Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[ "I'd like the new examples scripts to stay fairly focused on one problem (at the cost of potentially have some duplicate codes) so they're easy to understand (and tweak) by users. We don't support any kind of datasets either (if your QA dataset has fields with names slightly different than SQUAD for instance), users are supposed to adapt the relevant lines in the code to their needs.\r\n\r\nSo with that in mind, I'd definitely prefer a separate script :-)" ]

[]

[]

[ "@patrickvonplaten might be able to help", "Seq2Seq models are a bit special - they also need `decoder_input_ids` as the error message states. Since torchscript however does not allow keyword arguments we need to provide positional arguments and therefore it's mandatory to also provide the 2nd argument being the `attention_mask` (for the encoder).\r\n\r\nThe following is what you are looking for (I think):\r\n\r\n```python\r\nfrom transformers import T5Tokenizer, T5ForConditionalGeneration\r\nimport torch\r\n\r\ntokenizer = T5Tokenizer.from_pretrained('t5-small')\r\nmodel = T5ForConditionalGeneration.from_pretrained('t5-small', torchscript = True)\r\ninput_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids\r\nattention_mask = input_ids.ne(model.config.pad_token_id).long()\r\ndecoder_input_ids = tokenizer('<pad> <extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids\r\n\r\ntraced_model = torch.jit.trace(model, (input_ids, attention_mask, decoder_input_ids))\r\ntorch.jit.save(traced_model, \"traced_t5.pt\")\r\n```", "Thank you for the solution, your mentioned code works perfectly fine for creating a `torchscript `model. \r\nBut I have one more question, the generated `traced_t5.pt` model doesn't seem to have the `model.generate()` method. \r\nhow to get the `token ids` output from this newly created model? (only from model()). \r\nAlso, In General, how can we get the output (token ids) from a t5 model without using the `generate()` method?", "Yeah, I don't think our `generate()` method is torchscriptable yet :-/ You should take a look at the `greedy_search` method to see how the `generate()` method can be implemented by hand :-) \r\n\r\nGreedy search: https://github.com/huggingface/transformers/blob/df311a5ccf50be3031474e289b43b1be43111144/src/transformers/generation_utils.py#L622\r\n\r\nGenerate: \r\nhttps://github.com/huggingface/transformers/blob/df311a5ccf50be3031474e289b43b1be43111144/src/transformers/generation_utils.py#L296\r\n\r\n", "thank you, I'll look into it. ", "@Ki6an : Were you able to figure out how to make use of `greedy_search` to do the work which `generate` does? If so can I request you to share that as a gist?", "@karrtikiyerkcm have a look at [FastT5](https://github.com/Ki6an/fastT5) library, it implements both greedy and beam search for T5. ", "Thanks @Ki6an , I was trying something similar for Pegasus Models for the summarisation task.", "@Ki6an Hello, the input_id I inputed is 64*100 (Batch_size,max_sequence), why the size of \r\n T5ForConditionalGeneration.generate result is 100?where is the batch_size\r\n\r\n", "anyone know if it's still not possible to use torchscript with generate?", "@patrickvonplaten @Ki6an Hi, what should ```decoder_input_ids``` be if my input text is ```translate English to German: Thank you!```? I'm using this for inference. For decoder models like BERT and GPT, all I need to do is use Tokenizer to get the ```input_ids``` which will be passed into the models. But I'm not sure how that works for encoder-decoder models like T5 here.", "> @patrickvonplaten @Ki6an Hi, what should `decoder_input_ids` be if my input text is `translate English to German: Thank you!`? I'm using this for inference. For decoder models like BERT and GPT, all I need to do is use Tokenizer to get the `input_ids` which will be passed into the models. But I'm not sure how that works for encoder-decoder models like T5 here.\r\n\r\nHi, I wanted to follow up on this. I have the same question.", "> Seq2Seq models are a bit special - they also need `decoder_input_ids` as the error message states. Since torchscript however does not allow keyword arguments we need to provide positional arguments and therefore it's mandatory to also provide the 2nd argument being the `attention_mask` (for the encoder).\r\n> \r\n> The following is what you are looking for (I think):\r\n> \r\n> ```python\r\n> from transformers import T5Tokenizer, T5ForConditionalGeneration\r\n> import torch\r\n> \r\n> tokenizer = T5Tokenizer.from_pretrained('t5-small')\r\n> model = T5ForConditionalGeneration.from_pretrained('t5-small', torchscript = True)\r\n> input_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids\r\n> attention_mask = input_ids.ne(model.config.pad_token_id).long()\r\n> decoder_input_ids = tokenizer('<pad> <extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids\r\n> \r\n> traced_model = torch.jit.trace(model, (input_ids, attention_mask, decoder_input_ids))\r\n> torch.jit.save(traced_model, \"traced_t5.pt\")\r\n> ```\r\n\r\nHi can you please provide example of how to use this t5 jit traced model for inference..\r\nI tried using it but it requires decoder_input_ids.. is there any way of doing inference without the decoder_input_ids?" ]

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

[]

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

[]

[ { "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 } ]

[ "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! Indeed, model outputs cannot be unpacked this way. It is mentioned in the [documentation](https://huggingface.co/transformers/main_classes/output.html#transformers.file_utils.ModelOutput). You can retrieve the items by unpacking them like this if you use the `.to_tuple()` method.", "Oh, ok. Thanks and sorry for missing that. ", "No problem, happy to help!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Have you tried restarting the kernel after installing `sentencepiece`?", "Yes, I did, with:\r\n!pip install --upgrade sentencepiece", "Is it possible for you to share your notebook so that I may take a look?", "I've just relaunched my notebook now, I don't have any issue now.\r\n\r\nThank you for your help\r\nRegards", "Glad it works for you now!" ]

[]

[]

[]

[]

[]

[ "[Tooling comment] For some reason `python utils/check_copies.py --fix_and_overwrite` fails on my Python with following:\r\n```\r\n(.env) ipro:transformers gibbon$ python utils/check_copies.py --fix_and_overwrite\r\nTraceback (most recent call last):\r\n File \"utils/check_copies.py\", line 432, in <module>\r\n check_model_table(args.fix_and_overwrite)\r\n File \"utils/check_copies.py\", line 414, in check_model_table\r\n new_table = get_model_table_from_auto_modules()\r\n File \"utils/check_copies.py\", line 328, in get_model_table_from_auto_modules\r\n spec = importlib.util.spec_from_file_location(\r\nAttributeError: module 'importlib' has no attribute 'util'\r\n\r\n```", "I think it's better at the top rather than the end of the list (I don't think a user will read to until the end of the list TBH). We could even put it further up the README!" ]

[]

[]

[ "`run_squad.py` is more complete right now, as `run_squad_trainer.py` can't do evaluation (yet! it will be possible in a few days).\r\n\r\nWe try to keep the github issues for bugs/feature requests.\r\nFor next time, could you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!", "Okay sure. I don't know about the forum - sorry for that. Thank you @LysandreJik " ]

[ { "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" } ]

[]

[ "This doesn't seem correct" ]

[]

[ { "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 } ]

[ "Hmmm, this is weird, I can't reproduce with a very similar environment:\r\n\r\n```py\r\n- `transformers` version: 4.0.0\r\n- Platform: Linux-5.9.11-arch2-1-x86_64-with-glibc2.10\r\n- Python version: 3.8.3\r\n- PyTorch version (GPU?): 1.7.0 (True)\r\n- Tensorflow version (GPU?): 2.3.1 (False)\r\n```\r\n\r\nIt outputs the following:\r\n\r\n```\r\n{'score': 0.26648467779159546, 'start': 90, 'end': 106, 'answer': 'peintre français'}\r\n```\r\n\r\nDiffering element here seems to be Windows vs Linux. Do you mind trying to install the library from `master` and telling me if that works?\r\n\r\nYou can try it with:\r\n\r\n```\r\npip install git+https://github.com/huggingface/transformers\r\n```", "For what it's worth, I ran into the same issue using a GitHub Actions workflow with the following environment:\r\n\r\n```python\r\n- `transformers` version: 4.0.0\r\n- Platform: Windows Server 2019\r\n- Python version: 3.6.8\r\n- PyTorch version (GPU?): 1.6.0 (No)\r\n```\r\n\r\nI modified the workflow to run against the master branch and the issue appears to be resolved there.", "Yes this command solve my issue, thank you :)" ]

[]

[]

[ "Could you provide the environment information as mentioned in the issue template, alongside the a reproducible that outputs this so that we may check what's going on? Thank you.", "Hi @BryanWBear , \r\n\r\nI am facing this issue too. In the meantime, did you find a solution?\r\n\r\nThank you so much in advance!", "the default `logging_steps` in `TrainingArguments` is set to `500` steps, so no loss is reported before 500 steps", "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." ]

[]

[]

[ "It should work - we'll run some experiments soon :-) https://github.com/huggingface/transformers/pull/8358 @TevenLeScao", "Ok, great!\r\n\r\nFor the moment i'm running some experiments myself with FlaxBertModel but i'm getting unexpected behaviors: it seems that the flax implementation is slower than the torch one; I tried to run this simple code in google colab with GPU environment \r\n\r\n```python\r\n!pip install --upgrade pip\r\n!pip install --upgrade jax jaxlib==0.1.57+cuda101 -f https://storage.googleapis.com/jax-releases/jax_releases.html\r\n!pip install flax\r\n!pip install transformers\r\n\r\nimport flax\r\nimport jax\r\nfrom transformers import BertModel\r\nfrom transformers import FlaxBertModel\r\nfrom transformers import AutoTokenizer\r\nimport time\r\n\r\nbaseModel ='bert-base-uncased'\r\ntorchBert = BertModel.from_pretrained(baseModel)\r\nflaxBert = FlaxBertModel.from_pretrained(baseModel)\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(baseModel)\r\ntorchInput = tokenizer.encode(\"Random sentence\", truncation=True, padding=True, return_tensors='pt')\r\nflaxInput = tokenizer.encode(\"Random sentence\", truncation=True, padding=True, return_tensors='jax')\r\n\r\nstart_time = time.time()\r\nfor _ in range(10):\r\n torchBert(torchInput)\r\nprint(\"torch - \", time.time()-start_time)\r\n\r\nstart_time = time.time()\r\nfor _ in range(10):\r\n flaxBert(flaxInput)\r\nprint(\"flax - \", time.time()-start_time)\r\n```\r\n\r\nand i'm getting the following output\r\n\r\n```\r\ntorch - 0.6615538597106934\r\nflax - 5.129613161087036\r\n```\r\n\r\nWhat am i missing?", "You should probably use `jax.jit` to speed it up", "Indeed, as @patrickvonplaten says, using `jax.jit` on `flaxBert` will speed things up considerably. This will first shape-compile the function using XLA, and every time you call the function again (provided the shapes are the same), it will run the compiled version directly. I've demonstrated in this Colab: https://colab.research.google.com/drive/1davNsnV34KDZOyJ9i8zZfvxAVjBJC4dp?usp=sharing\r\n\r\nMake sure you set the accelerator to GPU/TPU! (Runtime -> Change runtime type)\r\n\r\nHere's a summary:\r\n\r\n```\r\n>>> %timeit torchBert(torchInput)\r\n1 loop, best of 5: 75.3 ms per loop\r\n>>> %timeit flaxBert(flaxInput)\r\n1 loop, best of 5: 1.41 s per loop\r\n>>> %timeit jitted_flax(flaxInput)\r\n100 loops, best of 5: 11.2 ms per loop\r\n```\r\n\r\nNote that this excluded the compilation time for the first time we called `jitted_flax`. Including this will increase the overall execution time, but since it has to be done only once this is negligible as you execute this function more often.\r\n\r\nTo learn more about JAX's jit, this quickstart is quite useful: https://jax.readthedocs.io/en/latest/notebooks/quickstart.html", "Thank you for you comment @marcvanzee. Indeed after @patrickvonplaten's reply I checked Flax and Jax documentation more carefully confirming that JIT compilation could solve performance issues. It still surprises me though that pytorch is quite fast even without JIT compilation while the same is not true for Flax. Frankly i didn't even know that JIT existed in pytorch so i'd be curious too to se how it compares to Flax.", "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." ]

[]

[]

[ "I think this is probably so because the `cls.predictions.decoder` is a linear layer, which is tied to the embeddings layer. They're pointing to the same weights, so freezeing one of those would result in freezing the other one." ]

[]

[]

[ "Thank you very much for this correction, @devangi2000 " ]

[]

[]

[]

[]

[]

[ "@sgugger you might be interested in that issue given that you're refactoring the squad example!", "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." ]

[]

[]

[ "A model trained for sequence classification can definitely be loaded with a different head. Here's an example:\r\n\r\n```py\r\nfrom transformers import DistilBertForSequenceClassification, DistilBertForMaskedLM\r\nsequence_classifier = DistilBertForSequenceClassification.from_pretrained(\"...\")\r\n# Do stuff with your model, train it, do what you like\r\n\r\n# Save the weights in a local directory\r\nsequence_classifier.save_pretrained(\"model-trained-on-xxx\")\r\n\r\n# Load the weights in the *ForMaskedLM model.\r\nlanguage_model = DistilBertForMaskedLM.from_pretrained(\"model-trained-on-xxx\")\r\n```\r\nThis `language_model` has kept all the weights of the base transformer model, has discarded the sequence classification layers, and has randomly initialized the new layers. This model can be loaded in an `AutoModelWithLMHead`.", "Oh wow, did not expect it to be this easy. Thanks very much!" ]

[]

[]

[ "Try the various Marian fine-tuning scripts. You should easily be able to get 22+ BLEU on wmt-en-ro with both `finetune_trainer.py` and `finetune.py` in < 30 minutes on brutasse.", "Speed / Memory benchmark of master vs. this PR is ok for me:\r\n\r\n\r\n", "@patrickvonplaten I would measure impact of running cnn summarization (which uses seq_len 1024) to fully acknowledge the trade off you are making:\r\n```bash\r\ncd examples/seq2seq\r\npython run_eval.py facebook/bart-large-cnn cnn_dm/test.source cnn_gens.txt \\\r\n\t--reference_path cnn_dm/test.target \\\r\n\t--score_path cnn_rouge.json --task summarization \\\r\n\t--n_obs 500 --fp16\r\n```\r\nThis should take ~5 mins per branch.\r\n\r\nOtherwise, LGTM! Thanks for cleaning up after me :)", "> @patrickvonplaten I would measure impact of running cnn summarization (which uses seq_len 1024) to fully acknowledge the trade off you are making:\r\n> \r\n> ```shell\r\n> cd examples/seq2seq\r\n> python run_eval.py facebook/bart-large-cnn cnn_dm/test.source cnn_gens.txt \\\r\n> \t--reference_path cnn_dm/test.target \\\r\n> \t--score_path cnn_rouge.json --task summarization \\\r\n> \t--n_obs 500 --fp16\r\n> ```\r\n> \r\n> This should take ~5 mins per branch.\r\n> \r\n> Otherwise, LGTM! Thanks for cleaning up after me :)\r\n\r\n\r\n\r\n> @patrickvonplaten I would measure impact of running cnn summarization (which uses seq_len 1024) to fully acknowledge the trade off you are making:\r\n> \r\n> ```shell\r\n> cd examples/seq2seq\r\n> python run_eval.py facebook/bart-large-cnn cnn_dm/test.source cnn_gens.txt \\\r\n> \t--reference_path cnn_dm/test.target \\\r\n> \t--score_path cnn_rouge.json --task summarization \\\r\n> \t--n_obs 500 --fp16\r\n> ```\r\n> \r\n> This should take ~5 mins per branch.\r\n> \r\n> Otherwise, LGTM! Thanks for cleaning up after me :)\r\n\r\nThanks for the command! What do you mean by \"per branch\"? ", "Also @sshleifer I didn't really manage to find a good marian command for fine-tuning. Can you by chance copy-paste a command that fine-tunes a marian model in ~30min to verify that fine-tuning works as expected?", "> @patrickvonplaten I would measure impact of running cnn summarization (which uses seq_len 1024) to fully acknowledge the trade off you are making:\r\n> \r\n> ```shell\r\n> cd examples/seq2seq\r\n> python run_eval.py facebook/bart-large-cnn cnn_dm/test.source cnn_gens.txt \\\r\n> \t--reference_path cnn_dm/test.target \\\r\n> \t--score_path cnn_rouge.json --task summarization \\\r\n> \t--n_obs 500 --fp16\r\n> ```\r\n> \r\n> This should take ~5 mins per branch.\r\n> \r\n> Otherwise, LGTM! Thanks for cleaning up after me :)\r\n\r\nI got this result:\r\n\r\n\r\n\r\n\r\non brutasse - does this look reasonable to you? took 2min30 ", "What I meant by \"per branch\" was to also run that command on master to facilitate comparison. Your `refactor-bart` output looks completely reasonable.\r\n\r\n#### Train Command\r\nreplace num_train_epochs=1 in this [./examples/seq2seq/builtin_trainer/train_distil_marian_enro.sh](https://github.com/huggingface/transformers/blob/master/examples/seq2seq/builtin_trainer/train_distil_marian_enro.sh). \r\n\r\n+ It should take 12-25 minutes on 1 GPU.\r\n+ I don't know the BLEU/timing to expect, you should again run on `master` and `bart-refactor` to compare.\r\n\r\n\r\n\r\n\r\n", "> @patrickvonplaten I would measure impact of running cnn summarization (which uses seq_len 1024) to fully acknowledge the trade off you are making:\r\n> \r\n> ```shell\r\n> cd examples/seq2seq\r\n> python run_eval.py facebook/bart-large-cnn cnn_dm/test.source cnn_gens.txt \\\r\n> \t--reference_path cnn_dm/test.target \\\r\n> \t--score_path cnn_rouge.json --task summarization \\\r\n> \t--n_obs 500 --fp16\r\n> ```\r\n> \r\n> This should take ~5 mins per branch.\r\n> \r\n> Otherwise, LGTM! Thanks for cleaning up after me :)\r\n\r\n\r\nApplied as much perf improvement as possible -> Time from master to this PR for the above command is reduced from ~2min10s to ~2min05s (ran three times) = 2.5% speed-up. Removed as many `contiguous()` operations as possible", "Training gives good/equal results to master. However, I see a 5% slow-down in training. `generation()` is as fast or faster than master, but training yields a slow-down of around 5% => so still investigating. Could be the masks", "Reran the fine-tuning script a couple of times on a gcp instance so that no other tasks can interfere and float masks are actually faster than boolean masks and and give more or less same results than previous bart model on master. Here the results: \r\n\r\n\r\nRefactor (no boolean masks)\r\n\r\n\r\n\r\nmaster\r\n\r\n\r\n", "good to merge for me" ]

[]

[]

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

[]

[]

[ "Installing transformers is also broken\r\n\r\n(test) rabeeh@gpu4:~/transformers/examples/seq2seq$ pip install git+https://github.com/huggingface/transformers.git\r\nCollecting git+https://github.com/huggingface/transformers.git\r\n Cloning https://github.com/huggingface/transformers.git to /tmp/pip-req-build-V7nNeF\r\n Installing build dependencies ... done\r\n Complete output from command python setup.py egg_info:\r\n Traceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/tmp/pip-req-build-V7nNeF/setup.py\", line 156\r\n entries = \"\\n\".join([f' \"{k}\": \"{v}\",' for k, v in deps.items()])\r\n ^\r\n SyntaxError: invalid syntax\r\n \r\n ----------------------------------------\r\nCommand \"python setup.py egg_info\" failed with error code 1 in /tmp/pip-req-build-V7nNeF/\r\n", "issue solved with python = 3.7" ]

[]

[]

[ "Also, could you add the command to run distributed training with GPUs with finetune_trainer in README? thanks ", "I tried with latest version of transformers on 4 gpu with distributed training \r\n\r\nrabeeh@gpu4:~/transformers/examples/seq2seq$ python -m torch.distributed.launch finetune.py --learning_rate=3e-5 --fp16 --gpus 4 --do_train --do_predict --n_val 1000 --val_check_interval 0.1 --data_dir wmt_en_ro --train_batch_size=1 --eval_batch_size=1 --output_dir=xsum_results --num_train_epochs 1 --model_name_or_path t5-smal\r\n\r\ngetting the following error, thanks \r\n\r\nfinetune.py: error: unrecognized arguments: --local_rank=0\r\nTraceback (most recent call last):\r\n File \"/opt/conda/envs/transformers/lib/python3.7/runpy.py\", line 193, in _run_module_as_main\r\n \"__main__\", mod_spec)\r\n File \"/opt/conda/envs/transformers/lib/python3.7/runpy.py\", line 85, in _run_code\r\n exec(code, run_globals)\r\n File \"/opt/conda/envs/transformers/lib/python3.7/site-packages/torch/distributed/launch.py\", line 260, in <module>\r\n main()\r\n File \"/opt/conda/envs/transformers/lib/python3.7/site-packages/torch/distributed/launch.py\", line 256, in main\r\n cmd=cmd)\r\nsubprocess.CalledProcessError: Command '['/opt/conda/envs/transformers/bin/python', '-u', 'finetune.py', '--local_rank=0', '--learning_rate=3e-5', '--fp16', '--gpus', '4', '--do_train', '--do_predict', '--n_val', '1000', '--val_check_interval', '0.1', '--data_dir', 'wmt_en_ro', '--train_batch_size=1', '--eval_batch_size=1', '--output_dir=xsum_results', '--num_train_epochs', '1', '--model_name_or_path', 't(trans(tr(transformers) \r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@rabeehk I also encountered the same problem as you. Have you solved it?" ]

[ { "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" } ]

[]

[ "Looks like this PR was unfortunately broken, so I'm going to close it. Also noting that the way to update a model card now is to update it directly in your model repo! see https://discuss.huggingface.co/t/announcement-all-model-cards-will-be-migrated-to-hf-co-model-repos/2755" ]

[ { "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" } ]

[]

[ "Looks like this PR was unfortunately broken, so I'm going to close it. Also noting that the way to update a model card now is to update it directly in your model repo! see https://discuss.huggingface.co/t/announcement-all-model-cards-will-be-migrated-to-hf-co-model-repos/2755" ]

[]

[]

[ "solved with implementing it inside the model of T5ForConditionalGeneration thanks" ]

[]

[]

[ "https://github.com/huggingface/transformers/issues/7675" ]

[]

[]

[ "Hi guys, most commits were from the previous PR (TFDPR). I do not know how to remove them, sorry!\r\nOnly `modeling_tf_rag.py` is new here.", "Awesome work @ratthachat!!!\r\n\r\nFor the input, I think the best way is to check how it is done in TF BERT for example and if you have difficulties to understand you can ask your questions here :)\r\n\r\nAbout the `.numpy()` as suggested you can use the `tf.make_ndarra()` like this `tf.make_ndarray(tf.make_tensor_proto(my_tensor))`.\r\n\r\nAbout the weights, I suggest you to check how it is done in TF BART there are similar weights naming.", "> Awesome work @ratthachat!!!\r\n> \r\n> For the input, I think the best way is to check how it is done in TF BERT for example and if you have difficulties to understand you can ask your questions here :)\r\n> \r\n> About the `.numpy()` as suggested you can use the `tf.make_ndarra()` like this `tf.make_ndarray(tf.make_tensor_proto(my_tensor))`.\r\n> \r\n> About the weights, I suggest you to check how it is done in TF BART there are similar weights naming.\r\n\r\nHi Julien @jplu , thanks for the reply!\r\nOn these points, I think I may miss something simple, but I could not solve the puzzle by myself at this moment.\r\n1) on `tf.make_ndarray(tf.make_tensor_proto(my_tensor))` , I could make it work **only** on eager mode too, so still have the problem in graph mode (I could not make it work inside @tf.function)\r\n\r\n2) about the input, yes I tried to replicate TF Bert & TF DPR (which is previously my implement) e.g. replace `inputs` with `input_ids` with/without default value (`None`) and use `input_processing` , but no matter what I tried , the simple call got error\r\n\r\n```\r\noutputs = model(inputs)\r\nValueError: The first argument to `Layer.call` must always be passed.\r\n```\r\n\r\nI really missed something simple here, I will try to play around again.\r\n\r\n3) about the weights, yes, I tried to replicate other TF models, so all weights loading works in `.from_pretrained_question_encoder_generator` and mostly properly loaded in `.from_pretrained`. \r\nOnly two aforementioned weights could not have the correct name . I tried various fixing to these minor cases but could not, except my very ugly manual load.", "1. Arf, I thought that this line would automatically deactivate the graph execution, but it is not the case. So to make it short, convert a graph tensor to numpy array is not possible because the graph does not execute in Python - so there is no numpy at graph execution. The only one work around would be to play with [tf.py_function](https://www.tensorflow.org/api_docs/python/tf/py_function) but you will literally kill the perf with this, even though it is your only way to go to.\r\n\r\n2. This error means that you are not passing the first argument to a call method (basically don't pass any `input_ids`), you always have to pass it. Where in the code the error is raised?\r\n\r\n3. What did you try more precisely?", "Hi again Julien,\r\n\r\n1. I see! Let us work only on eager mode for now and come back later.\r\n\r\n2. Here's 4.1.0 colab where I try to modify 4.1.0 input on `TFRagModel` (Cell 6) \r\nhttps://colab.research.google.com/drive/1RvtOxUIravWEkwMnj48pedv2mFnlWYkY?usp=sharing\r\nPlease see Cell 9, to see various ways I try to pass the first argument. Really sorry I think I overlooked some simple things here.\r\n\r\n3. I tried to adjust `base_prefix_name`, and also setting module `name` as discussed with Sam [here](https://discuss.huggingface.co/t/solved-issue-on-translating-dpr-to-tfdpr-on-loading-pytorch-weights-to-tf-model/1764/2): \r\n", "1. The problem is that you cannot do anything with the model if it cannot be run in graph mode (no serving, no training, no optimization, very slow) 😢 \r\n2. Please, see how it is done in all the other TF implementations, the way you handle the inputs in `modeling_tf_rag.py` is wrong.\r\n3. I don't really have time this week to go deeper in this, but I will take some time on Monday to do it!\r\n\r\nA test file is missing as well, having it might help you to detect what has to be updated :)", "Hi Julien!\r\n\r\n> 1. The problem is that you cannot do anything with the model if it cannot be run in graph mode (no serving, no training, no optimization, very slow) 😢\r\n\r\nI got it. I think we can do some work around on TFRag model training in graph mode. Instead of fitting the model with `input_ids` which will need `.numpy` and `retriever` , we can do all retriever stuffs offline (or with tf.Dataset) first to get the `context_input_ids` and then feed them directly to the trainnig loop. I will test this idea.\r\n\r\n> 2. Please, see how it is done in all the other TF implementations, the way you handle the inputs in `modeling_tf_rag.py` is wrong.\r\n\r\nFinally, I was able to find a single line that's wrong :) I will update the file in my local.\r\n\r\n> 3. I don't really have time this week to go deeper in this, but I will take some time on Monday to do it!\r\n> A test file is missing as well, having it might help you to detect what has to be updated :)\r\n\r\nThanks Julien. I will write the full test file soon. My reason is that I need some suggestions (as commented in the modeling file) for my current draft at this stage , so I have not written the full-fledge test file yet. The colab I posted did have some 10+ basic tests, which I still think I need to (cleanly) pass all these tests first. \r\n", "> I got it. I think we can do some work around on TFRag model training in graph mode. Instead of fitting the model with input_ids which will need .numpy and retriever , we can do all retriever stuffs offline (or with tf.Dataset) first to get the context_input_ids and then feed them directly to the trainnig loop. I will test this idea.\r\n\r\nI was thinking the same :) to do the search offline, might be a solution. I think the very long term solution would be to use something more adapted to TensorFlow than FAISS such as [SCANN](https://github.com/google-research/google-research/tree/master/scann). (Pinging @lhoestq to know what he is thinking about this :) )\r\n\r\n> Finally, I was able to find a single line that's wrong :) I will update the file in my local.\r\n\r\nNice!! Just to let you know that Friday we have also updated the way the handle the booleans, so be careful to integrate this as well.\r\n\r\n> Thanks Julien. I will write the full test file soon. My reason is that I need some suggestions (as commented in the modeling file) for my current draft at this stage , so I have not written the full-fledge test file yet. The colab I posted did have some 10+ basic tests, which I still think I need to (cleanly) pass all these tests first.\r\n\r\nIt is ok, we are not in a hurry, take your time ^^", "@ratthachat - I think you're on the correct track here :-)\r\n\r\nTFRag will actually be the first TF composite model, so I'm quite certain you'll run into problems here that we haven't seen before. \r\nIn general I think:\r\n\r\n1) We should try to make the `from_pretrained` method work in a nice way (this will actually also show us how `TFEncoderDecoder` could be implemented). I'll take a look at this :-) \r\n\r\n2) Make `TFRagTokenForGeneration` work with integration tests that the model behaves the same way as PT using the faiss index. It's fine for me if it works only in eager mode for now. Maybe we can think about a different solution at a later stage if it's impossible to have RAG + Faiss in graph mode. I think it's a bit out-of-the-scope to integrate SCANN here with RAG.\r\n\r\n3) Add the other functionalities. \r\n\r\nI'll try to help you with 1) here - will add some commits to your PR.", "> We should try to make the from_pretrained method work in a nice way (this will actually also show us how TFEncoderDecoder could be implemented). I'll take a look at this :-)\r\n\r\nI would like to remove all the `from_pretrained` calls from the model implementation, it will raises issues for some usage, such as training.\r\n\r\n> Make TFRagTokenForGeneration work with integration tests that the model behaves the same way as PT using the faiss index. It's fine for me if it works only in eager mode for now. Maybe we can think about a different solution at a later stage if it's impossible to have RAG + Faiss in graph mode. I think it's a bit out-of-the-scope to integrate SCANN here with RAG.\r\n\r\nThe problem here is that if the model runs only in eager mode, the model won't be able to be served properly and then becomes useless. I don't see the point to have a model that runs only in your console locally :( The best solution IMHO would be to run the FAISS search offline.", "Hey @ratthachat, \r\n\r\nI think the `from_pretrained()` functionality now works as expected. I removed some hacks and we shouldn't have to define any `from_pretrained()` method actually.\r\n\r\nI've added two tests. 1 already passes (great job! - I didn't really change anything here...), the other one (a very difficult one based on `generate()` does not pass yet). It'll be quite difficult to make the other one pass, but if you manage you'll certainly have an in-depth knowledge of how `generate()` works.\r\n\r\nI would recommend the following next steps for the PR:\r\n\r\n1) Implement @jplu's simplified handling of the inputs as proposed in this PR: https://github.com/huggingface/transformers/pull/8602 . This should remove a lot of boiler plate code. If some params are not supported yet, I'm sure @jplu can help. \r\n\r\n2) Make the generate test pass\r\n\r\nAfter this I'm happy to take another look :-) \r\n\r\nLemme know if you have any problems with the weight loading. It all worked nicely for me", "> Implement @jplu's simplified handling of the inputs as proposed in this PR: #8602 . This should remove a lot of boiler plate code. I some params are not supported yet, I'm sure @jplu can help.\r\n\r\nI will be happy to help. Which ones are the \"not supported yet\"?", "> > Implement @jplu's simplified handling of the inputs as proposed in this PR: #8602 . This should remove a lot of boiler plate code. If some params are not supported yet, I'm sure @jplu can help.\r\n> \r\n> I will be happy to help. Which ones are the \"not supported yet\"?\r\n\r\n\r\n\r\n> > Implement @jplu's simplified handling of the inputs as proposed in this PR: #8602 . This should remove a lot of boiler plate code. If some params are not supported yet, I'm sure @jplu can help.\r\n> \r\n> I will be happy to help. Which ones are the \"not supported yet\"?\r\n\r\nI think it should all work perfectly fine! Sorry, this came across a bit bad - meant to say \"just in case\" something doesn't work don't hesitate to ping you ;-) Wasn't sure if int inputs like `n_docs` are supported, but I think this was added as well - so it should all work fine :-) ", "Thanks so much for your great help, Patrick! I will carefully look in each point you made. Full addressing will take a while, but I will be back. For now I have some initial responses: \r\n\r\n- (Need help the most) Unfortunately ,there's still a bug in weight loading if removing the hack (please see details in below thread)\r\n\r\n- About graph mode & training, I think we can consistently combine we three's thoughts here by (a) Finish the code in eager mode first -- (b) Make minimal changes to support offline-mode for retriever (or maybe not change anything at all) -- (c) Make a community notebook to guide this offline-retrieved training in graph mode. -- (d) SCANN will be an interesting long-term solution we can discuss after all these stuffs.\r\n\r\n- I think there is similar graph-retrieval problem also in TFDPR training (which we haven't tested) , so I will also try make some example notebook to train TFDPR in graph mode using this offline principle.\r\n\r\n- May I ask what is the meaning of these original Pytorch's 3 lines? (in `def generate()` )\r\n```\r\n # retrieved_doc_embeds = retrieved_doc_embeds.to(question_hidden_states)\r\n # context_input_ids = context_input_ids.to(input_ids)\r\n # context_attention_mask = context_attention_mask.to(input_ids)\r\n```\r\n\r\n- About test on `generate`, I will try give it a shot. BTW, I previously test `Bart` vs. `TFBart` and found out that they produce **\"different\"** `generate` results as well. \r\nDo you have the same experience, and will this affect RAG `generate` test ??\r\n", "> I think the `from_pretrained()` functionality now works as expected. I removed some hacks and we shouldn't have to define any `from_pretrained()` method actually.\r\n> \r\n> Lemme know if you have any problems with the weight loading. It all worked nicely for me\r\n\r\nHi Patrick, @patrickvonplaten \r\nUnfortunately, I found the same bug prior to my hack. \r\n\r\n- `from_pretrained_question_encoder_generator` <-- Work great\r\n\r\n- `from_pretrained` <-- **BUG** (only on **_2 weights_**) : `['model.shared.weight', 'final_logits_bias']`\r\n\r\n```\r\nSome weights or buffers of the TF 2.0 model TFRagTokenForGeneration were not initialized from the PyTorch model and are newly initialized: ['model.shared.weight', 'final_logits_bias']\r\n```\r\n\r\n- (New found) local loading `from_pretrained(\"./rag\")` <-- **BUG** on **all weights**\r\n\r\nPlease (please :) take a look at this new colab which provides \"minimal\" code to show the bugs (just 8 cells).\r\nhttps://colab.research.google.com/drive/1s-j9PB9yzrFsL6q5rZUQyf8_Lt6jDAkL?usp=sharing\r\nBugs only in the last 2 cells.", "> > I think the `from_pretrained()` functionality now works as expected. I removed some hacks and we shouldn't have to define any `from_pretrained()` method actually.\r\n> > Lemme know if you have any problems with the weight loading. It all worked nicely for me\r\n> \r\n> Hi Patrick, @patrickvonplaten\r\n> Unfortunately, I found the same bug prior to my hack.\r\n> \r\n> * `from_pretrained_question_encoder_generator` <-- Work great\r\n> * `from_pretrained` <-- **BUG** (only on **_2 weights_**) : `['model.shared.weight', 'final_logits_bias']`\r\n\r\nThis is not a bug. It's fine actually. Those weights are handled differently in TF and PT, so this message is expected.\r\n\r\n> \r\n> ```\r\n> Some weights or buffers of the TF 2.0 model TFRagTokenForGeneration were not initialized from the PyTorch model and are newly initialized: ['model.shared.weight', 'final_logits_bias']\r\n> ```\r\n> \r\n> * (New found) local loading `from_pretrained(\"./rag\")` <-- **BUG** on **all weights**\r\n\r\nLet me look into this!\r\n\r\n> \r\n> Please (please :) take a look at this new colab which provides \"minimal\" code to show the bugs (just 8 cells).\r\n> https://colab.research.google.com/drive/1s-j9PB9yzrFsL6q5rZUQyf8_Lt6jDAkL?usp=sharing\r\n> Bugs only in the last 2 cells.", "@ratthachat,\r\n\r\nI think you can ignore those warnings for now. A good next step to make sure that the `from_pretrained()` methods work correctly is to add tests that verify that after saving/loading the model yields the same output as before: \r\n- https://github.com/huggingface/transformers/blob/9d7d0005b046a95d9d59354714bb6c3547a612fe/tests/test_modeling_rag.py#L900\r\n\r\nI checked and the following code works fully as expected:\r\n\r\n```\r\nfrom transformers import RagRetriever, TFRagTokenForGeneration\r\n\r\nretriever = RagRetriever.from_pretrained(\"facebook/rag-token-nq\", index_name=\"exact\", use_dummy_dataset=True)\r\nmodel = TFRagTokenForGeneration.from_pretrained(\"facebook/rag-token-nq\", from_pt=True, retriever=retriever)\r\n\r\nmodel.save_pretrained(\"./rag\")\r\nmodel = TFRagTokenForGeneration.from_pretrained(\"./rag\", retriever=retriever)\r\n```\r\n\r\nAll those commands work as they should, so I think we're good for now with the `from_pretrained()`. I think the next step should be to concentrate on removing the TF input boilerplate code and then making the generation 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." ]

[]

[]

[ "Trainer does indeed not work in distributed fashion with iterative datasets. You need to convert your iterative dataset to a regular dataset for the time being.", "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 @ZhaoQianfeng,\r\n\r\nThanks a lot for making the PR. I thought about this a bit and I think we don't have to change anything actually.\r\n\r\nThe reason is the following. Let's say you want to generate up to a length of 5.\r\n\r\nBOS is the start token which should be counted as part of the input length. However EOS should not be counted towards the sequence length for the length penalty IMO since it's the trigger to finish generation. \r\n\r\nSo the input:\r\n\r\n[BOS, hey, there, EOS] -> is ok for me to be counted as a sequence length of 3 ([BOS, Hey, there]) for the length penalty. I don't think the EOS token itself should penalize.\r\nHowever, un unfinished generation, such as [BOS, hey, there, how, are] should be counted to have a sequence length of 5 since it isn't finished.\r\n\r\nIf we would merge this PR, this would mean that [BOS, hey, there, peter, EOS] would receive the same length penalty as [BOS, hey, there, how, are], but IMO they should not. The first sequence is finished (*i.e.* shorter) than the second one.\r\n\r\nSo I'd prefer to leave it as it is. I think it's the right approach. Thanks a lot for looking into this however :-) ", "Hey @patrickvonplaten ,\r\nI think you are right!\r\n\r\nFor your example, my reason why ` [BOS, hey, there, peter, EOS]` and `[BOS, hey, there, how, are] `should have same length penalty is that the former probability is calculated by `log(P(hey))+log(P(there))+log(P(peter))+log(P(EOS))`, and the latter probability is calculated by `log(P(hey))+log(P(there)+log(P(how)+log(P(are))`, both 4 elements.So I used to think that they should be divided by same length.\r\n\r\nBut I think your explanation is more reasonable and convincing, the former sentence is actually shorter than the latter sentence!It should be what **length peanalty** really means. Thank you for taking the time to discuss this issue!:-)" ]

[]

[]

[ "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 would guess this is a memory error. Have you tried monitoring the memory available on your GPUs while the training is running?", "A CUDA device-side assert triggered means a bad index error somewhere, and persists until you restart your kernel. The code you provide does not allow us to reproduce the bug because it uses a tokenizer and a datasets we don't have access to. There are thus multiple reasons for a bad index error. If you want us to help, you'll need to give a reproducer using a pretrained tokenizer of the hub and a dataset on the hub (for instance GLUE MRPC is great since it's tiny).\r\n\r\nTo debug your problem locally:\r\n```\r\nfor batch in trainer.get_train_dataloader():\r\n break\r\nmodel.cpu()(**batch)\r\n```\r\nas on the CPU you will get a clear indication of where the index error is.", "@sgugger I'll work on making the datasets public and will post here. In the meantime, I'll run your snippet. \r\n\r\n@LysandreJik , it's not a memory issue - all four GPUs are at ~87% volatile util for the duration.", "@sgugger @LysandreJik I've made our bucket public, and the relevant material is in gs://bao-ai/transfer; you should be able to pull stuff down in jupyter via:\r\n`!gsutil -m cp -r gs://bao-ai/transfer .`\r\n... though I haven't tried that command on an unauthenticated computer.\r\n\r\nAlso, note I'm having the same issue in Colab ([notebook here](https://colab.research.google.com/drive/1y-Tgl_zPJzjrzsq3WeYeUFL9A6hGhLhD?usp=sharing)), so I suspect it's an issue with the dataset as suggested above. @sgugger could you elaborate on what sort of issues you mean by 'bad index'? Would rebuilding the dataset from our source files help? If so, are there steps I can take to make sure indexing issues don't arise?", "Note I also periodically get the following error messages at train time:\r\n```\r\n/home/b/anaconda3/envs/transformers_3/lib/python3.7/site-packages/torch/nn/parallel/_functions.py:61: UserWarning: Was asked to gather along dimension 0, but all input tensors were scalars; will instead unsqueeze and return a vector.\r\n warnings.warn('Was asked to gather along dimension 0, but all '\r\n```\r\nThat's copied directly - it cuts off like that in Jupyter.", "I rebuilt the dataset using the following script, reloaded, and trained - still breaking with the same error message. Here's the rebuild code:\r\n```\r\nbalanced = pickle.load(open('./balanced_ds.pickle', 'rb'))\r\nbal_df = datasets.Dataset.from_pandas(pd.DataFrame.from_records(balanced, columns = ['txt', 'labels']))\r\n\r\nBLOCK_SIZE = 512\r\ntok = RobertaTokenizerFast.from_pretrained(\"./art_tok_onefile_roberta_tuned/\")\r\n\r\nds_tokenized_no_special = bal_df.map(lambda example: tok(example['txt'], \r\n padding='max_length', \r\n max_length=BLOCK_SIZE, \r\n truncation=True,\r\n add_special_tokens = False), batched=True)\r\n\r\nds_tokenized_no_special.save_to_disk('./art_unit_tokenized_balanced_rebuild')\r\n```\r\n\r\nThis uses the same imports (probably redundantly) as the main script. You can access all the data using `!gsutil -m cp -r gs://bao-ai/transfer .`, same as above.\r\n\r\n\r\nI'm going to loop through all the data in the dataloader and see if it's returning anything janky. We're expecting tensors of size BATCHxBLOCK_SIZE for attention_mask and input_ids, and a tensor of size BATCHx1 for labels, right? Our labels are currently just 0 or 1, depending on whether a tokenized json document falls within a certain document class (art unit 3600, to be precise - this work is for patent law analysis).", "Note, I found experimentally that special tokens have to be removed from the tokenizer in order to be properly passed through the RobertaForSequenceClassification model; otherwise, we get an index error in the torch.nn.embedding step, due to the vocabulary exceeding the Roberta vocab size by 2.", "On Colab, before the trainer crashes, I get lots of these messages in the runtime logs:\r\n```/pytorch/aten/src/ATen/native/cuda/Indexing.cu:658: indexSelectLargeIndex: block: [374,0,0], thread: [19,0,0] Assertion `srcIndex < srcSelectDimSize` failed.```", "Ran the following checks on the entire dataset - it passed. Training still fails. This would suggest to me that the dataset is not actually the issue:\r\n```\r\ndef no_nans(t):\r\n return bool(t.flatten().isnan().sum() == 0)\r\n\r\ndef check_batch(ex):\r\n masks = ex['attention_mask']\r\n masks_expected_shape = torch.Size([BATCH_SIZE, BLOCK_SIZE])\r\n masks_are_expected_shape = (masks.shape == masks_expected_shape)\r\n\r\n allowed_masks = [0,1]\r\n only_allowed_masks = all(i in allowed_masks for i in masks.flatten())\r\n\r\n masks_not_nan = no_nans(masks)\r\n\r\n masks_valid = (masks_are_expected_shape and only_allowed_masks and masks_not_nan)\r\n\r\n ids = ex['input_ids']\r\n ids_expected_shape = torch.Size([BATCH_SIZE, BLOCK_SIZE])\r\n ids_are_expected_shape = (ids.shape == ids_expected_shape)\r\n\r\n ids_within_vocab_range = ((ids.max() < tok.vocab_size + 4) and (ids.min() >= 0))\r\n\r\n ids_not_nan = no_nans(ids)\r\n\r\n ids_valid = (ids_are_expected_shape and ids_within_vocab_range)\r\n\r\n\r\n labels = ex['labels']\r\n allowed_labels = [0,1]\r\n only_allowed_labels = all(i in allowed_labels for i in labels.flatten())\r\n\r\n labels_are_expected_shape = labels.shape == torch.Size([BATCH_SIZE])\r\n\r\n labels_not_nan = no_nans(labels)\r\n\r\n labels_valid = (only_allowed_labels and labels_are_expected_shape and labels_not_nan)\r\n\r\n failures = {\r\n 'masks_are_expected_shape': masks_are_expected_shape,\r\n 'only_allowed_masks': only_allowed_masks,\r\n 'masks_not_nan': masks_not_nan,\r\n #'masks_valid': masks_valid,\r\n 'ids_are_expected_shape': ids_are_expected_shape,\r\n 'ids_within_vocab_range': ids_within_vocab_range,\r\n 'ids_not_nan': ids_not_nan,\r\n #'ids_valid': ids_valid,\r\n 'only_allowed_labels': only_allowed_labels,\r\n 'labels_are_expected_shape': labels_are_expected_shape,\r\n 'labels_not_nan': labels_not_nan,\r\n #'labels_valid': labels_valid\r\n }\r\n\r\n \r\n return ((masks_valid and ids_valid and labels_valid), ex, failures)\r\n\r\nfailed = []\r\nfor idx, ex in tqdm(enumerate(iter(loader)), total=len(loader)):\r\n passed, ex, fail = check_batch(ex)\r\n if not passed:\r\n print(f'{idx} Failed!')\r\n failed.append((idx, ex, fail))\r\n```\r\n\r\n(This, again, uses all the above code as a base).", "I also deactivated the train test split and deleted the cache files in the dataset - also fails.", "Your check for ids in the proper change seems incorrect:\r\n```\r\nids_within_vocab_range = ((ids.max() < tok.vocab_size + 4) and (ids.min() >= 0))\r\n```\r\nwill allow for the indices `tok.vocab_size` to `tok.vocab_size+3` which are all going to generate an index error given the fact your model as `vocab_size = tok.vocab_size`.\r\n", "@sgugger vocab indices +1 thru +4 are the special tokens, though, right? And the model should be able to accept them, right?", "*or +0 thru +3 if we're being pythonic with our indexing", "The highest token index in the entire dataset is the pad token. That shouldn't throw an indexing error. Can you reproduce on your end?", "> Your check for ids in the proper change seems incorrect:\r\n> \r\n> ```\r\n> ids_within_vocab_range = ((ids.max() < tok.vocab_size + 4) and (ids.min() >= 0))\r\n> ```\r\n> \r\n> will allow for the indices `tok.vocab_size` to `tok.vocab_size+3` which are all going to generate an index error given the fact your model as `vocab_size = tok.vocab_size`.\r\n\r\nWhat should the proper check for this step be?", "And regardless of the checks - what would token indices have to do with the ultimate error, which is in clip_grad.py? \r\n\r\nTruncated from above:\r\n```\r\nRuntimeError Traceback (most recent call last)\r\n<ipython-input-11-3435b262f1ae> in <module>()\r\n----> 1 trainer.train()\r\n\r\n1 frames\r\n/usr/local/lib/python3.6/dist-packages/torch/nn/utils/clip_grad.py in clip_grad_norm_(parameters, max_norm, norm_type)\r\n 36 total_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]), norm_type)\r\n 37 clip_coef = max_norm / (total_norm + 1e-6)\r\n---> 38 if clip_coef < 1:\r\n 39 for p in parameters:\r\n 40 p.grad.detach().mul_(clip_coef.to(p.grad.device))\r\n\r\nRuntimeError: CUDA error: device-side assert triggered\r\n```", "I faced similar issue while running on colab with Linux OS . Ttried restarting and resetting the kernal error disappeared . ", "@shivaraj1994 can you define 'kernel'? Do you mean the Jupyter Kernel, the python Kernel, or the Linux/Mac/Windows kernel? \r\n\r\nThis problem was replicated both on Linux and on a Colab account; I don't think it's an issue with the operating system of a given computer. \r\n\r\n**NOTE:** I was able to fix the problem by ditching the datasets library and using the older pytorch dataset paradigm. Full code below (split between three files - note, I'm not importing the HF datasets library; rather, I'm importing a custom module called datasets.py, from the same directory in which I'm running train.py):\r\n\r\n# train.py\r\n```\r\nfrom transformers import (RobertaTokenizerFast,\r\n RobertaForSequenceClassification,\r\n RobertaConfig)\r\n\r\nfrom transformers import Trainer, TrainingArguments\r\n\r\nimport pickle\r\n\r\nfrom datasets import BalancedDataset\r\nfrom collators import DataCollatorForDocumentClassificationBATCH\r\n\r\n\r\n'''\r\nTODO:\r\nUse custom tok (w/ special tokens removed)\r\nTrain more layers\r\nUse longformer as drop-in for Roberta\r\nMake nicer dataset (with train/test split, etc...)\r\n'''\r\n\r\nBLOCK_SIZE = 512\r\nBATCH_SIZE = 32\r\n\r\nbalanced = pickle.load(open('./balanced_ds.pickle', 'rb'))\r\n\r\n\r\ntok = RobertaTokenizerFast.from_pretrained('roberta-base') #\"./art_tok_onefile_roberta_tuned/\")\r\n\r\nbal_ds = BalancedDataset(tok, balanced, BLOCK_SIZE)\r\n\r\ncollator = DataCollatorForDocumentClassificationBATCH()\r\n\r\n\r\nconfig = RobertaConfig.from_pretrained(\"roberta-base\",\r\n vocab_size=tok.vocab_size,\r\n max_position_embeddings=514,\r\n num_labels = 2)\r\n\r\nmodel = RobertaForSequenceClassification.from_pretrained('roberta-base',\r\n config=config)\r\n\r\n\r\n#Disable training on all but the Classification Head!\r\nfor param in model.base_model.parameters():\r\n param.requires_grad = False\r\n\r\ntraining_args = TrainingArguments(\r\n output_dir=\"./roberta_train_test\",\r\n overwrite_output_dir=True,\r\n num_train_epochs=5,\r\n per_device_train_batch_size=BATCH_SIZE,\r\n save_steps=150,\r\n save_total_limit=2,\r\n logging_steps=20,\r\n max_grad_norm = 5,\r\n dataloader_num_workers = 15,\r\n #fp16 = True #Enable low-precision via AMP\r\n)\r\n\r\n\r\ntrainer = Trainer(\r\n model = model,\r\n args = training_args,\r\n data_collator = collator,\r\n train_dataset = bal_ds\r\n)\r\n\r\ntrainer.train()\r\n```\r\n\r\n# datasets.py\r\n```\r\nimport torch\r\nfrom torch.utils.data.dataset import Dataset\r\nfrom tqdm import tqdm\r\n\r\n\r\n\r\nclass BalancedDataset(Dataset):\r\n def __init__(self, tokenizer, data, block_size: int, limit=None):\r\n self.block_size = block_size\r\n self.tok = tokenizer\r\n\r\n print('Ingesting data!')\r\n # Load Data\r\n self.txt = [i[0] for i in tqdm(data[:limit])]\r\n self.labels = torch.tensor([i[1] for i in tqdm(data[:limit])])\r\n\r\n def __len__(self):\r\n return len(self.txt)\r\n\r\n def __getitem__(self, item):\r\n d = self.tok(self.txt[item], padding='max_length',\r\n truncation=True, max_length=self.block_size,\r\n return_tensors='pt')\r\n d['labels'] = self.labels[item]\r\n return d\r\n```\r\n\r\n# collators.py\r\n```\r\nfrom dataclasses import dataclass\r\nfrom typing import Dict, List, Union\r\n\r\nimport torch\r\n\r\n\r\n@dataclass\r\nclass DataCollatorForDocumentClassificationBATCH:\r\n def __call__(\r\n self, examples: List[Union[List[int], torch.Tensor, Dict[str, torch.Tensor]]]\r\n ) -> Dict[str, torch.Tensor]:\r\n return {\r\n 'input_ids': torch.stack([e['input_ids'] for e in examples]).squeeze(),\r\n 'attention_mask': torch.stack([e['attention_mask'] for e in examples]).squeeze(),\r\n 'labels': torch.stack([e['labels'] for e in examples]),\r\n }\r\n\r\n```", "I will keep the bao-ai bucket open to the public for a bit longer so y'all can attempt to replicate the original issue. We still need to figure out what was causing the clip_grad_norm issue in the first place.\r\n\r\n", "> Also, note I'm having the same issue in Colab ([notebook here](https://colab.research.google.com/drive/1y-Tgl_zPJzjrzsq3WeYeUFL9A6hGhLhD?usp=sharing)), so I suspect it's an issue with the dataset as suggested above. @sgugger could you elaborate on what sort of issues you mean by 'bad index'? Would rebuilding the dataset from our source files help? If so, are there steps I can take to make sure indexing issues don't arise?\r\n\r\n**Replication note:** I continued to debug in that Colab notebook; if you want to replicate the original issue, you'll need to use the old code, not the code that currently exists at the end of that link.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "maybe related to #8793 , hope could help.", "> maybe related to #8793 , hope could help.\r\n\r\nseems to be related. I get high variance in accuracy, I guessed it was probably because of the random initialization of those two weights.", "If you're using the `run_mlm.py`, then you're doing masked language modeling with the `BertForMaskedLM` model. This model does not make use of the pooler, hence why those two layers are randomly initialized. They're not used for predictions or training.", "@LysandreJik would it make more sense to load the saved model using BertModel.load_pretrained(saved_mlm_model) or would it be better to use BertModel.load(\"bert-base-uncased\") and copy the weights over from the saved model? ", "I also met this problem. Have you solved it?", "@wenHK It's actually not really relevant. The BertForMaskedLM model doesn't use the pooler layer, so thus why there are no weights assigned. You don't really need to worry about the warning.", "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." ]

[]

[]

[ "Unfortunately, if simply printing the string is impossible, this is out of our expertise. You have probably already seen those threads, but they may help you debug what's going on:\r\n\r\nhttps://stackoverflow.com/questions/27366479/python-3-os-walk-file-paths-unicodeencodeerror-utf-8-codec-cant-encode-s\r\nhttps://stackoverflow.com/questions/38147259/how-to-work-with-surrogate-pairs-in-python\r\n\r\nLet us know if you find an answer!", "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." ]

[]

[]

[ "There has been no traction so far on the circleci forums, I filed a support ticket with cirlceci.", "So `pipeline.git.base_revision` is consistently undefined when making a PR via direct file edit on github.\r\n", "So far so good.\r\n\r\nAnd while monitoring I discovered an interesting thing. In this particular PR my check doesn't actually do what I thought it did. It doesn't check the range of commits from the beginning of PR. The range it checks is actually just for the last commit. That `pipeline.git.base_revision` is very unruly.\r\n \r\nYou can see a good example of it here: https://github.com/huggingface/transformers/pull/8918\r\n\r\nIf you look at the checks for the last few commits which are doc-only commits - the jobs are skipped, whereas any commit that had code in it is not skipped.\r\n\r\nSo actually this is better than what I intended. Since if we check the full range and there are code files and then there is a subsequent commit that has only docs changed in my vision it'd run the jobs normally. But this is better! Since this checks each commit and decides whether to run the jobs or not based on just that commit - this is much more efficient than my original intention.\r\n\r\nI hope I explained it clearly.\r\n\r\n**edit** Hmm, but what happens if several commits are pushed at once - which file range will it check - since normally it checks just the last commit - this I'm not sure about. `pipeline.git.base_revision` is a wild card it seems.", "Mmmm, that does mean that if a PR changes code then has a last commit that only changes the doc, it will appear green to us, correct?\r\nIf so, we should fine a way to correct this behavior as it will lull us (and the user) in a false sense that everything is alright.", "I will run tests once github works again and adjust accordingly.\r\n\r\nI'm also in touch with an engineer at circleCI via their support - so hopefully we will get some solid answers rather than needing to validate all the different circumstances.", "I wasn't able to reproduce it, but it's very clear that it happened, and this is not what we want.\r\n\r\nAnd while what I wrote here https://github.com/huggingface/transformers/pull/8885#issuecomment-738583812 is super-cool, it can't work since github relies on the last check for the overall status. So, we can only skip a job if *all* files in PR were docs.\r\n\r\nSo I merged a change which disabled that struggling new feature, but added a log instead to continue monitoring it while waiting for circleCI support to get back to me." ]

[]

[]

[]

[]

[]

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

[]

[]

[ "Thank you for adding this new property, @sgugger - it has indeed improved the readability!" ]

[]

[]

[ "Thanks!" ]

[]

[]

[ "MobileBERT does this in the `tie_weights` function. Should we do the same here?", "ALBERT also does it in the `_resize_token_embeddings`:\r\n\r\nhttps://github.com/huggingface/transformers/blob/a7d46a060930242cd1de7ead8821f6eeebb0cd06/src/transformers/models/albert/modeling_albert.py#L635-L639\r\n\r\nIt probably should have been done in that method for MobileBERT as well", "Fine by me :-) Should we do the mobileBERT change in this PR?", "> Fine by me :-) Should we do the mobileBERT change in this PR?\r\n\r\nwill do!", "@patrickvonplaten Not 100% sure that this fix works, forked and implemented the fixes as shown and receive the following error upon training:\r\n```\r\n Expected tensor for argument #1 'indices' to have scalar type Long; but got torch.cuda.FloatTensor instead (while checking arguments for embedding). \r\n```\r\nIf I revert back to transformers-4.0.0 and run exactly the same script, with the same data then T5 trains fine. Could be user error, although I hope not! Did look back through the script and try and diagnose, but no luck as of yet.", "> @patrickvonplaten Not 100% sure that this fix works, forked and implemented the fixes as shown and receive the following error upon training:\r\n> \r\n> ```\r\n> Expected tensor for argument #1 'indices' to have scalar type Long; but got torch.cuda.FloatTensor instead (while checking arguments for embedding). \r\n> ```\r\n> \r\n> If I revert back to transformers-4.0.0 and run exactly the same script, with the same data then T5 trains fine. Could be user error, although I hope not! Did look back through the script and try and diagnose, but no luck as of yet.\r\n\r\nCan you attach a simple code snippet showing what code produces your error? It's for T5 no? ", "> \r\n> \r\n> > @patrickvonplaten Not 100% sure that this fix works, forked and implemented the fixes as shown and receive the following error upon training:\r\n> > ```\r\n> > Expected tensor for argument #1 'indices' to have scalar type Long; but got torch.cuda.FloatTensor instead (while checking arguments for embedding). \r\n> > ```\r\n> > \r\n> > \r\n> > If I revert back to transformers-4.0.0 and run exactly the same script, with the same data then T5 trains fine. Could be user error, although I hope not! Did look back through the script and try and diagnose, but no luck as of yet.\r\n> \r\n> Can you attach a simple code snippet showing what code produces your error? It's for T5 no?\r\n\r\nSure:\r\n\r\n```python\r\nfrom transformers import T5TokenizerFast, T5ForConditionalGeneration\r\ndev = \"cuda\"\r\n\r\nMODEL_NAME = 'google/t5-v1_1-base'\r\ntokenizer = T5TokenizerFast.from_pretrained('t5-base')\r\nspecial_tokens_dict = {'additional_special_tokens': ['<ORG>','<PERSON>']}\r\nnum_added_tokens = tokenizer.add_special_tokens(special_tokens_dict)\r\nprint(f'ADDED TOKENS: {num_added_tokens}')\r\nmodel = T5ForConditionalGeneration.from_pretrained(MODEL_NAME)\r\nmodel.resize_token_embeddings(len(tokenizer))\r\nmodel.to(dev)\r\nBATCH_SIZE = 8\r\n```\r\n\r\n```python\r\n#Sets the module in training mode\r\nfrom IPython.display import HTML, display\r\ndef progress(loss,value, max=100):\r\n return HTML(\"\"\" Batch loss :{loss} <progress \r\nvalue='{value}'max='{max}',style='width: 100%'>{value}\r\n </progress> \"\"\".format(loss=loss,value=value, max=max))\r\n\r\nmodel.train()\r\nnum_of_batches= int(len(train_df) / BATCH_SIZE)\r\nprint(num_of_batches)\r\nNUM_EPOCHS = 1\r\nloss_per_10_steps=[]\r\nloss_values = []\r\nfor epoch in range(1,NUM_EPOCHS+1):\r\n print('Running epoch: {}'.format(epoch))\r\n \r\n running_loss=0\r\n\r\n out = display(progress(1, num_of_batches+1), display_id=True)\r\n for i in range(num_of_batches):\r\n inputbatch=[]\r\n labelbatch=[]\r\n new_df=train_df[i*BATCH_SIZE:i*BATCH_SIZE+BATCH_SIZE]\r\n for indx,row in new_df.iterrows():\r\n input = 'Product: '+row['product_name']\r\n labels = row['product_description']\r\n inputbatch.append(input)\r\n labelbatch.append(labels)\r\n inputbatch=tokenizer.batch_encode_plus(inputbatch,padding=True, max_length=512,return_tensors='pt')[\"input_ids\"]\r\n labelbatch=tokenizer.batch_encode_plus(labelbatch,padding=True, max_length=512,return_tensors=\"pt\") [\"input_ids\"]\r\n inputbatch=inputbatch.to(dev)\r\n labelbatch=labelbatch.to(dev)\r\n\r\n # clear out the gradients of all Variables \r\n optimizer.zero_grad()\r\n\r\n # Forward propogation\r\n outputs = model(input_ids=inputbatch, labels=labelbatch)\r\n loss = outputs.loss\r\n loss_num=loss.item()\r\n logits = outputs.logits\r\n running_loss+=loss_num\r\n if i%10 ==0: \r\n loss_per_10_steps.append(loss_num)\r\n out.update(progress(loss_num,i, num_of_batches+1))\r\n\r\n # calculating the gradients\r\n loss.backward()\r\n\r\n #updating the params\r\n optimizer.step()\r\n \r\n loss_values.append(loss_num)\r\n running_loss=running_loss/int(num_of_batches)\r\n```\r\n\r\n", "> > > @patrickvonplaten Not 100% sure that this fix works, forked and implemented the fixes as shown and receive the following error upon training:\r\n> > > ```\r\n> > > Expected tensor for argument #1 'indices' to have scalar type Long; but got torch.cuda.FloatTensor instead (while checking arguments for embedding). \r\n> > > ```\r\n> > > \r\n> > > \r\n> > > If I revert back to transformers-4.0.0 and run exactly the same script, with the same data then T5 trains fine. Could be user error, although I hope not! Did look back through the script and try and diagnose, but no luck as of yet.\r\n> > \r\n> > \r\n> > Can you attach a simple code snippet showing what code produces your error? It's for T5 no?\r\n> \r\n> Sure:\r\n> \r\n> ```\r\n> MODEL_NAME = 'google/t5-v1_1-base'\r\n> tokenizer = T5TokenizerFast.from_pretrained('t5-base')\r\n> special_tokens_dict = {'additional_special_tokens': ['<ORG>','<PERSON>']}\r\n> num_added_tokens = tokenizer.add_special_tokens(special_tokens_dict)\r\n> print(f'ADDED TOKENS: {num_added_tokens}')\r\n> model = T5ForConditionalGeneration.from_pretrained(MODEL_NAME, return_dict=True)\r\n> model.resize_token_embeddings(len(tokenizer))\r\n> model.to(dev)\r\n> BATCH_SIZE = 8\r\n> ```\r\n> \r\n> ```\r\n> #Sets the module in training mode\r\n> from IPython.display import HTML, display\r\n> def progress(loss,value, max=100):\r\n> return HTML(\"\"\" Batch loss :{loss} <progress \r\n> value='{value}'max='{max}',style='width: 100%'>{value}\r\n> </progress> \"\"\".format(loss=loss,value=value, max=max))\r\n> \r\n> model.train()\r\n> num_of_batches= int(len(train_df) / BATCH_SIZE)\r\n> print(num_of_batches)\r\n> NUM_EPOCHS = 1\r\n> loss_per_10_steps=[]\r\n> loss_values = []\r\n> for epoch in range(1,NUM_EPOCHS+1):\r\n> print('Running epoch: {}'.format(epoch))\r\n> \r\n> running_loss=0\r\n> \r\n> out = display(progress(1, num_of_batches+1), display_id=True)\r\n> for i in range(num_of_batches):\r\n> inputbatch=[]\r\n> labelbatch=[]\r\n> new_df=train_df[i*BATCH_SIZE:i*BATCH_SIZE+BATCH_SIZE]\r\n> for indx,row in new_df.iterrows():\r\n> input = 'Product: '+row['product_name']\r\n> labels = row['product_description']\r\n> inputbatch.append(input)\r\n> labelbatch.append(labels)\r\n> inputbatch=tokenizer.batch_encode_plus(inputbatch,padding=True, max_length=512,return_tensors='pt')[\"input_ids\"]\r\n> labelbatch=tokenizer.batch_encode_plus(labelbatch,padding=True, max_length=512,return_tensors=\"pt\") [\"input_ids\"]\r\n> inputbatch=inputbatch.to(dev)\r\n> labelbatch=labelbatch.to(dev)\r\n> \r\n> # clear out the gradients of all Variables \r\n> optimizer.zero_grad()\r\n> \r\n> # Forward propogation\r\n> outputs = model(input_ids=inputbatch, labels=labelbatch)\r\n> loss = outputs.loss\r\n> loss_num=loss.item()\r\n> logits = outputs.logits\r\n> running_loss+=loss_num\r\n> if i%10 ==0: \r\n> loss_per_10_steps.append(loss_num)\r\n> out.update(progress(loss_num,i, num_of_batches+1))\r\n> \r\n> # calculating the gradients\r\n> loss.backward()\r\n> \r\n> #updating the params\r\n> optimizer.step()\r\n> \r\n> loss_values.append(loss_num)\r\n> running_loss=running_loss/int(num_of_batches)\r\n> ```\r\n\r\nthanks! `dev` would be equal to `\"cuda\"` I suppose? ", "> \r\n> \r\n> > > > @patrickvonplaten Not 100% sure that this fix works, forked and implemented the fixes as shown and receive the following error upon training:\r\n> > > > ```\r\n> > > > Expected tensor for argument #1 'indices' to have scalar type Long; but got torch.cuda.FloatTensor instead (while checking arguments for embedding). \r\n> > > > ```\r\n> > > > \r\n> > > > \r\n> > > > If I revert back to transformers-4.0.0 and run exactly the same script, with the same data then T5 trains fine. Could be user error, although I hope not! Did look back through the script and try and diagnose, but no luck as of yet.\r\n> > > \r\n> > > \r\n> > > Can you attach a simple code snippet showing what code produces your error? It's for T5 no?\r\n> > \r\n> > \r\n> > Sure:\r\n> > ```\r\n> > MODEL_NAME = 'google/t5-v1_1-base'\r\n> > tokenizer = T5TokenizerFast.from_pretrained('t5-base')\r\n> > special_tokens_dict = {'additional_special_tokens': ['<ORG>','<PERSON>']}\r\n> > num_added_tokens = tokenizer.add_special_tokens(special_tokens_dict)\r\n> > print(f'ADDED TOKENS: {num_added_tokens}')\r\n> > model = T5ForConditionalGeneration.from_pretrained(MODEL_NAME, return_dict=True)\r\n> > model.resize_token_embeddings(len(tokenizer))\r\n> > model.to(dev)\r\n> > BATCH_SIZE = 8\r\n> > ```\r\n> > \r\n> > \r\n> > ```\r\n> > #Sets the module in training mode\r\n> > from IPython.display import HTML, display\r\n> > def progress(loss,value, max=100):\r\n> > return HTML(\"\"\" Batch loss :{loss} <progress \r\n> > value='{value}'max='{max}',style='width: 100%'>{value}\r\n> > </progress> \"\"\".format(loss=loss,value=value, max=max))\r\n> > \r\n> > model.train()\r\n> > num_of_batches= int(len(train_df) / BATCH_SIZE)\r\n> > print(num_of_batches)\r\n> > NUM_EPOCHS = 1\r\n> > loss_per_10_steps=[]\r\n> > loss_values = []\r\n> > for epoch in range(1,NUM_EPOCHS+1):\r\n> > print('Running epoch: {}'.format(epoch))\r\n> > \r\n> > running_loss=0\r\n> > \r\n> > out = display(progress(1, num_of_batches+1), display_id=True)\r\n> > for i in range(num_of_batches):\r\n> > inputbatch=[]\r\n> > labelbatch=[]\r\n> > new_df=train_df[i*BATCH_SIZE:i*BATCH_SIZE+BATCH_SIZE]\r\n> > for indx,row in new_df.iterrows():\r\n> > input = 'Product: '+row['product_name']\r\n> > labels = row['product_description']\r\n> > inputbatch.append(input)\r\n> > labelbatch.append(labels)\r\n> > inputbatch=tokenizer.batch_encode_plus(inputbatch,padding=True, max_length=512,return_tensors='pt')[\"input_ids\"]\r\n> > labelbatch=tokenizer.batch_encode_plus(labelbatch,padding=True, max_length=512,return_tensors=\"pt\") [\"input_ids\"]\r\n> > inputbatch=inputbatch.to(dev)\r\n> > labelbatch=labelbatch.to(dev)\r\n> > \r\n> > # clear out the gradients of all Variables \r\n> > optimizer.zero_grad()\r\n> > \r\n> > # Forward propogation\r\n> > outputs = model(input_ids=inputbatch, labels=labelbatch)\r\n> > loss = outputs.loss\r\n> > loss_num=loss.item()\r\n> > logits = outputs.logits\r\n> > running_loss+=loss_num\r\n> > if i%10 ==0: \r\n> > loss_per_10_steps.append(loss_num)\r\n> > out.update(progress(loss_num,i, num_of_batches+1))\r\n> > \r\n> > # calculating the gradients\r\n> > loss.backward()\r\n> > \r\n> > #updating the params\r\n> > optimizer.step()\r\n> > \r\n> > loss_values.append(loss_num)\r\n> > running_loss=running_loss/int(num_of_batches)\r\n> > ```\r\n> \r\n> thanks! `dev` would be equal to `\"cuda\"` I suppose?\r\n\r\nYeah \"cuda\" sorry.", "Could you also attach some code for `train_df` and `optimizer`? So that I can fully reproduce :-) ", "> \r\n> \r\n> Could you also attach some code for `train_df` and `optimizer`? So that I can fully reproduce :-)\r\n\r\nSure!\r\n```\r\noptimizer = Adafactor(model.parameters(),lr=1e-3,\r\n eps=(1e-30, 1e-3),\r\n clip_threshold=1.0,\r\n decay_rate=-0.8,\r\n beta1=None,\r\n weight_decay=0.0,\r\n relative_step=False,\r\n scale_parameter=False,\r\n warmup_init=False)\r\n```\r\n\r\ntrain_df is just a dataframe containing something like the following:\r\n product_name\tproduct_description\r\n37245\tTest Product 1\tTest Description 1\r\n23451\tTest Product 2 Test Description 2 \r\n\r\nNot sure how to attach a file via GitHub my apologies.\r\n", "> @patrickvonplaten Not 100% sure that this fix works, forked and implemented the fixes as shown and receive the following error upon training:\r\n> \r\n> ```\r\n> Expected tensor for argument #1 'indices' to have scalar type Long; but got torch.cuda.FloatTensor instead (while checking arguments for embedding). \r\n> ```\r\n> \r\n> If I revert back to transformers-4.0.0 and run exactly the same script, with the same data then T5 trains fine. Could be user error, although I hope not! Did look back through the script and try and diagnose, but no luck as of yet.\r\n\r\nI tried with some dummy training data and it works for me...not sure what the problem is. Also the error message hints at a wrong `dtype` of either `input_ids` or `labels`...\r\n\r\nCould you try to do the following:\r\n\r\n```python\r\n inputbatch=inputbatch.to(dev).to(torch.long)\r\n labelbatch=labelbatch.to(dev).to(torch.long)\r\n```\r\n\r\nand see if the error persists?", "> ```python\r\n> .to(torch.long)\r\n> ```\r\n\r\nnever mind, I can reproduce! Thanks for the message! Will see how to fix it -> weird error", "> \r\n> \r\n> > ```python\r\n> > .to(torch.long)\r\n> > ```\r\n> \r\n> never mind, I can reproduce! Thanks for the message! Will see how to fix it -> weird error\r\n\r\nGreat thank you :) \r\n\r\nPrevious issue describing the same error #7026 . Gave me some guidance but couldn't quite work it out.", "> > > ```python\r\n> > > .to(torch.long)\r\n> > > ```\r\n> > \r\n> > \r\n> > never mind, I can reproduce! Thanks for the message! Will see how to fix it -> weird error\r\n> \r\n> Great thank you :)\r\n> \r\n> Previous issue describing the same error #7026 . Gave me some guidance but couldn't quite work it out.\r\n\r\nShould be good now - was 100% introduces by this PR -> thanks a lot for spotting it!", "> \r\n> \r\n> > > > ```python\r\n> > > > .to(torch.long)\r\n> > > > ```\r\n> > > \r\n> > > \r\n> > > never mind, I can reproduce! Thanks for the message! Will see how to fix it -> weird error\r\n> > \r\n> > \r\n> > Great thank you :)\r\n> > Previous issue describing the same error #7026 . Gave me some guidance but couldn't quite work it out.\r\n> \r\n> Should be good now - was 100% introduces by this PR -> thanks a lot for spotting it!\r\n\r\nAmazing thank you will run through a test this evening. ", "@sgugger @LysandreJik - I updated the PR description. It's good to merge for me. Let me know what you think." ]

[]

[]

[ "having the **same problem**, what is happening, it was working just fine for the past like 90 days!!\r\n\r\n", "Hello! It would be very helpful if you could complete the information related to your environment. If you could have a reproducible code example, that would really be great as well.\r\n\r\nIt is possible you were affected by the breaking changes from v3.x to v4.x. If this is the case, I invite you to read the [migration notes](https://huggingface.co/transformers/migration.html), or to pin the transformers library to the major version 3: `pip install transformers==3`", "https://github.com/mosh98/Swedish_Sentiment_BERTIL/blob/main/Swe_Bert_Training_Bigger_dataset.ipynb", "I am afraid I see no error in your notebook.", "okej thanks for the tip, pinning it to version 3 did the trick!", "Thanks @LysandreJik, \r\n\r\nIt works but creates a new error here that says: \r\n\r\nError(s) in loading state_dict for SentimentClassifier:\r\n\tUnexpected key(s) in state_dict: \"bert.embeddings.position_ids\". \r\n[Notebook \r\n](https://colab.research.google.com/drive/1fEXY3IQ82u41KvwoDOg-oY95RDD1OpKg?usp=sharing)\r\n\r\n**After running the code below**: \r\n\r\n```\r\nsaved_model = torch.load('selective_stock_dataset_state-2.bin')\r\nmodel = SentimentClassifier(len(class_names))\r\nmodel.load_state_dict(saved_model)\r\nmodel = model.to(device)\r\n\r\n```\r\n\r\n\r\n\r\n```\r\nclass SentimentClassifier(nn.Module): \r\n def __init__(self, n_classes):\r\n super(SentimentClassifier, self).__init__()\r\n self.bert = BertModel.from_pretrained(PRE_TRAINED_MODEL_NAME)\r\n self.drop = nn.Dropout(p=0.3)\r\n self.out = nn.Linear(self.bert.config.hidden_size, n_classes)\r\n self.softmax = nn.Softmax(dim=1)\r\n\r\n def forward(self, input_ids, attention_mask):\r\n _, pooled_output = self.bert(\r\n input_ids=input_ids,\r\n attention_mask=attention_mask\r\n )\r\n output = self.drop(pooled_output)\r\n output = self.out(output)\r\n return self.softmax(output)\r\n```\r\n\r\nCould you advise. Please and thanks ", "as he mentioned earlier, try using `pip install transformers==3`", "@mosh98 , I have tried with `pip install transformers==3` and it removed the first error. \r\n\r\nBut i then get a new error that I was no getting before that says \r\n\r\n`Error(s) in loading state_dict for SentimentClassifier:\r\nUnexpected key(s) in state_dict: \"bert.embeddings.position_ids\".`\r\n\r\nsee my notebook here: [notebook](https://colab.research.google.com/drive/1fEXY3IQ82u41KvwoDOg-oY95RDD1OpKg?usp=sharing#scrollTo=iQ93LDzMXO58l)", "Ahh it was solved by changing \r\n\r\n`model.load_state_dict(saved_model)`\r\n\r\nto \r\n\r\n`model.load_state_dict(saved_model, strict=False)`\r\n", "Hi, indeed, this is a different error. We recommend using the `from_pretrained` method (your custom model would need to inherit from `PreTrainedModel` rather than `nn.Module`) rather than using `load_state_dict` to ensure maximum compatibility between checkpoints and architectures, otherwise the state dicts might not be 100% loadable on each custom architecture.\r\n\r\nYour workaround using `strict=False` also works!", "\"\"\"pip install transformers==3\"\"\" doesnt seem to work\r\n", "No need to downgrade the transformers. Just do the following - it's from the migration guide.\r\n\r\n```\r\nmodel = BertModel.from_pretrained(\"bert-base-cased\")\r\noutputs = model(**inputs, return_dict=False)\r\n```\r\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "`outputs = model(**inputs, return_dict=False)`\r\n\r\nor\r\n\r\n`model = BertModel.from_pretrained(\"bert-base-cased\",return_dict=False)`", "> `outputs = model(**inputs, return_dict=False)`\r\n> \r\n> or\r\n> \r\n> `model = BertModel.from_pretrained(\"bert-base-cased\",return_dict=False)`\r\n\r\ncool, it works.", "> > `outputs = model(**inputs, return_dict=False)`\r\n> > or\r\n> > `model = BertModel.from_pretrained(\"bert-base-cased\",return_dict=False)`\r\n> \r\n> cool, it works.\r\n\r\ngreat! it worked for me too, thanks a million :) ", "Still does not work for me transformers 4.24", "hi \r\ni am also facing the same issue. i am applying transformer on image data. i did the training. i haven't face any issues while training. but in prediction it throws the error\r\nTypeError: dropout(): argument 'input' (position 1) must be Tensor, not NoneType\r\n\r\nplease resolve if possible", "Hey @dishamohini could you make sure to do the following:\r\n- check that you are using the latest version of `transformers`\r\n- check that you are correctly running the model in eval mode\r\nIf you still have an issue:\r\n- open a new issue on `transformers`, following the contribution guidelines with the output of `transformers-cli envs`, and a full minimal reproducer as well as the trace-back \r\n- ping me there" ]

[]

[]

[]

[]

[]

[ "Given our discussion yesterday, I'm not sure `distributed_env` is fitting. As you convinced me that DP is not distributed when it comes to pytorch conventions, `if self.distributed_env == \"dp\"` is back to being confusing.\r\n\r\nGiven that with the exception of tpu, all dp/ddp/mp/pp are SomethingParallel, should it be called `parallel_mode`?\r\n\r\nI don't know anything about tpu, so it's hard for me to know where it fits. But it's probably not distributed either. And not parallel either.\r\n\r\nSo perhaps we call it `compute_env`", "LGTM, @sgugger!\r\n" ]

[]

[]

[ "It is expected that this would take some time, since it has to skip through `204,516` batches before continuing training. It will continue progressing after that skip is done.", "In the PR mentioned above, I'm adding a flag to ignore that step if you're prepared to pay the price of having the training be slightly different from a training from scratch to go faster.", "That would do for my case, thanks!", "With that PR everything worked as expected, thanks for the very quick turnaround!", "Happy to hear!", "@sgugger Sorry to bother, but I am wondering why skipping steps takes computing. I mean, the random_seed is specified, so the trainer just need to find the breakpoint of an epoch and resume, I shouldn't take much time.\r\n\r\nSo does any parts of my understand are wrong?", "Yes, but there is no way to be in the exact sample place in the dataloaders (that have randomness with the shuffling) without going through the first epochs and then batches.", "@sgugger thanks for the information, sorry to revive this issue. How long does it usually take to go through the first epochs and then batches? half of what it took to train until that point or less?", "It depends on your data and the time needed for your preprocessing. Note that there is a progress bar in the newer versions of Transformers so you can get a sense of the remaining time. You can also skip this with the [flag `ignore_data_skip`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments.ignore_data_skip) though the model will train on already seen data in this case.", "I have the same issue,\r\n```\r\n***** Running training *****\r\n Num examples = 2,560,000\r\n Num Epochs = 9,223,372,036,854,775,807\r\n Instantaneous batch size per device = 8\r\n Total train batch size (w. parallel, distributed & accumulation) = 16\r\n Gradient Accumulation steps = 2\r\n Total optimization steps = 160,000\r\n Number of trainable parameters = 332,891,919\r\n Continuing training from checkpoint, will skip to saved global_step\r\n Continuing training from epoch 0\r\n Continuing training from global step 77000\r\n Will skip the first 0 epochs then the first 154000 batches in the first epoch.\r\n 0%| | 0/160000 [00:00<?, ?it/s]\r\n The following columns in the training set don't have a corresponding argument in `Wav2Vec2ForCTC.forward` and have been ignored: locale, audio, input_length, sentence. If locale, audio, input_length, sentence are not expected by `Wav2Vec2ForCTC.forward`, you can safely ignore this message.\r\nThere seems to be not a single sample in your epoch_iterator, stopping training at step 77000! This is expected if you're using an IterableDataset and set num_steps (160000) higher than the number of available samples.\r\n\r\n\r\nTraining completed. Do not forget to share your model on huggingface.co/models =)\r\n\r\n\r\n{'train_runtime': 1869.2894, 'train_samples_per_second': 1369.504, 'train_steps_per_second': 85.594, 'train_loss': 0.0, 'epoch': 43.01} \r\n 0%| | 0/160000 [31:09<?, ?it/s]Saving model checkpoint to /usr/local/bin/source/output\r\nConfiguration saved in /usr/local/bin/source/output/config.json\r\nModel weights saved in /usr/local/bin/source/output/pytorch_model.bin\r\nFeature extractor saved in /usr/local/bin/source/output/preprocessor_config.json\r\ntokenizer config file saved in /usr/local/bin/source/output/tokenizer_config.json\r\nSpecial tokens file saved in /usr/local/bin/source/output/special_tokens_map.json\r\nadded tokens file saved in /usr/local/bin/source/output/added_tokens.json\r\ntrainer save model!\r\nmetric: train_runtime\r\n***** train metrics *****\r\n epoch = 43.01\r\n train_loss = 0.0\r\n train_runtime = 0:31:09.28\r\n train_samples_per_second = 1369.504\r\n train_steps_per_second = 85.594\r\n06/19/2023 08:28:34 - INFO - __main__ - *** Evaluate ***\r\n***** Running Evaluation *****\r\n Num examples: Unknown\r\n Batch size = 8\r\nThe following columns in the evaluation set don't have a corresponding argument in `Wav2Vec2ForCTC.forward` and have been ignored: locale, audio, input_length, sentence. If locale, audio, input_length, sentence are not expected by `Wav2Vec2ForCTC.forward`, you can safely ignore this message.\r\nDropping the following result as it does not have all the necessary fields:\r\n{'task': {'name': 'Automatic Speech Recognition', 'type': 'automatic-speech-recognition'}, 'metrics': [{'name': 'Wer', 'type': 'wer', 'value': 1.0006485084306096}]}\r\n```\r\n\r\nmax_steps= 160000\r\nlast checkpoint=77000\r\n\r\nif **ignore_data_skip**=True is set, it can resume training correctly." ]

[]

[]

[ "Thanks!", "Sorry again we let this sit for so long!\r\n\r\nSo since it's been a long time, the diff has gotten quite messy. Would you mind closing and re-opening a clean PR @noise-field ? Ping me on it and we'll expedite the review. Sorry again.", "Closing as requested" ]

[]

[]

[ "Hi there. The results are slightly different because your dataloader has some randomization (the train dataloader has `shuffle=True` ) and the `Trainer` currently does not go through your dataloader for the past epochs when resuming training. So it trains starting from the global step 282 with the data of the epoch 0 of the initial training (hope that makes sense).\r\n\r\nLet me see if we can support full reproducibility without a big drop in performance (cause we don't want to loop through that epoch 0 without doing anything either).", "Hej\r\nthanks for the fast answer!\r\nI tried it out, but the results still differ slightly (leaving out: `--ignore_data_skip`).\r\nYour change in the code should make the results exactly the same when continue training, right?\r\nIs it because of the training data sampler (`RandomSampler`)?\r\n\r\nCheers\r\n", "> Your change in the code should make the results exactly the same when continue training, right?\r\n\r\nYes, and this is enforced by tests in the CI. If your results still differ slightly, there might be another source of randomness not properly seeded that is responsible for those changes." ]

[]

[]

[ "I think you would need to compile your model with both the tokens tensor and the attention mask. Given that the attention mask is the second argument, you can pass it directly when tracing the model:\r\n\r\n```py\r\ntraced_model = torch.jit.trace(model, [tokens_tensor, attention_mask])\r\n```\r\n\r\nthen you can do:\r\n```py\r\nmodel(tokens_tensor, attention_mask)\r\n```", "@LysandreJik It works. Thanks for your help! " ]

[]

[ { "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 } ]

[ "@LysandreJik any update on this?", "@yaysummeriscoming To get sub words instead of numbers, you can call `tokenizer.gpt2_tokenizer.decode(tokens)`. Please take a look at [our code](https://github.com/huggingface/transformers/blob/52c9e842854a701a7d1b608600a614278b4407d3/src/transformers/tokenization_deberta.py#L396) for reference.", "That did the trick, thanks!" ]

[]

[]

[ "Hey @iliaschalkidis,\r\n\r\nthanks for your issue! The memory usage in Longformer does not decrease linearly when reducing the attention_window...but I'm a bit surprised that you are experiencing OOM in your set-up...Does the same happen for your in eager mode? I'll try to look into it a bit next week. One thing that would be of great help is if you find time to benchmark the memory usage of `TFLongformer`, for:\r\n\r\n- eager mode\r\n- compiled\r\n\r\nfor different settings of the window size", "Hi @patrickvonplaten, how do you recommend to perform this benchmarking? Any suggestion (best practice/ tool)? I also had the impression that TF2 is in eager execution by default...", "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." ]

[]

[]

[ "@mfuntowicz might have an idea. I'm guessing that what it does is copy each layer, while all the layers have shared weights and should point to the same tensor.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

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

[]

[]

[ "Facing the same issue. A reply on this is highly appreciated.", "can [this](https://github.com/huggingface/transformers/issues/8524#issuecomment-753876838) be your solution? Hope it helps...", "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 had the same issue for `GPT2LMHeadModel`. In my case, I found the solution. Let `hf3_model_dir` be the Hugging Face v3.x model directory that you give to `load_pretrained`. Inside this directory is the saved pytorch model file called `pytorch_model.bin`. Let's load this file directly using pytorch:\r\n\r\n```\r\nstate_dict = torch.load('pytorch_model.bin')`\r\n```\r\n\r\nNow check the values of these two entries:\r\n\r\n```\r\nstate_dict['transformer.wte.weight']\r\nstate_dict['lm_head.weight']\r\n```\r\n\r\nI found that they were different. However, they should be the same `vocab_size x embedding_size` matrix. Indeed, let's actually load the model: \r\n\r\n```\r\nmodel = transformers.GPT2LMHeadModel.from_pretrained(hf3_model)`\r\n```\r\nAnd check the following values:\r\n```\r\nmodel.transformer.wte.weight\r\nmodel.lm_head.weight\r\n```\r\nYou will find that they are the same. However,\r\nin Hugging Face v3.x, they are both equal to `state_dict['lm_head.weight']`\r\nin Hugging Face v4.x, they are both equal to `state_dict['transformer.wte.weight']`.\r\n\r\nSo that's the cause of the problem. To get the same behavior in Hugging Face v4.x as you get in Hugging Face v3.x, I manually set both equal to `state_dict['lm_head.weight']`.", "As a further comment, for models saved under Hugging Face v4.x, `state_dict['transformer.wte.weight']` and `state_dict['lm_head.weight']` are both equal as they should be.\r\n\r\nFor models saved under Hugging Face v3.x, `state_dict['transformer.wte.weight']` ends up being (I believe) just random garbage that is harmless if reloaded using Hugging Face v3.x but can be very harmful if reloaded using Hugging Face v4.x" ]

[]

[]

[ "2 resources in the thread linked by valhalla: https://discuss.huggingface.co/t/how-to-apply-pruning-on-a-bert-model/1658/5", "> 2 resources in the thread linked by valhalla: https://discuss.huggingface.co/t/how-to-apply-pruning-on-a-bert-model/1658/5\r\n\r\nThank you very much for your warm reply, I will study it." ]

[]

[]

[ "Same here a couple of hours ago", "1. Hi, could you please provide the information related to your environment? \r\n\r\n2. When you say it was working yesterday but was working before, do you mean to say you've upgraded to version v4.0.0 released yesterday? If this is so, you may be obtaining the following error message: `AttributeError: 'NoneType' object has no attribute 'from_pretrained'`. This would be because you do not have `sentencepiece` installed.\r\n\r\n3. Are you sure this worked previously? This should never have worked, as `AutoTokenizer` cannot be initialized like this, but has to be instantiated from the `from_pretrained` method:\r\n\r\n```py\r\nfrom transformers import AutoTokenizer\r\nAutoTokenizer.from_pretrained(\"Helsinki-NLP/opus-mt-en-fr\")\r\n```\r\n\r\nwhich works on v4.0.0 and on `master`, as long as you have SentencePiece installed.", "Putting a better error message in #8881.", "Right, I was using\r\n```py\r\nAutoTokenizer.from_pretrained(\"Helsinki-NLP/opus-mt-en-fr\")\r\n```\r\n\r\nThanks, `pip install sentencepiece` fixed the issue!\r\n\r\nIt looks that previously the tokenizer outputted torch tensors and now lists. Is this intended? It breaks existing code.", "Yes, this was a bug. Tokenizers are framework-agnostic and should not output a specific framework's tensor. The implementation of the Marian tokenizer was not respecting the API in that regard.\r\n\r\nTokenizers can still handle torch tensors, you need to specify that you want them though:\r\n\r\n```py\r\ntokenizer(xxx, return_tensors=\"pt\")\r\n```\r\n\r\nI guess in your situation it has to do with the `prepare_seq2seq_batch`:\r\n\r\n```py\r\ntokenizer.prepare_seq2seq_batch(xxx, return_tensors=\"pt\")\r\n```", "Thanks!" ]

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

[]

[ "@patrickvonplaten or @lhoestq might have an idea.", "Hey @JamesDeAntonis yes this is expected. The same behavior would occur for GPT2 if only one token is provided as the labels. You should at least add an EOS token at the end to `labels` (so that you have two labels tokens) to make sure the loss is not zero.\r\n\r\nIf you cannot do this you will have to fork the repo and manually change the `get_nll()` function.", "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." ]

[]

[]

[ "Hello, it seems you've upgraded your library from a 3.x version to 4.0.0. I invite you to consult the migration guide [here (deprecated attributes)](https://huggingface.co/transformers/migration.html#removed-some-deprecated-attributes) or to pin your `transformers` on version 3: `transformers==3`.", "Thank you!" ]

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

[]

[ "Question: the `generate()` documentation says that `topk` etc params will default to values in the `pretrainedconfig`, but `top_k`, for example is never read from the config file. This differs from, e.g., `num_beams, max_length` which are read from config if they are not passed in as params to the generate function.\r\n\r\nAnd since, for example, `T5PreTrainedModel` never overwrites the generate function, I don't see how the defaults in the config (for params like `top_k`) could actually end up being passed to the generate function?", "> Question: the `generate()` documentation says that `topk` etc params will default to values in the `pretrainedconfig`, but `top_k`, for example is never read from the config file. This differs from, e.g., `num_beams, max_length` which are read from config if they are not passed in as params to the generate function.\r\n> \r\n> And since, for example, `T5PreTrainedModel` never overwrites the generate function, I don't see how the defaults in the config (for params like `top_k`) could actually end up being passed to the generate function?\r\n\r\nRegarding reading from config - am I missing something or do these never get checked? \r\n\r\n@patrickvonplaten ", "> > Question: the `generate()` documentation says that `topk` etc params will default to values in the `pretrainedconfig`, but `top_k`, for example is never read from the config file. This differs from, e.g., `num_beams, max_length` which are read from config if they are not passed in as params to the generate function.\r\n> > And since, for example, `T5PreTrainedModel` never overwrites the generate function, I don't see how the defaults in the config (for params like `top_k`) could actually end up being passed to the generate function?\r\n> \r\n> Regarding reading from config - am I missing something or do these never get checked?\r\n> \r\n> @patrickvonplaten\r\n\r\nhttps://github.com/huggingface/transformers/blob/693ac3594b96e86dd282fdf8e413f3a48b176892/src/transformers/generation_utils.py#L240", "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." ]

[]

[]

[ "black complained about the style after the update, so I fixed it and squashed the commits again.", "Thank you @Craigacp!" ]