Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[ { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "This is the intended behavior.\r\nIt is very counterintuitive to me as well, but changing `prepare_seq2seq_batch` or `shift_tokens_right` to end up with `decoder_start_token_id` at the 0th position of `decoder_input_ids` seems to lead to worse fine-tuning performance.\r\n\r\nIf you have evidence to the contrary, I would be happy to change it.", "I'm running the experiment now on wmt-en-ro, we'll see how it goes!", "Thanks! I was using the pre-trained model for scoring (src, tgt) pairs and didn't actually get a chance to check the impact on finetuning yet. ", "I ran this last night, and finetuning loss was identical with and without the change.\r\nBleu was within 0.1 (`master` was slightly higher).\r\nHere is the branch if you want to play with it: https://github.com/sshleifer/transformers_fork/tree/hack-batches-v2" ]

[]

[]

[]

[]

[]

[]

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

[ { "login": "julien-c", "id": 326577, "node_id": "MDQ6VXNlcjMyNjU3Nw==", "avatar_url": "https://avatars.githubusercontent.com/u/326577?v=4", "gravatar_id": "", "url": "https://api.github.com/users/julien-c", "html_url": "https://github.com/julien-c", "followers_url": "https://api.github.com/users/julien-c/followers", "following_url": "https://api.github.com/users/julien-c/following{/other_user}", "gists_url": "https://api.github.com/users/julien-c/gists{/gist_id}", "starred_url": "https://api.github.com/users/julien-c/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/julien-c/subscriptions", "organizations_url": "https://api.github.com/users/julien-c/orgs", "repos_url": "https://api.github.com/users/julien-c/repos", "events_url": "https://api.github.com/users/julien-c/events{/privacy}", "received_events_url": "https://api.github.com/users/julien-c/received_events", "type": "User", "site_admin": false }, { "login": "mfuntowicz", "id": 2241520, "node_id": "MDQ6VXNlcjIyNDE1MjA=", "avatar_url": "https://avatars.githubusercontent.com/u/2241520?v=4", "gravatar_id": "", "url": "https://api.github.com/users/mfuntowicz", "html_url": "https://github.com/mfuntowicz", "followers_url": "https://api.github.com/users/mfuntowicz/followers", "following_url": "https://api.github.com/users/mfuntowicz/following{/other_user}", "gists_url": "https://api.github.com/users/mfuntowicz/gists{/gist_id}", "starred_url": "https://api.github.com/users/mfuntowicz/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/mfuntowicz/subscriptions", "organizations_url": "https://api.github.com/users/mfuntowicz/orgs", "repos_url": "https://api.github.com/users/mfuntowicz/repos", "events_url": "https://api.github.com/users/mfuntowicz/events{/privacy}", "received_events_url": "https://api.github.com/users/mfuntowicz/received_events", "type": "User", "site_admin": false } ]

[ "@julien-c @mfuntowicz any insights or updates for this issue ?", "I see the issue, but not sure how to best fix it to be honest – as it seems a very specific problem (token classification models that classify the special tokens as non-`O`)\r\n\r\nWhen we run fast tokenizers by default we'll get token alignment offsets into the original inputs, so that might solve this issue elegantly.\r\n\r\nMay I ask what's your use case here and do you need this use case supported by the inference widget (and at which horizon)?", "Our main project called [ProtTrans](https://github.com/agemagician/ProtTrans), which trains various language modelling models for protein sequences at large scale.\r\n\r\nThis specific use case predicts the secondary structure for protein sequences. It is one step behind predicting the 3D structure of protein sequences (like Google AlphaFold) that allows companies to find a drug or a cure for a virus like Covid-19.\r\n\r\nFor us we want to use the inference widget to show a live example for the prediction power of our fine-tuned models on different tasks. Later, companies or researchers might need to use it at large scale to make this prediction using your APIs.\r\n\r\nHopefully, this anwer your question 😄 \r\n\r\nReferences:\r\nhttps://blogs.nvidia.com/blog/2020/07/16/ai-reads-proteins-covid/\r\nhttps://www.youtube.com/watch?v=04E3EjsQLYo&t=89s", "👍 Oh yes I know (and love) your project and general goal/use case. I was referring to the specific use of the inference widget.\r\n\r\nI'll see what we can do. Out of curiosity, any specific reason you trained with special tokens (vs. just the raw sequence)? To be able to also do document-level classification from the same pretrained model?", "The original pretrained model [ProtBert-BFD](prot_bert_bfd) was trained using [Google Bert script](https://github.com/google-research/bert) on TPU, which automatically add these special tokens.\r\n\r\nThis allows us to perform also document-level classification as you mentioned. Like [ProtBert-BFD-MS](https://huggingface.co/Rostlab/prot_bert_bfd_membrane?text=M+G+L+P+V+S+W+A+P+P+A+L+W+V+L+G+C+C+A+L+L+L+S+L+W+A+L+C+T+A+C+R+R+P+E+D+A+V+A+P+R+K+R+A+R+R+Q+R+A+R+L+Q+G+S+A+T+A+A+E+A+S+L+L+R+R+T+H+L+C+S+L+S+K+S+D+T+R+L+H+E+L+H+R+G+P+R+S+S+R+A+L+R+P+A+S+M+D+L+L+R+P+H+W+L+E+V+S+R+D+I+T+G+P+Q+A+A+P+S+A+F+P+H+Q+E+L+P+R+A+L+P+A+A+A+A+T+A+G+C+A+G+L+E+A+T+Y+S+N+V+G+L+A+A+L+P+G+V+S+L+A+A+S+P+V+V+A+E+Y+A+R+V+Q+K+R+K+G+T+H+R+S+P+Q+E+P+Q+Q+G+K+T+E+V+T+P+A+A+Q+V+D+V+L+Y+S+R+V+C+K+P+K+R+R+D+P+G+P+T+T+D+P+L+D+P+K+G+Q+G+A+I+L+A+L+A+G+D+L+A+Y+Q+T+L+P+L+R+A+L+D+V+D+S+G+P+L+E+N+V+Y+E+S+I+R+E+L+G+D+P+A+G+R+S+S+T+C+G+A+G+T+P+P+A+S+S+C+P+S+L+G+R+G+W+R+P+L+P+A+S+L+P) fine-tuned model.\r\n\r\nWe found out that using also the special tokens during fine-tuning [ProtBert-BFD-SS3](https://huggingface.co/Rostlab/prot_bert_bfd_ss3) model perform better than not using it. I would assume because: 1)The positional encoding. 2) It matches the original Bert training method. 3) you recommended to use it in your token classification example :)\r\n\r\nThanks in advance for looking into this issue.", "not to keep pushing my own PR https://github.com/huggingface/transformers/pull/5970 but this solves some existing problems related to NER pipelines. The current hold-up is whether or not this provides a general enough solution for various models/langs [*](https://github.com/huggingface/transformers/pull/5970#discussion_r504519659).\r\nIf fast tokenizers are supported by all I can switch to a better implementation on the pipeline too, but at the current state I don't have an alternative. (suggestions are welcome)", "@julien-c Any progress for this issue ?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1108649053, "node_id": "MDU6TGFiZWwxMTA4NjQ5MDUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Help%20wanted", "name": "Help wanted", "color": "008672", "default": false, "description": "Extra attention is needed, help appreciated" }, { "id": 1834088753, "node_id": "MDU6TGFiZWwxODM0MDg4NzUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Tests", "name": "Tests", "color": "a6fcca", "default": false, "description": "Related to tests" }, { "id": 2373468354, "node_id": "MDU6TGFiZWwyMzczNDY4MzU0", "url": "https://api.github.com/repos/huggingface/transformers/labels/rag", "name": "rag", "color": "e58e85", "default": false, "description": "" }, { "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" } ]

[]

[ "sign me up, Sam", "@sshleifer \r\n\r\nexamples/rag/finetune.py is not that stable. Seems like it depends on the pytorch_lightning version also. It would be nice if we can test it properly. ", "I think this is still waiting for: https://github.com/huggingface/transformers/issues/8284 to complete the missing info and perhaps some tests were added since then?\r\n", "This issue has been stale for 1 month.", "As the required to implement this info was never provided and I since then moved to work on other things I removed self-assignment to this ticket...", "Should this issue be closed looks like [rag](https://github.com/huggingface/transformers/tree/main/examples/research_projects/rag) now has some tests [here](https://github.com/huggingface/transformers/blob/main/examples/research_projects/rag/_test_finetune_rag.py) and [here](https://github.com/huggingface/transformers/blob/main/examples/research_projects/rag/test_distributed_retriever.py)." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "@stas00 does this look reasonable? Am I missing anything you did for `examples/seq2seq`?", "yes, except you shouldn't need \r\n```\r\nsys.path.append(os.path.join(os.getcwd())) # noqa: E402 # noqa: E402 # isort:skip\r\n```\r\nthat's what `__init__.py` already did for you.\r\n\r\nand once removed, all those subsequent import `# noqa:` comments can be removed too. (except PL import)\r\n\r\nbesides using `cwd` is a bad idea - who knows where the script is invoked from and not just from the same dir as the script itself. Use `__file__` instead, which is deterministic. but it's not needed here.\r\n", "Scripts work, will merge on CI success.", "btw, you can also remove most of the `# noqa: E402 # isort:skipq` as they are no longer needed." ]

[]

[ { "login": "LysandreJik", "id": 30755778, "node_id": "MDQ6VXNlcjMwNzU1Nzc4", "avatar_url": "https://avatars.githubusercontent.com/u/30755778?v=4", "gravatar_id": "", "url": "https://api.github.com/users/LysandreJik", "html_url": "https://github.com/LysandreJik", "followers_url": "https://api.github.com/users/LysandreJik/followers", "following_url": "https://api.github.com/users/LysandreJik/following{/other_user}", "gists_url": "https://api.github.com/users/LysandreJik/gists{/gist_id}", "starred_url": "https://api.github.com/users/LysandreJik/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/LysandreJik/subscriptions", "organizations_url": "https://api.github.com/users/LysandreJik/orgs", "repos_url": "https://api.github.com/users/LysandreJik/repos", "events_url": "https://api.github.com/users/LysandreJik/events{/privacy}", "received_events_url": "https://api.github.com/users/LysandreJik/received_events", "type": "User", "site_admin": false } ]

[ "Duplicate of https://github.com/huggingface/transformers/issues/7565\r\n\r\nWorking on it in https://github.com/huggingface/transformers/pull/7645" ]

[]

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

[ "@sshleifer - thanks a lot for the issue! \r\n\r\nThis error seems related to https://github.com/huggingface/transformers/issues/7690#issuecomment-707382445. " ]

[ { "id": 2039044877, "node_id": "MDU6TGFiZWwyMDM5MDQ0ODc3", "url": "https://api.github.com/repos/huggingface/transformers/labels/marian", "name": "marian", "color": "30cc95", "default": false, "description": "" } ]

[]

[]

[]

[]

[ "Take a look at the official tutorial on fine-tuning a model on your own dataset here: https://huggingface.co/transformers/custom_datasets.html\r\n\r\nWhat's your dataset about? Is it text classification, question answering?\r\n\r\nBTW, please post any questions which are not bugs/new features you would like to see added on the [forum](https://discuss.huggingface.co/) rather than here. ", "Thanks @NielsRogge for the reply. \r\n\r\nMy dataset contains two clinical sentences (S1, S2) and its corresponding relation ('Entailment'/'Contradiction'/'Neutral'). So it basically contains three columns. ", "Ok so that's sentence pair classification. There's an example notebook on that [here](https://github.com/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb).", "Oh that's really helpful. Thank you again @NielsRogge ." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "@jplu might be interested in that issue." ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Ugh and I copied that docstrings to all other models... Thanks for fixing!" ]

[]

[]

[ "Or add this check in `on_evaluate`\r\n```python\r\nif self.prediction_bar is not None:\r\n self.prediction_bar.close()\r\n```\r\n", "I think the second solution is better (since it should be there in any case). We can add the events you suggest in the first one if we find other use cases that need them.\r\n\r\nDo you want to tackle this in a PR?", "I was thinking we can move the opening and closing of `prediction_bar` into separate events, so we don't need the `if` statements. \r\nBut it can be hard to debug if someone miss the opening, so it's probably unnecessary.\r\n\r\nI will create a PR with the second solution." ]

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

[]

[ "The way `xxxForMultipleChoice` models work is actually a bit tricky. It works as follows (based on the [original explanation by the author of BERT](https://github.com/google-research/bert/issues/38)):\r\n\r\nGiven a question, and several options, the question + options are processed by the model independently. So they will look as follows: `[CLS] question [SEP] option 1 [SEP]` first, then `[CLS] question [SEP] option 2 [SEP]` second, and so on. \r\n\r\nSo when you're using the tokenizer to encode the input, it should be used as follows:\r\n```\r\n# my multiple choice question has 4 options.\r\nquestion = \"this is a question\"\r\noption1 = \"option 1\"\r\noption2 = \"option 2\"\r\noption3 = \"option 3\"\r\noption4 = \"option 4\"\r\n\r\nencoded_input = longformer_tokenizer([question, question, question, question], \r\n [option1, option2, option3, option4], \r\n return_tensors='pt', \r\n padding='max_length')\r\n```\r\nwe need to `unsqueeze` the values of that dictionary, so to make sure they are all of shape (batch_size, num_choices, seq_len), or thus (1, 4, 4096). Also, the answer should be a tensor having shape (batch_size,), so in our case this is just a tensor containing a single element, containing the index of the correct option. Suppose the correct option is 3, then `answer` will be `torch.tensor([2])` (since indexing starts at zero). Next, we can run the forward pass as follows:\r\n\r\n```\r\nmc_labels = torch.tensor([2])\r\noutputs = model(**{k: v.unsqueeze(0) for k,v in encoding.items()}, labels=mc_labels,\r\n return_dict=True) # batch size is 1\r\n```\r\n\r\nThe `outputs` will be a `MultipleChoiceOutput` containing the loss and the logits. The logits are of shape (batch_size, number of choices), so (1, 4). \r\n\r\nIf you want to train the model, simply get the loss using `outputs.loss` and perform `loss.backward()`. If you want to get the predictions of the model, convert the logits into predictions by typing `outputs.logits.argmax(-1)`. \r\n\r\nUPDATE: fixed the fact that the answer shouldn't be unsqueezed, since it's just a tensor of shape (batch_size,). \r\nUPDATE: fix indexing of answer.\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[ "Happy to get a PR on this!", "@noise-field is remote tracking with remote server uri with authentication also enabled as part of this feature request ?", "@RahulKulhari well, this is not part of the feature request, but you can certainly do remote tracking with mlflow callback. However, you will need to use environment variables (MLFLOW_TRACKING_URI , MLFLOW_TRACKING_USERNAME, MLFLOW_TRACKING_PASSWORD) in advance to configure your connection to the remote server." ]

[]

[]

[ "Also stumbled upon this package. I was surprised we were forced to add the RC version of tokenizers. I would expect this to be pinned to \"0.8.1\" or \"0.9.1\"", "https://github.com/huggingface/transformers/pull/7794 to update to release version 0.9.1", "Hi, we have a strict requirement on `tokenizers==0.8.1rc2`. We're updating it in https://github.com/huggingface/transformers/pull/7659 but the current `transformers` `master` branch will stay pinned until that PR is merged.\r\n\r\nBoth libraries evolve quickly and generally evolve together, so having a strict `==` dependency is necessary until tokenizers version 1.0.0 is released." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thanks!" ]

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

[]

[ "It depends of what you understand as \"embedding\", as it can be ambiguous with transformer models. \r\n\r\nEmbeddings can be the embedding matrix, which returns context-less embedding of tokens, that you can obtain with `model.get_input_embeddings()`.\r\n\r\nEmbeddings can also be understood as the features generated by the base model, which are token embeddings with context (depends on the tokens surrounding the token you're studying). You can simply do a forward pass through the base models (e.g., `BertModel`, `GPT2Model`, etc.) to get these embeddings.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Might be of interest to @jplu ", "Hello!\r\n\r\nCan you give more detail on how to reproduce your issue, otherwise we cannot help you.", "> Hello!\r\n> \r\n> Can you give more detail on how to reproduce your issue, otherwise we cannot help you.\r\n\r\nThanks for your replay. I followed the instruction [here](https://github.com/huggingface/transformers/blob/master/examples/text-classification/README.md) with my own datasets (I can not provide my dataset due to confidentiality.). My script is below:\r\n\r\n> python3 run_tf_text_classification.py \\\r\n--train_file $data_dir/train.csv \\\r\n--dev_file $data_dir/dev.csv \\\r\n--test_file $data_dir/test.csv \\\r\n--label_column_id 0 \\\r\n--model_name_or_path distilbert-base-uncased \\\r\n--cache_dir $cache_dir \\\r\n--output_dir $output_dir \\\r\n--num_train_epochs 4 \\\r\n--per_device_train_batch_size 12 \\\r\n--per_device_eval_batch_size 12 \\\r\n--do_train \\\r\n--do_eval \\\r\n--do_predict \\\r\n--logging_steps 10 \\\r\n--evaluate_during_training \\\r\n--save_steps 10 \\\r\n--overwrite_output_dir \\\r\n--max_seq_length 128\r\n\r\nI can see the training started and run for a while, as there were checkpoints saved in the `output_dir`.", "Sorry, I tried with one of my dataset with the exact same command line on 4/2/1 GPUs and CPU and cannot reproduce your error. The only one thing I can tell you to do is to be sure to use the master version of the script. Otherwise without more information I cannot really help you more sorry :(", "> Sorry, I tried with one of my dataset with the exact same command line on 4/2/1 GPUs and CPU and cannot reproduce your error. The only one thing I can tell you to do is to be sure to use the master version of the script. Otherwise without more information I cannot really help you more sorry :(\r\n\r\nThanks for your help. It could be related to tensorflow or transformers versions. I will try a few of them and see how it would solve my problem.", "@lkluo I have the exact same problem. Did you solve this?", "> @lkluo I have the exact same problem. Did you solve this?\r\n\r\nI have tried many ways without any luck, so I gave up.\r\nYou may open a new issue and seek help from @jplu.", "@lkluo Thanks for your response. I think I got it fixed. I was saving the model after each epoch with `tf.saved_model.save(self.model, self.args.output_dir)`. However, when using the model for evaluation after saving it once with this method, I got the error you described. I changed it to using `self.model.ckpt_manager.save()` which is a bit inconvenient since I want .pb files, but at least the code runs fine now. If your error is also related to storing the model, this might help you.\r\n", "> @lkluo Thanks for your response. I think I got it fixed. I was saving the model after each epoch with `tf.saved_model.save(self.model, self.args.output_dir)`. However, when using the model for evaluation after saving it once with this method, I got the error you described. I changed it to using `self.model.ckpt_manager.save()` which is a bit inconvenient since I want .pb files, but at least the code runs fine now. If your error is also related to storing the model, this might help you.\r\n\r\nGood to know, thanks for letting me know. I will definitely give a try with your method.", "## Environment info\r\nPlatform: Jupyter notebook on Ubuntu 2004\r\nTF version: 2.3.1\r\nTransformers version: 3.5.0\r\nPython version: 3.6.9\r\nSingle GPU: RTX2080TI\r\n\r\n## Issue\r\nI am encountering the same error during evaluation using TFTrainer.train(). It is not reproducible, and it seems to happen randomly. I installed the latest docker image from the Tensorflow website (docker pull tensorflow/tensorflow:latest-gpu-jupyter). It seems everyone is downgrading Tensorflow to avoid this issue? What is the lowest possible version of Tensorflow for using Transformers 3.5.0?", "I met the same problem. It caused by TFTrainer.train() when excute to line 573 'self.distributed_training_steps(batch)' in trainer_tf.py. And it throws \r\n```\r\n2020-11-21 19:34:31.165454: I tensorflow/core/kernels/data/shuffle_dataset_op.cc:172] Filling up shuffle buffer (this may take a while): 9713 of 16392\r\n2020-11-21 19:34:38.101580: I tensorflow/core/kernels/data/shuffle_dataset_op.cc:221] Shuffle buffer filled.\r\n```\r\nI tried with colab gpu, it is not work. And I searched the same issue \"Shuffle buffer filled.\" in tensorflow, it is still not solved.", "I reopen this issue because many others encountered the same problem.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "Using tiny was very smart.\r\nWe upgraded to pytorch_lightning 0.9.0 (`pip install -r examples/requirements.txt`), does that fix your issue?", "It worked! But.... I had to update pyarrow from 0.14.1 to 0.17.1 because I was getting the following error:\r\n`AttributeError: module 'pyarrow' has no attribute 'PyExtensionType'`\r\n\r\nWhich I am guessing is due to y'all epic datasets library requiring pyarrow=>0.17.1:\r\n`ERROR: datasets 1.1.2 has requirement pyarrow>=0.17.1, but you'll have pyarrow 0.14.1 which is incompatible.`\r\n\r\nI opened a PR to add this dependency on pyarrow 0.17.1 to the `examples/requirements.txt`: https://github.com/huggingface/transformers/pull/7750#issue-501958657\r\nIf the PR can be accepted, I'd say this issue can be fully closed.\r\n\r\nThanks for your help with this!" ]

[]

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

[ "Just to follow up on this, look like special tokens are loaded for the RAG generator [here](https://github.com/huggingface/transformers/blob/master/src/transformers/tokenization_utils_base.py#L1635), but it is not converted to `AddedTokens` [here](https://github.com/huggingface/transformers/blob/master/src/transformers/tokenization_utils_base.py#L1592) and hence not compatible with downstream operations. ", "When I run the examples from: \r\nhttps://huggingface.co/transformers/model_doc/rag.html \r\n\r\nI get exactly the same error:\r\n\r\n", "Hey @dzorlu - thanks for your error, I will take a look tomorrow!", "> Hey @dzorlu - thanks for your error, I will take a look tomorrow!\r\n\r\nthanks @patrickvonplaten . Appreciate all the hard work :+1: ", "Should be solved now - let me know if you still experience problems @dzorlu ", "Thank you!" ]

[]

[]

[]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "Thanks!", "Thanks @julien-c!" ]

[]

[]

[]

[]

[]

[ "It was probably because I repeatly downloaded the indexes and the previous incomplete files still exist. Issue closed", "I met the same bug." ]

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

[]

[ "Hi! This should have been solved in `master`. Can you install from source and let us know if you're facing the same issue?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "This doesn't seem to be an issue with `transformers` but with your TPU and its communication with your VM. You would probably have more help if you asked over at https://github.com/pytorch/xla", "@LysandreJik Thanks for your comment. The problem rooted in the TPU software version. Setting it to `PyTorch-1.6` resolved the issue.\r\n" ]

[]

[]

[ "This is great, thanks for working on it! There's something that I had forgotten in the initial GPT-2 implementation, which was to add it to the auto-models. I did it in this PR: https://github.com/huggingface/transformers/pull/7630.\r\n\r\nCould you apply the fix to GPT-1 as well, before we merge?", "> This is great, thanks for working on it! There's something that I had forgotten in the initial GPT-2 implementation, which was to add it to the auto-models. I did it in this PR: #7630.\r\n> \r\n> Could you apply the fix to GPT-1 as well, before we merge?\r\n\r\nDone! 😊", "Thanks @fmcurti!" ]

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

[]

[ "Both pre-training and fine-tuning involve _training_ a model (i.e., updating the weights of the model using backpropagation).\r\n\r\n* In a first step, a model is (pre-)trained on a very large dataset (such as all English Wikipedia articles). BERT for example, is pre-trained on 2 tasks: masked language modeling (MLM) and next sentence prediction (NSP). Masked language modeling is, for example, given the sentence \"I went to the [MASK] to buy a bread\", the model must predict the word \"bakery\" in this case. For sentence prediction, given 2 sentences A and B, the model must predict whether sentence B follows sentence A in the dataset, or is just a random sentence. Note that both MLM and NSP are self-supervised learning (we don't need to manually annotate the dataset, because we can just use Wikipedia and mask out some words or randomize the sentences). \r\n\r\nThe reason they call it pre-training is because it's training a model before you train it on another, second, dataset. \r\n\r\n* In a second step, a model can be fine-tuned on a task of interest, usually called the downstream task. This can be text classification for example, or named-entity recognition, question-answering, summarization,... Fine-tuning is also training a model, but we start with the model that was already (pre-)trained on the large dataset in the first step. The reason this is done is because the model has already learned a lot about language in general (just by predicting masked words and interpreting the order of sentences). So the weights of these model are already quite good, they contain some \"knowledge\". We can now just use this model, and train it further on our own (usually way smaller) dataset. Note that this is supervised learning (we need to collect a labelled dataset). In case the downstream task is text classification (for example determining whether a movie review is positive or negative), then we need to collect a dataset of movie reviews and label each individual review with either \"positive\" or \"negative\". \r\n\r\nThis picture by [Jay Allamar](http://jalammar.github.io/illustrated-bert/) illustrates this very well (note that in the figure below, the downstream task is also text classification):\r\n\r\n\r\n\r\nBTW, please post any questions which are not bugs/new features you would like to see added on the [forum](https://discuss.huggingface.co/) rather than here.\r\n\r\n", "> Both pre-training and fine-tuning involve _training_ a model (i.e., updating the weights of the model using backpropagation).\r\n> \r\n> * In a first step, a model is (pre-)trained on a very large dataset (such as all English Wikipedia articles). BERT for example, is pre-trained on 2 tasks: masked language modeling (MLM) and next sentence prediction (NSP). Masked language modeling is, for example, given the sentence \"I went to the [MASK] to buy a bread\", the model must predict the word \"bakery\" in this case. For sentence prediction, given 2 sentences A and B, the model must predict whether sentence B follows sentence A in the dataset, or is just a random sentence. Note that both MLM and NSP are self-supervised learning (we don't need to manually annotate the dataset, because we can just use Wikipedia and mask out some words or randomize the sentences).\r\n> \r\n> The reason they call it pre-training is because it's training a model before you train it on another, second, dataset.\r\n> \r\n> * In a second step, a model can be fine-tuned on a task of interest, usually called the downstream task. This can be text classification for example, or named-entity recognition, question-answering, summarization,... Fine-tuning is also training a model, but we start with the model that was already (pre-)trained on the large dataset in the first step. The reason this is done is because the model has already learned a lot about language in general (just by predicting masked words and interpreting the order of sentences). So the weights of these model are already quite good, they contain some \"knowledge\". We can now just use this model, and train it further on our own (usually way smaller) dataset. Note that this is supervised learning (we need to collect a labelled dataset). In case the downstream task is text classification (for example determining whether a movie review is positive or negative), then we need to collect a dataset of movie reviews and label each individual review with either \"positive\" or \"negative\".\r\n> \r\n> This picture by [Jay Allamar](http://jalammar.github.io/illustrated-bert/) illustrates this very well (note that in the figure below, the downstream task is also text classification):\r\n> \r\n> \r\n> \r\n> BTW, please post any questions which are not bugs/new features you would like to see added on the [forum](https://discuss.huggingface.co/) rather than here.\r\n\r\nthanks @NielsRogge for the splendid information", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

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

[]

[ "I see the following requirements:\r\n\r\n1. automatic batching\r\n2. maybe smart batching\r\n3. multi GPU support\r\n4. Tokenization with multiple processes in parallel to the prediction\r\n5. `max_length` and `truncation` support\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Hello!\r\n\r\nIt looks to be an XLA error due to a precision issue. Did you train your model with mixed precision?", "No, It was trained with the official Bert script on TPU without mixed precision.", "By official Bert script you mean this one ? https://github.com/tensorflow/models/blob/master/official/nlp/bert/run_pretraining.py\r\n\r\nTo be sure it is coming from the trainer can you try with the `bert-base-cased` without the `from_pt=True` and let us know if the training finally starts.", "By the official Bert I mean script, I mean:\r\nhttps://github.com/google-research/bert\r\n\r\nWith the `bert-base-cased` model without the `from_pt=True`, I get another error:\r\n```\r\nINFO:tensorflow:Initializing the TPU system: grpc://10.112.10.242:8470\r\nINFO:tensorflow:Clearing out eager caches\r\nINFO:tensorflow:Clearing out eager caches\r\nINFO:tensorflow:Finished initializing TPU system.\r\nINFO:tensorflow:Finished initializing TPU system.\r\nWARNING:absl:`tf.distribute.experimental.TPUStrategy` is deprecated, please use the non experimental symbol `tf.distribute.TPUStrategy` instead.\r\nINFO:tensorflow:Found TPU system:\r\nINFO:tensorflow:Found TPU system:\r\nINFO:tensorflow:*** Num TPU Cores: 8\r\nINFO:tensorflow:*** Num TPU Cores: 8\r\nINFO:tensorflow:*** Num TPU Workers: 1\r\nINFO:tensorflow:*** Num TPU Workers: 1\r\nINFO:tensorflow:*** Num TPU Cores Per Worker: 8\r\nINFO:tensorflow:*** Num TPU Cores Per Worker: 8\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:CPU:0, CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:CPU:0, CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:localhost/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:CPU:0, CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:CPU:0, CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:0, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:0, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:1, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:1, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:2, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:2, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:3, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:3, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:4, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:4, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:5, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:5, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:6, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:6, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:7, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU:7, TPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU_SYSTEM:0, TPU_SYSTEM, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:TPU_SYSTEM:0, TPU_SYSTEM, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)\r\nINFO:tensorflow:*** Available Device: _DeviceAttributes(/job:worker/replica:0/task:0/device:XLA_CPU:0, XLA_CPU, 0, 0)\r\nDownloading: 100%\r\n527M/527M [00:07<00:00, 74.8MB/s]\r\n\r\nSome weights of the model checkpoint at bert-base-cased were not used when initializing TFBertForSequenceClassification: ['nsp___cls', 'mlm___cls']\r\n- This IS expected if you are initializing TFBertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).\r\n- This IS NOT expected if you are initializing TFBertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of TFBertForSequenceClassification were not initialized from the model checkpoint at bert-base-cased and are newly initialized: ['dropout_37', 'classifier']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\n---------------------------------------------------------------------------\r\nInvalidArgumentError Traceback (most recent call last)\r\n<ipython-input-15-e4258565c051> in <module>()\r\n 21 )\r\n 22 \r\n---> 23 trainer.train()\r\n\r\n4 frames\r\n/usr/local/lib/python3.6/dist-packages/transformers/trainer_tf.py in train(self)\r\n 472 Train method to train the model.\r\n 473 \"\"\"\r\n--> 474 train_ds = self.get_train_tfdataset()\r\n 475 \r\n 476 if self.args.debug:\r\n\r\n/usr/local/lib/python3.6/dist-packages/transformers/trainer_tf.py in get_train_tfdataset(self)\r\n 135 \r\n 136 self.total_train_batch_size = self.args.train_batch_size * self.args.gradient_accumulation_steps\r\n--> 137 self.num_train_examples = tf.data.experimental.cardinality(self.train_dataset).numpy()\r\n 138 \r\n 139 if self.num_train_examples < 0:\r\n\r\n/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py in numpy(self)\r\n 1061 \"\"\"\r\n 1062 # TODO(slebedev): Consider avoiding a copy for non-CPU or remote tensors.\r\n-> 1063 maybe_arr = self._numpy() # pylint: disable=protected-access\r\n 1064 return maybe_arr.copy() if isinstance(maybe_arr, np.ndarray) else maybe_arr\r\n 1065 \r\n\r\n/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/ops.py in _numpy(self)\r\n 1029 return self._numpy_internal()\r\n 1030 except core._NotOkStatusException as e: # pylint: disable=protected-access\r\n-> 1031 six.raise_from(core._status_to_exception(e.code, e.message), None) # pylint: disable=protected-access\r\n 1032 \r\n 1033 @property\r\n\r\n/usr/local/lib/python3.6/dist-packages/six.py in raise_from(value, from_value)\r\n\r\nInvalidArgumentError: Unable to parse tensor proto\r\n```\r\n\r\nIt works fine with Pytorch but not with tensorflow for some reason.", "Ok, then the error is normal. The TF part of transformers don't take into account a model that comes straight from the official BERT script. You have to:\r\n\r\n1. Convert your checkpoint with this [script](https://github.com/huggingface/transformers/blob/master/src/transformers/convert_bert_original_tf_checkpoint_to_pytorch.py) if you are using the TF1 checkpoint or this [one](https://github.com/huggingface/transformers/blob/master/src/transformers/convert_bert_original_tf2_checkpoint_to_pytorch.py) if you are using the TF2 checkpoint.\r\n2. Once you get the checkpoints in PyTorch use this [script](https://github.com/huggingface/transformers/blob/master/src/transformers/convert_pytorch_checkpoint_to_tf2.py) to get your model in a proper transformers format and use it in your pipeline.\r\n\r\nWith `bert-base-cased` the error you get is because you have to apply a cardinality to your dataset with:\r\n\r\n```\r\nmy_dataset.apply(tf.data.experimental.assert_cardinality(number_of_examples_in_my_dataset))\r\n```", "Thanks a lot @jplu , that did solve my issue.\r\n\r\nThe problem was in the second step which is converting the Pytorch to tf2 checkpoint.\r\n\r\nBy the way, there is a bug in the conversion script:\r\n\r\n```\r\n2020-10-09 21:45:28.508061: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1\r\n====================================================================================================\r\n Converting model type 1/1: bert\r\n====================================================================================================\r\nTraceback (most recent call last):\r\n File \"/usr/lib/python3.6/runpy.py\", line 193, in _run_module_as_main\r\n \"__main__\", mod_spec)\r\n File \"/usr/lib/python3.6/runpy.py\", line 85, in _run_code\r\n exec(code, run_globals)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/convert_pytorch_checkpoint_to_tf2.py\", line 437, in <module>\r\n only_convert_finetuned_models=args.only_convert_finetuned_models,\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/convert_pytorch_checkpoint_to_tf2.py\", line 322, in convert_all_pt_checkpoints_to_tf\r\n config_class, model_class, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]\r\nValueError: not enough values to unpack (expected 5, got 4)\r\n```\r\n\r\nI had to call directly the \"convert_pt_checkpoint_to_tf\" function inside the file, because \"MODEL_CLASSES[model_type]\" has only 4 objects while in \"convert_all_pt_checkpoints_to_tf\" it tries to extract 5 objects .", "Happy it worked.\r\n\r\nCan you open a new issue about the error you got with the converting script. Thanks." ]

[ { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "Just for reference, there is two open-source MIT implementations in pytorch.\r\n\r\nhttps://github.com/lucidrains/performer-pytorch\r\nAnd\r\nhttps://github.com/idiap/fast-transformers", "This could prove particularly important for longer sequences like protein sequences and long texts.\r\nHigh level overview at https://ai.googleblog.com/2020/10/rethinking-attention-with-performers.html.", "if this could be implemented it would be dope!", "It would be nice to make it possible to use FAVOR+ in combination with the pretrained models that use softmax attention— at least the popular ones like BERT. Or even better, someone would just do the fine tuning for the common pretrained models and then we could make those available out of the box. I should be able to do that for DistilBERT since I plan to be using DistilBERT + FAVOR for a project soon.", "Just started a fork to work on this at https://github.com/norabelrose/transformers-plus-performers. Is it okay with everyone if I implement it by creating a new file implementing FAVOR+ multihead attention (maybe one file for the PyTorch implementation and one for the TF implementation), then adding an option to BertConfig and DistilBertConfig (and maybe other model config classes) allowing the user to select FAVOR+ as the attention implementation?\r\n\r\nIt just seems sort of silly and wasteful to create multiple entirely new models for this when FAVOR+ has backwards compatibility.\r\n\r\nAlso since FAVOR+ is an unbiased estimator of full softmax attention, it should be possible to have an option that would tell the model to dynamically switch between FAVOR+ and full attention at test time depending on the sequence length. This would be desirable since FAVOR+ is slower than softmax attention when the sequence is shorter than O(d*log(d)), where d is the number of dimensions per attention head. Implementing such dynamic switching would be easier and more elegant if FAVOR+ is just a config option and not a new model class.", "Any update on the implementation of this new architecture? @norabelrose", "@marcoabrate The initial implementation is complete at https://github.com/norabelrose/transformers-plus-performers/blob/performers/src/transformers/modeling_performer_attention.py. Haven't been able to test it yet because getting my hands on the right datasets for finetuning DistilBERT with Performer attention, preprocessing the data, etc. has proven to be a huge ordeal. Should hopefully be able to do it today though.", "UPDATE: The most recent commit on my transformers-plus-performers repo is now up and running. Right now I only changed DistilBertModel and DistilBertConfig to enable them to use Performer attention (just set attention_type='performer'), but it should be quite trivial to add the feature to other models.\r\n\r\nAs I type this I'm fine-tuning the distilbert-base-uncased pretrained model to work with Performer attention by distilling it against bert-base-uncased. You should be able to just directly fine-tune it with MLM but I figured that distillation might get you better results. It seems to be converging rather quickly but I haven't been running it for long and I only have one GPU to work with.\r\n\r\nI would welcome other people taking a look at my repo and submitting pull requests to it.", "> FAVOR+ is slower than softmax attention when the sequence is shorter than O(d*log(d)), where d is the number of dimensions per attention head\r\n\r\nWhat are those numbers for DistilBERT, BERT-base and BERT-large?\r\n\r\nDid you compare real speed?", "I haven't had a chance to compare the difference on actual models yet, but I should be able to do that in the next day or two.\r\n\r\nI have, however, tested the speed difference between softmax attention and FAVOR+ on random Gaussian matrices. FAVOR+ really starts to get faster when the sequence length is ~18 times larger than d*ln(d), at least on my GPU. With BERT settings (d_model = 768, num_heads = 12) that means about 5000 tokens.\r\n\r\n\r\n\r\n\r\nThis is basically because you have to matrix-multiply Q and K by the random feature matrix, which you don't have to do for softmax attention. You get better results with Performer when (d_model / num_heads) is smaller:\r\n\r\n\r\n\r\nI should mention that while FAVOR+ might be slower than softmax for some of these \"medium\" sequence lengths, it should still be using less _memory_ than softmax, since it isn't allocating that L x L attention matrix. So there's somewhat of a memory-time tradeoff here.\r\n\r\nThe numbers I show above are from my own implementation of FAVOR+, but I also tried it with the performer_pytorch implementation and got almost identical results. Really, FAVOR+ is an attention mechanism for long sequences. It's got this great unique property that it's an unbiased estimator of softmax attention. That means that you can easily use it with models that were pretrained on softmax attention, and you can switch between FAVOR+ and softmax at inference time. And that's why it should be part of Huggingface.", "UPDATE: While I have Performer up and running with DistilBertModel, I've run into a problem that I didn't even think about when I started. DistilBERT, BERT, RoBERTa, and several other models use _learned_ positional embeddings, which impose a fixed 512-token max sequence length. In order to process sequences longer than 512 tokens, and thereby get the benefits of Performer attention, we'll need to use some other type of positional embeddings; for maximum flexibility, probably fixed sinusoidal embeddings with some large max sequence length. We could also try using relative position embeddings, although AFAIK no one has actually tried doing that with Performer attention and I would need to think about it a bit to figure out if that's actually feasible. DistilBertModel actually already comes with a sinusoidal_pos_embds option, but this option is overridden when you load the weights from a pretrained model.\r\n\r\nIt's not clear how hard it would be to finetune a pretrained model that was trained with learned positional embeddings to use fixed sinusoidal ones, or if it would even be worth it— it may be necessary to just train them from scratch, especially since we are _also_ trying to swap out the attention mechanism. I'll try finetuning soon and see what happens. But it's looking more likely that we won't be able to just plug in the already existing checkpoints like we initially hoped. If that turns out to be the case, it would be really great if someone with access to more than one GPU could do the training from scratch and upload the models :)\r\n\r\nPS: After @djstrong's comment about FAVOR+'s performance on relatively short sequences, I wanted to get to the bottom of why FAVOR+ was so much slower until you get up to around 5000 tokens. Oddly enough, it turns out that the torch.max() operation which is used to generate the numerical stabilizer for the exp() kernel was the main culprit. When you don't use a stabilizer, Performer attention starts beating softmax attention at much shorter sequence lengths. So I added an option in PerformerAttentionConfig to turn off the stabilizer.", "https://github.com/huggingface/transformers/issues/8893\r\n\r\nTensorflow code, not jax. Thank you.", "@guotong1988 as of about half an hour ago, my fork now has a TensorFlow implementation: https://github.com/norabelrose/transformers-plus-performers/blob/performers/src/transformers/modeling_tf_performer_attention.py.\r\n\r\nI have not had a chance to test it at all. If someone else could at least try getting it working on their own system that would be great. Pull requests are welcome.", "Hey @norabelrose , I'm part of the Performer team at Google, it's great to see this getting added to huggingface! Would you be open to meeting so we can discuss how we can work together on this? If anyone else is interested in joining the meeting please comment here and I'll reach out to coordinate.", "@tomweingarten Sure! Send me an email at [email protected] and we can set up a time to talk in the next couple weeks. As I mentioned above, the basic implementation in PyTorch and TensorFlow is done but we need to write unit tests and make sure everything is working properly. \r\n\r\nAlso, in my fork at transformers-plus-performers I had to make a few minor changes to other parts of HuggingFace in order to get training to run smoothly on my machine— in particular, the distillation example program, since I initially tested PerformerAttention by continuing distillation of a pretrained DistilBERT model with Performer attention against bert-base. The implementation of distillation on master loads all the training data at once into RAM, which blows up on my meager hardware. I changed it so that you can load the training data incrementally. That's probably a good thing to add to the master branch, but arguably it should be put in a separate pull request. So we'll have to change that, a long with a couple other little things.\r\n\r\nI'd recommend you check out my fork at https://github.com/norabelrose/transformers-plus-performers/. The relevant files are /src/transformers/configuration_performer_attention.py, /src/transformers/modeling_performer_attention.py, and /src/transformers/modeling_tf_performer_attention.py. I also changed the BERT and DistilBERT model and config files so the user can use Performer attention with them. I'll accept pull requests on that repo.\r\n\r\nPS: Also just realizing that the definition of short_sequence_behavior on PerformerAttentionConfig in the last commit is defined variously as Union[str, dict], Union[str, Callable], or Union[str, tuple]— sorry about that, I wasn't really sure how best to implement that feature. Right now the actual implementation in PerformerAttention assumes it's a str or Callable.", "@tomweingarten @norabelrose I would like to participate in the meeting too, if possible. I am working with long sequences for summarization. I have not had the chance to go through the code thoroughly yet, but I am ready to help soon.\r\n\r\nEdit: you can reach me at [email protected]", "@norabelrose Is there any plan to support unidirectional attention ?", "Hi guys, thanks to @kchoro and @ValeryTyumen on the Performers team, we've open-sourced the Tensorflow version of FAVOR+ here: https://github.com/google-research/google-research/tree/master/performer/fast_attention/tensorflow\r\n\r\nBTW, we've edited the folder name and code to be `fast_attention` now rather than `fast_self_attention`.\r\n\r\nPlease let us know how well it works in your pipelines!", "UPDATE: The new default branch (\"clean\") on my fork at https://github.com/norabelrose/transformers-plus-performers/ now has all the extraneous changes I made to the upstream removed. I also merged in all new commits from upstream.\r\n\r\n@TwinMooon Yes, we should be able to add causal attention. I was under the impression that it would be necessary to include a custom CUDA kernel from the fast-transformers library to compute the prefix sums— since that's what the performer_pytorch implementation does, which I used as a template for my implementation— but now looking at the Google code in both Jax and TensorFlow I realize that they just compute the prefix sums in Python code and then use a custom gradient. So it looks like it's not necessary, although it's possible that using the CUDA kernel gives you a noticeable speed boost.", "I'd like to set a goal of making an official pull request to add this to master by the end of the year. I haven't been able to do that yet because I've been busy with school and other projects, and I haven't gotten any help from other contributors. Key things that need to be done are:\r\n- Add causal attention\r\n- Translate the unit tests from the Google implementation and add them to the fork (and make sure we pass those tests, obviously)\r\n- Clean up the short_sequence_behavior feature (or just get rid of it)\r\n\r\nAs always, any and all help with these tasks is welcome.", "@TwinMooon Update: I got causal attention working by translating the Google implementation, but as I feared, it's very slow since it doesn't use a custom CUDA kernel. On my GPU, it's 19-20 times slower than noncausal attention. But there might be a way around this; I'll have to think about it.\r\n\r\nIn the meantime, I think I'm going to add an optional dependency on the fast_transformers package (just wrapping the import statement in a try... except block) to get access their custom CUDA kernel. I'll include a warning if the user doesn't have it installed that causal attention might have bad performance without the package. That's what the performer_pytorch package does.", "@norabelrose For the past two days, I have implemented a version of causal attention by just translating Google's TensorFlow implementation. After reading your code, I found that our implementation is quite similar. However, The causal version runs a little faster than the non-casual version in my machine. \r\nMy PyTorch version is 1.5.0 and run it in a 2080Ti with CUDA 10.0", "@TwinMooon Ok cool! If you wouldn’t mind submitting a pull request to my fork or just copy and pasting the relevant block of code here then I could check to see if your version is faster. It’s possible that I’m making some silly mistake.\r\n\r\nI’m running it on a GeForce GTX 1080 with PyTorch 1.4.0 and CUDA 10.0.0. It was also noticeably a lot slower than noncausal attention on my CPU only laptop which has PyTorch 1.7.\r\n\r\nPS: Is it possible that you got the tensor shapes mixed up? The Google implementation expects tensors of shape [length, batch, heads, random features/embedding dim] while everywhere else it's usually [batch, heads, length, random features/embedding dim], so you have to permute the tensor dimensions. The code will actually run if you give it tensors with the [B, H, L, D] shape though, so I got tripped up on that when I first translated the Google code and it made it look like it was faster than it actually was. If you're using a small batch size of say, 1 or 5, it'll be a lot faster to compute prefix sums over the batch dimension than doing it over the sequence length dimension of size 512 (which is what it's actually supposed to do).", "@norabelrose You can review my implementation [here](https://github.com/TwinMooon/transformers-plus-performers/commit/c17d6473deb5316363f60bb2ddd1007d4364abe4). I permuted the tensor shape before stuff into the casual attention. ", "@TwinMooon In your code, you spell the word \"causal\" two different ways: \"causal\" and \"casual\". You use the \"causal\" spelling in the forward() method where short_sequence_behavior indicates to use softmax attention, and then you use casual everywhere else.\r\n\r\nIs it possible that you're initializing the PerformerAttention object sort of like this:\r\n`PerformerAttention(PerformerAttentionConfig(d_model=768, num_heads=12), causal=True)`\r\nso that the \"casual\" attribute remains its default value of False, and none of the causal attention code ever actually gets called? I should probably change `__init__` so it that it always throws an error when you include a nonexistent attribute in kwargs.\r\n\r\nIn other news, I figured out a sort of clever way of making causal attention like 2x faster, and that's in my latest commit.", "Mark Zakharov made a Colab where he successfully finetuned a DistilBERT model with the most recent version of my fork, which you can check out here: https://colab.research.google.com/drive/1BUYk4qxdt1b3d5mx6_t0nnX5jP9KwVAv?usp=sharing\r\n\r\nI think the project is almost ready to be made into a formal pull request.", "@norabelrose cool! I'll try it now.", "This is really great work guys! We are currently running some experiments on the flax version of Performer internally and looking into how to best integrate the model into Transformers. @norabelrose a PR in PyTorch and or Tensorflow would be amazing!", "Excited to see the progress here! Just wanted to give a heads-up that we fixed a [significant bug](https://github.com/google-research/google-research/commit/b09ac837cd5720bc60f1c16b472a7ab462b0ddb8) in our TF implementation of Performer fast attention.", "Pull request finally submitted: #9325 " ]

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

[]

[ "This would be still interesting.\r\n\r\nFor tokenizing a list of pairs I get \r\n\r\n```\r\ninput_ids = tokenizer(pairs, max_length=50, padding=\"max_length\",truncation=True, return_tensors=\"tf\")\r\n\r\n 'token_type_ids': <tf.Tensor: shape=(1, 50), dtype=int32, numpy=\r\narray([[0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,\r\n 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,\r\n 0, 0, 0, 0, 0, 0],\r\n....\r\n```\r\nSo I wonder I am doing right as it seems the padded values are connetced to the 1. sentences (0 id)\r\nSO token type IDS are 0 for the padded places, is that right?", "Hey @datistiquo ,\r\n\r\nAs one can see in the following script:\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nfrom transformers import BertTokenizer\r\n\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\r\n\r\ninput_ids = tokenizer([[\"hey hello there\", \"what is going on\"], [\"peter is in the\", \"what is going on\"]], max_length=20, padding=\"max_length\", truncation=True, return_tensors=\"tf\")\r\n\r\nprint(\"List of pairs\", input_ids)\r\n\r\ninput_ids = tokenizer(\"hey hello there\", \"what is going on\", max_length=20, padding=\"max_length\", truncation=True, return_tensors=\"tf\")\r\n\r\nprint(\"Pair\", input_ids)\r\n```\r\n\r\ntokenizing a list of pairs should be done exactly as proposed by you. Regarding the token_type_ids it is also correct that padded places should have a value of 0. In general if a model does not make use of `token_type_ids`, we return a 0 for such a model, see: https://github.com/huggingface/transformers/blob/6b034309ca4ca2ec6e5c3cacda92a448fa10b921/src/transformers/models/roberta/tokenization_roberta.py#L233 . So for padded tokens that should be discarded in the model, 0 seems like the most sensible choice to me.", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "@dineshggaonkar Did you fix it?\r\n" ]

[]

[]

[]

[]

[]

[ "(tagging @eltoto1219 for information)" ]

[]

[]

[]

[]

[]

[]

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

[]

[ "`facebook/bart-large-mnli` is `mit` like other pretrained models initially released in [fairseq](https://github.com/pytorch/fairseq#license)\r\n\r\nFor `dbmdz` I'll let @stefan-it chime in, with the caveat that lineage/inheritance of licenses in fine-tuned ML models is (AFAIK) uncharted territory so if you have more info on that subject, please feel free to share it.\r\n\r\nFinally, for models where the license isn't indicated in the model card, please feel free to open a PR to add it.", "Hi @julien-c @ankane,\r\n\r\nI am also very interested in clarifying the licenses of default models. In particular, I'd like to know the license of `dbmdz/bert-large-cased-finetuned-conll03-english`.\r\n\r\nCheers,\r\n\r\nAlex Combessie", "Hey @julien-c, thanks for the quick response and `facebook/bart-large-mnli` info. PR submitted.\r\n\r\nRe fine-tuning licensing: Seems like it may fit the definition of \"Derivative Works\" in the Apache 2.0 license, but I don't have any special knowledge here, so will defer further discussion to someone that does.", "Hi guys,\r\n\r\nsorry for the late reply! I have no strong opinion on that topic, so I would just say that license of our `dbmdz` models will be MIT, because we're usually use this kind of license for both software and our pre-trained LMs :) ", "Great, thanks @stefan-it! That makes it clear that the model is open source :tada:\r\n\r\nIt'd be good to add a model card with the license. I personally think the most accurate summary of the model license is MIT + Apache-2.0 (unless it wasn't derived from Apache-2.0 work), but will leave it to you and the Transformers team to decide how you want to represent it.\r\n", "On the technical side, just took a look at the code and our YAML parser would support an array of licenses, so feel free to open a PR with \r\n\r\n```\r\nlicense:\r\n- mit\r\n- apache-2.0\r\n```\r\n\r\nOn the legal side, 🤷‍♂️", "Thanks @julien-c, good to know 👍 \r\n\r\nWill wait to hear thoughts from @stefan-it.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1834088753, "node_id": "MDU6TGFiZWwxODM0MDg4NzUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Tests", "name": "Tests", "color": "a6fcca", "default": false, "description": "Related to tests" } ]

[]

[ "You're correct, it is not currently running as there were some issues setting up both the PT/TF test suites. Will look into it this afternoon.", "The slow tests in TF take an absurdly long time. I had to stop them from running after ~3.5 hours as it was holding the whole test suite back. Will investigate more on a separate machine and try to skin it down.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1834054694, "node_id": "MDU6TGFiZWwxODM0MDU0Njk0", "url": "https://api.github.com/repos/huggingface/transformers/labels/TensorFlow", "name": "TensorFlow", "color": "FF6F00", "default": false, "description": "Anything TensorFlow" } ]

[]

[ "Yeah, that's a know failure and I didn't manage to make it work yet with the `cast_bool_to_primite(...)` function", "This should be fixed in the next big TF rework.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "I'm not sure how we can expect this to work with the encoding and the usage of the `generate` method directly in the forward method.\r\n\r\n@patrickvonplaten can chime in if I'm wrong, but I believe the `generate` method can not be back-propagated through as it is right now.", "@LysandreJik I came across one issue (#6105) with `generate` being used in `forward`. My temporary workaround was to introduce `is_finetuning_current_model` into `generate` that will call `generate_text_while_finetuning` instead of `forward` again to avoid the recursion.\r\n\r\nI'm still learning pytorch so I might be wrong on this, and correct me if I am. I checked the `grad_fn` for `input_ids`, `full_generated_gpt2_ids`, and each of them were set to `None`. `tmp_losses`, `losses`, and `loss` all had their `grad_fn` set. My naive assumption is that backpropagation will run up to `tmp_losses`, skip over the `generate` part, and then continue on through the gpt2 model.\r\n\r\nAnother interesting point is that I get the error on different training examples. I set the batch size to 1 and it would produce the error on, say, example 5. Removing example 5 from the training set and rerunning would cause the error on example 3, etc.", "yes, the `generate()` cannot be used for backpropagation at the moment . ", "@patrickvonplaten Would that explain why I'm encountering the above error? Do you also mind elaborating on why `generate()` cannnot be usef for backpropagation? I'm interested to hear the details for the sake of my own knowledge.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

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

[]

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

[]

[]

[ "Ok ready for review @LysandreJik @sgugger.\r\n\r\nIt's pretty big sorry.\r\n\r\nFor now sentencepiece is still in the requirements as removing it has some effect on the pipeline tests and I think it's probably good to study this in a separate future PR.\r\n\r\nThere is no breaking change apart from the fact that importing the `**Fast` tokenizer directly from the `transformers.tokenization_xxx` is not possible anymore, they should be imported from `transformers` (the best and most robust choice) or from their new respective location at `transformers.tokenization_xxx_fast`.", "Ok the `examples/seq2seq/test_seq2seq_examples.py::test_finetune[stas/tiny-wmt19-en-de]` is working now.\r\nI'll address the other comments and we can merge on Monday.", "@thomwolf, could you please assign defaults that are different from \"stas/tiny-wmt19-en-de\" entry and its contents? Otherwise it defeats the purpose of testing with this model, since defaults are used instead.\r\n\r\nAlternatively, I will need to create a new tiny model with different config and change tests to use that instead.\r\n\r\nOnce this is done let's add this test that I tried to add here: https://github.com/huggingface/transformers/pull/7860 - I expanded it below a bit to do better testing:\r\n\r\n```\r\ndiff --git a/tests/test_tokenization_fsmt.py b/tests/test_tokenization_fsmt.py\r\nindex c3e08d56..833b1742 100644\r\n--- a/tests/test_tokenization_fsmt.py\r\n+++ b/tests/test_tokenization_fsmt.py\r\n@@ -24,6 +24,7 @@ from transformers.tokenization_fsmt import VOCAB_FILES_NAMES, FSMTTokenizer\r\n\r\n from .test_tokenization_common import TokenizerTesterMixin\r\n\r\n+FSMT_TINY = \"stas/tiny-wmt19-en-de\"\r\n\r\n class FSMTTokenizationTest(TokenizerTesterMixin, unittest.TestCase):\r\n tokenizer_class = FSMTTokenizer\r\n@@ -86,6 +87,13 @@ class FSMTTokenizationTest(TokenizerTesterMixin, unittest.TestCase):\r\n def tokenizer_en_ru(self):\r\n return FSMTTokenizer.from_pretrained(\"facebook/wmt19-en-ru\")\r\n\r\n+ def test_online_tokenizer_config(self):\r\n+ \"\"\"this just tests that the online tokenizer files get correctly fetched and\r\n+ loaded via its tokenizer_config.json and it's not slow so it's run by normal CI\r\n+ \"\"\"\r\n+ tokenizer = FSMTTokenizer.from_pretrained(FSMT_TINY)\r\n+ self.assertListEqual([tokenizer.src_lang, tokenizer.tgt_lang], [\"en\", \"de\"])\r\n+\r\n def test_full_tokenizer(self):\r\n \"\"\" Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt \"\"\"\r\n tokenizer = FSMTTokenizer(self.langs, self.src_vocab_file, self.tgt_vocab_file, self.merges_file)\r\n\r\n```\r\n\r\nThanks.\r\n", "Yes feel free to create another model for fsmt @stas00.\r\n\r\nOk this big PR is ready for merge as soon as possible (with regards to other PR merges not absolute time) so it doesn't drift too much.", "There are some `~transformers.tokenization_utils_base.PreTrainedTokenizer` left (and same with fast) but that's an easy pattern to search for a subsequent PR.", "Ok then I'm merging this PR and continuing in another one to:\r\n- add CI tests for the package without sentencepiece and tokenizer\r\n- remove sentencepiece as a required dependency\r\n- switch to fast tokenizers by default\r\n- fix the remaining doc patterns that you mentioned\r\n\r\nOn the topic of `from_pretrained` logic, we could (should probably be another PR):\r\n- add a test that the config of the tokenizers is used as mentioned by @stas00\r\n- we could probably remove the hard-coded configs at the same time\r\n- switch to the cloud-front links like the models for faster dowloads", "**edited**: thanks to @sshleifer - I needed to `pip install -e \".[dev]\"` to update the new dependencies. that fixed the issues.\r\n\r\n----------\r\n\r\nI'm getting a massive amount of identical failures after this merge, primarily:\r\n\r\n```\r\n_____________________________________________ XLNetTokenizationTest.test_num_special_tokens_to_add_equal _____________________________________________\r\n[gw1] linux -- Python 3.8.5 /home/stas/anaconda3/envs/main-38/bin/python\r\n\r\nself = <tests.test_tokenization_xlnet.XLNetTokenizationTest testMethod=test_num_special_tokens_to_add_equal>\r\n\r\n def test_num_special_tokens_to_add_equal(self):\r\n for tokenizer, pretrained_name, kwargs in self.tokenizers_list:\r\n with self.subTest(\"{} ({})\".format(tokenizer.__class__.__name__, pretrained_name)):\r\n> tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)\r\n\r\ntests/test_tokenization_common.py:1896: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\nsrc/transformers/tokenization_utils_base.py:1588: in from_pretrained\r\n return cls._from_pretrained(\r\nsrc/transformers/tokenization_utils_base.py:1661: in _from_pretrained\r\n tokenizer = cls(*init_inputs, **init_kwargs)\r\nsrc/transformers/tokenization_xlnet_fast.py:142: in __init__\r\n super().__init__(\r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\n\r\nself = <[AttributeError(\"'XLNetTokenizerFast' object has no attribute 'name_or_path'\") raised in repr()] XLNetTokenizerFast object at 0x7f2ccdaaee80>\r\nargs = (), kwargs = {'additional_special_tokens': ['<eop>', '<eod>'], 'bos_token': '<s>', 'cls_token': '<cls>', 'do_lower_case': False, ...}\r\nslow_tokenizer = None\r\nfast_tokenizer_file = '/home/stas/.cache/torch/transformers/d152c146766f0a31888c4c9c0dcf82e42e42d09bf818bb74e126f2420cbd36c4.ecf1d38c0b94010f431264b9ded85217342f84c7bdae79b0472f7cd20b94052d'\r\n\r\n def __init__(self, *args, **kwargs):\r\n slow_tokenizer = kwargs.pop(\"__slow_tokenizer\", None)\r\n fast_tokenizer_file = kwargs.pop(\"tokenizer_file\", None)\r\n \r\n if fast_tokenizer_file is not None:\r\n # We have a serialization from tokenizers which let us directly build the backend\r\n> fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file)\r\nE Exception: data did not match any variant of untagged enum PyNormalizerTypeWrapper at line 1 column 318041\r\n```\r\ndo I need to remove cache or something? I won't test this until you tell me to in case you need someone with the old cache to test that it can recover from this.\r\n\r\nA total of 90 failed tests with this error.", "You should update `tokenizers` to the main PyPi version @stas00 \r\n```\r\npip install tokenizers --update\r\n```" ]

[]

[]

[ "Thanks a lot!" ]

[]

[]

[ "Yes, unfortunately that link will only be live at the next release (for now squeezeBERT is only in master, so only in the master documentation).\r\n@LysandreJik not sure if there is a way to properly fix this unless we add \"Check all the links in the README to remove the master\" in our release check list.", "You're right, I don't think there's any other way without over-engineering a feature." ]

[]

[]

[ "Hey @marcoabrate, \r\n\r\nplease make sure that you use the correct parameter name `num_beams` instead of `num_beans`. \r\nWhen using `num_beams`, I cannot reproduce your error.", "of course it was that!\r\n\r\nthank you" ]