Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "and when I upgrade transformers version to 4.3.3, onnx version is 1.8.1,another error occurred:\r\nTraceback (most recent call last):\r\n File \"test_segment.py\", line 104, in <module>\r\n session = onnxruntime.InferenceSession(output_model_path, sess_options, providers=['CPUExecutionProvider'])\r\n File \"/usr/lib64/python3.6/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py\", line 206, in __init__\r\n self._create_inference_session(providers, provider_options)\r\n File \"/usr/lib64/python3.6/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py\", line 226, in _create_inference_session\r\n sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model)\r\nonnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Load model from onnx_tf/segment.onnx failed:Fatal error: BroadcastTo is not a registered function/op", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

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

[]

[ "Hi @gchhablani \r\n\r\nThis is great! You can ping me if you need any help with this model.\r\n\r\n\r\nAlso, we have now added a step-by-step doc for how to add a model, you can find it here \r\nhttps://huggingface.co/transformers/add_new_model.html\r\n\r\nAlso have a look at the `cookiecutter` tool, which will help you generate lots of boilerplate code.\r\nhttps://github.com/huggingface/transformers/tree/master/templates/adding_a_new_model", "Hi @patil-suraj\r\n\r\nThanks a lot.\r\n\r\nI wanted to add this model in 2020, but I faced a lot of issues and got busy with some other work. Seems like the code structure has changed quite a bit after that. I'll check the shared links and get back.", "Hi @patil-suraj,\r\n\r\nI have started adding some code. For comparison, you can see: https://github.com/uclanlp/visualbert/blob/master/pytorch_pretrained_bert/modeling.py.\r\n\r\nPlease look at the `VisualBertEmbeddings` and `VisualBertModel`. I have checked these using dummy tensors. I am adding/fixing other kinds of down-stream models. Please tell me if you think this is going in the right direction, and if there are any things that to need to be kept in mind.\r\n\r\nI skipped testing the original repository for now, they don't have an entry point, and require a bunch of installs for loading /handling the dataset(s) which are huge.\r\n\r\nThere are several different checkpoints, even for the pre-trained models (with different embedding dimensions, etc.). For each we'll have to make a separate checkpoint, I guess.\r\n\r\nIn addition, I was wondering if we want to provide encoder-decoder features (`cross-attention`, `past_key_value`, etc.). I don't think it has been used in the original code, but it will certainly be a nice feature to have in case we have some task in the future which involves generation of text given an image and a text (probably there is something already).\r\n\r\nThanks :)", "Hi @patil-suraj\r\n\r\nI would appreciate some feedback :) It'll help me know if I am going in the right direction,\r\n\r\nThanks,\r\nGunjan", "Hey @gchhablani \r\n\r\nI'm thinking about this and will get back to you tomorrow :)", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Unstale", "Hi @LysandreJik \r\nThanks for reviewing.\r\n\r\nI'll make the suggested changes in a few (4-5) hours.\r\n\r\nThe `position_embeddings_type` is not exactly being used by the authors. They use absolute embeddings. They do have an `embedding_strategy_type` argument but it is unused and kept as `'plain'`.\r\n\r\nYes, almost all of it is copied from BERT. The new additions are the embeddings and the model classes.\r\n\r\nInitially, I did have plans to add actual examples and notebooks initially in the same PR :stuck_out_tongue: I guess I'll work on it now.", "@LysandreJik Do you wanna take a final look?", ":') Thanks a lot for all your help @patil-suraj @LysandreJik @sgugger @patrickvonplaten\r\n\r\nEdit:\r\nI will be working on another PR soon to add more/better examples, and to use the `SimpleDetector` as used by the original authors. Probably will also attempt to create `TF`/`Flax` models.", "The PR looks awesome! I still think however that we should add `get_visual_embeddings` with a Processor in this PR to have it complete. @gchhablani @patil-suraj - do you think it would be too time-consuming to add a `VisualBERTFeatureExtractor` here? Just don't really think people will be able to run an example after we merge it at the moment", "@patrickvonplaten here `visual_embeddings` comes from a cnn backbone (ResNet) or object detection model. So I don't think\r\n we can add `VisualBERTFeatureExtractor`. The plan is to add the cnn model and a demo notebook in `research_projects` dir in a follow-up a PR.", "> @patrickvonplaten here `visual_embeddings` comes from a cnn backbone (ResNet) or object detection model. So I don't think\r\n> we can add `VisualBERTFeatureExtractor`. The plan is to add the cnn model and a demo notebook in `research_projects` dir in a follow-up a PR.\r\n\r\nOk! Do we have another model that could give us those embeddings? ViT maybe?", "@patrickvonplaten, I am not sure about ViT as I haven't used or read about it yet. The VisualBERT authors used Detectron-based (MaskRCNN with FPN-ResNet-101 backbone) features for 3 out of 4 tasks. Each \"token\" is actually an object detected and the token features/embeddings from the detectron classifier layer. In case of VCR task, they use a ResNet on given box coordinates. \r\n\r\nUnless ViT has/is an extractor similar to this, if we could use ViT, it'd be very different from the original and might not work with the provided pre-trained weights. :/\r\n", "Adding a common Fast/Faster/MaskRCNN feature extractor, however, will help with LXMERT/VisualBERT and other models I'm planning to contribute in the future - ViLBERT (#11986), VL-BERT, (and possibly MCAN). \r\n\r\n**Edit**: There's already an example for LXMERT: https://github.com/huggingface/transformers/blob/master/examples/research_projects/lxmert/modeling_frcnn.py which I'll build upon.", "Ok! Good to merge for me then", "maybe not promote it yet", "@patrickvonplaten Yes, won't be promoted before adding an example.\r\n\r\nThe plan forward is to add a detector and example notebook in `research_projects` dir in a follow-up PR.\r\n\r\nVerified that all slow tests are passing :)\r\n\r\nMerging!", "Great model addition @gchhablani. Small note: can you fix the code examples of the HTML page?\r\n\r\nCurrently they look like this:\r\n\r\n\r\n\r\nThe Returns: statement should be below the Example: statement in `modeling_visual_bert.py`. Sorry for nitpicking ;) " ]

[]

[]

[ "I don't know what those .nfs are used for in Ray, is it safe to remove them @amogkam ?" ]

[]

[]

[ "Could you try to add `\"max_length\": 200` to your payload ? (also cc @Narsil )", "Hi @gstranger ,\r\n\r\nCan you reproduce the problem locally ? It could be that your model simply produces EOS token with high probability (leading to having the exact same prompt as output)\r\n\r\nIf not, do you mind telling us your username by DM `[email protected]` so we can investigate this issue ?", "Hello @Narsil ,\r\n\r\nWhen I run the same model locally, without either `max_length` or `min_length` I receive additional output, typically it will generate about 20-40 tokens. Also @patrickvonplaten when I add either that parameter or the `min_length` parameter the model still returns in less than a second with the same input text. I've sent an email with additional information. ", "Hi @gstranger,\r\n\r\nIt does seem like a `max_length`. Your config defines it as `20` which is not long enough for the prompt that gets automatically added (because it's a transfo-xl model): https://github.com/huggingface/transformers/blob/master/src/transformers/pipelines/text_generation.py#L23\r\n\r\nYou can override both `max_length` and `prefix` within the config to override the default `transfo-xl` behavior (depending on how it was trained it might lead to significant perf boost, or loss).\r\n\r\nBy default the API will read the config first. Replying also with more information by email with more information. Just sharing here so that the community can get help too.\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\n\r\ncould you please ask questions related to training of models on the [forum](https://discuss.huggingface.co/)?\r\n\r\nAll questions related to fine-tuning a model for summarization on CNN can be found [here](https://discuss.huggingface.co/search?q=summarization%20cnn) for example. \r\n\r\n", "> Hi,\r\n> \r\n> could you please ask questions related to training of models on the [forum](https://discuss.huggingface.co/)?\r\n> \r\n> All questions related to fine-tuning a model for summarization on CNN can be found [here](https://discuss.huggingface.co/search?q=summarization%20cnn) for example.\r\n\r\n@NielsRogge Oh, thank you for your reminder, that really helped!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "That's correct! Do you want to open a PR to fix it?" ]

[]

[]

[ "Hello! This is a duplicate of https://github.com/huggingface/transformers/issues/4875, https://github.com/huggingface/transformers/issues/10144 and https://github.com/huggingface/transformers/issues/9247\r\n\r\n@patrickvonplaten, maybe we could do something about this in the docs? In the docs we recommend doing this:\r\n```py\r\nmodel.resize_token_embedding(len(tokenizer))\r\n```\r\nbut this is unfortunately false for T5!", "@LysandreJik @cih9088 , actually I think doing: \r\n\r\n```python\r\nmodel.resize_token_embedding(len(tokenizer))\r\n```\r\n\r\nis fine -> it shouldn't throw an error & logically it should also be correct...Can you try it out?\r\n", "```python\r\nmodel.resize_token_embedding(len(tokenizer))\r\n```\r\nThis works perfectly fine but here is the thing.\r\n\r\nOne might add the `model.resize_token_embedding(len(tokenizer))` in their code and use other configuration packages such as `hydra` from Facebook to train models with additional tokens or without them dynamically at runtime.\r\nHe would naturally think that `vocab_size` of tokenizer (no tokens added) and `vocab_size` of model are the same because other models are.\r\nEventually, he fine-tunes the `google/mt5-base` model without added tokens but because of `model.resize_token_embedding(len(tokenizer))`, model he will fine-tune is not the same with `google/mt5-base`.\r\nAfter training, he wants to load the trained model to test but the model complains about inconsistent embedding size between a loaded model which is `google/mt5-base`, and the trained model which has a smaller size of token embedding.\r\n\r\nOf course, we could resize token embedding before loading model, but what matters is inconsistency with other models I think.\r\nI reckon that people would not very much care about how the dictionary is composed in tokenizer. Maybe we add some dummy tokens to the tokenizer to keep consistency with other huggingface models or add documentation about it (I could not find any).\r\n\r\nWhat do you think?", "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.", "\r\n> Hello! This is a duplicate of #4875, #10144 and #9247\r\n> \r\n> @patrickvonplaten, maybe we could do something about this in the docs? In the docs we recommend doing this:\r\n> \r\n> ```python\r\n> model.resize_token_embedding(len(tokenizer))\r\n> ```\r\n> \r\n> but this is unfortunately false for T5!\r\n\r\nWhat is the correct way to resize_token_embedding for T5/mT5?" ]

[]

[]

[ "...The test pasts on my local machine, I ran make test, style, quality, fixup. I dont know why this failed..", "I just rebased the PR to ensure that the tests pass. We'll merge if all is green!", "Thanks guys" ]

[]

[ { "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 } ]

[ "Is this part correct?\r\n\r\n> Recommended T5 finetuning settings:\r\n> - Scheduled LR warm-up to fixed LR\r\n> - disable relative updates\r\n> - use clip threshold: https://arxiv.org/abs/2004.14546\r\n\r\nIn particular:\r\n- are we supposed to do scheduled LR? adafactor handles this no?\r\n- we *should not* disable relative updates\r\n- i don't know what clip threshold means in this context", "@sshleifer can you accept this documentation change?", "No but @stas00 can!", "@jsrozner, thank you for the PR. \r\n\r\nReading https://github.com/huggingface/transformers/issues/7789 it appears that the `Recommended T5 finetuning settings` are invalid.\r\n\r\nSo if we are fixing this, in addition to changing the example the prose above it should be synced as well. \r\n\r\nI don't know where the original recommendation came from - do you by chance have a source we could point to for the corrected recommendation? If you know that is, if not, please don't worry.\r\n\r\nThank you.\r\n\r\n", "I receive the following error when using this the \"recommended way\":\r\n```{python}\r\nTraceback (most recent call last):\r\n File \"./folder_aws/transformers/examples/seq2seq/run_seq2seq.py\", line 759, in <module>\r\n main()\r\n File \"./folder_aws/transformers/examples/seq2seq/run_seq2seq.py\", line 651, in main\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n File \"/home/alejandro.vaca/SpainAI_Hackaton_2020/folder_aws/transformers/src/transformers/trainer.py\", line 909, in train\r\n self.create_optimizer_and_scheduler(num_training_steps=max_steps)\r\n File \"/home/alejandro.vaca/SpainAI_Hackaton_2020/folder_aws/transformers/src/transformers/trainer.py\", line 660, in create_optimizer_and_scheduler\r\n self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)\r\n File \"/home/alejandro.vaca/SpainAI_Hackaton_2020/folder_aws/transformers/src/transformers/optimization.py\", line 452, in __init__\r\n raise ValueError(\"warmup_init requires relative_step=True\")\r\nValueError: warmup_init requires relative_step=True\r\n```\r\n\r\nFollowing this example in documentation:\r\n\r\n```{python}\r\nAdafactor(model.parameters(), lr=1e-3, relative_step=False, warmup_init=True)\r\n```\r\n@sshleifer @stas00 @jsrozner ", "@alexvaca0, what was the command line you run when you received this error?\r\n\r\nHF Trainer doesn't set `warmup_init=True`. Unless you modified the script?\r\n\r\nhttps://github.com/huggingface/transformers/blob/21e86f99e6b91af2e4df3790ba6c781e85fa0eb5/src/transformers/trainer.py#L649-L651\r\n\r\nIt is possible that the whole conflict comes from misunderstanding how this optimizer has to be used?\r\n\r\n> To use a manual (external) learning rate schedule you should set `scale_parameter=False` and `relative_step=False`.\r\n\r\nwhich is what the Trainer does at the moment. \r\n\r\nand:\r\n> relative_step (:obj:`bool`, `optional`, defaults to :obj:`True`):\r\n> If True, time-dependent learning rate is computed instead of external learning rate\r\n\r\nIs it possible that you are trying to use both an external and the internal scheduler at the same time?\r\n\r\nIt'd help a lot of you could show us the code that breaks, (perhaps on colab?) and how you invoke it. Basically, help us reproduce it.\r\n\r\nThank you.", "Hi, @stas00 , first thank you very much for looking into it so fast. I forgot to say it, but yes, I changed the code in Trainer because I was trying to use the recommended settings for training T5 (I mean, setting an external learning rate with warmup_init = True as in the documentation. `Training without LR warmup or clip threshold is not recommended. Additional optimizer operations like gradient clipping should not be used alongside Adafactor.` From your answer, I understand that Trainer was designed for using Adam, as it uses the external learning rate scheduler and doesn't let you pass None as learning rate. Is there a workaround to be able to use the Trainer class with Adafactor following Adafactor settings recommended in the documentation? I'd also like to try using Adafactor without specifying the learning rate, would that be possible? I think maybe this documentation causes a little bit of confusion, because when you set the parameters specified in it `Adafactor(model.parameters(), lr=1e-3, relative_step=False, warmup_init=True)` it breaks. \r\n\r\n\r\n", "OK, so first, as @jsrozner PR goes and others commented, the current recommendation appears to be invalid.\r\n\r\nSo we want to gather all the different combinations that work and identify which of them provides the best outcome. I originally replied to this PR asking if someone knows of an authoritative paper we could copy the recommendation from - i.e. finding someone who already did the quality studies, so that we won't have it. So I'm all ears if any of you knows of such source.\r\n\r\nNow to your commentary, @alexvaca0, while HF trainer has Adam as the default it has `--adafactor` which enables it in the mode with using the external scheduler. Surely, we could change the Trainer to skip the external scheduler (or perhaps simpler feeding it some no-op scheduler) and instead use this recommendation if @sgugger agrees with that. But we first need to see that it provides a better outcome in the general case. Or alternatively to make `--adafactor` configurable so it could support more than just one way.\r\n\r\nFor me personally I want to understand first the different combinations, what are the impacts and how many of those combinations should we expose through the Trainer. e.g. like apex optimization level, we could have named combos `--adafactor setup1`, `--adafactor setup2` and would activate the corresponding configuration. But first let's compile the list of desirable combos.\r\n\r\nWould any of the current participants be interested in taking a lead on that? I'm asking you since you are already trying to get the best outcome with your data and so are best positioned to judge which combinations work the best for what situation.\r\n\r\nOnce we compiled the data it'd be trivial to update the documented recommendation and potentially extend HF Trainer to support more than one setting for Adafactor.\r\n\r\n", "I only ported the `--adafactor` option s it was implemented for the `Seq2SeqTrainer` to keep the commands using it working as they were. The `Trainer` does not have vocation to support all the optimizers and all their possible setups, just one sensible default that works well, that is the reason you can:\r\n- pass an `optimizer` at init\r\n- subclass and override `create_optimizer`.\r\n\r\nIn retrospect, I shouldn't have ported the Adafactor option and it should have stayed just in the script using it.", "Thank you for your feedback, @sgugger. \r\n\r\nSo let's leave the trainer as it is and let's then solve this for Adafactor as just an optimizer and then document the best combinations. ", "Per my comment on #7789, I observed that \r\n> I can confirm that `Adafactor(lr=1e-3, relative_step=False, warmup_init=False)` seems to break training (i.e. I observe no learning over 4 epochs, whereas `Adafactor(model.parameters(), relative_step=True, warmup_init=True, lr=None)` works well (much better than adam)\r\n\r\nGiven that relative_step and warmup_init must take on the same value, it seems like there is only one configuration that is working?\r\n\r\nBut, this is also confusing (see my comment above): https://github.com/huggingface/transformers/pull/10526#issuecomment-791012771\r\n\r\n> > Recommended T5 finetuning settings:\r\n> > - Scheduled LR warm-up to fixed LR\r\n> > - disable relative updates\r\n> > - use clip threshold: https://arxiv.org/abs/2004.14546\r\n> \r\n> In particular:\r\n> - are we supposed to do scheduled LR? adafactor handles this no?\r\n> - we *should not* disable relative updates\r\n> - i don't know what clip threshold means in this context\r\n", "I first validated that HF `Adafactor` is 100% identical to the latest [fairseq version](https://github.com/pytorch/fairseq/blob/5273bbb7c18a9b147e3f0cfc97121cc945a962bd/fairseq/optim/adafactor.py).\r\n\r\nI then tried to find out the source of these recommendations and found:\r\n\r\n1. https://discuss.huggingface.co/t/t5-finetuning-tips/684/3\r\n```\r\nlr=0.001, scale_parameter=False, relative_step=False\r\n```\r\n2. https://discuss.huggingface.co/t/t5-finetuning-tips/684/22 which is your comment @jsrozner where you propose that the opposite combination works well:\r\n\r\n```\r\nlr=None, relative_step=True, warmup_init=True\r\n```\r\n\r\nIf both found to be working, I propose we solve this conundrum by documenting this as following:\r\n```\r\n Recommended T5 finetuning settings (https://discuss.huggingface.co/t/t5-finetuning-tips/684/3):\r\n\r\n - Scheduled LR warm-up to fixed LR\r\n - disable relative updates\r\n - scale_parameter=False\r\n - use clip threshold: https://arxiv.org/abs/2004.14546\r\n\r\n Example::\r\n\r\n Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)\r\n\r\n Others reported the following combination to work well::\r\n\r\n Adafactor(model.parameters(), scale_parameter=False, relative_step=True, warmup_init=True, lr=None)\r\n\r\n - Training without LR warmup or clip threshold is not recommended. Additional optimizer operations like\r\n gradient clipping should not be used alongside Adafactor.\r\n```\r\n\r\nhttps://discuss.huggingface.co/t/t5-finetuning-tips/684/22 Also highly recommends to turn `scale_parameter=False` - so I added that to the documentation and the example above in both cases. Please correct me if I'm wrong.\r\n\r\nAnd @jsrozner's correction in this PR is absolutely right to the point.\r\n\r\nPlease let me know if my proposal makes sense, in particular I'd like your validation, @jsrozner, since I added your alternative proposal. And don't have any other voices to agree or disagree with it.\r\n\r\nThank you!\r\n\r\n\r\n\r\n\r\n", "> Is this part correct?\r\n> \r\n> > ```\r\n> > Recommended T5 finetuning settings:\r\n> > - Scheduled LR warm-up to fixed LR\r\n> > - disable relative updates\r\n> > - use clip threshold: https://arxiv.org/abs/2004.14546\r\n> > ```\r\n> \r\n> In particular:\r\n> \r\n> * are we supposed to do scheduled LR? adafactor handles this no?\r\n\r\nsee my last comment - it depends on whether we use the external LR scheduler or not.\r\n\r\n> * we _should not_ disable relative updates\r\n\r\nsee my last comment - it depends on `warmup_init`'s value\r\n\r\n> * i don't know what clip threshold means in this context\r\n\r\nthis?\r\n```\r\n def __init__(\r\n [....]\r\n clip_threshold=1.0,\r\n```", "I'm running some experiments, playing around with Adafactor parameters. I'll post here which configuration has best results. From T5 paper, they used the following parameters for fine-tuning: Adafactor with *constant* lr 1e-3, with batch size 128, if I understood the paper well. Therefore, I find it appropriate the documentation changes mentioned above, leaving the recommendations from the paper while mentioning other configs that have worked well for other users. In my case, for example, the configuration from the paper doesn't work very well and I quickly overfit. ", "Finally, I'm trying to understand the confusing:\r\n```\r\n - use clip threshold: https://arxiv.org/abs/2004.14546\r\n [...]\r\n gradient clipping should not be used alongside Adafactor.\r\n```\r\n\r\nAs the paper explains these are 2 different types of clipping.\r\n\r\nSince the code is:\r\n```\r\nupdate.div_((self._rms(update) / group[\"clip_threshold\"]).clamp_(min=1.0))\r\n```\r\nthis probably means that the default `clip_threshold=1.0` is in effect disables clip threshold. \r\n\r\nI can't find any mentioning of clip threshold in https://arxiv.org/abs/2004.14546 - is this a wrong paper? Perhaps it needed to link to the original paper https://arxiv.org/abs/1804.04235 where clipping is actually discussed? I think it's the param `d` in the latter paper and it proposes to get the best results with `d=1.0` without learning rate warmup:\r\n\r\npage 5 from https://arxiv.org/pdf/1804.04235:\r\n> We added update clipping to the previously described fast-\r\n> decay experiments. For the experiment without learning rate\r\n> warmup, update clipping with d = 1 significantly amelio-\r\n> rated the instability problem – see Table 2 (A) vs. (H). With\r\n> d = 2, the instability was not improved. Update clipping\r\n> did not significantly affect the experiments with warmup\r\n> (with no instability problems).\r\n\r\nSo I will change the doc to a non-ambiguous:\r\n\r\n`use clip_threshold=1.0 `", "OK, so here is the latest proposal. I re-organized the notes:\r\n\r\n```\r\n Recommended T5 finetuning settings (https://discuss.huggingface.co/t/t5-finetuning-tips/684/3):\r\n\r\n - Training without LR warmup or clip_threshold is not recommended. \r\n * use scheduled LR warm-up to fixed LR\r\n * use clip_threshold=1.0 (https://arxiv.org/abs/1804.04235)\r\n - Disable relative updates\r\n - Use scale_parameter=False\r\n - Additional optimizer operations like gradient clipping should not be used alongside Adafactor\r\n\r\n Example::\r\n\r\n Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)\r\n\r\n Others reported the following combination to work well::\r\n\r\n Adafactor(model.parameters(), scale_parameter=False, relative_step=True, warmup_init=True, lr=None)\r\n```\r\n\r\nI added these into this PR, please have a look.\r\n", "Let's just wait to hear from both @jsrozner and @alexvaca0 to ensure that my edits are valid before merging.", "I observed that `Adafactor(lr=1e-3, relative_step=False, warmup_init=False)` failed to lead to any learning. I guess this is because I didn't change `scale_parameter` to False? I can try rerunning with scale_param false.\r\n\r\nAnd when I ran with `Adafactor(model.parameters(), relative_step=True, warmup_init=True, lr=None)`, I *did not* set `scale_parameter=False`. Before adding the \"others seem to have success with ...\" bit, we should check on the effect of scale_parameter.\r\n\r\nRegarding clip_threshold - just confirming that the comment is correct that when using adafactor we should *not* have any other gradient clipping (e.g. `nn.utils.clip_grad_norm_()`)?\r\n\r\nSemi-related per @alexvaca0, regarding T5 paper's recommended batch_size: is the 128 recommendation agnostic to the length of input sequences? Or is there a target number of tokens per batch that would be optimal? (E.g. input sequences of max length 10 tokens vs input sequences of max length 100 tokens -- should we expect 128 to work optimally for both?)\r\n\r\nBut most importantly, shouldn't we change the defaults so that a call to `Adafactor(model.paramaters())` == `Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)`\r\n\r\ni.e, we default to what we suggest?", "> I observed that Adafactor(lr=1e-3, relative_step=False, warmup_init=False) failed to lead to any learning. I guess this is because I didn't change scale_parameter to False? I can try rerunning with scale_param false.\r\n\r\nYes, please and thank you!\r\n\r\n\r\n\r\n\r\n> Regarding clip_threshold - just confirming that the comment is correct that when using adafactor we should not have any other gradient clipping (e.g. nn.utils.clip_grad_norm_())?\r\n\r\nThank you for validating this, @jsrozner \r\n\r\nIs the current incarnation of the doc clear wrt this subject matter or should we add an explicit example?\r\n\r\nOne thing I'm concerned about is that the Trainer doesn't validate this and will happily run `clip_grad_norm` with Adafactor\r\n\r\nMight we need to add to:\r\n\r\nhttps://github.com/huggingface/transformers/blob/9f8fa4e9730b8e658bcd5625610cc70f3a019818/src/transformers/trainer.py#L649-L651\r\n\r\n```\r\nif self.args.max_grad_norm:\r\n raise ValueError(\"don't use max_grad_norm with adafactor\")\r\n```\r\n\r\n\r\n> But most importantly, shouldn't we change the defaults so that a call to Adafactor(model.paramaters()) == `Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)`\r\n\r\nSince we copied the code verbatim from fairseq, it might be a good idea to keep the defaults the same? I'm not attached to either way. @sgugger what do you think?\r\n\r\nedit: I don't think we can/should since it may break people's code that relies on the current defaults.", "> > Regarding clip_threshold - just confirming that the comment is correct that when using adafactor we should not have any other gradient clipping (e.g. nn.utils.clip_grad_norm_())?\r\n> \r\n> Thank you for validating this, @jsrozner\r\n\r\nSorry, I didn't validate this. I wanted to confirm with you all that this is correct.\r\n\r\n> > But most importantly, shouldn't we change the defaults so that a call to Adafactor(model.paramaters()) == `Adafactor(model.parameters(), scale_parameter=False, relative_step=False, warmup_init=False, lr=1e-3)`\r\n> \r\n> Since we copied the code verbatim from fairseq, it might be a good idea to keep the defaults the same? I'm not attached to either way. @sgugger what do you think?\r\n\r\nAlternative is to not provide defaults for these values and force the user to read documentation and decide what he/she wants. Can provide the default implementation as well as Adafactor's recommended settings\r\n\r\n", "> I'm running some experiments, playing around with Adafactor parameters. I'll post here which configuration has best results. From T5 paper, they used the following parameters for fine-tuning: Adafactor with _constant_ lr 1e-3, with batch size 128, if I understood the paper well. Therefore, I find it appropriate the documentation changes mentioned above, leaving the recommendations from the paper while mentioning other configs that have worked well for other users. In my case, for example, the configuration from the paper doesn't work very well and I quickly overfit.\r\n\r\n@alexvaca0 \r\nWhat set of adafactor params did you find work well when you were finetuning t5?", "> Sorry, I didn't validate this. I wanted to confirm with you all that this is correct.\r\n\r\nI meant validating as in reading over and checking that it makes sense. So all is good. Thank you for extra clarification so we were on the same page, @jsrozner \r\n \r\n> Alternative is to not provide defaults for these values and force the user to read documentation and decide what he/she wants. Can provide the default implementation as well as Adafactor's recommended settings\r\n\r\nas I appended to my initial comment, this would be a breaking change. So if it's crucial that we do that, this would need to happen in the next major release.\r\n", "Or maybe add a warning message that indicates that default params may not be optimal? It will be logged only a single time at optimizer init so not too annoying.\r\n\r\n`log.warning('Initializing Adafactor. If you are using default settings, it is recommended that you read the documentation to ensure that these are optimal for your use case.)`", "But we now changed it propose two different ways - which one is the recommended one? The one used by the Trainer?\r\n\r\nSince it's pretty clear that there is more than one way, surely the user will find their way to the doc if they aren't happy with the results.\r\n\r\n", "I ran my model under three different adafactor setups:\r\n\r\n```python\r\n optimizer = Adafactor(self.model.parameters(),\r\n relative_step=True,\r\n warmup_init=True)\r\n```\r\n```python\r\noptimizer = Adafactor(self.model.parameters(),\r\n relative_step=True,\r\n warmup_init=True,\r\n scale_parameter=False)\r\n```\r\n```python\r\n optimizer = Adafactor(self.model.parameters(),\r\n lr=1e-3,\r\n relative_step=False,\r\n warmup_init=False,\r\n scale_parameter=False)\r\n```\r\n\r\nI track exact match and NLL on the dev set. Epochs are tracked at the bottom. They start at 11 because of how I'm doing things. (i.e. x=11 => epoch=1)\r\n\r\nNote that I'm training a t5-small model on 80,000 examples, so maybe there's some variability with the sort of model we're training?\r\n\r\n\r\n(Image works if you navigate to the link, but seems not to appear?)\r\n\r\npurple is (1)\r\nblue is (2) and by far the worst (i.e. shows that scale_param should be set to True if we are using relative_step)\r\nbrown is (3)\r\n\r\nIn particular, it looks like scale_param should be True for the setting under \" Others reported the following combination to work well::\"\r\n\r\nOn the other hand, it looks like for a t5-large model, (3) does better than (1) (although I also had substantially different batch sizes). ", "Thank you, @jsrozner, for running these experiments and the analysis.\r\n\r\nSo basically we should adjust \" Others reported the following combination to work well::\" to `scale_param=True`, correct? \r\n\r\nInterestingly we end up with 2 almost total opposites.", "@jsrozner Batch size and learning rate configuration go hand in hand, therefore it's difficult to know about your last reflexion, as having different different batch sizes lead to different gradient estimations (in particular, the lower the batch size, the worse your gradient estimation is), the larger your batch size, the larger your learning rate can be without negatively affecting performance. \r\nFor the rest of the experiments, thanks a lot, the comparison is very clear and this will be very helpful for those of us who want to have some \"default\" parameters to start from. I see a clear advantage to leave the learning rate as None, as when setting an external learning rate we typically have to make experiments to find the optimal one for that concrete problem, so if it provides results similar or better to the ones provided by the paper's recommended parameters, I'd go with that as default. \r\nI hope I can have my experiments done soon, which will be with t5-large probably, to see if they coincide with your findings.", "OK, let's merge this and if we need to make updates for any new findings we will do it then. ", "Although I didn't really run an experiment, I have found that my settings for adafactor (relative step, warmup, scale all true) do well when training t5-large, also.\r\n\r\n@alexvaca0 please post your results when you have them!" ]

[]

[]

[ "Hi,\r\n\r\ncould you please ask this question on the [forum](https://discuss.huggingface.co/)? We're happy to help you there!\r\n\r\nQuestions regarding training of models are a perfect use case for the forum :) for example, [here](https://discuss.huggingface.co/search?q=pegasus) you can find all questions related to fine-tuning PEGASUS.", "@NielsRogge thanks a lot! I will post the question in forum 😄🤝", "I think this issue can be closed - I used another input text and the generated text is different - I guess the fine-tuned model is different but for some input text, the generated result is exactly the same as the large model - interesting to know why though :)." ]

[]

[]

[ "@sgugger Can you review this extremely simple PR", "Hi there. The default is actually the same as the legacy script and I don't see any reason to change it. It provides the results given in the README which are consistent with the paper.\r\n\r\nAlso a default of 512 won't work with models that have a smaller max_length like distilbert.", "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.", "Oh. Yeah, I didn’t know distilbert had a smaller max_len. I’d seen max_len at least greater than or equal to 512 until now. " ]

[]

[]

[ "Hello! In the documentation you pointed to, you'll see that there is the `kwargs` argument, which accepts any keyword argument. The doc says:\r\n\r\n```\r\n(optional) Remaining dictionary of keyword arguments. Keyword arguments come in two flavors:\r\n\r\nWithout a prefix which will be input as **encoder_kwargs for the encoder forward function.\r\n\r\nWith a decoder_ prefix which will be input as **decoder_kwargs for the decoder forward function.\r\n```\r\n\r\nHave you tried passing `position_ids` to the `__call__` method of your model instantiated with the `EncoderDecoderModel.from_encoder_decoder_pretrained`? It should work!", "Thanks so much!" ]

[]

[]

[ "`torchscript` does not support anything else than tuple outputs, so you can't rely on the attributes when using it and transformers automatically sets `return_dict=False` in this case.\r\n\r\nYou need to access to the output fields with indices.", "@sgugger Thanks for the explanations.", "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 @lecidhugo \r\n\r\nThank you for reporting this. Is this for one particular example or is it happening for all examples?", "Hi @patil-suraj,\r\nThank you for your reply.\r\nIndeed, it is for all examples and for many languages", "Thanks. I'll look into it next week.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Unstale\r\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

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

[]

[ "Previous mention of this idea\r\nhttps://github.com/huggingface/transformers/issues/4476#issuecomment-677823688", "Here's my workaround (from the previous issue): https://github.com/huggingface/transformers/issues/4476#issuecomment-951445067", "May be of interest to @SaulLu @NielsRogge ", "This is indeed a feature that is requested and would make sense to have!\r\n\r\nTo my knowledge @NielsRogge has started a [PR to add this feature](https://github.com/huggingface/transformers/pull/12913). Unfortunately, the feature requires development also in the Tokenizers library and nobody has yet got the bandwidth to support [the dedicated issue](https://github.com/huggingface/tokenizers/issues/779) for this feature.", "> the feature requires development\r\n\r\nLinks provided show that only fast tokenizers are stopping this (\"requested and would make sense to have\" feature) to be implemented (\"since july.... July, Karl!\"). May this be implemented for ordinary tokenizers faster, and for faster tokenizers later, when they solve dedicated issues?\r\n\r\nIn my opinion, having left truncation for \"slow\" tokenizers now, is better than not having one at all (or in distant future releases). It is awaited (https://github.com/huggingface/transformers/issues/4476#event-3364051747) for more than a year now at least.", "Seems this was added in #14947 about a month ago, so hopefully it will be in a near-future release!" ]

[ { "id": 1834083927, "node_id": "MDU6TGFiZWwxODM0MDgzOTI3", "url": "https://api.github.com/repos/huggingface/transformers/labels/External", "name": "External", "color": "fbca04", "default": false, "description": "Using the library with external tools (onnx, tflite, ...)" } ]

[]

[ "Hello!\r\n\r\nThanks for reporting this issue! Did you try to convert your H5 file to be able to use it with `tensorflowjs_converter --input_format keras path/to/my_model.h5 path/to/tfjs_target_dir`? You can also have a SavedModel version with:\r\n\r\n1. `model = TFAutoModel.from_pretrained(\"distilbert-base-uncased\")`\r\n2. `model.save_pretrained(path, saved_model=True)`\r\n\r\nH5 and SavedModel conversion process are nicely explained in https://www.tensorflow.org/js/tutorials/conversion/import_keras and https://www.tensorflow.org/js/tutorials/conversion/import_saved_model", "Hi @jplu thanks for the fast reply ! \r\n\r\nYes I used the `tensorflowjs_converter` to convert the H5 huggingFace model to a TFJS layerModel (.json).\r\n\r\nIf not using a h5 but a savedModel I don't know how to then use it in tensorflowJS.... all my researches seems to indicate that a savedModel can only do inference but i would like to finetune it. \r\n\r\nMy goal is just to finetune Bert in nodejs (could be with other framework than tfjs if you know some (like pytorch node bindings or some fancy framework but I found nothing that allow training))\r\n\r\n## Attempts : \r\n\r\n### HuggingFace : \r\n1. Load model with huggingFace : `model = TFAutoModel.from_pretrained(model_name)`\r\n2. Save it (H5 format) : `model.save_pretrained(path_saved)`\r\n3. convert it to tfjs.json Layer format : \r\n```python\r\ndispatch_keras_h5_to_tfjs_layers_model_conversion(\r\n h5_path='path_saved/tf_model.h5',\r\n output_dir=my_tfjs_output_dir\r\n)\r\n```\r\n(`dispatch_keras_h5_to_tfjs_layers_model_conversion` is the function called when using `tensorflowjs_convert`, i'm calling it from python code)\r\n4. try to load it in nodejs : `const bert = await tf.loadLayersModel('file://${my_tfjs_output_dir}/model.json')`\r\n5. Get the error `UnhandledPromiseRejectionWarning: TypeError: Cannot read property 'model_config' of null`\r\n\r\n### Keras or tfjs saving (to have the `model_config`)\r\n1. Load model with huggingFace : `model = TFAutoModel.from_pretrained(model_name)`\r\n2. Save it (H5 format) : \r\n```python\r\ntf.keras.models.save_model(\r\n model=model,\r\n filepath=path_saved,\r\n signatures=tf.function(model.call).get_concrete_function([\r\n tf.TensorSpec([1, 512], tf.int32, name=\"input_ids\"),\r\n tf.TensorSpec([1, 512], tf.int32, name=\"attention_mask\")\r\n ]),\r\n save_format='h5'\r\n)\r\n```\r\nor \r\n`tfjs.converters.save_keras_model( model, path_saved)`\r\n3. Get the error `Saving the model to HDF5 format requires the model to be a Functional model or a Sequential model.`\r\n\r\n### Other formats : \r\n1. Load model with huggingFace : `model = TFAutoModel.from_pretrained(model_name)`\r\n2. Save it (SavedModel format) : `model.save_pretrained(path_saved)`\r\n4. try to load it in nodejs : `const bert = await tf.node.loadSavedModel(path_saved, ['serve'], \"serving_default\")`\r\n5. Cannot train it or use it in another model with `this._clf = tf.sequential(); this._clf.add(bert);`\r\n\r\n#### GraphModel or onnx model are for inference only \r\n\r\nSorry for this (too) long reply. \r\n\r\nThanks for your help (I guess many people try to attempt the same), \r\n\r\nHave a great day ", "Indeed you cannot train or fine tune a SavedModel in other envs than Python.\r\n\r\nCurrently to load a model you are forced to have the config file because it is required as an argument to init a model. Then the H5 as well. Furthermore, the TF implementations are using subclass models which are not compliant with most of the internal TF process (such as what you are trying to do in your described process when saving a model).\r\n\r\nThen, I suggest you to fine tune your model in Python, create a SavedModel, and then use it in JS as described in the link I shared. Sorry for the inconvenience.", "Okay so SOA is training custom models in tfjs but finetunning is almost impossible. It's what I thought but I'm glad you confirmed it. \r\n\r\nDo you know any other solutions that would allow me to finetune a model in a webstack (nodejs / typescript) ?\r\nI'm still hoping that one day pytorch will have node bindings... \r\n\r\nThanks again for your time. \r\nI guess we can close the issue.", "> Do you know any other solutions that would allow me to finetune a model in a webstack (nodejs / typescript) ?\r\n\r\nNo, sorry this is not really my domain. And sorry again for the inconvenience on this." ]

[]

[]

[]

[]

[]

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

[]

[]

[ "Hey @NielsRogge \r\n\r\n\r\n> Add and improve tests. Currently I have defined the following tests: test_modeling_vit.py, test_feature_extraction_vit.py. However, for the former, since ViT does not use input_ids/input_embeds, some tests are failing, so I wonder whether it should use all tests defined in test_modeling_common.py. For the latter, I also need some help in creating random inputs to test the feature extractor on.\r\n\r\n\r\nSome common modeling test depend on the specific parameter names, (`input_ids`, `input_embeds`). You could just override such tests in your test class and use the correct parameter names. For example the `test_forward_signature` test\r\nexpects `inputs_ids`, so it should be overridden in your class to expect `input_values`. \r\n\r\nAlso, the tests for `input_embeds` (for example `test_inputs_embeds`) can be skipped since `ViT` does not use those. Agin just overrides the test and use `pass` in the method body.\r\n\r\nYou could use the modeling tests of `Wav2Vec2` and `Speech2Text` for reference since those models also use different parameter names.", "I like the overall design `ViTFeatureExtractor`. Regrading the import `ViTFeatureExtractor`\r\nI think it should be always imported in the __init__ files, and instead, `ViTFeatureExtractor` could check for `torchvision` and raise if it’s not installed. Otherwise, the TF tests on CI will fail because they won’t be able to import `ViTFeatureExtractor` as we don’t install `torchvision` in TF tests.\r\n\r\nWe should also add the `torchvision` and `PIL` dependency in the setup.py file as `extras[\"vision\"]` and also add it in `config.yaml` for CI", "Thanks for the reviews, addressed most of the comments. To do:\r\n\r\n- [x] rename `self.self` to `self.attention` and update conversion script accordingly\r\n- [x] convert more models, place them under the google namespace\r\n- [x] add model cards\r\n- [x] add 1,000 ImageNet class names to config", "I've addressed all comments. Most important updates:\r\n\r\n* moved the ImageNet id to classes dict to a new file under transformers.utils named `imagenet_classes`.py.\r\n* added a warning to the `__call__` method of `ViTFeatureExtractor` to indicate that NumPy arrays and PyTorch tensors are converted to PIL images when resizing, so it's most efficient to pass in PIL images. \r\n\r\nThe remaining comments which are still open have to do with styling. I seem to have some issues with `make style`. The max_length is set to 119, so not sure what's causing this." ]

[]

[]

[ "Pinging @patil-suraj and @sgugger ", "Hi @clang88 \r\n\r\nThe goal of the examples scripts is to keep them minimal and simple and I'm not sure if we want to support this immediately. \r\n\r\nFor now, you could use the `--group_by_length` argument which will group the long sequences together to avoid varying lengths and minimize the number of padding tokens.\r\n\r\nAlso to train large models, I would recommend you take a look at `fairscale/deepspeed` integration. Check this [blog post](https://huggingface.co/blog/zero-deepspeed-fairscale) for how to use `fairscale/deepspeed` with `Trainer` \r\n\r\n@sgugger think we could add a `MaxTokensSampler` for this in case we want to support this. ", "Hi @patil-suraj,\r\n\r\nthank you for the quick reply! \r\n\r\nI will take a look at the `fairscale/deepspeed` integration!\r\n\r\nAs for `--group_by_length`: this will only work correctly if I use trainer with a `data_collator`, am I correct? I have been already experimenting with that approach, but am having some trouble during the evaluation phase with `custom_metrics`. For whatever reason, the labels passed to the function by the trainer appear to be padded with the default of -100, even though I am passing `label_pad_token_id=` of 0 (for mT5) or 1 (for mBART) in the collator. I am aware this is a whole other issue, but maybe you are aware of any potential solutions for this?\r\n\r\nThat said, I am sure `max_tokens_per_batch` would a be a great asset, as `group_by_length` does not fix the underlying issue of having batches with very long sentences that go OOM. For now I am just truncating my dataset with `max_length`, but that clearly leads to less than ideal performance of the fine-tuned model.", "> think we could add a MaxTokensSampler for this in case we want to support this\r\n\r\nIt's a whole batch sampler that would be needed, since it results in batch size not being constant. And it would need the same version as a distributed batch sampler. This is a lot of work for a very specific use case, so we could accept a PR on an example in a research project first.\r\n\r\nOf course there is still the possibility of one user using the implementation from FAIR as was done in the old `finetune` script.", "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." ]

[]

[]

[ "BERT uses absolute position embeddings by default. The sin/cos embeddings are from the original Transformer paper IRRC. ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

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

[]

[]

[ "The `GITHUB_TOKEN` got rate-limited, unfortunately. The PR should be good to go, I'll try to run it again tomorrow night." ]

[]

[]

[ "I noticed that the size of pytorch_model.bin that was downloaded into .cache is of size 442791751. When I save_pretrained() the model, the size is 442792154. \r\nIf I copy the first model into the model directory, I get valid results....", "Hello! You're using `'google/tapas-base'` in order to initialize weights, why don't you use the variant fine-tuned on WTQ like in the configuration and tokenizer? \r\n\r\nUnless you're using a fine-tuned version, you won't benefit from the best possible predictions, like you have seen here.", "@LysandreJik , I fixed that typo and changed it to \r\nmodel = TapasForQuestionAnswering.from_pretrained('google/tapas-base-finetuned-wtq', config=config), but I get the same discrepancy in results.\r\nI am trying to figure out why I get no results when I load the model from a directory(that was generated as above).", "I'm running your code and I get some sensible results:\r\n\r\n```py\r\nfrom transformers import TapasConfig,TapasTokenizer,TapasForQuestionAnswering\r\nimport torch\r\nconfig = TapasConfig.from_pretrained('google/tapas-base-finetuned-wtq',from_pt=True)\r\nmodel = TapasForQuestionAnswering.from_pretrained('google/tapas-base-finetuned-wtq', config=config)\r\ntokenizer=TapasTokenizer.from_pretrained(\"google/tapas-base-finetuned-wtq\",from_pt=True)\r\nimport sys\r\n\r\noutdir = \"tmp\"\r\n\r\nmodel.save_pretrained(outdir)\r\ntokenizer.save_pretrained(outdir)\r\nconfig.save_pretrained(outdir)\r\n\r\nfrom transformers import pipeline\r\n\r\nnlp = pipeline(task=\"table-question-answering\",framework=\"pt\",model=outdir, tokenizer=outdir)\r\n#nlp = pipeline(task=\"table-question-answering\")\r\n\r\nimport pandas as pd\r\n\r\ndata= { \"actors\": [\"brad pitt\", \"leonardo di caprio\", \"george clooney\"],\r\n \"age\": [\"56\", \"45\", \"59\"],\r\n \"number of movies\": [\"87\", \"53\", \"69\"],\r\n \"date of birth\": [\"7 february 1967\", \"10 june 1996\", \"28 november 1967\"]\r\n}\r\n\r\nimport numpy as np\r\ntable = pd.DataFrame.from_dict(data)\r\nprint(np.shape(table))\r\nresult = nlp(query=[\"How many movies has Brad Pitt acted in\",\"What is Leonardo di caprio's age\"],table=table)\r\n\r\nprint(result)\r\n```\r\n\r\nResults in:\r\n```\r\n[{'answer': 'SUM > 87', 'coordinates': [(0, 2)], 'cells': ['87'], 'aggregator': 'SUM'}, {'answer': 'AVERAGE > 45', 'coordinates': [(1, 1)], 'cells': ['45'], 'aggregator': 'AVERAGE'}]\r\n```\r\n\r\nThe aggregator are a bit off but the results are correct. Brad Pitt has played in 87 movies and Leonardo di Caprio is 45.", "I did additionally specify `tokenizer=outdir`, could that be the source of your issue?", "@LysandreJik , that was it ! \r\nI will look into the code, but any insight into why tokenizer=outdir needs to be specified (for this pipeline task only) ?", "@LysandreJik , what is the size of your pytorch_model.bin in outdir ?\r\n", "The tokenizer needs to be specified for every pipeline tasks. You should always specify the checkpoint for the tokenizer as well as for the model. The size is `442.79 MB`!" ]

[]

[]

[ "Hi, this is a duplicate, see #10446 ", "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." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Tested on TPUs and all worked well (training, checkpointing, reloading from checkpoint, evaluating a saved model), so I will merge this." ]

[]

[]

[ "Hi! I'm sorry, I don't understand what you mean exactly. A pipeline uses a model and a tokenizer under the hood, do you mean you want to fine-tune the model used underneath, and use that fine-tuned model?\r\n\r\nIf that is so, it is simple: any model/tokenizer can be loaded in the pipeline via local path/hub identifier. I invite you to read the [documentation of the pipelines](https://huggingface.co/transformers/main_classes/pipelines.html#transformers.pipeline), especially the `model` and `tokenizer` arguments.", "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! Actually I think you touched the base of the issue for MNLI, I would guess it has to do with the number of labels. When you do the following:\r\n```py\r\n>>> from transformers import RobertaForMaskedLM, RobertaConfig, RobertaTokenizer\r\n>>> tok = RobertaTokenizer.from_pretrained('roberta-base')\r\n>>> config = RobertaConfig.from_pretrained('roberta-base')\r\n>>> model = RobertaForMaskedLM(config=config)\r\n>>> model.save_pretrained('tmp_model')\r\n>>> from transformers import RobertaForSequenceClassification\r\n>>> model = RobertaForSequenceClassification.from_pretrained(\"tmp_model\")\r\n```\r\n\r\nYou'll see that the sequence classification model you've loaded has 2 labels, whereas MNLI is a 3-way classification:\r\n```py\r\n>>> model.classifier.out_proj\r\nLinear(in_features=768, out_features=2, bias=True)\r\n```\r\n\r\nI don't get any such errors when changing the configuration initialization to the following:\r\n```py\r\n>>> config = RobertaConfig.from_pretrained('roberta-base', num_labels=3)\r\n```\r\n\r\nRegarding the STS-B issue, I think this is an issue that was recently solved on `master`. Can you try pulling the `master` branch once again and letting me know if it fixes your issue? If not, I'll try and check what's happening.", "> Hi! Actually I think you touched the base of the issue for MNLI, I would guess it has to do with the number of labels. When you do the following:\r\n> \r\n> ```python\r\n> >>> from transformers import RobertaForMaskedLM, RobertaConfig, RobertaTokenizer\r\n> >>> tok = RobertaTokenizer.from_pretrained('roberta-base')\r\n> >>> config = RobertaConfig.from_pretrained('roberta-base')\r\n> >>> model = RobertaForMaskedLM(config=config)\r\n> >>> model.save_pretrained('tmp_model')\r\n> >>> from transformers import RobertaForSequenceClassification\r\n> >>> model = RobertaForSequenceClassification.from_pretrained(\"tmp_model\")\r\n> ```\r\n> \r\n> You'll see that the sequence classification model you've loaded has 2 labels, whereas MNLI is a 3-way classification:\r\n> \r\n> ```python\r\n> >>> model.classifier.out_proj\r\n> Linear(in_features=768, out_features=2, bias=True)\r\n> ```\r\n> \r\n> I don't get any such errors when changing the configuration initialization to the following:\r\n> \r\n> ```python\r\n> >>> config = RobertaConfig.from_pretrained('roberta-base', num_labels=3)\r\n> ```\r\nI didn't have this problem on branch `v4.0.1-release` because the `num_labels` parameter was set automatically. I didn't have to set the exact number of label for each GLUE task. In fact, the script computes them automatically, but for some reason the error still happens.\r\n\r\n> Regarding the STS-B issue, I think this is an issue that was recently solved on `master`. Can you try pulling the `master` branch once again and letting me know if it fixes your issue? If not, I'll try and check what's happening.\r\n\r\nError still present on master (`4.4.0.dev0`).", "cc @sgugger for knowledge regarding the MNLI issue." ]

[]

[]

[]

[]

[]

[ "Hi! I think this is more of a question for `pyannote` rather than for `transformers`?", "cc'ing @hbredin for visibility :)", "Thanks @julien-c for the ping.\r\n\r\n@karthikgali please open an issue or discussion in [pyannote.audio Github repo](https://github.com/pyannote/pyannote-audio) instead.", "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 there, \r\n\r\nYou should use the `AutoModelForQuestionAnswering` class to load a QA model, the `AutoModel` class just loads the base model and doesn't load the task-specific head, which is the reason for this error.\r\n\r\nIn general, always use the task-specific auto classes to load task-specific architectures.", "@patil-suraj confirmed, it works \r\n\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(\r\n 'mrm8488/bert-italian-finedtuned-squadv1-it-alfa',\r\n cache_dir=os.getenv(\"cache_dir\", \"model\"))\r\nmodel = AutoModelForQuestionAnswering.from_pretrained(\r\n 'mrm8488/bert-italian-finedtuned-squadv1-it-alfa',\r\n cache_dir=os.getenv(\"cache_dir\", \"model\"))\r\n\r\nnlp_qa_bert = pipeline(\r\n 'question-answering',\r\n model=model,\r\n tokenizer=tokenizer)\r\n```" ]

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

[]

[ "Hi @brandenchan ,\r\n\r\nI think it should be easier with version 2 of DeBERTa, because they use a \"normal\" sentence piece model now:\r\n\r\nhttps://github.com/huggingface/transformers/pull/10018\r\n\r\nSo having a fast alternative would be great. \r\n\r\n(The new 128k vocab size should really boost performance on QA tasks!)\r\n\r\n", "Indeed, this would be a very nice addition and way easier to implement than for the first DeBERTa. I'm adding the `Good Second Issue` label so that a community member may work on it. @brandenchan or @stefan-it feel free to take it too if you feel like it!", "Hi, I am looking for my first open source contribution. May I take this if its still available?", "Yes, of course! Thank you!", "@ShubhamSanghvi Maybe wait until #10703 is merged.", "Hi, as far as I understand I will have to add tokenizer files for debarta_v2 to implement the fast tokenizer? \r\n\r\nMay I know how could I get the tokenizer files for deberta_v2 models and how to upload them to the intended destinations, which I believe should be (for deberta-v2-xlarge) :\r\n\r\nhttps://huggingface.co/microsoft/deberta-v2-xlarge/resolve/main/\r\n\r\nThanks, Shubham", "@ShubhamSanghvi Do you only want to implement the fast tokenizer for DebertaV2 or also for Deberta? \r\n\r\n> May I know how could I get the tokenizer files for deberta_v2 models \r\n\r\nI think this is what you have to figure out. I would check the other models that have a slow sentencepiece tokenizer.\r\n\r\n> how to upload them to the intended destinations, which I believe should be (for deberta-v2-xlarge) \r\n\r\nYou can not upload them there. Upload them to some kind of a public cloud and request an upload.\r\n\r\n\r\n\r\n\r\n", "@ShubhamSanghvi Are you planning to create a PR for this issue soon?\r\n", "Hi @mansimane, I am currently working on it. I am hoping to get it done by next week. " ]

[]

[]

[ "Hey @idanmoradarthas, \r\n\r\nI will soon release a notebook, that will explain in-detail how to fine-tune a Wav2Vec2 model (~1week). \r\n\r\nIt's quite time consuming for me to debug user-specific code, such as `convert_to_dataset_torch`, so I can only give you some tips here:\r\n\r\n- Try to convert your dataset to PyTorch tensors istead of `np.ndarray`'s. This means you should change all of your lines that do:\r\n```processor(row[\"speech\"], sampling_rate=16000)``` to ```processor(row[\"speech\"], sampling_rate=16000, return_tensors=\"pt\")```", "Thanks =)", "with self.processor.as_target_processor():\r\n labels_batch = self.processor.pad(\r\ni use this command in order to encode labels but i got number 3 instead of all letters in sentence\r\nHow can i solve this issue?", "Hey @kasrasehat,\r\n\r\nCould you please open a new issue? ", "@patrickvonplaten sure" ]

[]

[]

[ "> Looks good to me!\r\n> \r\n> Some of the other library models also use this trick (BART-like models), we should also investigate those.\r\n\r\nGood point - yeah, let me fix this in this PR actually" ]

[]

[]

[ "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." ]

[]

[]

[ "We need to run benchmark tests to see by how much the speed improved both on CPU and GPU", "@patrickvonplaten Can you please help with above comments ?", "Hey @punitvara,\r\n\r\nAt the moment, I sadly don't have the time to handle the big chunk of the PR. It would be great if you could try to:\r\n\r\n1) Find a way to benchmark your new function on GPU and show that it yields a speed-up in the forward pass compared to the old function\r\n\r\n2) Try out some advanced PyTorch indexing to replace the for loops.\r\n\r\nTaking a look at those PRs should help you: https://github.com/huggingface/transformers/pull/9600, https://github.com/huggingface/transformers/pull/9453, https://github.com/huggingface/transformers/pull/6064", "Closing due to inactivity. Sorry @punitvara, I saw a lot of interest from other people to open a PR and this one seems to have stalled. Feel free to re-open it and give it a second shot if you want :-) ", "I got busy into some other work. I will try to work on different issue. If you get any PR, feel free to merge it " ]

[]

[]

[ "Hi, I tried to keep the code changes to a minimum. Thus, I avoided adding another argument for returning cross-attention weights and used `output_attentions` to check instead. \r\nAlso in docstrings, for `decoder_attentions` the shape is mentioned as `(batch_size*num_return_sequences, num_heads, generated_length,\r\n sequence_length)`, although it seems that the shape is always `(batch_size*num_return_sequences, num_heads, 1, 1)`. `generated_length` is 1 as the code returns tuples per generated token and `sequence_length` is 1 since it's decoder self-attentions.\r\nSimilarly for `cross_attentions` the shape should be `(batch_size, num_heads, 1, input_sequence_length)`. Are there any examples where this is not True?\r\n\r\nPlease let me know what you think. Thanks!", "> Hi, I tried to keep the code changes to a minimum. Thus, I avoided adding another argument for returning cross-attention weights and used `output_attentions` to check instead.\r\n> Also in docstrings, for `decoder_attentions` the shape is mentioned as `(batch_size*num_return_sequences, num_heads, generated_length, sequence_length)`, although it seems that the shape is always `(batch_size*num_return_sequences, num_heads, 1, 1)`. `generated_length` is 1 as the code returns tuples per generated token and `sequence_length` is 1 since it's decoder self-attentions.\r\n> Similarly for `cross_attentions` the shape should be `(batch_size, num_heads, 1, input_sequence_length)`. Are there any examples where this is not True?\r\n> \r\n> Please let me know what you think. Thanks!\r\n\r\nRegarding the `decoder_attentions` shape -> it's just the first attentions that is of shape `batch_size, num_heads, 1, 1`. Then it goes up to `..., 2, 2`, `..., 3, 3` etc. However if `use_cache` is enabled, then the shape is `...1, 1`, `...1, 2`, `....1, 3`.\r\n\r\nCheck this example:\r\n\r\n```python\r\nfrom transformers import BartForConditionalGeneration\r\nimport torch\r\n\r\nbart = BartForConditionalGeneration.from_pretrained(\"facebook/bart-large-cnn\")\r\nbart.config.max_length = 5 \r\nbart.config.num_beams = 1\r\n\r\noutputs_use_cache = bart.generate(torch.tensor([ 10 * [0]]), return_dict_in_generate=True, output_attentions=True)\r\noutputs_no_cache = bart.generate(torch.tensor([ 10 * [0]]), return_dict_in_generate=True, output_attentions=True, use_cache=False)\r\n\r\n# outputs_no_cache.decoder_attentions[1][0].shape gives (1, 16, 2, 2)\r\n# outputs_use_cache.decoder_attentions[1][0].shape gives (1, 16, 1, 2)\r\n```", "> Regarding the decoder_attentions shape -> it's just the first attentions that is of shape batch_size, num_heads, 1, 1. Then it goes up to ..., 2, 2, ..., 3, 3 etc. However if use_cache is enabled, then the shape is ...1, 1, ...1, 2, ....1, 3.\r\n\r\nThanks for the example! It makes sense now.", "Thanks for the great work. But I am still a bit confused about getting the cross-attention or encode-decoder attention if I am not mistaken the cross attention. Any hint would be great. Maybe it would be like \r\n\r\n```\r\n# outputs_no_cache.cross_attentions[1][0]\r\n# outputs_use_cache.cross_attentions[1][0]\r\n```\r\n\r\nThanks! @Mehrad0711 ", "hi @xwuShirley \r\n\r\nYes, you are right. You can access the cross attentions using \r\n`outputs. cross_attentions`" ]

[]

[]

[ "Hello!\n\nYou cannot load a model inside a namescope. This is the expected behavior because all the names are forced insides the h5 file. To use the model you have to load it outside your defined namescope and then use it inside.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

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

[]

[]

[ "Just to be sure I understand completely the problem since I don't see the F1s for the full results obtained via pipeline, which is producing the best results? `pipeline` or `run_qa`? Or are they different but overall comparable?", "`run_qa.py` produces the better results. They are different on about 3000 records, 2700 of those are off by a character or so, 300 are off by a larger margin, sometimes different answers, 12 records Null (pipeline) vs Not-Null (run_qa) (false negative). And that is on SQuAD. I see a wider gap on custom datasets. \"Comparable\" is a subjective word so I perhaps should not comment on that. In my opinion they need to match. What we get out of run_qa.py in batch for a dataset should be the same for record i, meaning when we pass the same query and context to the model, the script and the module should deterministically return the same answer. The problem might not be exactly in the predictions but rather in the post prediction functions. ", "@sgugger Hi, any update on this? Did you all get a chance to look into the problem? by the way, after I wrote the message above, I realized \"better\" is also subjective. I would prefer a 5 token answer to a 1 token answer for instance depending on the question and context. As I mentioned pipeline does not clean up the answer as well, punctuation gets carried over. Looking at custom validation sets I can tell the F1 score is lower for Pipelines, which I did it pretty much manually by writing custom code as I had a smaller set. If there is a way to take the pipeline predictions and calculate a F1 score against the ground truths in SQuAD, I would be happy to do that and return you a result. But again, that would be just SQuAD. better or worse, the answers don't match. Maybe the problem is in run_qa.py and this is an opportunity to get that fixed. Thanks in advance. ", "The problem is that neither is easily fixable. For backward compatibility reason we can't change the way the qa pipeline works out of the blue. We are trying to think of having some \"pipeline configs\" to be able to switch behaviors in pre or post processing but it won't be added overnight.\r\n\r\nOverall this is going to take some time to solve.", "@sgugger thank you for the acknowledgement and clarification that there indeed is a difference. I was not at least dreaming stuff up. \r\n\r\nHere is what I ended up doing: I gutted out `run_qa.py` and created a custom inference code using the evaluation functions in that script. Now my model answers match obviously but I ended up with a different problem. When I pass **a single** query/context pair to my custom question answerer, the latency increased by about 10x. I profiled my code and lo and behold the problem is in the data prep part that is written with `datasets` and seems like is optimized for batch processing during training, meaning there is a cost incurred in the upstream processing before the prediction.\r\n\r\nSo now I am caught between a rock and a hard place as the model does not produce the same answers if I use the `pipeline` and I have a latency issue for single inference with the `run_qa.py` code.. Is it possible to summarize the problem with `pipeline` so that maybe I can write some code to patch that configurable part into it? That would go a long way. Merci beaucoup!\r\n\r\nPS: I also noted 2 other main differences between these two inference codes. \r\n\r\n1) I don't see where we can incorporate a null threshold in the `pipeline` function. I actually benefit up to 2-3 points for F1 on my custom dataset with a threshold. \r\n2) The pipeline can take multiple query/context pairs but it loops through them one at a time, so if I pass a batch of 16, I get 16x latency. \r\n\r\nI am thinking about entering separate issues as I can't see any open issues regarding them. Please let me know if I have overlooked or you all have incorporated some of this recently. ", "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 @sgugger @oersoy1,\r\n\r\nI am facing the same issue. The pipeline predictions are not the same as run_qa.py predictions. Please let me know if there is any update or work around for the same. " ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Are you sure `wandb` is installed in the container you are using for training? There can't be any report if it's not installed and initialized.", "Yup, it's installed. Here is the requirements file that goes into the container\r\n\r\n```\r\nboto3==1.16.32\r\npeewee==3.13.3\r\npandas==1.0.5\r\ntorch==1.6.0\r\nnumpy==1.18.2\r\ntransformers==4.3.0\r\nwandb==0.10.20\r\nbotocore==1.19.37\r\nsrsly==1.0.2\r\npsycopg2_binary==2.8.5\r\nscikit_learn==0.23.2\r\nuvloop==0.14.0\r\n./prodigy-1.10.5-cp36.cp37.cp38-cp36m.cp37m.cp38-linux_x86_64.whl\r\nhttps://github.com/explosion/spacy-models/releases/download/en_core_web_md-2.3.1/en_core_web_md-2.3.1.tar.gz\r\n```\r\nI also upload my API key from a separate file. \r\n\r\nAlso, if I manually call \r\n`wandb.log({\"Some metric\":\"This will appear\"})`\r\n\r\nThen it shows up in the run's dashboard. The only information that doesn't show up is the automatic logging that the Trainer is supposed to do. ", "Could you share the training script you are using?", "I can't share the full script, but it's something like \r\n\r\n```\r\nimport wandb\r\nfrom transformers import (\r\n AutoConfig,\r\n AutoModelForSequenceClassification,\r\n HfArgumentParser,\r\n Trainer,\r\n TrainingArguments,\r\n)\r\n\r\nclass ModelArguments:\r\n#Declare arguments here\r\n\r\nclass DataTrainingArguments:\r\n#Declare more arguments\r\n\r\nparser = HfArgumentParser((ModelArguments, DataTrainingArguments))\r\nmodel_args, data_args, training_args = parser.parse_args_into_dataclasses()\r\n\r\n#Assuming I have the config already\r\nmodel = AutoModelForSequenceClassification.from_pretrained(\r\n model_args.model_name_or_path,\r\n config=config\r\n )\r\n\r\n#train_data, eval_data = get data from a remote host\r\n\r\ntrainer = Trainer(\r\n model=model,\r\n args=training_args,\r\n train_dataset=train_data,\r\n eval_dataset=eval_data,\r\n )\r\n\r\ntrainer.train()\r\n\r\nwandb.log({\"Fake Metric\":\"This will show up\"})\r\n```\r\nWhen I run on SageMaker, I can see from the console logs and the post-training evaluation that it has been training. \r\n\r\n", "It would be interesting to have the result of \r\n```\r\nfrom transformers.integrations import get_available_reporting_integrations\r\nprint(get_available_reporting_integrations())\r\n```\r\nto check whether or not `wandb` is listed.\r\n\r\nAnother thing you can do is force the reporting to wandb with adding `report_to = [\"wand\"]` in your hyperparameters.", "Got this from the integrations print\r\n`['wandb']`\r\nAlso, tried to force the reporting to WandB. It's a bit tricky to pass lists into the SageMaker Estimator as a hyperparam, so I just tried to force the reporting in my script\r\n\r\n```\r\n parser = HfArgumentParser((ModelArguments, DataTrainingArguments))\r\n model_args, data_args, training_args = parser.parse_args_into_dataclasses()\r\n training_args.report_to = [\"wandb\"]\r\n data_args.report_to = [\"wandb\"]\r\n```\r\nStill getting the same result", "Curiouser and curiouser. So it is detected an added but doesn't log anything. @borisdayma do you have some idea of why theWandbCallback would not log anything in a training launched on SageMaker?", "I'm not sure but now I'm feeling the curiousest!\r\nThere is actually some W&B documentation with specific setup instructions for Sagemaker: see [here](https://docs.wandb.ai/integrations/sagemaker)", "I'm feeling pretty curious too...already followed the instructions with the secrents.env file. ", "Could you add somewhere in the methods of [`WandbCallback`](https://github.com/huggingface/transformers/blob/395ffcd757103ed2ccc888e48d90fd2ccb4d506f/src/transformers/integrations.py#L510) some `print` statements just to see if they are being called at all?\r\n\r\nYou could just copy that file in your local folder and import from it.", "I cloned the repo, added the` print `statements, imported the `WandbCallback` and added it to the `Trainer` callbacks.\r\n\r\nI am seeing some `print` statements, but...no logs to WandB dashboard :( ", "That's progress!\r\nCan you add some maybe just after the calls to `wandb.log` and maybe also print `wandb.run.id`?", "Got some interesting results.\r\n\r\nI added\r\n\r\n```\r\n def on_train_begin(self, args, state, control, model=None, **kwargs):\r\n print(\"TESTING WANDB: BEGINNING TRAINING\")\r\n wandb.log({\"Begining training\": \"Please appear\"})\r\n ...\r\n```\r\nand ran it locally (in the CLI on a Notebook instance). I got:\r\n\r\n`NameError: name 'wandb' is not defined`\r\n\r\nIf I run this code in the callbacks init function:\r\n\r\n\r\n```\r\n def __init__(self):\r\n print(\"TESTING WANDB: INITIALIZING CALLBACK\")\r\n has_wandb = is_wandb_available()\r\n assert has_wandb, \"WandbCallback requires wandb to be installed. Run `pip install wandb`.\"\r\n if not has_wandb:\r\n print(\"NO WEIGHTS AND BIASES!!!!!!!!\")\r\n if has_wandb:\r\n import wandb\r\n\r\n wandb.ensure_configured()\r\n if wandb.api.api_key is None:\r\n has_wandb = False\r\n logger.warning(\r\n \"W&B installed but not logged in. Run `wandb login` or set the WANDB_API_KEY env variable.\"\r\n )\r\n self._wandb = None\r\n else:\r\n self._wandb = wandb\r\n self._initialized = False\r\n # log outputs\r\n self._log_model = os.getenv(\"WANDB_LOG_MODEL\", \"FALSE\").upper() in ENV_VARS_TRUE_VALUES.union({\"TRUE\"})\r\n```\r\n`\"NO WEIGHTS AND BIASES!!!!!!!!\"` does not print.\r\n\r\nIf `wandb` gets imported, shouldn't the log statements work? Not sure if this is a separate issue from the SageMaker one. ", "Very interesting!\r\nWhat if you ignore the second check and just set `self._wandb = wandb`", "Same error! ", "Does [this example](https://github.com/wandb/examples/tree/master/examples/pytorch/pytorch-cifar10-sagemaker) work for you?", "Yup", "Maybe can you try to log directly something to even before using the `Trainer`?\r\n\r\n```python\r\nimport wandb\r\nwandb.init()\r\nwandb.log({'val': 1})\r\n```\r\n\r\nThat way we will see if it comes from your wandb setup in Sagemaker or HF.", "I do already do that actually, and it does appear! I successfully log a couple of charts before and after training. ", "In your code where you set `self._wandb`, can you try to do something like `self._wandb.log({'test_value': 1})`", "This works! \r\n\r\nI have regular training logs and the test_value logged when the callback's int function is like this:\r\n\r\n```\r\n def __init__(self):\r\n print(\"TESTING WANDB: INITIALIZING CALLBACK\")\r\n has_wandb = is_wandb_available()\r\n assert has_wandb, \"WandbCallback requires wandb to be installed. Run `pip install wandb`.\"\r\n if has_wandb:\r\n import wandb\r\n\r\n wandb.ensure_configured()\r\n '''\r\n if wandb.api.api_key is None:\r\n has_wandb = False\r\n logger.warning(\r\n \"W&B installed but not logged in. Run `wandb login` or set the WANDB_API_KEY env variable.\"\r\n )\r\n self._wandb = None\r\n else:\r\n '''\r\n self._wandb = wandb\r\n self._wandb.log({'test_value': 1})\r\n self._initialized = False\r\n # log outputs\r\n self._log_model = os.getenv(\"WANDB_LOG_MODEL\", \"FALSE\").upper() in ENV_VARS_TRUE_VALUES.union({\"TRUE\"})\r\n```\r\n\r\nHowever, if I do not pass in the modified version of WandbCallback the problem resumes. ", "This is very strange...\r\nSo just adding the `log` statement makes it work and it stops working if you remove it?", "It looks like what makes the difference is commenting out this portion\r\n\r\n```\r\n '''\r\n if wandb.api.api_key is None:\r\n has_wandb = False\r\n logger.warning(\r\n \"W&B installed but not logged in. Run `wandb login` or set the WANDB_API_KEY env variable.\"\r\n )\r\n self._wandb = None\r\n else:\r\n '''\r\n```\r\nAnd then passing the modified callback in explicitly in the Trainer. \r\n\r\nI believe that the NameError above was because I was calling `wandb.log`, instead of `self._wandb.log`", "This should now be fixed on the master branch of `transformers`. Let me know if you still have any issue @alexf-a ", "It is still happening! It looks like the WandB API key is not getting properly loaded from the secrets.env. If I run this code, it works\r\n\r\n``` \r\n with open(\"./secrets.env\", \"r\") as secrets_f:\r\n wandb_api_key = secrets_f.read().replace('\\n', '').split(\"=\")[1]\r\n os.environ[\"WANDB_API_KEY\"] = wandb_api_key \r\n\r\n```", "Hey @alexf-a we'll work on a patch in the next release that handles the sagemaker case. Until then if you just add this code before you instantiate your trainer it should work.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "I think you should take a look at `prefix_allowed_tokens_fn` on the [`generate` method](https://huggingface.co/transformers/main_classes/model.html?highlight=generate#transformers.generation_utils.GenerationMixin.generate). \r\n\r\nIt let's you create a function to constrain the generation based on previously generated tokens. It is inspired by this paper: [Autoregressive Entity Retrieval](https://arxiv.org/abs/2010.00904).", "thanks! Is there a way to have more generic callbacks that can possibly access the beams (for instance, if I want to do custom scoring/tracking/modification to the entries in the beam)?\r\n\r\nIf not, would that be a useful PR to submit?", "So you mean to access the different beams so that generations is conditioned to not just the current beam so far but also the others? \r\nYou probably should look at transformers.LogitsProcessor in:\r\nhttps://github.com/huggingface/transformers/blob/1750e629006bb6989aef5b4e141f3477f891a098/src/transformers/generation_utils.py#L555-L629\r\n\r\n which deals with the constraints and scoring of tokens at generation. Perhaps what you described could be introduced in a similar fashion as `prefix_allowed_tokens_fn`.\r\n\r\nRegarding a PR I am not the best to say, I would first make sure if what you aim for can be done within the existing functionality. ", "I did take a look at `LogitsProcessor` and might be able to work with that - thanks!", "@kailashkarthiks Hi, I am also interested in performing constrained decoding with seq2seq language models like BART. I was wondering if you made any progress on this, and if yes, would you be able to share your findings or code? Thank you!" ]

[]

[ { "login": "ydshieh", "id": 2521628, "node_id": "MDQ6VXNlcjI1MjE2Mjg=", "avatar_url": "https://avatars.githubusercontent.com/u/2521628?v=4", "gravatar_id": "", "url": "https://api.github.com/users/ydshieh", "html_url": "https://github.com/ydshieh", "followers_url": "https://api.github.com/users/ydshieh/followers", "following_url": "https://api.github.com/users/ydshieh/following{/other_user}", "gists_url": "https://api.github.com/users/ydshieh/gists{/gist_id}", "starred_url": "https://api.github.com/users/ydshieh/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/ydshieh/subscriptions", "organizations_url": "https://api.github.com/users/ydshieh/orgs", "repos_url": "https://api.github.com/users/ydshieh/repos", "events_url": "https://api.github.com/users/ydshieh/events{/privacy}", "received_events_url": "https://api.github.com/users/ydshieh/received_events", "type": "User", "site_admin": false } ]

[ "Uploaded full dataset and trained model: https://drive.google.com/drive/u/1/folders/1A7PIG1E98uuGUi8mDA2m_6T_oQp8XDhF\r\n\r\nYou can reproduce the issue by simply evaluating the test set using the trained model and observe the behavior with the aforementioned sets of decoder input ids. I suspect the issue is the same during the training process (which makes it converge to zero). I don't think I am doing anything wrong in the code, but please let me know. Thanks!", "Hey @Slash0BZ,\r\n\r\nHmm, this might actually be very difficult to debug since 2.11 is quite outdated by now :-/.\r\n\r\n2 things:\r\n\r\n1) I'm very confident that in the decoder the causal mask is always enabled, so that tokens have **no** access to future tokens -> they should not be able to learn to \"cheat\". See this line (in 2.11 version): https://github.com/huggingface/transformers/blob/b42586ea560a20dcadb78472a6b4596f579e9043/src/transformers/modeling_t5.py#L707 if you follow the function definition you see that a causal mask is generated if the model is a decoder `self.config.is_decoder is True` - see: https://github.com/huggingface/transformers/blob/b42586ea560a20dcadb78472a6b4596f579e9043/src/transformers/modeling_utils.py#L192\r\n\r\n2) There was a bug in the relative positional encoding that was fixed in this PR: https://github.com/huggingface/transformers/pull/8518 . In this PR I also made sure that the original T5 and our T5 implementation give the exact same results.", "Hi @patrickvonplaten, thank you for the quick response! Sorry about the version issue, 2.11.0 was the latest when I conducted all experiments for a paper under review.\r\n\r\nI understand how the causal mask is created, and I can confirm it is working, but it cannot explain what I see. Below is what I did (recap: 2024 and 2062 are two vocab ids I used for the binary tag, 1525 and 10 represent \"answer:\")\r\n\r\nwith `decoder_input_ids = [0, 1525, 10, 2062]`, inside the decoder (T5Stack), I printed input_ids, which is of size [16, 31] and of content `[0, 1525, 10, 2062, 0, ... 0]`. The extended_attention_mask is of size [16, 1, 31, 31] and of content (at position 2) `[0, 0, 0, -10000, -10000, ... -10000]`. Is everything here behaving as expected (i.e., should the first few masks be 0?) Under this, the prediction of an instance using a trained model at position 2 is 2062. \r\n\r\nHowever, if I change the decoder_input_ids to `[0, 1525, 10, 2024]` (different binary vocab), the **same** model's prediction on the **same** instance at position 2 becomes 2024, showing that it sees what the input is at position 3, or at least it changed with different position 3 inputs. \r\n\r\nBelow is how I got the prediction at position 2, using the lm_logits directly from the forward() function in a T5ForConditionalGeneration. Please let me know if you spot any issues with it.\r\n\r\n outputs = model(\r\n input_ids=inputs['input_ids'],\r\n attention_mask=inputs['attention_mask'],\r\n decoder_input_ids=inputs['decoder_input_ids'],\r\n # lm_labels=inputs['lm_labels'],\r\n decoder_attention_mask=inputs['decoder_attention_mask'],\r\n use_cache=False,\r\n )[0].cpu().numpy()\r\n ids = []\r\n for output in outputs:\r\n arr = []\r\n binary_tags = [2024, 2062]\r\n for val in binary_tags:\r\n arr.append(output[2][val])\r\n argmax_idx = int(np.argmax(np.array(arr)))\r\nThanks again for your help. I understand how difficult it is to look at previous versions, but I need to figure out if all experiments need to be re-done. ", "Hmm, the `extended_attention_mask` looks correct to me. Position 2 is allowed to attend to itself and to position 0 & 1.\r\n\r\nAlso, I ran the following code snippet both on current master and on 2.11 and it passes -> showing that the attention mask works correctly:\r\n\r\n```python\r\nfrom transformers import T5ForConditionalGeneration\r\nimport torch\r\n\r\nmodel = T5ForConditionalGeneration.from_pretrained('t5-small')\r\n\r\ninput_ids = torch.tensor([list(range(30))], dtype=torch.long)\r\ndecoder_input_ids = torch.ones((1, 4), dtype=torch.long)\r\n\r\n\r\n# take output at position 2\r\nlogits_at_2 = model(input_ids, decoder_input_ids=decoder_input_ids)[0][:, 2]\r\n\r\ndecoder_input_ids[:, 3] = 10\r\n\r\n# take output at position 2 having changed the decoder_input_ids\r\nlogits_at_2_same = model(input_ids, decoder_input_ids=decoder_input_ids)[0][:, 2]\r\n\r\nassert abs(logits_at_2.sum().item() - logits_at_2_same.sum().item()) < 1e-3, \"Error\"\r\n```", "Thanks, @patrickvonplaten . Following your snippet, this is how you can reproduce my issue (please give it a try, it has been bugging me for weeks): \r\n\r\n```\r\nfrom transformers import T5ForConditionalGeneration\r\nimport torch\r\nimport numpy as np\r\n\r\nmodel = T5ForConditionalGeneration.from_pretrained(\"trained_model\")\r\n\r\ninput_ids = torch.tensor([list(range(30))], dtype=torch.long)\r\ndecoder_input_ids = torch.tensor([[0, 1525, 10, 2024]])\r\n\r\nlogits_at_2 = model(input_ids, decoder_input_ids=decoder_input_ids)[0][:, 2]\r\nprint(np.argmax(logits_at_2[0].detach().cpu().numpy()))\r\n\r\ndecoder_input_ids = torch.tensor([[0, 1525, 10, 2062]])\r\n\r\nlogits_at_2_same = model(input_ids, decoder_input_ids=decoder_input_ids)[0][:, 2]\r\nprint(np.argmax(logits_at_2_same[0].detach().cpu().numpy()))\r\n\r\nassert abs(logits_at_2.sum().item() - logits_at_2_same.sum().item()) < 1e-3, \"Error\"\r\n```\r\n\r\nWhere trained_model can be downloaded here: https://drive.google.com/drive/u/1/folders/1A7PIG1E98uuGUi8mDA2m_6T_oQp8XDhF\r\nIt has the same config as a T5-large, just different learned weights.\r\n\r\nUnder my local env (2.11.0), it prints 2024 and 2062, and triggers the assertion error.\r\n\r\nGiven this, it seems that something corrupted during the training process, and somehow the learned weights let the model to look at future inputs. Do you have any suggestions?", "@patrickvonplaten I just tried the latest Huggingface version with the snippet above using my trained model, and it also triggers the assertion error. Seems like something interesting is going on with certain model weights. ", "Interesting, so our causal mask actually doesn't fully force \"attending-to-next-tokens\" to be impossible -> it just gives it a very large negative number before softmax (-10000) so that after softmax this value should be zero. Maybe your model has incredibly high activations that can somehow overturn the -10000. Could you maybe try the following:\r\n\r\n- Replace our setting of -10000 with `-float(\"inf\")`. Then the script above should definitely not yield an assertion error anymore", "Thanks @patrickvonplaten . I tried what you said, -float(\"inf\") doesn't work because it makes the logits \"NaN\". So I tried -999999 and the predictions are now valid. Now that we know what the issue is, here are some of my concerns:\r\n\r\n- Does this only affect some \"extreme\" experiments or it affects all experiments more or less? i.e., is that attention value fine as long as it stays below zero, or it's later used continuously?\r\n- In my personal view, something has to be done more than this method to create the masks. I found this issue initially when doing a valid experiment studying numerical relations, and the model easily found this \"cheating\" way out within the first 1k steps. I suspect this issue might have affected people's experiments such as studying scrambled word orders, adding noises, etc., just without them knowing. This applies to both pre-trained Google T5 and random-initialized T5 (tried both). \r\n\r\nPlease let me know. Thanks again for your help!", "Thanks for trying it out! Hmm, yeah I've never heard of such an issue before, so I assume that it will only affect the \"extreme\" experiments. But T5 tends to have very extreme values, which is also why we (so far) managed to run T5 only partly in fp16 mode. \r\n\r\nWe usually like to use -10000 as the masking value because it makes the model `fp16` compatible...Not really sure what to do here -> we could change the masking values to `-inf` in general if such errors occur more often. Also pinging @LysandreJik @sgugger @patil-suraj here. Have you guys heard of a case before where the model learned to cheat the `-10000` masking value? ", "We use https://github.com/allenai/allennlp/blob/f091cb9cd92e767f55659b2b59f0ffb75bc613be/allennlp/nn/util.py#L239, which ultimately boils down to using this value: `torch.finfo(tensor.dtype).min`.", "@patrickvonplaten, yes, the `-10000` can totally cheat the value. We've seen that in the past in cases where the output values are passed through an argmax while the probability distribution is very uniform.\r\n\r\nWe've kept `-10000` to stay as close as possible to the original BERT implementation, and we recommend to use as few padding tokens as possible for this not to have an effect (while keeping in mind that the -10000 should keep values very very small and should have a minimal impact).\r\n\r\n@dirkgr's solution is definitely more robust and I don't think switching the -10000 value to be lower would change anyone's workflow, so I wouldn't be opposed to switching.", "@patrickvonplaten I never faced this issue in my T5 experiments but it does seem possible that -10000 can cause some issues because while investigating the fp16 issue we have seen that T5 produces large activation values.\r\n\r\nAnd I agree with @dirkgr solution.", "Hi @patil-suraj @patrickvonplaten @sgugger I am experiencing similar issues with mt5, and I am getting nan always with fp16 mode, you mentioned you partly made T5 work with fp16, do you mind telling me how you managed it? I am having really a hard time with mT5 model + fp16 thanks a lot all ", "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.", "Putting this on my ToDo-List as it seems to be quite important actually", "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": 1936351150, "node_id": "MDU6TGFiZWwxOTM2MzUxMTUw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Examples", "name": "Examples", "color": "d4c5f9", "default": false, "description": "Which is related to examples in general" } ]

[ { "login": "bhadreshpsavani", "id": 26653468, "node_id": "MDQ6VXNlcjI2NjUzNDY4", "avatar_url": "https://avatars.githubusercontent.com/u/26653468?v=4", "gravatar_id": "", "url": "https://api.github.com/users/bhadreshpsavani", "html_url": "https://github.com/bhadreshpsavani", "followers_url": "https://api.github.com/users/bhadreshpsavani/followers", "following_url": "https://api.github.com/users/bhadreshpsavani/following{/other_user}", "gists_url": "https://api.github.com/users/bhadreshpsavani/gists{/gist_id}", "starred_url": "https://api.github.com/users/bhadreshpsavani/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/bhadreshpsavani/subscriptions", "organizations_url": "https://api.github.com/users/bhadreshpsavani/orgs", "repos_url": "https://api.github.com/users/bhadreshpsavani/repos", "events_url": "https://api.github.com/users/bhadreshpsavani/events{/privacy}", "received_events_url": "https://api.github.com/users/bhadreshpsavani/received_events", "type": "User", "site_admin": false } ]

[ "I think we should have it on all scripts except the language-modeling ones -> it doesn't make much sense there.", "@bhadreshpsavani, please feel free to make this part of your project or not. Please do not feel obliged as you can see a small need quickly expands into a much bigger one.\r\n\r\nIf not, then completing https://github.com/huggingface/transformers/issues/10437 would be a fantastic contribution. And I will then label it to be open to anybody else who would like to contribute.\r\n\r\nThank you!", "Sure @stas00,\r\nI will take this. Actually, I would love to contribute more. I really enjoy contributing to this community.", "Excellent! If you run into any puzzles please don't hesitate to ask.\r\n\r\nThank you!", "Hi @stas00,\r\nIn the QA examples for the test dataset, should we keep The preprocessing same as evaluate dataset preprocessing?\r\nDo we need to apply any additional post-processing as well at the end?", "Yes, that would be needed.", "Hi, @stas00 and @sgugger, \r\n\r\nAdding predict function for the `run_qa` example is slightly complicated. \r\nIn the eval section itself, we are generating two files `predictions.json` and `nbest_predictions.json` using `postprocess_qa_predictions` \r\n from `utils_qa`. In Predict function also the same file will be generated and override the same files which will not be very good behavior.\r\n\r\nIn the predict function currently `trainer.predict()` calculates metrics but metrics calculation might not require in case of `predict()` right? \r\n\r\nThis issue might take a longer time for me to complete may be few weeks, is it fine?", "> In Predict function also the same file will be generated and override the same files which will not be very good behavior.\r\n\r\nThis is simple: add `eval_` prefix for eval outputs and `test_` for predict.\r\n\r\nrelated: eventually we will rename the latter prefix, but for now this is the convention used in all examples. Please see: https://github.com/huggingface/transformers/issues/10165\r\n\r\n> In the predict function currently trainer.predict() calculates metrics but metrics calculation might not require in case of predict() right?\r\n\r\nWe always want the metrics for each stage, since now we report speed and memory usage, so for the quality metrics see what makes sense to report there - typically similar to eval.\r\n\r\n> This issue might take a longer time for me to complete may be few weeks, is it fine?\r\n\r\nYes, of course, thank you for giving us heads up and the clarity of your needs, @bhadreshpsavani ", "Hi @stas00,\r\n\r\nWhen I run the evaluation on [example for question answering](https://github.com/bhadreshpsavani/UnderstandingNLP/blob/master/Check_Pretrained_Model_on_SQUAD2.ipynb),\r\n\r\nI was getting below error,\r\n```\r\nTraceback (most recent call last):\r\n File \"./transformers/examples/question-answering/run_qa.py\", line 546, in <module>\r\n main()\r\n File \"./transformers/examples/question-answering/run_qa.py\", line 531, in main\r\n metrics = trainer.evaluate()\r\n File \"/content/transformers/examples/question-answering/trainer_qa.py\", line 63, in evaluate\r\n metrics = self.compute_metrics(eval_preds)\r\n File \"./transformers/examples/question-answering/run_qa.py\", line 492, in compute_metrics\r\n return metric.compute(predictions=p.predictions, references=p.label_ids)\r\n File \"/usr/local/lib/python3.7/dist-packages/datasets/metric.py\", line 403, in compute\r\n output = self._compute(predictions=predictions, references=references, **kwargs)\r\n File \"/root/.cache/huggingface/modules/datasets_modules/metrics/squad/c0855591f1a2c2af8b7949e3146b9c86a6b7f536b4154019b03472639d310181/squad.py\", line 109, in _compute\r\n score = evaluate(dataset=dataset, predictions=pred_dict)\r\n File \"/root/.cache/huggingface/modules/datasets_modules/metrics/squad/c0855591f1a2c2af8b7949e3146b9c86a6b7f536b4154019b03472639d310181/evaluate.py\", line 68, in evaluate\r\n exact_match += metric_max_over_ground_truths(exact_match_score, prediction, ground_truths)\r\n File \"/root/.cache/huggingface/modules/datasets_modules/metrics/squad/c0855591f1a2c2af8b7949e3146b9c86a6b7f536b4154019b03472639d310181/evaluate.py\", line 53, in metric_max_over_ground_truths\r\n return max(scores_for_ground_truths)\r\nValueError: max() arg is an empty sequence\r\n```\r\nWhile calculating metrics at the end,\r\nI could not found anything to resolve this!\r\n\r\n\r\n", "I am able to reproduce it - will have a look shortly. Thank you for the notebook - makes it super-easy to reproduce!\r\n\r\nBTW, I recommend you use `--max_val_samples 10` or similar to make your testing much faster ;)", "OK, so we have a case of a broken dataset and the metrics evaluation code that doesn't check if the input data is valid.\r\n\r\nWhile ideally the eval code should be robust against an occasional corrupted input, the dataset should not have broken entries, observe the kind of data it evaluates against (`ground_truths`) \r\n\r\n```\r\n[{'answers': {'answer_start': [159, 159, 159, 159], 'text': ['France', 'France', 'France', 'France']}, 'id': '56ddde6b9a695914005b9628'},\r\n{'answers': {'answer_start': [94, 87, 94, 94], 'text': ['10th and 11th centuries', 'in the 10th and 11th centuries', '10th and 11th centuries', '10th and 11th centuries']}, 'id': '56ddde6b9a695914005b9629'},\r\n\r\n {'answers': {'answer_start': [], 'text': []}, 'id': '5ad39d53604f3c001a3fe8d1'},\r\n {'answers': {'answer_start': [], 'text': []}, 'id': '5ad39d53604f3c001a3fe8d2'},\r\n {'answers': {'answer_start': [], 'text': []}, 'id': '5ad39d53604f3c001a3fe8d3'},\r\n {'answers': {'answer_start': [], 'text': []}, 'id': '5ad39d53604f3c001a3fe8d4'}, \r\n```\r\n- and it has a ton of those - so this is a big problem on the dataset-level.\r\n\r\nHaving a quick look at the viewer:\r\nhttps://huggingface.co/datasets/viewer/?dataset=squad_v2\r\nI don't immediately see missing answers (scroll horizontally to the right to see it), but it doesn't mean anything.\r\n\r\nWhich most likely means that either the dataset conversion from the original to `datasets` format is borked, or there is some odd bug in the dataloader. But most likely it's the former. that is I'd debug this first and validate that it generates the answers correctly:\r\nhttps://github.com/huggingface/datasets/blob/b51cb81e736b86103089a584daa6e43db3c88bb5/datasets/squad_v2/squad_v2.py#L101\r\n\r\n\r\nif so then proceed to where they are loaded in the script and debug there, and so on - until you find a place where some of the records disappear in what it appears more than half of samples. There you will find the bug.\r\n\r\nPlease let me know if you'd like to try to investigate this and whether my suggestion at how to potentially approach this is clear. If it sounds too complicated please don't hesitate to say so and we will find another way to resolve this. Either way works.\r\n\r\nMeanwhile you can also try with a different model/dataset pair and see if it works there, which would also help isolate the problem. (if another dataset works then we know for sure the issue lies with `squad_v2`.", "Hi @stas00,\r\nFor `squad_v1` = `squad` it is working fine.\r\n\r\nBut when i run the following script,\r\n```\r\n!python ./transformers/examples/question-answering/run_qa.py \\\r\n--model_name_or_path distilbert-base-uncased \\\r\n--train_file ./transformers/tests/fixtures/tests_samples/SQUAD/sample.json \\\r\n--validation_file ./transformers/tests/fixtures/tests_samples/SQUAD/sample.json \\\r\n--do_eval \\\r\n--max_val_sample 10 \\\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\nwhich again data, I got the same error\r\n\r\nIn the earlier error logs it is giving like this for this issue,\r\n```\r\nFile \"/root/.cache/huggingface/modules/datasets_modules/metrics/squad/c0855591f1a2c2af8b7949e3146b9c86a6b7f536b4154019b03472639d310181/squad.py\", line 109, in _compute\r\n```\r\nBut for the squad_v2 dataset, it should be like this, \r\n```\r\n\"/root/.cache/huggingface/modules/datasets_modules/metrics/squad_v2/c0855591f1a2c2af8b7949e3146b9c86a6b7f536b4154019b03472639d310181/squad_v2.py\", line 109, in _compute\r\n```\r\ni mean it is picking wrong metrics somehow", "OK, I found this in `run_qa.py`:\r\n```\r\n version_2_with_negative: bool = field(\r\n default=False, metadata={\"help\": \"If true, some of the examples do not have an answer.\"}\r\n )\r\n```\r\n\r\nThat is add: `--version_2_with_negative` to the cl args and it does:\r\n```\r\n metric = load_metric(\"squad_v2\" if data_args.version_2_with_negative else \"squad\")\r\n```\r\n\r\nExcept it's broken too:\r\n```\r\nexamples/question-answering/run_qa.py --model_name_or_path ktrapeznikov/albert-xlarge-v2-squad-v2 --dataset_name squad_v2 --do_eval --per_device_train_batch_size 12 --learning_rate 3e-5 --num_train_epochs 2 --max_seq_length 384 --doc_stride 128 --output_dir /tmp/debug_squad/ --max_val_samples 10 --version_2_with_negative\r\nTraceback (most recent call last): | 0/11873 [00:00<?, ?it/s]\r\n File \"examples/question-answering/run_qa.py\", line 553, in <module>\r\n main()\r\n File \"examples/question-answering/run_qa.py\", line 538, in main\r\n metrics = trainer.evaluate()\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/examples/question-answering/trainer_qa.py\", line 62, in evaluate\r\n eval_preds = self.post_process_function(eval_examples, eval_dataset, output.predictions)\r\n File \"examples/question-answering/run_qa.py\", line 475, in post_processing_function\r\n predictions = postprocess_qa_predictions(\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/examples/question-answering/utils_qa.py\", line 159, in postprocess_qa_predictions\r\n null_score = min_null_prediction[\"score\"]\r\nTypeError: 'NoneType' object is not subscriptable\r\n```\r\nthis fails with the same error:\r\n```\r\npython ./examples/question-answering/run_qa.py --model_name_or_path distilbert-base-uncased --train_file tests/fixtures/tests_samples/SQUAD/sample.json --validation_file tests/fixtures/tests_samples/SQUAD/sample.json --do_eval --max_val_sample 10 --per_device_train_batch_size 12 --learning_rate 3e-5 --num_train_epochs 2 --max_seq_length 384 --doc_stride 128 --output_dir /tmp/debug_squad/ --version_2_with_negative\r\n```\r\n\r\nSo it looks like this feature is half-baked, and ideally we should have a test for it.\r\n\r\nLet me know if you're ok to investigate this new issue.\r\n\r\nPlus ideally `run_qa.py` should signal the user that if the dataset has missing answers it should bail and recommend using `--version_2_with_negative`.\r\n\r\n(and the choice of this flag's name could probably be improved too to be more intuitive at what it does, but that's another story)\r\n", "Sure @stas00,\r\n\r\nI can investigate this issue.\r\n\r\nOne more thing currently `run_qa.py` only supports `squad_v1` and `squad_v2` datasets completely, right? Shouldn't it support all the different question-answering datasets of the reading comprehension task? I mean preprocessing and postprocessing might be different for all the tasks but the concept is the same for all the tasks. Please correct me if I am wrong.\r\n\r\nI tried narrativeQA it was not supporting that one.", "Awesome!\r\n\r\n> One more thing currently run_qa.py only supports squad_v1 and squad_v2 datasets completely, right? Shouldn't it support all the different question-answering datasets of the reading comprehension task? I mean preprocessing and postprocessing might be different for all the tasks but the concept is the same for all the tasks. Please correct me if I am wrong.\r\n\r\nHonestly, I have no idea as I didn't write it. \r\n\r\n@sgugger, @patrickvonplaten, do we want `run_qa.py` to support more than `squad_v1` and `squad_v2` datasets? Thank you!", "Like all other examples, the script is given as just that, an example. As said in the main README under \"Why shouldn't I use Transformers?\":\r\n```\r\nWhile we strive to present as many use cases as possible, the scripts in our examples folder are just that: examples. It is expected that they won't work out-of-the box on your specific problem and that you will be required to change a few lines of code to adapt them to your needs.\r\n```\r\n\r\nSo like all other example scripts, the `run_qa` script will support any dataset that is structured the same way as the original dataset that was used with it (squad) but if the user wants the script to work on another dataset structured differently they will need to tweak it to their needs.", "Hi @stas00 and @sgugger,\r\n\r\nI figured the cause for that squad2 issue while using `max_sample_*` arguments, I fixed it locally. The cause of the error was in the below line\r\nhttps://github.com/huggingface/transformers/blob/5f19c07a704eca4db376b56f950b729dcaa73039/examples/question-answering/run_qa.py#L493\r\n\r\nIt uses entire data for reference while it should use only `max_sample_*` data only.\r\n\r\nI tried to fix it by below code,\r\n```python\r\nif training_args.do_eval:\r\n if \"validation\" not in datasets:\r\n raise ValueError(\"--do_eval requires a validation dataset\")\r\n eval_examples = datasets[\"validation\"]\r\n if data_args.max_val_samples is not None:\r\n # We will select sample from whole data\r\n eval_examples = eval_examples.select(range(data_args.max_val_samples))\r\n # Validation Feature Creation\r\n eval_dataset = eval_examples.map(\r\n prepare_validation_features,\r\n batched=True,\r\n num_proc=data_args.preprocessing_num_workers,\r\n remove_columns=column_names,\r\n load_from_cache_file=not data_args.overwrite_cache,\r\n )\r\n if data_args.max_val_samples is not None:\r\n # During Feature creation dataset samples might increase, we will select required samples again\r\n eval_dataset = eval_dataset.select(range(data_args.max_val_samples))\r\n```\r\nand I used `eval_examples ` instead `datasets[\"validation\"]` in the above-mentioned line, This is working fine for data I tested, but as we know after applying `prepare_validation_features`, `eval_dataset ` and `eval_examples` might not have the same data (length will be same but because of sliding window example and feature might not be representing the same item)\r\n\r\nFor `max_val_samples=10` these changes are working fine so I added predict/test method, It seems to be working fine.\r\n\r\nhttps://github.com/huggingface/transformers/blob/5f19c07a704eca4db376b56f950b729dcaa73039/examples/question-answering/utils_qa.py#L214-L226\r\n\r\ni modified in below lines,\r\n```python\r\n prediction_file = os.path.join(\r\n output_dir, \"predictions.json\" if prefix is None else f\"{prefix}_predictions.json\"\r\n )\r\n nbest_file = os.path.join(\r\n output_dir, \"nbest_predictions.json\" if prefix is None else f\"{prefix}_nbest_predictions.json\"\r\n )\r\n if version_2_with_negative:\r\n null_odds_file = os.path.join(\r\n output_dir, \"null_odds.json\" if prefix is None else f\"{prefix}_null_odds_{prefix}.json\"\r\n```\r\nbecause while passing prefix and running it I was getting error like `string can't have json attribute`\r\nit will save test and eval files like `test_predictions.json` and `eval_predictions.json`\r\n ", "Good job getting to the root of the issue, I hadn't thought of that when you added the `max_sample_xxx` but this task uses a subclass of the main `Trainer` that does require the original `eval_examples`.\r\n\r\nThe fix you propose appears good to me and you should definitely make a PR with it as soon as you can :-)\r\n\r\nFor the predicting stage, note that the subclass of the `Trainer` will require a test_dataset and test_examples to work (basically to interpret the predictions of the model as spans of the original texts, the `Trainer` needs the original texts). I do think adding a `--do_predict` to `run_qa` is going to be a bit complex so should be treated separately so my advise would be to:\r\n\r\n1. make a PR with the fix for evaluation in run_qa/run_qa_beam_search when `max_val_samples` is passed\r\n2. make a PR to add predict in all but run_qa/run_qa_beam_search scripts (when it makes sense of course)\r\n3. make a PR to add predict in run_qa/run_qa_beam_search\r\n\r\nLet me know if that makes sense to you and if you need any help along the way (or don't want to do one of those steps yourself).", "Really awesome, @bhadreshpsavani! Glad you were able to find out the cause!", "I gone through the traceback(stack trace) by running example with different stage combination and figured the root. Since I wrote earlier code about max_sample and the error was related I was able to find it!\r\n\r\nThanks @stas00 and @sgugger for your constant guidance, Now opensource contribution don't seems very difficult like earlier!\r\n\r\nNow I need to figure out for two more examples `run_swag.py` and `run_xlni.py`, all the other examples has predict if we ignore language modeling examples", "Hi @sgugger,\r\n\r\nFor `run_swag.py` and `run_xlni.py` changes are still remaining", "Hi @stas00,\r\n\r\nI was working to update `run_swag.py` and I got logs like for test metrics, \r\n```\r\nINFO|trainer_pt_utils.py:656] 2021-03-21 20:56:40,939 >> ***** test metrics *****\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,939 >> eval_accuracy = 1.0\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,939 >> eval_loss = 0.2582\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,939 >> eval_runtime = 4.1585\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,941 >> eval_samples_per_second = 2.405\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,941 >> test_mem_cpu_alloc_delta = 0MB\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,941 >> test_mem_cpu_peaked_delta = 0MB\r\n[INFO|trainer_pt_utils.py:661] 2021-03-21 20:56:40,941 >> test_samples = 10\r\n```\r\nI think we need to write `trainer_swag.py` because even predictions files need to be saved like `test_prediction.json` / `eval_prediction.json`.", "If I understood your correctly you suggest to write `MultipleChoiceTrainer` subclass.\r\n\r\nIf so unlike the `question-answering` folder that has 2 scripts, `multiple-choice` has only one script so one way is to add it directly in `run_swag.py` or if there is a convention then as you suggest as `trainer_mc.py`. I think follow you instinct and then we can decide at PR point whether to have it in a separate file. ", "Hi, @stas00 @sgugger,\r\n\r\nThe changes are ready for both examples, `run_xlni.py` works perfectly.\r\n\r\n`run_swag.py` has an issue after adding changes of predict,\r\n```\r\npython ./examples/multiple-choice/run_swag.py --model_name_or_path distilbert-base-uncased --do_train --do_eval --do_predict --max_train_samples 5 --max_val_samples 5 --max_test_samples 5 --learning_rate 5e-5 --num_train_epochs 3 --output_dir D:/tmp/swag_base --per_gpu_eval_batch_size=16 --per_device_train_batch_size=16 --overwrite_output\r\n```\r\nit gives the following error while prediction,\r\n```\r\nTraceback (most recent call last):\r\n File \"./examples/multiple-choice/run_swag.py\", line 481, in <module>\r\n main()\r\n File \"./examples/multiple-choice/run_swag.py\", line 466, in main\r\n predictions, labels, metrics = trainer.predict(test_dataset=test_dataset)\r\n File \"d:\\transformers\\src\\transformers\\trainer.py\", line 1762, in predict\r\n output = self.prediction_loop(\r\n File \"d:\\transformers\\src\\transformers\\trainer.py\", line 1829, in prediction_loop\r\n loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)\r\n File \"d:\\transformers\\src\\transformers\\trainer.py\", line 1943, in prediction_step\r\n loss, outputs = self.compute_loss(model, inputs, return_outputs=True)\r\n File \"d:\\transformers\\src\\transformers\\trainer.py\", line 1504, in compute_loss\r\n outputs = model(**inputs)\r\n File \"C:\\Users\\Bhadr\\Miniconda3\\envs\\trans-env2\\lib\\site-packages\\torch\\nn\\modules\\module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"d:\\transformers\\src\\transformers\\models\\distilbert\\modeling_distilbert.py\", line 929, in forward\r\n loss = loss_fct(reshaped_logits, labels)\r\n File \"C:\\Users\\Bhadr\\Miniconda3\\envs\\trans-env2\\lib\\site-packages\\torch\\nn\\modules\\module.py\", line 889, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"C:\\Users\\Bhadr\\Miniconda3\\envs\\trans-env2\\lib\\site-packages\\torch\\nn\\modules\\loss.py\", line 1047, in forward\r\n return F.cross_entropy(input, target, weight=self.weight,\r\n File \"C:\\Users\\Bhadr\\Miniconda3\\envs\\trans-env2\\lib\\site-packages\\torch\\nn\\functional.py\", line 2690, in cross_entropy\r\n return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)\r\n File \"C:\\Users\\Bhadr\\Miniconda3\\envs\\trans-env2\\lib\\site-packages\\torch\\nn\\functional.py\", line 2385, in nll_loss\r\n ret = torch._C._nn.nll_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index)\r\nIndexError: Target -1 is out of bounds.\r\n```\r\nThe problem is with test dataset,\r\n```python\r\ndataset = load_dataset(\"swag\", \"regular\")\r\npd.Series(dataset['test']['label']).value_counts()\r\n\"\"\"\r\nOutput:\r\n-1 20005\r\ndtype: int64\r\n\"\"\"\r\n```\r\nIt should have label In range(0, 3) like eval dataset not -1,\r\n```python\r\npd.Series(dataset['validation']['label']).value_counts()\r\n\"\"\"\r\nOutput:\r\n2 5038\r\n1 5029\r\n3 5006\r\n0 4933\r\ndtype: int64\r\n\"\"\"\r\n```\r\nI am not sure how can we fix this since it's with the dataset and this example is for only swag. when we pass the sample data present in `tests/fixtures/tests_samples/swag/sample.json` it works fine since it has label 0.\r\n\r\nThere is another small issue that needs to be fixed as well in the `trainer.py`\r\nhttps://github.com/huggingface/transformers/blob/a8d4d6776dd8a759324d0f57c60e8a738e7977a4/src/transformers/trainer.py#L1724-L1726\r\nBecause of this current test files are being saved with eval prefix, the issue that I mentioned in my previous comment.\r\nIt should be `metric_key_prefix: str = \"test\"`.\r\n\r\nPlease correct me if I am wrong.\r\n", "I'm not sure we need the predict stage for `run_swag`, especially if the test dataset is \"broken\".", "Filing an Issue with https://github.com/huggingface/datasets/issues and meanwhile using the local test file for the \r\nREADME.md?\r\n\r\n> There is another small issue that needs to be fixed as well in the trainer.py\r\n\r\nIndeed. In the default value and also the docstring - I'd say please make a separate PR for that as it's a standalone bug.\r\n\r\nAnd then we will eventually take care of https://github.com/huggingface/transformers/issues/10165 but for now let's use `test`.", "Hi @stas00,\r\nThe original [dataset](https://github.com/rowanz/swagaf/tree/master/data) doesn't have a label in the test dataset so it's using `-1` as a label. \r\n\r\n`meanwhile using the local test file for the\r\nREADME.md` - I don't understand this part, can you please tell me a bit more about this?\r\n\r\nI will create a PR with `run_xlni.py`, Please let me know if we need to add changes for `run_swag` (On sample test data predict is working fine)", "I see, they made a real \"test\" split where the target is unknown ;)\r\n\r\n> meanwhile using the local test file for the README.md - I don't understand this part, can you please tell me a bit more about this?\r\n\r\nI meant that perhaps the example in `README.md` (and test if we add any) could use a test dataset that we know has the labels, (like tests/fixtures/tests_samples/swag/sample.json it) does it make sense? And indicate that the swag dataset's `test` split can't be used here because it doesn't have labels.\r\n\r\nThis is just an idea.\r\n\r\nAlternatively, we could create a new dataset derived from the original swag, but whose test does have the labels. So basically take train+eval splits, merge them, re-splice them into train+eval+test. But perhaps this is a much larger project and perhaps it's not worth the effort.\r\n\r\nThat's why I thought that for the exemplification feeding it a known to have labels test dataset split would be sufficient.\r\n\r\nThe other approach suggested by @sgugger is not to have the predict stage in the first place. and perhaps documenting in the script why it's missing.\r\n\r\nBottom line - if you can think of a way that works for you to make it happen, then please go for it. If not, let's just leave a comment in place of what could be the predict stage.", "Hi @stas00,\r\n\r\nI made a typo in `qa_util.py` \r\nhttps://github.com/huggingface/transformers/blob/1c06240e1b3477728129bb58e7b6c7734bb5074e/examples/question-answering/utils_qa.py#L225\r\nI forget to remove the second prefix, I will fix it.\r\n\r\n> Alternatively, we could create a new dataset derived from the original swag, but whose test does have the labels. So basically take train+eval splits, merge them, re-splice them into train+eval+test. But perhaps this is a much larger project and perhaps it's not worth the effort.\r\n\r\nI like this idea, I don't think this will take much time, It will hardly take few minutes. I worked on the dataset today at my work and it's really cool! The modifying dataset is very easy. Perhaps we can add one modified version of mrpc with your suggestion and add it to the huggingface datasets. \r\n", "@sgugger, do you think it's a reasonable approach to re-make the dataset as I suggested so that it has a test split that has labels and then we can easier support the `predict` section in this example." ]

[]

[]

[]

[]

[]

[ "Hi @voidful \r\n\r\nThe reason we need to shift labels in roberta because the `labels` start with the `decoder_start_token_id` (`pad` or `bos`),\r\nwhich are then passed directly to the decoder as `decoder_input_ids`, which is the reason we need to shift the `labels` when calculating the loss\r\n\r\nNow in BART, `decoder_input_ids` are prepared inside the model by pretending the `labels` with `decoder_start_token_id`, so we don't need to shift the labels there\r\nhttps://github.com/huggingface/transformers/blob/b013842244df7be96b8cc841491bd1e35e475e36/src/transformers/models/bart/modeling_bart.py#L1274\r\n\r\nHope this clears the difference.", "> Hi @voidful\r\n> \r\n> The reason we need to shift labels in roberta because the `labels` start with the `decoder_start_token_id` (`pad` or `bos`),\r\n> which are then passed directly to the decoder as `decoder_input_ids`, which is the reason we need to shift the `labels` when calculating the loss\r\n> \r\n> Now in BART, `decoder_input_ids` are prepared inside the model by pretending the `labels` with `decoder_start_token_id`, so we don't need to shift the labels there\r\n> \r\n> https://github.com/huggingface/transformers/blob/b013842244df7be96b8cc841491bd1e35e475e36/src/transformers/models/bart/modeling_bart.py#L1274\r\n> \r\n> Hope this clears the difference.\r\n\r\nI got your point, but it seems to be apply on BartForConditionalGeneration, \r\nhttps://github.com/huggingface/transformers/blob/b013842244df7be96b8cc841491bd1e35e475e36/src/transformers/models/bart/modeling_bart.py#L1272\r\n\r\nMaybe we can apply same strategy on BartForCausalLM:\r\nhttps://github.com/huggingface/transformers/blob/b013842244df7be96b8cc841491bd1e35e475e36/src/transformers/models/bart/modeling_bart.py#L1594", "Oh, I missed your point, great catch!\r\n\r\nFeel free to open PR to add the same strategy to `BartForCausalLM`, i.e prepare `decoder_input_ids` using the `shift_tokens_right` and pass them as the `input_ids` to the decoder.\r\n\r\ncc @patrickvonplaten ", "Hmm, I'm not 100% whether everybody is on the same page here. `BartForCausalLM` was mostly created to be used in combination with `EncoderDecoderModel` and not as a standalone model. Also, Roberta requires both `input_ids` and `labels` as an input to correctly calculate the loss - the difference is just that that `input_ids` should be equal to `labels` with the labels being shifted under-the-hood. This is not the same thing as the `shift_tokens_right` function, which fully generates the `decoder_input_ids` from the labels...\r\n\r\nI think I would be fine with changing the behavior of `BartForCausalLM` so that `labels==input_ids` can be input to the function, even if this would be a slight breaking change. It would align `BartForCausalLM` closer with `RobertaForCausalm, GPT2LMHeadModel, ...` which would then also allow `EncoderDecoderModel` to have a general `shift_tokens` function.\r\n\r\nDoes this make sense?", "`BartForCausalLM` does accept `labels==input_id`, in general, all the decoders in `EncoderDecoder` accept that and that's what we have documented, pass the same input as `labels` and `decoder_input_ids`.\r\n\r\nThe reason I suggested using `shift_tokens_right`, because BART uses `eos` as `decoder_start_token` which the `shift_tokens_right` function handles. This is different from `RobertaForCausalm, GPT2LMHeadModel ...`" ]

[]

[]

[ "~Please ask on discuss.huggingface.co so that others can see your answer!~", "Thanks for the prompt reply, @sshleifer. I wasn't aware of the discussion page. \r\nIn that case, should I close the issue since it is just a query question?\r\n\r\nThanks, \r\nNaman", "- I got confused. This is a reasonable place for this post. Sorry.\r\n- Bart fine-tuned with no weights frozen, but this can be very slow.\r\n`self.lm_head` is tied (the same parameter as) to self.encoder.embed_tokens and self.decoder.embed_tokens.\r\n- I would recommend fine-tuning on your dataset with only the encoder frozen.\r\n- `final-logits-bias` doesn't matter it's all 0s and frozen.", "Thanks for your quick response. This is super-helpful to me. \r\n\r\nI have one more question related to the training process. \r\n\r\nMy understanding is that BartModel (bart-base) was trained with two input sequences just like Google Bert (sadly, details are not given in the original paper. They only mention the difference in the objective function). On the other hand, BartForConditionalGeneration was trained (fine-tuned) for the summarization task i.e. single input and single input. \r\n\r\nIn my current task, I am trying, to do *constraint summarization using multi-document*. Ideally, I would want my model to take **two different inputs** \r\n```\r\ninput_encodings = self.tokenizer.batch_encode_plus(list(zip(example_batch['inp1'], example_batch['inp2'])), padding='max_length', truncation=True)\r\n```\r\nbut I feel this kind of fine-tuning will be harder given my smaller dataset and the change in training regime. \r\n\r\nThe other option is simply **concatenating the two documents** and passing it as one sequence \r\n```\r\n## comb_inp is a list of concatenated inputs from example_batch\r\ninput_encodings = self.tokenizer.batch_encode_plus(comb_inp, padding='max_length', truncation=True)\r\n``` \r\n\r\nIn your opinion, which one makes more sense? Your feedback and comments would be hugely appreciated in this case?\r\n\r\nP.S. - I am using your distilled-bart model for fine-tuning since it is smaller version 😇\r\n\r\nThanks, \r\nNaman\r\n", "- I would try concatenating. \r\n- I would also grid search evaluation parameters (min_length, max_length, num_beam, length_penalty). \r\n- I would evaluate a few distillbart/distill-pegasus variants before any fine-tuning to decide which to start from.\r\n", "Hey @sshleifer, \r\n\r\nI was thinking of using the default hyper-parameters used during the training as a starting point. I could find the default settings for BART model ([here](https://github.com/pytorch/fairseq/blob/master/examples/bart/README.summarization.md)) but not for distilled BART. I even looked at the paper but the values are missing over there as well. Would it be possible for you to provide me/refer me with/to the default settings that you used to train the model. \r\n\r\nThanks, \r\nNaman ", "all the scripts have been moved to the [research_projects](https://github.com/huggingface/transformers/blob/master/examples/research_projects/seq2seq-distillation/) directory.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hey @sshleifer, I was trying to fine-tune the `distill-pegasus-cnn-16-4` [model](https://huggingface.co/sshleifer/distill-pegasus-cnn-16-4) provided by you but I am not sure of the hyper-parameters and the corresponding file is also not avaialble in the [`research_projects/seq2seq_distillation`](https://github.com/huggingface/transformers/tree/master/examples/research_projects/seq2seq-distillation) directory. Could you please share the hyper-parameters that you used to train this model (and achieve the results shown in Table 5 from your [paper](https://arxiv.org/pdf/2010.13002.pdf))?\r\n\r\nThanks a lot!\r\nNaman", "To clarify, you are trying to reproduce `distill-pegasus-cnn-16-4`, rather than finetune it further?\r\n\r\n\r\n1) Make a student using make_student.py or get one from model hub.\r\n2) guessing the train command from memory.\r\nI think it should be a combination of `train_distilbart_cnn.sh` and `train_pegasus_xsum.sh`.\r\n- You definitely want `max_target_length=142, --adafactor, --freeze-embeds --freeze-encoder`. \r\n- I think fp16 will not work.\r\n\r\n\r\nSo roughly...\r\n```bash\r\npython finetune.py \\\r\n --learning_rate=1e-4 \\\r\n --do_train \\\r\n --do_predict \\\r\n --n_val 1000 \\\r\n --val_check_interval 0.25 \\\r\n --max_target_length 142 \\\r\n --freeze_embeds --label_smoothing 0.1 --adafactor \\\r\n\t--eval_beams 2 \\\r\n\t--freeze-encoder \\\r\n\t--sortish_sampler \\\r\n\t--model_name_or_path $YOUR_STUDENT \\\r\n \"$@\"\r\n```\r\n3) Copy the `config.json` for `distill-pegasus-cnn-16-4` to increase rouge score of trained student model.\r\n\r\n\r\nHope that helps, and sorry for not having a definitive answer.\r\n", "Hey @sshleifer , \r\n\r\nSorry, I think I phrased the question wrongly. I am trying to finetune the `distill-pegasus` model on my own dataset since the original pegasus model is huge and is taking a lot of time. I was simply hoping if you could provide me the hyper-parameters like you have provided for `distill-bart` models for CNN/DailyMail dataset ([here](https://github.com/huggingface/transformers/blob/master/examples/research_projects/seq2seq-distillation/train_distilbart_cnn.sh)).\r\nWould I be able to use the hyper-parameters that you provided above?\r\n\r\nThanks, \r\nNaman", "Yes", "Hello, \r\n\r\n@[sshleifer](https://github.com/sshleifer) how can I freeze only the first two layer from the encoder (I am using BART), and how can I change the dropout of some layers also in BART?\r\n\r\nThank you", "Hey @JessicaLopezEspejel , \r\n\r\n`model.get_encoder().layers` will give you a list (`torch.nn.modules.container.ModuleList` to be precise) of layers in encoder, and you can freeze the required layers using the `freeze_params` function provided in the `utils.py` file. I have included a small code snippet for your reference. Hope this helps!\r\n\r\n```\r\nfrom torch import nn\r\nfrom transformers import AutoTokenizer, AutoModel\r\n\r\ndef freeze_params(model: nn.Module):\r\n \"\"\"Set requires_grad=False for each of model.parameters()\"\"\"\r\n for par in model.parameters():\r\n par.requires_grad = False\r\n\r\nmodel = AutoModel.from_pretrained(\"facebook/bart-large\")\r\nenc_layers = model.get_encoder().layers\r\nfreeze_params(enc_layers[0]) # freeze layer 0\r\ndropout = enc_layers[0].dropout # return dropout value for layer 0\r\nenc_layers[0].dropout = 0.5 # set dropout value for layer 0\r\n```\r\n\r\nThanks, \r\nNaman\r\n\r\n", "Than you so much @bnaman50 , I will try it. " ]

[]

[]

[ "The issue was discussed at https://github.com/huggingface/transformers/pull/10552, and this is expected behaviour. If one wants to generate using `decoder_input_ids` with different lengths, the suggested approach is to use `padding_side` as `left` on the tokenizer.\r\nFor a custom solution if one wants to preserve the same results as if there was no padding, see code changes in https://github.com/huggingface/transformers/pull/10552.", "I hope it is okay, that I ask a follow up question here, because I feel it might be highly related (also to [this](https://github.com/huggingface/transformers/pull/10552#issuecomment-801246652) comment about non-identical outputs with & without padding). If not I would be happy over any pointers were to ask this question.\r\n\r\nI am using an BERT encoder-decoder model as per the example [here](https://huggingface.co/transformers/model_doc/encoderdecoder.html#encoderdecodermodel) and want to condition the decoder output on some know sequence. The known sequence is of varying length, tokenized and `left` padded, exactly as described above.\r\n\r\nTo exactly match the training regime (which never contains padding before the first token), I was wondering whether there is a way to pass an additional padding mask for the `decoder_input_ids`, e.g. a `decoder_attention_mask` of size `(batch x max_known_seq_len)`?", "Hi @l-salewski \r\n\r\nAs I mentioned in the PR comment, it would be better to group the sequence of similar length into a batch then pass them to generate otherwise call `generate` for each example.\r\n\r\nAlso it's possible to pass `decoder_attention_mask`, its shape need to be `[bath, current_seq_length]`", "Hi @patil-suraj,\r\nthank you for getting back to me! Grouping of similar length seems to be an approximate solution, but e.g. my test set is relatively small and exhibits sentences of many different lengths. Running each example individually on the other hand may be quite slow.\r\n\r\nUsing `decoder_attention_mask` is what I did in the end. I had to overwrite the `prepare_inputs_for_generation` function, such that if a `decoder_attention_mask` is passed, it overrules the generated one:\r\n```python\r\ndef prepare_inputs_for_generation(\r\n self, input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs\r\n):\r\n decoder_inputs = self.decoder.prepare_inputs_for_generation(input_ids, past=past)\r\n decoder_attention_mask = decoder_inputs[\"attention_mask\"] if \"attention_mask\" in decoder_inputs else None\r\n\r\n # if we have been passed an attention mask, use it to overrule the generated one\r\n if \"decoder_attention_mask\" in kwargs:\r\n initial_decoder_attention_mask = kwargs.pop(\"decoder_attention_mask\")\r\n initial_sequence_length = initial_decoder_attention_mask.size(1)\r\n decoder_attention_mask[:,:initial_sequence_length] = initial_decoder_attention_mask\r\n\r\n\r\n input_dict = {\r\n \"attention_mask\": attention_mask,\r\n \"decoder_attention_mask\": decoder_attention_mask,\r\n \"decoder_input_ids\": decoder_inputs[\"input_ids\"],\r\n \"encoder_outputs\": encoder_outputs,\r\n \"past_key_values\": decoder_inputs[\"past_key_values\"],\r\n \"use_cache\": use_cache,\r\n **kwargs,\r\n }\r\n return input_dict\r\n```\r\n Furthermore, I overwrite `_expand_inputs_for_generation` from the beam search such that the `decoder_attention_mask` is also expanded for each of the beams:\r\n```python\r\n@staticmethod\r\ndef _expand_inputs_for_generation(\r\n input_ids: torch.LongTensor,\r\n expand_size: int = 1,\r\n is_encoder_decoder: bool = False,\r\n attention_mask: torch.LongTensor = None,\r\n encoder_outputs: ModelOutput = None,\r\n **model_kwargs,\r\n) -> Tuple[torch.LongTensor, Dict[str, Any]]:\r\n expanded_return_idx = (\r\n torch.arange(input_ids.shape[0]).view(-1, 1).repeat(1, expand_size).view(-1).to(input_ids.device)\r\n )\r\n input_ids = input_ids.index_select(0, expanded_return_idx)\r\n\r\n if \"token_type_ids\" in model_kwargs:\r\n token_type_ids = model_kwargs[\"token_type_ids\"]\r\n model_kwargs[\"token_type_ids\"] = token_type_ids\r\n \r\n # this has been added to the original method\r\n if \"decoder_attention_mask\" in model_kwargs:\r\n model_kwargs[\"decoder_attention_mask\"] = model_kwargs[\"decoder_attention_mask\"].index_select(0, expanded_return_idx)\r\n\r\n if attention_mask is not None:\r\n model_kwargs[\"attention_mask\"] = attention_mask.index_select(0, expanded_return_idx)\r\n\r\n if is_encoder_decoder:\r\n assert encoder_outputs is not None\r\n encoder_outputs[\"last_hidden_state\"] = encoder_outputs.last_hidden_state.index_select(\r\n 0, expanded_return_idx.to(encoder_outputs.last_hidden_state.device)\r\n )\r\n model_kwargs[\"encoder_outputs\"] = encoder_outputs\r\n return input_ids, model_kwargs\r\n```\r\nTo exactly match the training setting, I tokenize the known inputs, prepend a `[CLS]` token to each and extend the `decoder_attention_mask` with a 1 column to the left such that it also attends to the `[CLS]` token:\r\n```python\r\n# Combine a CLS Column with the forced input\r\nkwargs[\"decoder_input_ids\"] = torch.cat([\r\n torch.zeros_like(forced_input.input_ids)[:,:1]+self.tokenizer.cls_token_id,\r\n forced_input.input_ids],\r\n dim=1)\r\n# Attend to the CLS column, but not the PAD tokens of the forced input\r\nkwargs[\"decoder_attention_mask\"] = torch.cat([\r\n torch.ones_like(forced_input.attention_mask)[:,:1],\r\n forced_input.attention_mask],\r\n dim=1)\r\n```\r\nThen `**kwargs` is passed to `generate`. Overall this approach works flawlessly as it reduces the overhead (e.g. no organizing batches or looping over batches) from the user perspective. I just tokenize the known inputs separately add the `[CLS]` token as described above and that is it." ]

[]

[]

[ "cc @sgugger ", "This seems like a problem in your environment install for NCCL: if your script can run on two GPUs there is nothing in the code to change to make it run on four GPUs so this is not a bug in the training script or transformers. I have never seen that particular NCCL error so I'm afraid I can't really help debugging it.", "Thanks for the quick reply. \r\nYeah, it’s strange that it works on 2 GPUs but not on 4. Will check again and let you know.", "@sgugger just to clarify:\r\nThe system has 4 GPUs. It’s only the nproc_per_node argument I’m changing (from 1 to 2,4,etc.). \r\nJust want to ensure I’ve not misunderstood the cause of the error. Right?", "Yes I understood that. The PyTorch launcher is going to spawn a different number of processes depending on the number your pass, which in turn will use the number of GPUs specified (and the others are idle).", "Thanks. Just wanted to confirm that. Will try reinstalling the environment and update if I find the solution.\r\n", "Hi @sgugger,\r\nGood news, the issue seems to have been an environment issue.\r\nThanks for the instant help", "I still meet the same problem， could you please tell me how to solve it？" ]

[]

[]

[ "Hi there, \r\n\r\nThe training scripts load the dataset using the `datasets` library, so this issue is related to the `datasets` lib, you can open it in that repo https://github.com/huggingface/datasets.", "Thank you!" ]

[]

[]

[ "Thanks a lot!" ]

[]

[]

[ "Hi there,\r\n\r\nthe `run_mlm` script expects an unlabeled text dataset, i.e a dataset with the column `text` in it, if there is no `text` column then it assumes that the first column is the text column.\r\n\r\nHere `sst2` is a classification dataset and the first column is `idx`. \r\nSo you could change the script and directly hardcode the `text_column_name`, which is `sentence` for `sst2`.\r\nhttps://github.com/huggingface/transformers/blob/b013842244df7be96b8cc841491bd1e35e475e36/examples/language-modeling/run_mlm.py#L304\r\n\r\n And also pass the `--line_by_line` argument.", "Thanks for your prompt and accurate reply.\r\nThis does help!\r\n\r\nBut I notice that the line_by_line argument is not needed as long as the `text_column_name` is hard-coded in that line.\r\nAnyways, thanks!" ]

[]

[]

[ "You've converted them successfully in PyTorch, so you should be left with a `pytorch_model.bin` alongside a `config.json`, is that right?\r\n\r\nI recommend reading the [quicktour entry related to using models](https://huggingface.co/transformers/quicktour.html#using-the-model) in order to get a sense of how one can load a model in PyTorch.", "Hi @LysandreJik, Thank you for replying on this thread. \r\n\r\n> You've converted them successfully in PyTorch, so you should be left with a pytorch_model.bin alongside a config.json, is that right? \r\n\r\nYes, I have pytorch_model.bin file but the point is: it is not single serialised file like [this](https://huggingface.co/bert-base-uncased/blob/main/pytorch_model.bin) but I can unzip that exported bin file which I have generated using [convert_bert_original_tf2_checkpoint_to_pytorch.py](https://github.com/huggingface/transformers/blob/master/src/transformers/models/bert/convert_bert_original_tf2_checkpoint_to_pytorch.py) script and I can see the three main components which I posted in snapshot earlier. Ideally, I should have single serialised pytorch_model.bin, right? \r\n\r\nI have checked the same with original BERT model and tried to convert it from TF2 to PyTorch there also I have bert-model.bin file but I can unzip that bert-model.bin and it also have three components present inside the bin. See the snapshots. \r\n\r\nSnapshot 1: My exported bin file\r\n\r\n\r\nSnapshot 2: I can unzip this bin file\r\n\r\n\r\nSnapshot 3: I can see the three components \r\n\r\n\r\nIs this expected or Am I missing something?\r\n\r\nBasically my goal here is to make sure that I'm converting the my own BERT model which is trained using TF2 to PyTorch in bug-free manner.\r\n\r\nThanks you again for sharing the documentation link I will try that and post my further questions. \r\n\r\nThanks,", "That is the way PyTorch works. See the documentation for [torch.save](https://pytorch.org/docs/stable/generated/torch.save.html).\r\n\r\nIt states:\r\n\r\n> The 1.6 release of PyTorch switched torch.save to use a new zipfile-based file format. torch.load still retains the ability to load files in the old format. If for any reason you want torch.save to use the old format, pass the kwarg _use_new_zipfile_serialization=False.", "Thank you so much @LysandreJik for your help now I have one single .bin file. Closing this issue. I will reopen if needed. " ]

[]

[]

[ "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." ]

[]

[]

[ "One more thing (maybe this is a bug): why are the datasets stuff cached in ~/.cache/huggingface/datasets when I have both the env var set up AND I specify a cache dir with --cache_dir when I run my scripts? The pretrained models go in the --cache_dir specified but the datasets don't. This is confusing (and maybe buggy?)", "I wonder if this has something to do with newer versions of the library. I remember at the last upgrade I had to install the package `datasets` separately, which I didn't have to do before. Is there a way to specify a cache for the `datasets` and it's separate than the model one?\r\n", "I'll have to look into this more, but I'm doing an educated guess that there is probably a way to send the cache_dir to the datasets api. In any case, I would like to learn how to change the cache dir globally, if at all possible. Thanks!", "I'm guessing HF_HOME is what I'm looking for... ", "Yes, that worked. Closing. " ]

[]

[]

[ "https://github.com/alexa/bort/issues/10" ]

[]

[]

[ "Ah, indeed! cc @sgugger" ]

[]

[]

[ "Hi @pranav-s \r\n\r\nWe only support the `json` and `csv` files in the examples. To convert your text files to the `json` format, you could use this `datasets` script as references, which converts the `conll` text files to the `datasets` format.\r\nhttps://github.com/huggingface/datasets/blob/master/datasets/conll2003/conll2003.py ", "Hi @pranav-s,\r\nI found this script from `test_examples.py` file,\r\n```\r\nrun_ner.py\r\n --model_name_or_path bert-base-uncased\r\n --train_file tests/fixtures/tests_samples/conll/sample.json\r\n --validation_file tests/fixtures/tests_samples/conll/sample.json\r\n --output_dir {tmp_dir}\r\n --overwrite_output_dir\r\n --do_train\r\n --do_eval\r\n --warmup_steps=2\r\n --learning_rate=2e-4\r\n --per_device_train_batch_size=2\r\n --per_device_eval_batch_size=2\r\n --num_train_epochs=2\r\n```\r\nat this path `tests/fixtures/tests_samples/conll/sample.json` you can find an example file which is having required input format", "Thank you @bhadreshpsavani and @patil-suraj . This was helpful.", "Thank you @bhadreshpsavani . You saved my life!" ]