w/ NVlink| 110 |\r\n| 2:DDP w/ NVlink| 101 |\r\n| 2:DDP w/o NVlink | 131 |\r\n\r\nI get the same bus report w/ and w/o NCCL_P2P_DISABLE=1 - I don't think `nvidia-smi` respects this env var:\r\n\r\n```\r\nNCCL_P2P_DISABLE=1 nvidia-smi topo -m\r\n\r\n GPU0 GPU1 CPU Affinity NUMA Affinity\r\nGPU0 X NV2 0-23 N/A\r\nGPU1 NV2 X 0-23 N/A\r\n```\r\n\r\nbut clearly the runtime is much slower w/o the NVlink as the benchmark shows, so pytorch/cuda does respect it.\r\n\r\nAnalysis:\r\n\r\n1. DP is ~10% slower than DDP w/ NVlink, but ~15% faster than DDP w/o NVlink\r\n2. DDP w/ NVLink doubles the speed of single gpu, so the communication overheard is almost nill in this particular experiment \r\n\r\nHere is the full benchmark code and outputs:\r\n\r\n```\r\n# 1 gpu\r\n\r\nrm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0 python run_clm.py --model_name_or_path gpt2 \\\r\n--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train --output_dir \\\r\n/tmp/test-clm --per_device_train_batch_size 4 --max_steps 400\r\n\r\n{'train_runtime': 204.8202, 'train_samples_per_second': 1.953, 'epoch': 0.69}\r\n\r\n# DP\r\n\r\nrm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 python run_clm.py --model_name_or_path gpt2 \\\r\n--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train --output_dir \\\r\n/tmp/test-clm --per_device_train_batch_size 4 --max_steps 200\r\n\r\n{'train_runtime': 110.5948, 'train_samples_per_second': 1.808, 'epoch': 0.69}\r\n\r\n# DDP\r\n\r\nrm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node 2 \\\r\nrun_clm.py --model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \\\r\n--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200\r\n\r\n{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}\r\n\r\n# DDP w/o NVlink\r\n\r\nrm -r /tmp/test-clm; NCCL_P2P_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \\\r\n--nproc_per_node 2 run_clm.py --model_name_or_path gpt2 --dataset_name wikitext \\\r\n--dataset_config_name wikitext-2-raw-v1 --do_train --output_dir /tmp/test-clm \\\r\n--per_device_train_batch_size 4 --max_steps 200\r\n\r\n{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}\r\n```", "Yes, apologies for the confusion; the ETA numbers above are from aborting early (after a few minutes) and noting the ETA. I actually did compile PyTorch from source with CUDA 11.2 and it doesn't seem to have changed the results (although I don't know if there are further changes PyTorch will make to take full advantage of 11.2).\r\n\r\nYour benchmark code is much more self-contained than mine, so I will give your benchmarks a shot with the RTX8000 and V100 nodes on our cluster, but it will probably be a few days before I can get time there as the ICML deadline is very close :)\r\n\r\nHere's nvidia-smi -m topo for the 3090 machine:\r\n\r\n```\r\nnvidia-smi topo -m\r\n GPU0 GPU1 CPU Affinity NUMA Affinity\r\nGPU0 X NV4 0-31 N/A\r\nGPU1 NV4 X 0-31 N/A\r\n```", "Note that the timing compare 200 training steps, so the numbers you reported wrong @stas00 in the sense that 2 GPUs have covered 400 samples instead of 200. Training on the full dataset would therefore go twice as fast as with one GPU.", "This is correct - that my report was incorrect. Thank you for validating my concern in https://github.com/huggingface/transformers/issues/9801, @sgugger \r\n\r\nThat's why I'm asking for a less confusing way to truncate the dataset.\r\n\r\nI need to find an easy-way to do it so I don't have to be in full thinking capacity if I do it late at night which was the case last night.\r\n\r\nI will revisit my benchmark with corrections hopefully today. \r\n\r\nBut it doesn't change the fact that nvlink gives 30% faster performance.\r\n", "> Yes, apologies for the confusion; the ETA numbers above are from aborting early (after a few minutes) and noting the ETA. \r\n\r\nThat's what I guessed - I am glad you didn't waste all that electricity to run these to completion! It was a smart move, since you waited a few minutes.\r\n\r\n> I actually did compile PyTorch from source with CUDA 11.2 and it doesn't seem to have changed the results (although I don't know if there are further changes PyTorch will make to take full advantage of 11.2).\r\n\r\nOh, thank you for validating that! \r\n\r\nBuilding pytorch from source is hard! Hat off to you!\r\n\r\nYes, we don't know whether everything has been put in place for 11.2 support. \r\n\r\n> Your benchmark code is much more self-contained than mine, so I will give your benchmarks a shot with the RTX8000 and V100 nodes on our cluster, but it will probably be a few days before I can get time there as the ICML deadline is very close :)\r\n\r\nplease note that I corrected a mistake in my benchmark as kindly pointed out by @sgugger:\r\nhttps://github.com/huggingface/transformers/issues/9371#issuecomment-767323420\r\n \r\n> Here's nvidia-smi -m topo for the 3090 machine:\r\n> \r\n> ```\r\n> nvidia-smi topo -m\r\n> GPU0 GPU1 CPU Affinity NUMA Affinity\r\n> GPU0 X NV4 0-31 N/A\r\n> GPU1 NV4 X 0-31 N/A\r\n> ```\r\n\r\nLooks very similar. Do you know what exactly:\r\n```\r\n NV# = Connection traversing a bonded set of # NVLinks\r\n```\r\nmeans? is NV4 better than NV2? since I get NV2. Why do you have 4? As I can see you only have 2 gpus.\r\n\r\n", "According to [this table](https://docs.nvidia.com/datacenter/nvtags/0.1/nvtags-user-guide/index.html#supported-link-names) NV4 means \"Connection traversing a bonded set of 4 NVLinks\".\r\n\r\nThere are some more details in the [GA102 whitepaper](https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/ampere/pdf/NVIDIA-ampere-GA102-GPU-Architecture-Whitepaper-V1.pdf):\r\n\r\n> GA102 GPUs utilize NVIDIA’s third-generation NVLink interface, which includes four x4 links, with each link providing 14.0625 GB/sec bandwidth in each direction between two GPUs. Four links provide 56.25 GB/sec bandwidth in each direction, and 112.5 GB/sec total bandwidth between two GPUs. Two RTX 3090 GPUs can be connected together for SLI using NVLink. ", "Super! Thank you for that insight, @moyix!\r\n\r\nI started compiling performance/scalability notes here: https://github.com/huggingface/transformers/issues/9824\r\n\r\nI summarized the useful insights from this thread. If you get a chance to validate the GPU inter-connectivity section that would be great!\r\n\r\nAnd if you have other insights to contribute I'm all ears. If you don't have time/inspiration to write something complete even a stab would be great and then over time we will fill it out with details and benchmarks.\r\n\r\nThe idea is to discuss in-depth the different hardware/software nuances to speed up training and fit larger models.\r\n\r\nThank you!", "Very nice, I will take a look at it!\r\n\r\nWhile I am waiting for HPC time, I ran your benchmark script on the 3090 system while varying two parameters: the model size (gpt2, gpt2-medium, and gpt2-large) and the block size (128, 256, 512).\r\n\r\nThe script:\r\n```\r\nfor MODEL in gpt2 gpt2-medium gpt2-large; do\r\n for BLOCK_SIZE in 128 256 512 ; do \r\n # Skip gpt2-large at block size 512 due to memory constraints\r\n if [ $MODEL = \"gpt2-large\" ] && [ $BLOCK_SIZE -eq 512 ] ; then continue ; fi\r\n # 1 gpu\r\n\r\n rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0 python run_clm.py --model_name_or_path $MODEL \\\r\n --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train --output_dir \\\r\n /tmp/test-clm --per_device_train_batch_size 4 --max_steps 400 --block_size $BLOCK_SIZE 2>&1 > /tmp/clm_bench.log\r\n result=$(grep train_runtime /tmp/clm_bench.log)\r\n echo $MODEL $BLOCK_SIZE \"1GPU\" $result >> clm_bench_results.log\r\n\r\n # DP\r\n\r\n rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 python run_clm.py --model_name_or_path $MODEL \\\r\n --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train --output_dir \\\r\n /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200 --block_size $BLOCK_SIZE 2>&1 > /tmp/clm_bench.log\r\n\r\n result=$(grep train_runtime /tmp/clm_bench.log)\r\n echo $MODEL $BLOCK_SIZE \"DP\" $result >> clm_bench_results.log\r\n\r\n # DDP\r\n\r\n rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node 2 \\\r\n run_clm.py --model_name_or_path $MODEL --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \\\r\n --do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200 --block_size $BLOCK_SIZE 2>&1 > /tmp/clm_bench.log\r\n\r\n result=$(grep train_runtime /tmp/clm_bench.log)\r\n echo $MODEL $BLOCK_SIZE \"DDP\" $result >> clm_bench_results.log\r\n\r\n # DDP w/o NVlink\r\n\r\n rm -r /tmp/test-clm; NCCL_P2P_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \\\r\n --nproc_per_node 2 run_clm.py --model_name_or_path $MODEL --dataset_name wikitext \\\r\n --dataset_config_name wikitext-2-raw-v1 --do_train --output_dir /tmp/test-clm \\\r\n --per_device_train_batch_size 4 --max_steps 200 --block_size $BLOCK_SIZE 2>&1 > /tmp/clm_bench.log\r\n\r\n result=$(grep train_runtime /tmp/clm_bench.log)\r\n echo $MODEL $BLOCK_SIZE \"DDP_no_NV\" $result >> clm_bench_results.log\r\n done\r\ndone\r\n```\r\n\r\nAnd the results:\r\n\r\n```\r\ngpt2 128 1GPU {'train_runtime': 19.5621, 'train_samples_per_second': 20.448, 'epoch': 0.09}\r\ngpt2 128 DP {'train_runtime': 16.6426, 'train_samples_per_second': 12.017, 'epoch': 0.09}\r\ngpt2 128 DDP {'train_runtime': 13.5368, 'train_samples_per_second': 14.775, 'epoch': 0.09}\r\ngpt2 128 DDP_no_NV {'train_runtime': 30.0181, 'train_samples_per_second': 6.663, 'epoch': 0.09}\r\ngpt2 256 1GPU {'train_runtime': 30.423, 'train_samples_per_second': 13.148, 'epoch': 0.17}\r\ngpt2 256 DP {'train_runtime': 22.6101, 'train_samples_per_second': 8.846, 'epoch': 0.17}\r\ngpt2 256 DDP {'train_runtime': 18.6943, 'train_samples_per_second': 10.698, 'epoch': 0.17}\r\ngpt2 256 DDP_no_NV {'train_runtime': 35.4208, 'train_samples_per_second': 5.646, 'epoch': 0.17}\r\ngpt2 512 1GPU {'train_runtime': 58.0856, 'train_samples_per_second': 6.886, 'epoch': 0.34}\r\ngpt2 512 DP {'train_runtime': 37.6376, 'train_samples_per_second': 5.314, 'epoch': 0.34}\r\ngpt2 512 DDP {'train_runtime': 32.3616, 'train_samples_per_second': 6.18, 'epoch': 0.34}\r\ngpt2 512 DDP_no_NV {'train_runtime': 49.1999, 'train_samples_per_second': 4.065, 'epoch': 0.34}\r\ngpt2-medium 128 1GPU {'train_runtime': 49.3823, 'train_samples_per_second': 8.1, 'epoch': 0.09}\r\ngpt2-medium 128 DP {'train_runtime': 40.5947, 'train_samples_per_second': 4.927, 'epoch': 0.09}\r\ngpt2-medium 128 DDP {'train_runtime': 33.4365, 'train_samples_per_second': 5.981, 'epoch': 0.09}\r\ngpt2-medium 128 DDP_no_NV {'train_runtime': 74.9924, 'train_samples_per_second': 2.667, 'epoch': 0.09}\r\ngpt2-medium 256 1GPU {'train_runtime': 79.6724, 'train_samples_per_second': 5.021, 'epoch': 0.17}\r\ngpt2-medium 256 DP {'train_runtime': 56.0446, 'train_samples_per_second': 3.569, 'epoch': 0.17}\r\ngpt2-medium 256 DDP {'train_runtime': 47.7543, 'train_samples_per_second': 4.188, 'epoch': 0.17}\r\ngpt2-medium 256 DDP_no_NV {'train_runtime': 89.3616, 'train_samples_per_second': 2.238, 'epoch': 0.17}\r\ngpt2-medium 512 1GPU {'train_runtime': 152.6255, 'train_samples_per_second': 2.621, 'epoch': 0.34}\r\ngpt2-medium 512 DP {'train_runtime': 92.4563, 'train_samples_per_second': 2.163, 'epoch': 0.34}\r\ngpt2-medium 512 DDP {'train_runtime': 82.1935, 'train_samples_per_second': 2.433, 'epoch': 0.34}\r\ngpt2-medium 512 DDP_no_NV {'train_runtime': 124.1163, 'train_samples_per_second': 1.611, 'epoch': 0.34}\r\ngpt2-large 128 1GPU {'train_runtime': 98.5939, 'train_samples_per_second': 4.057, 'epoch': 0.09}\r\ngpt2-large 128 DP {'train_runtime': 79.2193, 'train_samples_per_second': 2.525, 'epoch': 0.09}\r\ngpt2-large 128 DDP {'train_runtime': 65.7918, 'train_samples_per_second': 3.04, 'epoch': 0.09}\r\ngpt2-large 128 DDP_no_NV {'train_runtime': 152.2178, 'train_samples_per_second': 1.314, 'epoch': 0.09}\r\ngpt2-large 256 1GPU {'train_runtime': 154.5437, 'train_samples_per_second': 2.588, 'epoch': 0.17}\r\ngpt2-large 256 DP {'train_runtime': 106.7075, 'train_samples_per_second': 1.874, 'epoch': 0.17}\r\ngpt2-large 256 DDP [out of memory]\r\ngpt2-large 256 DDP_no_NV [out of memory]\r\ngpt2-large 512 1GPU [out of memory]\r\ngpt2-large 512 DP [out of memory]\r\ngpt2-large 512 DDP [out of memory]\r\ngpt2-large 152 DDP_no_NV [out of memory]\r\n```\r\n\r\nOne thing that I find interesting is that the behavior I originally observed where training on a single GPU could be slower than on multiple GPUs without NVLink only seems to be true for small block sizes like 128 or (sometimes) 256. So my hypothesis is that with smaller block sizes it is effectively using smaller batches and therefore synchronizing between GPUs more often?\r\n\r\nAs soon as I can get some time on our HPC I can update this with numbers for the 4xRTX8000 and the 4xV100, although the NVLink rows will no longer be applicable (since I don't have access to a machine with those cards in NVLink/NVSwitch configuration).", "Awesome! Thank you for more benchmarks, @moyix\r\n\r\nLet's apply some magic to your log:\r\n```\r\nperl -lne 'BEGIN{ print qq[|model|block|type|runtime|sample/sec|]; print \"|-\" x 5, \"|\"} $d=qr/([\\d\\.]+)/; m|^(\\S+) $d (\\S+) ..train_runtime.. $d, .train_samples_per_second.. $d| && print qq[|$1|$2|$3|$4|$5|]' log.txt\r\n```\r\nbut let's round it up to make reading easier:\r\n```\r\nperl -lne 'BEGIN{ print qq[|model|block|type|runtime|sample/sec|]; print \"|-\" x 5, \"|\"} $d=qr/([\\d\\.]+)/; m|^(\\S+) $d (\\S+) ..train_runtime.. $d, .train_samples_per_second.. $d| && print qq[|$1|$2|$3|] . int($4). \"|\". sprintf(\"%0.1f\", $5).\"|\"' log.txt\r\n```\r\n|model|block|type|runtime|sample/sec|\r\n|-|-|-|-|-|\r\n|gpt2|128|1GPU|19|20.4|\r\n|gpt2|128|DP|16|12.0|\r\n|gpt2|128|DDP|13|14.8|\r\n|gpt2|128|DDP_no_NV|30|6.7|\r\n|gpt2|256|1GPU|30|13.1|\r\n|gpt2|256|DP|22|8.8|\r\n|gpt2|256|DDP|18|10.7|\r\n|gpt2|256|DDP_no_NV|35|5.6|\r\n|gpt2|512|1GPU|58|6.9|\r\n|gpt2|512|DP|37|5.3|\r\n|gpt2|512|DDP|32|6.2|\r\n|gpt2|512|DDP_no_NV|49|4.1|\r\n|gpt2-medium|128|1GPU|49|8.1|\r\n|gpt2-medium|128|DP|40|4.9|\r\n|gpt2-medium|128|DDP|33|6.0|\r\n|gpt2-medium|128|DDP_no_NV|74|2.7|\r\n|gpt2-medium|256|1GPU|79|5.0|\r\n|gpt2-medium|256|DP|56|3.6|\r\n|gpt2-medium|256|DDP|47|4.2|\r\n|gpt2-medium|256|DDP_no_NV|89|2.2|\r\n|gpt2-medium|512|1GPU|152|2.6|\r\n|gpt2-medium|512|DP|92|2.2|\r\n|gpt2-medium|512|DDP|82|2.4|\r\n|gpt2-medium|512|DDP_no_NV|124|1.6|\r\n|gpt2-large|128|1GPU|98|4.1|\r\n|gpt2-large|128|DP|79|2.5|\r\n|gpt2-large|128|DDP|65|3.0|\r\n|gpt2-large|128|DDP_no_NV|152|1.3|\r\n|gpt2-large|256|1GPU|154|2.6|\r\n|gpt2-large|256|DP|106|1.9|\r\n\r\nDoing a quick scan it's clear that as the model grows in size and the block in its size they results start to diverge more and more, though proportions don't change much. Probably could pipe this to convert into relative sizes and then it'd very clear.\r\n\r\n> my hypothesis is that with smaller block sizes it is effectively using smaller batches and therefore synchronizing between GPUs more often?\r\n\r\nIt certainly has less data to communicate to the other gpus with smaller blocks", "ok, a quick hack to add ratios relative to 1gpu, so now it's easier to see the comparison.\r\n```\r\nperl -lne 'BEGIN{ print qq[|model|block|type|runtime|sample/sec|ratios]; print \"|-\" x 6, \"|\"} $d=qr/([\\d\\.]+)/; if (m|^(\\S+) $d (\\S+) ..train_runtime.. $d, .train_samples_per_second.. $d|) {if($3==\"1GPU\") {$s=$4; print \"| \" x 6, \"|\"}; print qq[|$1|$2|$3|] . int($4). \"|\". sprintf(\"%0.1f\", $5).\"|\".sprintf(\"%0.1f\", $4/$s).\"|\"}' log.txt\r\n```\r\n\r\nSo I added a new column runtime `ratios` and each 4 rows get recalculated wrt to their first runtime entry with 1gpu.\r\n\r\nedit: someone asked to explain the ratio and why the runtime is faster for DDP, but samples per second is smaller.\r\n\r\nHere is a puzzle to solve:\r\n\r\n1. one cake eater eats the cake at 60 sec/cake\r\n2. now a second cake eater joins and who eats at the same speed as the first one, but now after every bite they have to shout \"ML rocks\", which slows down both of them, so they are now eating 20% slower than when alone\r\n\r\nWill one cake eater finish the cake faster than two of them?\r\n\r\n(the answer is after the table, so you don't see it right away)\r\n\r\n|model|block|type|runtime|sample/sec|ratios\r\n|-|-|-|-|-|-|\r\n| | | | | | |\r\n|gpt2|128|1GPU|19|20.4|1.0|\r\n|gpt2|128|DP|16|12.0|0.9|\r\n|gpt2|128|DDP|13|14.8|0.7|\r\n|gpt2|128|DDP_no_NV|30|6.7|1.5|\r\n| | | | | | |\r\n|gpt2|256|1GPU|30|13.1|1.0|\r\n|gpt2|256|DP|22|8.8|0.7|\r\n|gpt2|256|DDP|18|10.7|0.6|\r\n|gpt2|256|DDP_no_NV|35|5.6|1.2|\r\n| | | | | | |\r\n|gpt2|512|1GPU|58|6.9|1.0|\r\n|gpt2|512|DP|37|5.3|0.6|\r\n|gpt2|512|DDP|32|6.2|0.6|\r\n|gpt2|512|DDP_no_NV|49|4.1|0.8|\r\n| | | | | | |\r\n|gpt2-medium|128|1GPU|49|8.1|1.0|\r\n|gpt2-medium|128|DP|40|4.9|0.8|\r\n|gpt2-medium|128|DDP|33|6.0|0.7|\r\n|gpt2-medium|128|DDP_no_NV|74|2.7|1.5|\r\n| | | | | | |\r\n|gpt2-medium|256|1GPU|79|5.0|1.0|\r\n|gpt2-medium|256|DP|56|3.6|0.7|\r\n|gpt2-medium|256|DDP|47|4.2|0.6|\r\n|gpt2-medium|256|DDP_no_NV|89|2.2|1.1|\r\n| | | | | | |\r\n|gpt2-medium|512|1GPU|152|2.6|1.0|\r\n|gpt2-medium|512|DP|92|2.2|0.6|\r\n|gpt2-medium|512|DDP|82|2.4|0.5|\r\n|gpt2-medium|512|DDP_no_NV|124|1.6|0.8|\r\n| | | | | | |\r\n|gpt2-large|128|1GPU|98|4.1|1.0|\r\n|gpt2-large|128|DP|79|2.5|0.8|\r\n|gpt2-large|128|DDP|65|3.0|0.7|\r\n|gpt2-large|128|DDP_no_NV|152|1.3|1.5|\r\n| | | | | | |\r\n|gpt2-large|256|1GPU|154|2.6|1.0|\r\n|gpt2-large|256|DP|106|1.9|0.7|\r\n\r\nand the answer to the puzzle posted at the beginning of this comment: 2 cake eaters will eat the cake faster together despite the slowdown, because they only have half a cake to finish each!\r\n\r\nSame here, while each of the GPUs in the DDP assembly performs slower due to the gradient syncing, but because it has to consume half the samples, overall the assembly will train faster.\r\n\r\nFurther, this benchmark is just for 2 GPUs\r\n\r\nSo going from 1GPU to 2GPUs, you create the overhead, and so you get some loss in performance, and some gain\r\n\r\nWhen you go from 2GPUs to 4GPUs (on the same node), it's pure performance doubling.\r\ni.e. 4GPUs will perform disproportionally faster than 2GPUs over 1 GPU.\r\n\r\n- 1 GPU has no inter-gpu communication to do\r\n- 2+ gpus have to average gradients\r\n\r\nso they add this overhead, but then they can parallelize the processing so the overhead becomes almost negligible as the number of GPUs grows\r\n\r\nThe next problem is once you outgrow a single node. So the next issue is inter-node connects, which can be blazing fast (Infiniband) or super-slow (ethernet hub). So to scale from 8GPUs to 10 (for 8-gpu node), you first lose performance, because now the inter-node connection is the slow component that slows everything down. But as you add more nodes, again that overhead becomes less and less significant.\r\n\r\nOf course when working with multi-node one often uses other parallelization techniques than DDP, so it's PP or TP (https://huggingface.co/transformers/parallelism.html#concepts), and there one generally performs TP only inside a node, and PP and DP over nodes.\r\n\r\n**It'd be amazing if someone re-did this table for 1, 2, 4 gpus, then 1, 2, 4 nodes.**", "OK, now we have some extensive benchmarks for the RTX8000 machine. This machine does not have NVLink, but it apparently can do P2P GPU-GPU communication via the PCI bus. However, this seems to be quite slow – slower, in fact, than disabling P2P altogether.\r\n\r\nHere's `nvidia-smi topo -m`:\r\n\r\n```\r\n GPU0 GPU1 GPU2 GPU3 mlx5_0 CPU Affinity NUMA Affinity\r\nGPU0 X SYS SYS SYS SYS 0-7 0-1\r\nGPU1 SYS X SYS SYS SYS 0-7 0-1\r\nGPU2 SYS SYS X SYS SYS 0-7 0-1\r\nGPU3 SYS SYS SYS X SYS 0-7 0-1\r\nmlx5_0 SYS SYS SYS SYS X \r\n\r\nLegend:\r\n\r\n X = Self\r\n SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\r\n NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\r\n PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\r\n PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\r\n PIX = Connection traversing at most a single PCIe bridge\r\n NV# = Connection traversing a bonded set of # NVLinks\r\n```\r\n\r\nI used the script from before (slightly expanded) and set `max-steps` to 800 for the single GPU case, 400 for two GPUs, and 200 for 4 GPUs. Here are the benchmarks (long!):\r\n\r\n|model|block|type|runtime|sample/sec|ratios\r\n|-|-|-|-|-|-|\r\n| | | | | | |\r\n|gpt2|128|1GPU|67|11.9|1.0|\r\n|gpt2|128|DP_2GPU|530|0.8|7.9|\r\n|gpt2|128|DDP_2GPU|350|1.1|5.2|\r\n|gpt2|128|DDP_no_P2P_2GPU|119|3.3|1.8|\r\n|gpt2|128|DP_4GPU|243|0.8|3.6|\r\n|gpt2|128|DDP_4GPU|159|1.3|2.4|\r\n|gpt2|128|DDP_no_P2P_4GPU|88|2.3|1.3|\r\n| | | | | | |\r\n|gpt2|256|1GPU|113|7.0|1.0|\r\n|gpt2|256|DP_2GPU|582|0.7|5.1|\r\n|gpt2|256|DDP_2GPU|376|1.1|3.3|\r\n|gpt2|256|DDP_no_P2P_2GPU|142|2.8|1.3|\r\n|gpt2|256|DP_4GPU|313|0.6|2.8|\r\n|gpt2|256|DDP_4GPU|174|1.1|1.5|\r\n|gpt2|256|DDP_no_P2P_4GPU|102|1.9|0.9|\r\n| | | | | | |\r\n|gpt2|512|1GPU|215|3.7|1.0|\r\n|gpt2|512|DP_2GPU|694|0.6|3.2|\r\n|gpt2|512|DDP_2GPU|426|0.9|2.0|\r\n|gpt2|512|DDP_no_P2P_2GPU|192|2.1|0.9|\r\n|gpt2|512|DP_4GPU|454|0.4|2.1|\r\n|gpt2|512|DDP_4GPU|201|1.0|0.9|\r\n|gpt2|512|DDP_no_P2P_4GPU|124|1.6|0.6|\r\n| | | | | | |\r\n|gpt2-medium|128|1GPU|183|4.4|1.0|\r\n|gpt2-medium|128|DP_2GPU|1476|0.3|8.0|\r\n|gpt2-medium|128|DDP_2GPU|863|0.5|4.7|\r\n|gpt2-medium|128|DDP_no_P2P_2GPU|280|1.4|1.5|\r\n|gpt2-medium|128|DP_4GPU|653|0.3|3.6|\r\n|gpt2-medium|128|DDP_4GPU|375|0.5|2.0|\r\n|gpt2-medium|128|DDP_no_P2P_4GPU|193|1.0|1.1|\r\n| | | | | | |\r\n|gpt2-medium|256|1GPU|306|2.6|1.0|\r\n|gpt2-medium|256|DP_2GPU|1600|0.2|5.2|\r\n|gpt2-medium|256|DDP_2GPU|919|0.4|3.0|\r\n|gpt2-medium|256|DDP_no_P2P_2GPU|339|1.2|1.1|\r\n|gpt2-medium|256|DP_4GPU|814|0.2|2.7|\r\n|gpt2-medium|256|DDP_4GPU|401|0.5|1.3|\r\n|gpt2-medium|256|DDP_no_P2P_4GPU|218|0.9|0.7|\r\n| | | | | | |\r\n|gpt2-medium|512|1GPU|573|1.4|1.0|\r\n|gpt2-medium|512|DP_2GPU|1884|0.2|3.3|\r\n|gpt2-medium|512|DDP_2GPU|1053|0.4|1.8|\r\n|gpt2-medium|512|DDP_no_P2P_2GPU|472|0.8|0.8|\r\n|gpt2-medium|512|DP_4GPU|1177|0.2|2.1|\r\n|gpt2-medium|512|DDP_4GPU|462|0.4|0.8|\r\n|gpt2-medium|512|DDP_no_P2P_4GPU|278|0.7|0.5|\r\n| | | | | | |\r\n|gpt2-large|128|1GPU|402|2.0|1.0|\r\n|gpt2-large|128|DP_2GPU|3181|0.1|7.9|\r\n|gpt2-large|128|DDP_2GPU|1760|0.2|4.4|\r\n|gpt2-large|128|DDP_no_P2P_2GPU|565|0.7|1.4|\r\n|gpt2-large|128|DP_4GPU|1361|0.1|3.4|\r\n|gpt2-large|128|DDP_4GPU|717|0.3|1.8|\r\n|gpt2-large|128|DDP_no_P2P_4GPU|349|0.6|0.9|\r\n| | | | | | |\r\n|gpt2-large|256|1GPU|642|1.2|1.0|\r\n|gpt2-large|256|DP_2GPU|3440|0.1|5.4|\r\n|gpt2-large|256|DDP_2GPU|1882|0.2|2.9|\r\n|gpt2-large|256|DDP_no_P2P_2GPU|686|0.6|1.1|\r\n|gpt2-large|256|DP_4GPU|1673|0.1|2.6|\r\n|gpt2-large|256|DDP_4GPU|770|0.3|1.2|\r\n|gpt2-large|256|DDP_no_P2P_4GPU|403|0.5|0.6|\r\n| | | | | | |\r\n|gpt2-large|512|1GPU|1168|0.7|1.0|\r\n|gpt2-large|512|DP_2GPU|3947|0.1|3.4|\r\n|gpt2-large|512|DDP_2GPU|2145|0.2|1.8|\r\n|gpt2-large|512|DDP_no_P2P_2GPU|952|0.4|0.8|\r\n|gpt2-large|512|DP_4GPU|2303|0.1|2.0|\r\n|gpt2-large|512|DDP_4GPU|902|0.2|0.8|\r\n|gpt2-large|512|DDP_no_P2P_4GPU|531|0.4|0.5|\r\n| | | | | | |\r\n|gpt2-xl|128|1GPU|770|1.0|1.0|\r\n|gpt2-xl|128|DP_2GPU|6391|0.1|8.3|\r\n|gpt2-xl|128|DDP_2GPU|3396|0.1|4.4|\r\n|gpt2-xl|128|DDP_no_P2P_2GPU|751|0.5|1.0|\r\n|gpt2-xl|128|DP_4GPU|2588|0.1|3.4|\r\n|gpt2-xl|128|DDP_4GPU|1356|0.1|1.8|\r\n|gpt2-xl|128|DDP_no_P2P_4GPU|635|0.3|0.8|\r\n| | | | | | |\r\n|gpt2-xl|256|1GPU|1210|0.7|1.0|\r\n|gpt2-xl|256|DP_2GPU|6826|0.1|5.6|\r\n|gpt2-xl|256|DP_4GPU|3130|0.1|2.6|\r\n", "Thank you for doing this immense work, @moyix!\r\n\r\nFrom a quick look it appears the model size doesn't matter, but the block-size makes a big difference to a faster outcome with the various DDP approaches - the larger the block the more benefits one gets, and for small blocks the performance is terrible.", "@JJack0812, your issue report won't get addresses here as we are talking about a totally different topic in this thread - I'd say post a separate issue - may be under pytorch or transformers, but first study [existing tickets](https://www.google.com/search?q=RuntimeError%3A+NCCL+error+in%3A+%2Fpytorch%2Ftorch%2Flib%2Fc10d%2FProcessGroupNCCL.cpp), e.g.: [this one](https://github.com/pytorch/pytorch/issues/39388) ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "I have met the same problem, thanks for the answers" ]

[]

[]

[ "Hi @BatMrE \r\n\r\nThe syntax of the command is wrong, there should be no spaces around `=` or you can also just remove the `=`\r\n\r\nSo it should be either like this\r\n```bash\r\npython run_qa.py \\\r\n --model_name_or_path bert-base-uncased \\\r\n --train_file=train-v1.1.json \\\r\n --validation_file=dev-v1.1.json \\\r\n```\r\n\r\nor this\r\n```bash\r\npython run_qa.py \\\r\n --model_name_or_path bert-base-uncased \\\r\n --train_file train-v1.1.json \\\r\n --validation_file dev-v1.1.json \\\r\n```", "removing the spaces worked for me, thoe I'm still not able to run that script getting:\r\n```\r\nTraceback (most recent call last):\r\n File \"run_qa.py\", line 469, in <module>\r\n main()\r\n File \"run_qa.py\", line 252, in main\r\n answer_column_name = \"answers\" if \"answers\" in column_names else column_names[2]\r\nIndexError: list index out of range\r\n\r\n```\r\n\r\nNote: I am using official training and dev json file to run the script\r\nplease see if someone can help.\r\n@patrickvonplaten / @stas00 / @vasudevgupta7", "@sgugger might know this.", "I had the same problem, here's what I found. \r\n\r\nIf you read through the script, you'll see it uses the `datasets.load_dataset()` function to load your data (line 211). As commented in the script check out [https://huggingface.co/docs/datasets/loading_datasets.html](https://huggingface.co/docs/datasets/loading_datasets.html) to learn more. \r\n\r\nI noticed it doesn't natively support squad style json files. \r\nHowever you can: \r\n- Use one of the supported formats;\r\n- create your own dataset loading script or [adapt an existing loading script](https://huggingface.co/docs/datasets/add_dataset.html#dataset-scripts-of-reference);\r\n- or use the [squad.py loading script](https://github.com/huggingface/datasets/blob/master/datasets/squad/squad.py).\r\n\r\nYou'll have to adapt the run_qa.py script a bit to use your loading script.\r\n", "@Jos1988 \r\nI am bit confused in how to use squad.py file for conversion of data\r\nI have tried this \r\n`dataset = load_dataset('squad', ...)`\r\n", "@BatMrE download the [squad.py](https://github.com/huggingface/datasets/blob/master/datasets/squad/squad.py) script and change the first few lines of _split_generators function(see the code below) to make `dl_manager` use your local QA dataset files instead of downloading the squad data. `self.config.data_files` uses the data_files you pass to `load_dataset` function.\r\n\r\n```python\r\n def _split_generators(self, dl_manager):\r\n if not self.config.data_files:\r\n raise ValueError(\r\n f\"At least one data file must be specified, but got data_files={self.config.data_files}\"\r\n )\r\n downloaded_files = dl_manager.download_and_extract(self.config.data_files)\r\n ......\r\n\r\n```\r\n\r\nAftter you do the above changes, just load your dataset using:\r\n\r\n```python\r\ndataset = load_dataset(<path to changed squad.py dataloader>, data_files={'train': <train-path>, 'validation': <validation-path>})\r\n```\r\nThe `data_files` contains the paths to your local train and dev QA datasets which are in squad format", "I have made all the expected changes\r\n\r\n- Made changes in squad.py file\r\n- datasets = load_dataset('squad.py', data_files={'train': 'train_custom.json', 'validation': 'dev_custom.json'})\r\n\r\npassing my custom file (which is different from orignal squad v1 files)\r\n**Note : code hits the custom file as if I pass irrelevant name it will throw error of file not found**\r\n\r\nI am getting the expected results but it is exactly same as the result I get on running\r\n```\r\npython run_qa.py \\\r\n --model_name_or_path bert-base-uncased \\\r\n --dataset_name squad \\\r\n --do_train \\\r\n --do_eval \\\r\n --per_device_train_batch_size 12 \\\r\n --learning_rate 3e-5 \\\r\n --num_train_epochs 2 \\\r\n --max_seq_length 384 \\\r\n --doc_stride 128 \\\r\n --output_dir /tmp/debug_squad/\r\n```\r\nmy current run script:\r\n```\r\n\r\npython run_qa.py \\\r\n --model_name_or_path bert-base-uncased \\\r\n --train_file=train_custom.json \\\r\n --validation_file=dev_custom.json \\\r\n --do_train \\\r\n --do_eval \\\r\n --per_device_train_batch_size 16 \\\r\n --learning_rate 3e-5 \\\r\n --num_train_epochs 2 \\\r\n --max_seq_length 384 \\\r\n --doc_stride 128 \\\r\n --output_dir/tmp/debug_squad2/\r\n```\r\nalso my _split_generators function in squad.py :\r\n```\r\n def _split_generators(self, dl_manager):\r\n if not self.config.data_files:\r\n raise ValueError(\r\n f\"At least one data file must be specified, but got data_files={self.config.data_files}\"\r\n )\r\n downloaded_files = dl_manager.download_and_extract(_URLS)\r\n\r\n return [\r\n datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={\"filepath\": downloaded_files[\"train\"]}),\r\n datasets.SplitGenerator(name=datasets.Split.VALIDATION, gen_kwargs={\"filepath\": downloaded_files[\"dev\"]}),\r\n ]\r\n```\r\n\r\nhow is it possible to get exactly same results for custom and official script, can someone please recommend something..\r\n\r\n@gowtham1997 @Jos1988", "@BatMrE\r\nInstead of `downloaded_files = dl_manager.download_and_extract(_URLS)` , use `downloaded_files = dl_manager.download_and_extract(self.config.data_files)`\r\n\r\n`downloaded_files = dl_manager.download_and_extract(_URLS)` downloads the squad dataset from _URLS specified in the squad.py loader file. You should instead use the local dataset files passed with `config.data_files`\r\n", "Thanks @gowtham1997 ,\r\nI have done some hardcoding in squad.py file to send my custom data files\r\n```\r\n_URLS = {\r\n \"train\": \"train_custom.json\",\r\n \"dev\": \"dev_custom.json\",\r\n}\r\n```\r\nJust one more thing.. I am able to use any custom data made on top of squad version 1, but I am not able to use squad version 2. As I am aware we need to use run_squad.py for squad version 2 and not run_qa, can some one add some comments on it.\r\n\r\n", "@sgugger can you lend a hand here?\r\n\r\nI have ran into the same problem but with a different error\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/builder.py\", line 434, in incomplete_dir\r\n yield tmp_dir\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/builder.py\", line 476, in download_and_prepare\r\n dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/builder.py\", line 553, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/builder.py\", line 897, in _prepare_split\r\n for key, table in utils.tqdm(generator, unit=\" tables\", leave=False, disable=not_verbose):\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/tqdm/std.py\", line 1130, in __iter__\r\n for obj in iterable:\r\n File \"/home/abashir/.cache/huggingface/modules/datasets_modules/datasets/json/fb88b12bd94767cb0cc7eedcd82ea1f402d2162addc03a37e81d4f8dc7313ad9/json.py\", line 75, in _generate_tables\r\n parse_options=self.config.pa_parse_options,\r\n File \"pyarrow/_json.pyx\", line 247, in pyarrow._json.read_json\r\n File \"pyarrow/error.pxi\", line 122, in pyarrow.lib.pyarrow_internal_check_status\r\n File \"pyarrow/error.pxi\", line 84, in pyarrow.lib.check_status\r\npyarrow.lib.ArrowInvalid: straddling object straddles two block boundaries (try to increase block size?)\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"/GW/Health-Corpus/work/UMLS/transformers/examples/question-answering/run_qa.py\", line 495, in <module>\r\n main()\r\n File \"/GW/Health-Corpus/work/UMLS/transformers/examples/question-answering/run_qa.py\", line 222, in main\r\n datasets = load_dataset(extension, data_files=data_files, field=\"data\")\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/load.py\", line 611, in load_dataset\r\n ignore_verifications=ignore_verifications,\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/builder.py\", line 483, in download_and_prepare\r\n self._save_info()\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/contextlib.py\", line 130, in __exit__\r\n self.gen.throw(type, value, traceback)\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/site-packages/datasets/builder.py\", line 440, in incomplete_dir\r\n shutil.rmtree(tmp_dir)\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/shutil.py\", line 498, in rmtree\r\n onerror(os.rmdir, path, sys.exc_info())\r\n File \"/home/abashir/anaconda3/envs/mpi/lib/python3.7/shutil.py\", line 496, in rmtree\r\n os.rmdir(path)\r\nOSError: [Errno 39] Directory not empty: '/home/abashir/.cache/huggingface/datasets/json/default-43dfe5d134316dba/0.0.0/fb88b12bd94767cb0cc7eedcd82ea1f402d2162addc03a37e81d4f8dc7313ad9.incomplete'\r\n```\r\n\r\n\r\nWhen tried the above fixes. altering the datasers line with loading the `squad.py` altered script I run into\r\n```\r\n30a174f57e692deb3b377336683/squad.py\", line 106, in _split_generators\r\n datasets.SplitGenerator(name=datasets.Split.VALIDATION, gen_kwargs={\"filepath\": downloaded_files[\"dev\"]}),\r\nKeyError: 'dev'\r\n\r\n```", "@thomwolf ", "@abdallah197 \r\nCan you please share the script you are trying, and squad.py file changes you have done ", "After using modified `squad.py` and converting data to JSON. It loads the data without error but when it starts training I got the following error message. @gowtham1997 \r\n\r\n```\r\n[INFO|trainer.py:837] 2021-03-04 01:19:16,915 >> ***** Running training *****\r\n[INFO|trainer.py:838] 2021-03-04 01:19:16,915 >> Num examples = 14842\r\n[INFO|trainer.py:839] 2021-03-04 01:19:16,916 >> Num Epochs = 5\r\n[INFO|trainer.py:840] 2021-03-04 01:19:16,916 >> Instantaneous batch size per device = 16\r\n[INFO|trainer.py:841] 2021-03-04 01:19:16,916 >> Total train batch size (w. parallel, distributed & accumulation) = 48\r\n[INFO|trainer.py:842] 2021-03-04 01:19:16,916 >> Gradient Accumulation steps = 1\r\n[INFO|trainer.py:843] 2021-03-04 01:19:16,916 >> Total optimization steps = 1550\r\n\r\n 0%| | 0/1550 [00:00<?, ?it/s]Traceback (most recent call last):\r\n File \"/okyanus/users/ctantug/transformers/examples/question-answering/run_qa.py\", line 507, in <module>\r\n main()\r\n File \"/okyanus/users/ctantug/transformers/examples/question-answering/run_qa.py\", line 481, in main\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/trainer.py\", line 940, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/trainer.py\", line 1304, in training_step\r\n loss = self.compute_loss(model, inputs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/trainer.py\", line 1334, in compute_loss\r\n outputs = model(**inputs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py\", line 161, in forward\r\n outputs = self.parallel_apply(replicas, inputs, kwargs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py\", line 171, in parallel_apply\r\n return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py\", line 86, in parallel_apply\r\n output.reraise()\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/_utils.py\", line 428, in reraise\r\n raise self.exc_type(msg)\r\nValueError: Caught ValueError in replica 0 on device 0.\r\nOriginal Traceback (most recent call last):\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py\", line 61, in _worker\r\n output = module(*input, **kwargs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py\", line 1793, in forward\r\n start_logits, end_logits = logits.split(1, dim=-1)\r\nValueError: too many values to unpack (expected 2)\r\n```", "> After using modified `squad.py` and converting data to JSON. It loads the data without error but when it starts training I got the following error message. @gowtham1997\r\n> \r\n> ```\r\n> [INFO|trainer.py:837] 2021-03-04 01:19:16,915 >> ***** Running training *****\r\n> [INFO|trainer.py:838] 2021-03-04 01:19:16,915 >> Num examples = 14842\r\n> [INFO|trainer.py:839] 2021-03-04 01:19:16,916 >> Num Epochs = 5\r\n> [INFO|trainer.py:840] 2021-03-04 01:19:16,916 >> Instantaneous batch size per device = 16\r\n> [INFO|trainer.py:841] 2021-03-04 01:19:16,916 >> Total train batch size (w. parallel, distributed & accumulation) = 48\r\n> [INFO|trainer.py:842] 2021-03-04 01:19:16,916 >> Gradient Accumulation steps = 1\r\n> [INFO|trainer.py:843] 2021-03-04 01:19:16,916 >> Total optimization steps = 1550\r\n> \r\n> 0%| | 0/1550 [00:00<?, ?it/s]Traceback (most recent call last):\r\n> File \"/okyanus/users/ctantug/transformers/examples/question-answering/run_qa.py\", line 507, in <module>\r\n> main()\r\n> File \"/okyanus/users/ctantug/transformers/examples/question-answering/run_qa.py\", line 481, in main\r\n> train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/trainer.py\", line 940, in train\r\n> tr_loss += self.training_step(model, inputs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/trainer.py\", line 1304, in training_step\r\n> loss = self.compute_loss(model, inputs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/trainer.py\", line 1334, in compute_loss\r\n> outputs = model(**inputs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n> result = self.forward(*input, **kwargs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py\", line 161, in forward\r\n> outputs = self.parallel_apply(replicas, inputs, kwargs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py\", line 171, in parallel_apply\r\n> return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py\", line 86, in parallel_apply\r\n> output.reraise()\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/_utils.py\", line 428, in reraise\r\n> raise self.exc_type(msg)\r\n> ValueError: Caught ValueError in replica 0 on device 0.\r\n> Original Traceback (most recent call last):\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py\", line 61, in _worker\r\n> output = module(*input, **kwargs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n> result = self.forward(*input, **kwargs)\r\n> File \"/okyanus/users/ctantug/.conda/envs/py37/lib/python3.7/site-packages/transformers/models/bert/modeling_bert.py\", line 1793, in forward\r\n> start_logits, end_logits = logits.split(1, dim=-1)\r\n> ValueError: too many values to unpack (expected 2)\r\n> ```\r\n\r\nSolved it. Turns out I have to change the config file to have only two labels (one for the first sentence and one for the second).", "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": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "@sgugger's 1st point is a good point and we should probably either:\r\n- recommend that users `git clone` the same version of `transformers` as is installed in the DLC image.\r\n- or even find a way to bundle the scripts themselves (or their capabilities) in the image itself, kinda like what I was suggesting before:\r\n\r\n```python\r\nestimator = HuggingFace(\r\n task_name=\"text-classification\",\r\n dataset=\"imdb\",\r\n from_model=\"distilbert-base-cased\",\r\n publish_model=\"my-fine-tuned-model\",\r\n huggingface_token=\"...\",\r\n)\r\n```\r\n\r\n(then there's not even a need to have a free-standing script. My question on whether this is SageMaker-idiomatic still stands)", "That´s true both of you are right. We must be able to ensure that the correct script version is used for the correct transformers & datasets version within the container image. \r\n\r\nI would not bundle them into the official DLC container since there is always that need to have an `entry_point`. My idea is maybe we still could use `task_name=\"text-classification\"` as \"entry_point\" and in the background, we can clone/get the correct script using the transformers version and the Github tags. \r\n\r\nSo for this version, we could use the script from https://github.com/huggingface/transformers/tree/v4.1.1. \r\n\r\n\r\n", "Closed by stale bot. If this shouldn't have been closed, let me know." ]

[]

[]

[ "The image is also on this link.\r\nhttps://img-blog.csdnimg.cn/20191025102941935.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9qYWNrY3VpLmJsb2cuY3Nkbi5uZXQ=,size_16,color_FFFFFF,t_70", "Hey @zhizeng8, \r\n\r\nIn the future, it would be nice if such questions are posted in the forum: https://discuss.huggingface.co/ as it is not about a bug. \r\nTo answer your question I'd use something like the following\r\n\r\n```python\r\n import torch\r\n\r\n tgt_len = 5\r\n # make causal mask\r\n mask = torch.full((tgt_len, tgt_len), float(\"-inf\"))\r\n mask_cond = torch.arange(mask.size(-1))\r\n mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)\r\n # attend to encoder part\r\n mask[:, :3] = 0\r\n```\r\n\r\nThis mask can however not just be input as an `attention_mask` to transformer models. Because BERT accepts 3d masks however with the 0-th index being the batch_size the above mask could be extended for all batches and input to BERT:\r\n\r\n```python\r\n3d_attention_mask = mask[None, :, :]\r\n\r\nbert = BertModel.from_pretrained(...)\r\nbert(input_ids, attention_mask=3d_attention_mask)\r\n```", "Thank you very much!\r\nI think in my case the 3d_attention_mask is different for each instance in a batch, due to the different length of source and target sequence.", "I have been reading this article recently. BertModel accepts 3d masks of dimensions [batch_size, from_seq_length, to_seq_length]. \r\n```\r\ntext='I love you' \r\nattention_mask = [[[1,0,0],[1,1,0],[1,1,1]]]\r\n```\r\nMaybe can help you.\r\nAlso I have a question.\r\n```\r\ntext='I love you' \r\nattention_mask = [[[1,1,1],[1,1,1],[1,1,1]]]\r\n```\r\n I found tensor of the last word 'you' are not same, I don't know the reason.", "I know. Last word's tensor in first attention layer is same, but there are 12 attention layer. Hiddenstate may change.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

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

[ "Hi @mailong25 for blenderbot the dialogs are separated by the newline `\\n`.\r\n\r\nSo the text should be \r\n`I am from Vietnam\\nI've never been there, but I've always wanted to go. How do you like it?\\npretty good actually , where you are from ?`\r\n\r\nwhich model are you using, 90M or 3B?\r\n\r\nAlso, could you post the `parlai` command that you used?", "Thanks for a quick response. I use the `facebook/blenderbot-1B-distill` model\r\n\r\nFor parlai, I use the cmd:\r\n\r\n`python parlai/scripts/interactive.py -t blended_skill_talk -mf zoo:blender/blender_1Bdistill/model --include_personas=False`", "I tried to use the `'\\n'` separator with `blenderbot-1B` and `blender_1Bdistill` and the results are still the same with ` </sep>`, which are different than parlai version.\r\n\r\n\r\nAlso, when I tried to move the model and the input sentence to \"cuda\", the following errors occur:\r\n\r\n```\r\nimport os\r\nos.environ[\"TRANSFORMERS_CACHE\"] = '/mnt/disks/blender/'\r\nfrom transformers import BlenderbotTokenizer, BlenderbotForConditionalGeneration\r\nmname = 'facebook/blenderbot-3B'\r\nmodel = BlenderbotForConditionalGeneration.from_pretrained(mname)\r\ntokenizer = BlenderbotTokenizer.from_pretrained(mname)\r\nmodel.to('cuda')\r\n\r\nimport torch\r\nwith torch.no_grad():\r\n UTTERANCE = []\r\n UTTERANCE.append(\"I am from Vietnam\")\r\n UTTERANCE.append(\"I've never been there, but I've always wanted to go. How do you like it?\")\r\n UTTERANCE.append(\"pretty good actually , where you are from ?\")\r\n \r\n UTTERANCE = '\\n'.join(UTTERANCE)\r\n print(UTTERANCE)\r\n inputs = tokenizer([UTTERANCE], return_tensors='pt')\r\n reply_ids = model.generate(**inputs)\r\n print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in reply_ids])\r\n\r\n---------------------------------------------------------------------------\r\nRuntimeError Traceback (most recent call last)\r\n<ipython-input-9-dff8c43ffc48> in <module>\r\n 10 print(UTTERANCE)\r\n 11 inputs = tokenizer([UTTERANCE], return_tensors='pt')\r\n---> 12 reply_ids = model.generate(**inputs)\r\n 13 print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in reply_ids])\r\n\r\n~/.local/lib/python3.7/site-packages/torch/autograd/grad_mode.py in decorate_context(*args, **kwargs)\r\n 13 def decorate_context(*args, **kwargs):\r\n 14 with self:\r\n---> 15 return func(*args, **kwargs)\r\n 16 return decorate_context\r\n 17 \r\n\r\n~/.local/lib/python3.7/site-packages/transformers/generation_utils.py in generate(self, input_ids, max_length, min_length, do_sample, early_stopping, num_beams, temperature, top_k, top_p, repetition_penalty, bad_words_ids, bos_token_id, pad_token_id, eos_token_id, length_penalty, no_repeat_ngram_size, num_return_sequences, decoder_start_token_id, use_cache, num_beam_groups, diversity_penalty, prefix_allowed_tokens_fn, **model_kwargs)\r\n 501 if self.config.is_encoder_decoder:\r\n 502 # add encoder_outputs to model_kwargs\r\n--> 503 model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(input_ids, model_kwargs)\r\n 504 \r\n 505 # set input_ids as decoder_input_ids\r\n\r\n~/.local/lib/python3.7/site-packages/transformers/generation_utils.py in _prepare_encoder_decoder_kwargs_for_generation(self, input_ids, model_kwargs)\r\n 84 argument: value for argument, value in model_kwargs.items() if not argument.startswith(\"decoder_\")\r\n 85 }\r\n---> 86 model_kwargs[\"encoder_outputs\"]: ModelOutput = encoder(input_ids, return_dict=True, **encoder_kwargs)\r\n 87 return model_kwargs\r\n 88 \r\n\r\n~/.local/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)\r\n 720 result = self._slow_forward(*input, **kwargs)\r\n 721 else:\r\n--> 722 result = self.forward(*input, **kwargs)\r\n 723 for hook in itertools.chain(\r\n 724 _global_forward_hooks.values(),\r\n\r\n~/.local/lib/python3.7/site-packages/transformers/models/bart/modeling_bart.py in forward(self, input_ids, attention_mask, inputs_embeds, output_attentions, output_hidden_states, return_dict)\r\n 750 \r\n 751 if inputs_embeds is None:\r\n--> 752 inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale\r\n 753 \r\n 754 embed_pos = self.embed_positions(input_shape)\r\n\r\n~/.local/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)\r\n 720 result = self._slow_forward(*input, **kwargs)\r\n 721 else:\r\n--> 722 result = self.forward(*input, **kwargs)\r\n 723 for hook in itertools.chain(\r\n 724 _global_forward_hooks.values(),\r\n\r\n~/.local/lib/python3.7/site-packages/torch/nn/modules/sparse.py in forward(self, input)\r\n 124 return F.embedding(\r\n 125 input, self.weight, self.padding_idx, self.max_norm,\r\n--> 126 self.norm_type, self.scale_grad_by_freq, self.sparse)\r\n 127 \r\n 128 def extra_repr(self) -> str:\r\n\r\n~/.local/lib/python3.7/site-packages/torch/nn/functional.py in embedding(input, weight, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse)\r\n 1812 # remove once script supports set_grad_enabled\r\n 1813 _no_grad_embedding_renorm_(weight, input, max_norm, norm_type)\r\n-> 1814 return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\r\n 1815 \r\n 1816 \r\n\r\nRuntimeError: Expected object of device type cuda but got device type cpu for argument #3 'index' in call to _th_index_select\r\n```", "The error occurs because the `inputs` is not on GPU, putting `inputs` on GPU should fix the error. ", "Hi. I tried to use the blenderbot example code from huggingface. I just copy pasted it in colab and ran it. But it is showing me the following error : \n\nTypeError: forward() got an unexpected keyword argument 'token_type_ids'\n\nPlease help me out. What can be done to solve this?", "According to the Parlai [documentation](https://parl.ai/docs/tutorial_task.html), the Parlai format is to use a `<\\t>` or 4 spaces as a separator of turns in a conversation.\r\nI ran your example using 4 spaces and got the Parlai response\r\n```\r\nfrom transformers import BlenderbotForConditionalGeneration, BlenderbotTokenizer\r\n\r\nMODEL_ID = \"facebook/blenderbot-400M-distill\"\r\nmodel = BlenderbotForConditionalGeneration.from_pretrained(MODEL_ID)\r\ntokenizer = BlenderbotTokenizer.from_pretrained(MODEL_ID)\r\n\r\ntext = [\"I am from Vietnam I've never been there, but I've always wanted to go. How do you like it? pretty good actually , where you are from ?\"]\r\n\r\ninputs = tokenizer(text, return_tensors='pt')\r\nres = model.generate(inputs['input_ids'])\r\ntokenizer.batch_decode(res)\r\n\r\n#[\"<s> I'm from the United States. I've heard it's a beautiful place to visit. </s>\"]\r\n```\r\n\r\n", "> According to the Parlai [documentation](https://parl.ai/docs/tutorial_task.html), the Parlai format is to use a `<\\t>` or 4 spaces as a separator of turns in a conversation.\r\n> I ran your example using 4 spaces and got the Parlai response\r\n> \r\n> ```\r\n> from transformers import BlenderbotForConditionalGeneration, BlenderbotTokenizer\r\n> \r\n> MODEL_ID = \"facebook/blenderbot-400M-distill\"\r\n> model = BlenderbotForConditionalGeneration.from_pretrained(MODEL_ID)\r\n> tokenizer = BlenderbotTokenizer.from_pretrained(MODEL_ID)\r\n> \r\n> text = [\"I am from Vietnam I've never been there, but I've always wanted to go. How do you like it? pretty good actually , where you are from ?\"]\r\n> \r\n> inputs = tokenizer(text, return_tensors='pt')\r\n> res = model.generate(inputs['input_ids'])\r\n> tokenizer.batch_decode(res)\r\n> \r\n> #[\"<s> I'm from the United States. I've heard it's a beautiful place to visit. </s>\"]\r\n> ```\r\n\r\nI did a whole bunch of testing with different turn separation tokens, and the only one that consistently separated the turns and created outputs that made sense was 4 spaces. \r\n\r\nEverything else would sometimes separate a turn, but sometimes not (tab, new line, `</s> <s>`, two spaces)\r\n\r\nThe documentation there mentioned that tabs are rendered as 4 spaces in the browser, but it still should be tab as a separator? \r\n\r\nAny new insights into why?", "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 there a finalized way to construct inputs for Blenderbot? According to this https://github.com/huggingface/transformers/blob/v4.5.1/src/transformers/models/blenderbot/modeling_blenderbot.py#L499 should be using `</s> <s>` not sure if I should be following this or `\\n` as indicated in parlai", "I have tried four spaces, \\n and \\t. Four spaces gave the best results.\r\n", "Should the first sentence in the encoder input be prepended with an extra space ' '? Because I note that the first token generated by the decoder has the space prefix (e.g., ' I' or ' yes').", "I noticed the space prefix in the generation too. But I didn't check if adding extra space to encoder input gives better results.", "Check out this example. This example is crafted in such a way that it could be a 1 turn or 2 turn depend on how you separate it. I am using the 3B model.\r\n```\r\nUsing \\n as seperator\r\nUTTERANCE = \"I am from tokyo\\nwhere you are from?\"\r\ninputs = tokenizer([UTTERANCE], return_tensors=\"pt\")\r\nreply_ids = model.generate(**inputs)\r\nprint(tokenizer.batch_decode(reply_ids,skip_special_tokens=True))\r\n#[' I was born and raised in Tokyo, the capital of Japan. How about you?']\r\n\r\n\r\nUsing four spaces as seperator\r\nUTTERANCE = \"I am from tokyo where you are from?\"\r\ninputs = tokenizer([UTTERANCE], return_tensors=\"pt\")\r\nreply_ids = model.generate(**inputs)\r\nprint(tokenizer.batch_decode(reply_ids,skip_special_tokens=True))\r\n#[\" I'm from Tokyo. It's the capital of Japan. It's a big city\"]\r\n\r\nUsing <t> as seperator\r\nUTTERANCE = \"I am from tokyo<t>where you are from?\"\r\ninputs = tokenizer([UTTERANCE], return_tensors=\"pt\")\r\nreply_ids = model.generate(**inputs)\r\nprint(tokenizer.batch_decode(reply_ids,skip_special_tokens=True))\r\n#[\" I'm from the United States. I've never been to Tokyo, but I've always wanted to go.\"]\r\n \r\n\r\n# Using </s> <s> as seperator\r\n​UTTERANCE = \"I am from tokyo</s> <s>where you are from?\"\r\ninputs = tokenizer([UTTERANCE], return_tensors=\"pt\")\r\nreply_ids = model.generate(**inputs)\r\nprint(tokenizer.batch_decode(reply_ids,skip_special_tokens=True))\r\n#[' I am also from Tokyo, the capital and most populous metropolitan area in Japan.']\r\n\r\n```\r\n\r\nIt looks like using ```\\n``` and four spaces, the model interpret it as 2 turn but 1 turn for using ```\\<t> and \\</s> \\<s>```" ]

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

[]

[ "Hi @Eymen3455 \r\n\r\nIn the `run_eval` command I see that you are setting `n_obs` to 100 while the `finetune_trainer` uses all test examples, could you maybe run eval again with all test examples and see if you get close results?", "First of all, thank you very much @patil-suraj for your reply. I did as you said and used all the test dataset for `n_obs`, but the result remained unchanged. In addition, when we examine the texts produced in `finetune_trainer.py` and `run_eval.py`, the texts produced in `finetune_trainer.py` are acceptable. While the texts produced in `run_eval.py` are different, it is worse and the rouge score is less than half of obtained scores in `finetune_trainer.py`.\r\n\r\nThen when I saw this issue (https://github.com/huggingface/transformers/issues/9236), I used Xsum dataset as dataset and rouge score increased in `run_eval.py`. Just changing the dataset caused to get logical text in `run_eval.py`, but still I could not understand the difference between the rouge score and generated texts in `finetune_trainer.py` and `run_eval.py`.\r\n\r\nI do not understand what has affected the dataset so much. When I use the mbart model, why can't I get the same success in `finetune_train.py` and `run_eval.py`, what could be the reason for this? By the way, when I try the `sshleifer/student_cnn_12_6` model instead of the mbart model, I can achieve exactly the same success in `finetune_trainer.py` and `run_eval.py`. I would appreciate if you could help.\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "I ran into the same problem and positioned it to the max_length parameters in Seq2SeqTrainer.evaluate() . Whether to set max_length parameters will result in different results.", "I have encountered the same issue:\r\n\r\nval_rouge2 obtained during training is different from scores I got with \"run_eval.py\" script\r\n\r\ndo you have any suggestions ? @patil-suraj " ]