Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Thank you for the invitation, @sshleifer. \r\n\r\nMy feeling is that a person in charge of any domain should have commit rights to that domain. Since I don't have those I'd be happy to be delegated to. (with the exception of fsmt since I wrote it)" ]

[]

[]

[ "Ha yeah I saw your message on Slack just afterwards, I'll just change the default :)", "It's changed !", "Merging now so that it's in v3.5.0!" ]

[ { "id": 1834067346, "node_id": "MDU6TGFiZWwxODM0MDY3MzQ2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Documentation", "name": "Documentation", "color": "77cc3b", "default": false, "description": "" } ]

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

[ "The docs are incorrect, sorry about that.\r\n\r\nTry\r\n```python\r\n from transformers import MBartForConditionalGeneration, MBartTokenizer\r\n model = MBartForConditionalGeneration.from_pretrained(\"facebook/mbart-large-en-ro\")\r\n tokenizer = MBartTokenizer.from_pretrained(\"facebook/mbart-large-en-ro\")\r\n article = \"UN Chief Says There Is No Military Solution in Syria\"\r\n batch = tokenizer.prepare_seq2seq_batch(src_texts=[article], src_lang=\"en_XX\")\r\n translated_tokens = model.generate(**batch, decoder_start_token_id=tokenizer.lang_code_to_id[\"ro_RO\"])\r\n translation = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]\r\n assert translation == \"Şeful ONU declară că nu există o soluţie militară în Siria\"\r\n```", "> The docs are incorrect, sorry about that.\r\n> \r\n> Try\r\n> \r\n> ```python\r\n> from transformers import MBartForConditionalGeneration, MBartTokenizer\r\n> model = MBartForConditionalGeneration.from_pretrained(\"facebook/mbart-large-en-ro\")\r\n> tokenizer = MBartTokenizer.from_pretrained(\"facebook/mbart-large-en-ro\")\r\n> article = \"UN Chief Says There Is No Military Solution in Syria\"\r\n> batch = tokenizer.prepare_seq2seq_batch(src_texts=[article], src_lang=\"en_XX\")\r\n> translated_tokens = model.generate(**batch, decoder_start_token_id=tokenizer.lang_code_to_id[\"ro_RO\"])\r\n> translation = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]\r\n> assert translation == \"Şeful ONU declară că nu există o soluţie militară în Siria\"\r\n> ```\r\n\r\nthank you for your reply, If I don't want to generate, I just want to train. How should I change it?\r\n\r\n```python\r\nexample_english_phrase = \"UN Chief Says There Is No Military Solution in Syria\"\r\nexpected_translation_romanian = \"Şeful ONU declară că nu există o soluţie militară în Siria\"\r\nbatch = tokenizer.prepare_seq2seq_batch(example_english_phrase, src_lang=\"en_XX\", tgt_lang=\"ro_RO\", tgt_texts=expected_translation_romanian)\r\ninput_ids = batch[\"input_ids\"]\r\ntarget_ids = batch[\"decoder_input_ids\"] # Error\r\ndecoder_input_ids = target_ids[:, :-1].contiguous()\r\nlabels = target_ids[:, 1:].clone()\r\nmodel(input_ids=input_ids, decoder_input_ids=decoder_input_ids, labels=labels) #forward\r\n```", "See this https://github.com/huggingface/transformers/blob/master/examples/seq2seq/finetune.py#L138\r\n\r\nthe `batch` argument to that fn is the same as your `batch` (the output of `prepare_seq2seq_batch`)" ]

[ { "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" } ]

[]

[]

[]

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

[ "Hi,\r\nI'm experiencing the same issue while fine-tuning on TPU for marianMT. Code works well on GPU. On TPU, it throws similar exception: \r\n```python\r\nRuntimeError: Cannot access data pointer of Tensor that doesn't have storage\r\n```\r\nThanks!", "Hi @chris-tng ,\r\n\r\ncould you post your env info, and the script/command that you are running ?", "Sure, I'm using \r\n```shell\r\ntransformers 4.0.1 \r\ntorch 1.7.0+cu101 \r\ntorch-xla 1.7 \r\ntorchsummary 1.5.1 \r\ntorchtext 0.3.1 \r\ntorchvision 0.8.1+cu101\r\n```\r\n- command\r\n\r\n```shell\r\n!python xla_spawn.py --num_cores 8 finetune_trainer.py \\\r\n --tokenizer_name \"Helsinki-NLP/opus-mt-es-en\" \\\r\n --model_name_or_path \"Helsinki-NLP/opus-mt-es-en\" \\\r\n --data_dir \"/content/data\" \\\r\n --output_dir \"/content/marian_es_en\" --overwrite_output_dir \\\r\n --learning_rate=3e-4 \\\r\n --warmup_steps 500 \\\r\n --per_device_train_batch_size=256 --per_device_eval_batch_size=256 \\\r\n --freeze_encoder --freeze_embeds \\\r\n --num_train_epochs=6 \\\r\n --save_steps 3000 --eval_steps 3000 \\\r\n --logging_first_step --logging_steps 200 \\\r\n --max_source_length 128 \\\r\n --max_target_length 128 --val_max_target_length 128 --test_max_target_length 128 \\\r\n --do_train --do_eval --do_predict \\\r\n --n_val 5000 --n_test 10000 --evaluation_strategy steps \\\r\n --prediction_loss_only \\\r\n --task translation --label_smoothing 0.1 \\\r\n \"$@\"\r\n```\r\n\r\nHere is the error\r\n```\r\n[INFO|trainer.py:666] 2020-12-11 05:56:11,550 >> Total train batch size (w. parallel, distributed & accumulation) = 2048\r\n[INFO|trainer.py:667] 2020-12-11 05:56:11,550 >> Gradient Accumulation steps = 1\r\n[INFO|trainer.py:668] 2020-12-11 05:56:11,550 >> Total optimization steps = 1830\r\n 0% 0/1830 [00:00<?, ?it/s]Exception in device=TPU:0: Cannot access data pointer of Tensor that doesn't have storage\r\nTraceback (most recent call last):\r\n File \"/usr/local/lib/python3.6/dist-packages/torch_xla/distributed/xla_multiprocessing.py\", line 330, in _mp_start_fn\r\n _start_fn(index, pf_cfg, fn, args)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch_xla/distributed/xla_multiprocessing.py\", line 324, in _start_fn\r\n fn(gindex, *args)\r\n File \"/content/transformers/examples/seq2seq/finetune_trainer.py\", line 309, in _mp_fn\r\n main()\r\n File \"/content/transformers/examples/seq2seq/finetune_trainer.py\", line 258, in main\r\n model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 747, in train\r\n tr_loss += self.training_step(model, inputs)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 1089, in training_step\r\n loss.backward()\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/tensor.py\", line 221, in backward\r\n torch.autograd.backward(self, gradient, retain_graph, create_graph)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/autograd/__init__.py\", line 132, in backward\r\n allow_unreachable=True) # allow_unreachable flag\r\nRuntimeError: Cannot access data pointer of Tensor that doesn't have storage\r\n 0% 0/1830 [03:34<?, ?it/s]\r\nTraceback (most recent call last):\r\n File \"xla_spawn.py\", line 72, in <module>\r\n main()\r\n File \"xla_spawn.py\", line 68, in main\r\n xmp.spawn(mod._mp_fn, args=(), nprocs=args.num_cores)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch_xla/distributed/xla_multiprocessing.py\", line 395, in spawn\r\n start_method=start_method)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/multiprocessing/spawn.py\", line 157, in start_processes\r\n while not context.join():\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/multiprocessing/spawn.py\", line 112, in join\r\n (error_index, exitcode)\r\nException: process 0 terminated with exit code 17\r\n```\r\n", "Hi @patil-suraj , any idea why it happens? ", "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 looks like the config was not saved in the checkpoint folder you are passing. Double-check its contents, but apparently, the model was not properly saved inside it.", "> It looks like the config was not saved in the checkpoint folder you are passing. Double-check its contents, but apparently, the model was not properly saved inside it.\r\n\r\nThanks for replying @sgugger \r\n\r\nI have the following files in my checkpoint folder:\r\n\r\n```\r\nconfig.json optimizer.pt pytorch_model.bin scheduler.pt trainer_state.json training_args.bin\r\n\r\n```\r\nand inside the \"config.json\" it looks like this:\r\n\r\n```\r\n{\r\n \"_name_or_path\": \"bert-base-cased\",\r\n \"architectures\": [\r\n \"BertForMaskedLM\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"gradient_checkpointing\": false,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 768,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"bert\",\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 12,\r\n \"pad_token_id\": 0,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 28996\r\n}\r\n\r\n```\r\n\r\nDo you have any idea where exactly I am going wrong?", "Oh, I think you just have a typo in your path:\r\n```\r\n/media/ai-students/Data/Nesara/Bert_MLM_fine_tune/new_results/result_dir/checkpoint_37000/\r\n```\r\nshould be\r\n```\r\n/media/ai-students/Data/Nesara/Bert_MLM_fine_tune/new_results/result_dir/checkpoint-37000/\r\n```\r\n(a dash instead of the underscore).", "Ahhhh my bad!\r\nThanks a lot @sgugger \r\nI had encountered one more problem (vocab.txt is missing):\r\n```\r\nTraceback (most recent call last):\r\n File \"run_language_modeling.py\", line 355, in <module>\r\n main()\r\n File \"run_language_modeling.py\", line 244, in main\r\n tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)\r\n File \"/home/ai-students/anaconda3/envs/env_nesara/lib/python3.6/site-packages/transformers/tokenization_auto.py\", line 336, in from_pretrained\r\n return tokenizer_class_py.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)\r\n File \"/home/ai-students/anaconda3/envs/env_nesara/lib/python3.6/site-packages/transformers/tokenization_utils_base.py\", line 1649, in from_pretrained\r\n list(cls.vocab_files_names.values()),\r\nOSError: Model name '/media/ai-students/Data/Nesara/Bert_MLM_fine_tune/result_dir/checkpoint-37000' was not found in tokenizers model name list (bert-base-uncased, bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, bert-base-multilingual-cased, bert-base-chinese, bert-base-german-cased, bert-large-uncased-whole-word-masking, bert-large-cased-whole-word-masking, bert-large-uncased-whole-word-masking-finetuned-squad, bert-large-cased-whole-word-masking-finetuned-squad, bert-base-cased-finetuned-mrpc, bert-base-german-dbmdz-cased, bert-base-german-dbmdz-uncased, TurkuNLP/bert-base-finnish-cased-v1, TurkuNLP/bert-base-finnish-uncased-v1, wietsedv/bert-base-dutch-cased). We assumed '/media/ai-students/Data/Nesara/Bert_MLM_fine_tune/result_dir/checkpoint-37000' was a path, a model identifier, or url to a directory containing vocabulary files named ['vocab.txt'] but couldn't find such vocabulary files at this path or url.\r\n\r\n```\r\n\r\nAfter adding the parameter --tokenizer_name I could resolve the issue and now the fine-tuning resumes as expected\r\n\r\nBelow is the command:\r\n\r\n```\r\npython run_language_modeling.py --output_dir=/media/ai-students/Data/Nesara/Bert_MLM_fine_tune/new_results --model_type=bert --model_name_or_path=/media/ai-students/Data/Nesara/Bert_MLM_fine_tune/result_dir/checkpoint-37000 --do_train --train_data_file=$TRAIN_FILE --do_eval --eval_data_file=$TEST_FILE --mlm --per_gpu_train_batch_size=4 --tokenizer_name=bert-base-cased\r\n```" ]

[]

[]

[]

[]

[]

[ "Similar error at [1662]https://github.com/huggingface/transformers/issues/1662\r\n\r\nThe problem could be solved by switching to multilingual tokenizer(if present), else it would require some hot fix. " ]

[]

[]

[ "Thanks for the report... I'll take a look at this.", "Again, if you find out how comet_ml got installed, that would be helpful.", "> Again, if you find out how comet_ml got installed, that would be helpful.\r\n\r\nYay, I found at least one of them `pytorch-lightning`. https://github.com/PyTorchLightning/pytorch-lightning/blob/master/environment.yml#L49\r\n\r\nI have a feeling `pipdeptree` misses packages not installed directly via pypi deps, but via a local `pip install -e .[dev]`\r\n\r\nI probably should just go ahead and create the key - though I have no use for it - please let me know whether it'd serve better if I didn't and continued reporting any related problems.", "Thanks for tracking down pytorch's dependencies. I see that that is for conda. What is weird though is that I can't figure out how `comet_ml.config` wouldn't be defined. In any event, this PR seems fine. (I'd like to get to the bottom of this at some point).\r\n\r\nRemember that you don't have to set a `COMET_API_KEY` (unless you really want to log stuff). You can also set `COMET_MODE=\"DISABLED\"`. ", "I think it could be some fragile error handling. As I mentioned the crash happened when I mistakenly provided a bogus value to one of the pytorch-lightening clargs - so it was supposed to fail telling me that the argument was wrong, but instead failed with the error posted in OP. Once the band-aid was added it reported the error properly w/o crashing. ", "I still have it re-producable w/o rebasing to this fix:\r\n\r\n```\r\ncd examples/seq2seq\r\nBS=2; PYTHONPATH=\"../../src\" python finetune.py --data_dir cnn_dm --do_predict --do_train --eval_batch_size $BS --fp16 --fp16_opt_level O1 --freeze_embeds --freeze_encoder --gpus 1 --gradient_accumulation_steps 1 --learning_rate 3e-5 --max_target_length 142 --model_name_or_path sshleifer/student_cnn_12_6 --n_val 500 --num_train_epochs 2 --output_dir distilbart-cnn-12-6 --tokenizer_name facebook/bart-large --train_batch_size $BS --val_check_interval 0.25 --val_max_target_length 142 --warmup_steps 500\r\n```\r\ngiving:\r\n```\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/util.py:36: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import namedtuple, Mapping, Sequence\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/vendor/graphql-core-1.1/graphql/type/directives.py:55: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n assert isinstance(locations, collections.Iterable), 'Must provide locations for directive.'\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/monkey_patching.py:19: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses\r\n import imp\r\nTraceback (most recent call last):\r\n File \"finetune.py\", line 18, in <module>\r\n from callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback, get_early_stopping_callback\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/examples/seq2seq/callbacks.py\", line 11, in <module>\r\n from utils import save_json\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/examples/seq2seq/utils.py\", line 22, in <module>\r\n from transformers import BartTokenizer, EvalPrediction, PreTrainedTokenizer, T5Tokenizer\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/src/transformers/__init__.py\", line 22, in <module>\r\n from .integrations import ( # isort:skip\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/src/transformers/integrations.py\", line 17, in <module>\r\n if comet_ml.config.get_config(\"comet.api_key\"):\r\nAttributeError: module 'comet_ml' has no attribute 'config'\r\n```\r\nSee if you get it yourself (pre-this PR)? or if you want me to try something let me know \r\n", "Hmm, actually the crash described in the OP happens all the time pre this PR. I even uninstalled and reinstalled `comet_ml` via `pip`.", "I figured it out:\r\n\r\n```\r\npython -c \"import comet_ml; print(comet_ml.config)\"\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/__init__.py\", line 34, in <module>\r\n from .api import API, APIExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/api.py\", line 28, in <module>\r\n from .experiment import CommonExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/experiment.py\", line 97, in <module>\r\n from .gpu_logging import (\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/gpu_logging.py\", line 27, in <module>\r\n import pynvml\r\nModuleNotFoundError: No module named 'pynvml'\r\n```\r\nIf I install `pynvml` the error disappers.\r\n\r\nSo basically `comet_ml` fails to load and the try: block ignores the exception.", "The bug is somewhere in seq2seq utils (**edit**: doesn't seem to be the case)\r\n\r\nFollowing the trace in https://github.com/huggingface/transformers/pull/8410#issuecomment-724281982\r\n\r\nThis does the right thing:\r\n```\r\ncd examples/seq2seq\r\nPYTHONPATH=\"../../src\" python -c \"import utils\"\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/examples/seq2seq/utils.py\", line 22, in <module>\r\n from transformers import BartTokenizer, EvalPrediction, PreTrainedTokenizer, T5Tokenizer\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/src/transformers/__init__.py\", line 22, in <module>\r\n from .integrations import ( # isort:skip\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/src/transformers/integrations.py\", line 16, in <module>\r\n import comet_ml # noqa: F401\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/__init__.py\", line 34, in <module>\r\n from .api import API, APIExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/api.py\", line 28, in <module>\r\n from .experiment import CommonExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/experiment.py\", line 97, in <module>\r\n from .gpu_logging import (\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/gpu_logging.py\", line 27, in <module>\r\n import pynvml\r\nModuleNotFoundError: No module named 'pynvml'\r\n```\r\nbut one above that imports `utils` eats the exception:\r\n```\r\nPYTHONPATH=\"../../src\" python -c \"import callbacks\"\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/util.py:36: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import namedtuple, Mapping, Sequence\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/vendor/graphql-core-1.1/graphql/type/directives.py:55: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n assert isinstance(locations, collections.Iterable), 'Must provide locations for directive.'\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/monkey_patching.py:19: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses\r\n import imp\r\n2020-11-09 14:15:06.573247: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1\r\n/mnt/nvme1/code/github/00nlp/fairseq/fairseq/optim/adam.py:8: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import Collection\r\n(main-38) /mnt/nvme1/code/huggingface/transformers-master-stas00/examples/seq2seq [stas00/transformers|patch-3|+1?8]> PYTHONPATH=\"../../src\" python -c \"import callbacks\"\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/util.py:36: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import namedtuple, Mapping, Sequence\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/vendor/graphql-core-1.1/graphql/type/directives.py:55: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n assert isinstance(locations, collections.Iterable), 'Must provide locations for directive.'\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/monkey_patching.py:19: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses\r\n import imp\r\n2020-11-09 14:15:31.366275: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1\r\n/mnt/nvme1/code/github/00nlp/fairseq/fairseq/optim/adam.py:8: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import Collection\r\n```\r\n", "Oh, great find! Thanks for this... we'll take a further look on our end and get something out soon.", "This is very odd, since I don't see any `try` blocks around this sequence of imports - except at the end inside `integrations.py` and even if I remove it there, the exception is still suppressed. `import utils` catches the error, but `import callbacks` which imports `utils` suppresses it (see https://github.com/huggingface/transformers/pull/8410#issuecomment-724312036).\r\n\r\nCould this somehow be related to this:\r\n\r\n```\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/monkey_patching.py:19: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses\r\n import imp\r\n```\r\n\r\nDo you by chance mess something up with the importer there? it feels like something overrides `import` because suddenly it ignores import errors. (**edit**: ruled that out too - see the next comment)", "It's possibly the doings of `PL`, observe this:\r\n```\r\nPYTHONPATH=\"../../src\" python -c \"import utils\"\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/examples/seq2seq/utils.py\", line 22, in <module>\r\n from transformers import BartTokenizer, EvalPrediction, PreTrainedTokenizer, T5Tokenizer\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/src/transformers/__init__.py\", line 22, in <module>\r\n from .integrations import ( # isort:skip\r\n File \"/mnt/nvme1/code/huggingface/transformers-master-stas00/src/transformers/integrations.py\", line 16, in <module>\r\n import comet_ml # noqa: F401\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/__init__.py\", line 34, in <module>\r\n from .api import API, APIExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/api.py\", line 28, in <module>\r\n from .experiment import CommonExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/experiment.py\", line 97, in <module>\r\n from .gpu_logging import (\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/gpu_logging.py\", line 27, in <module>\r\n import pynvml\r\nModuleNotFoundError: No module named 'pynvml'\r\n```\r\nand now let's add `import pytorch_lightning` first:\r\n```\r\nPYTHONPATH=\"../../src\" python -c \"import pytorch_lightning; import utils\"\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/util.py:36: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import namedtuple, Mapping, Sequence\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/vendor/graphql-core-1.1/graphql/type/directives.py:55: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n assert isinstance(locations, collections.Iterable), 'Must provide locations for directive.'\r\n/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/monkey_patching.py:19: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses\r\n import imp\r\n2020-11-09 14:35:24.485463: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1\r\n/mnt/nvme1/code/github/00nlp/fairseq/fairseq/optim/adam.py:8: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated since Python 3.3, and in 3.9 it will stop working\r\n from collections import Collection\r\n```\r\n\r\nThe exception was suppressed - so most likely it's PL's doing - most likely `import` functionality gets modified incorrectly.\r\n\r\nDo you have enough @dsblank to proceed from here? this is no longer `transformers`-related.", "Yes, I'll take it from here. We have a fix, and I'll test some more tomorrow and probably have a new comet_ml release out shortly. Thanks again!", "Oh, and I forgot to give the fully isolated command that reproduces the problem:\r\n```\r\npython -c \"import pytorch_lightning; import comet_ml\"\r\n```\r\ndoesn't fail and it should\r\n\r\nwhereas:\r\n```\r\npython -c \"import comet_ml\"\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/__init__.py\", line 34, in <module>\r\n from .api import API, APIExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/api.py\", line 28, in <module>\r\n from .experiment import CommonExperiment\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/experiment.py\", line 97, in <module>\r\n from .gpu_logging import (\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/comet_ml/gpu_logging.py\", line 27, in <module>\r\n import pynvml\r\nModuleNotFoundError: No module named 'pynvml'\r\n```\r\nas it should." ]

[]

[]

[]

[ { "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" } ]

[]

[ "Note that `exbert` integration is unfortunately not automatic. The authors from Exbert (@bhoov @HendrikStrobelt) need to re-deploy manually to add support for a new model. We've discussed with @Narsil @mfuntowicz hooking this into the hosted Inference API, so this might be automatic in the future.\r\n\r\ncc'ing @JetRunner \r\n\r\nTo help us understand, what's your use case for ExBERT @smanjil?", "> Note that `exbert` integration is unfortunately not automatic. The authors from Exbert (@bhoov @HendrikStrobelt) need to re-deploy manually to add support for a new model. We've discussed with @Narsil @mfuntowicz hooking this into the hosted Inference API, so this might be automatic in the future.\r\n> \r\n> cc'ing @JetRunner\r\n> \r\n> To help us understand, what's your use case for ExBERT @smanjil?\r\n\r\n\r\n@julien-c I did not know that. \r\n\r\nAs this model is a fine-tuned model for German medical domain texts, I wanted to see the attention distribution as done in German BERT. \r\n\r\nI believe this will be helpful for me as well as others to understand the effects of fine-tuning. Lastly, I tried with jessevig in colab, but, I have to fire it up in colab everytime. And, it was difficult to load the fine-tuned model there.\r\n\r\nSo, I am looking for a possibility here, and hope it will be done.", "@julien-c Oh I thought the model was already added to ExBERT. Good to know!" ]

[]

[]

[ "but when i use the same way to initializing `MyBert` which is based on the `BertModel`, it works and have a good result.", "This is probably because you put `self.bert = RobertaModel(config)`. It's looking for the identifier `roberta`, which isn't in your model.\r\n\r\nTry replacing `self.roberta = RobertaModel(config)`?", "> This is probably because you put `self.bert = RobertaModel(config)`. It's looking for the identifier `roberta`, which isn't in your model.\n> \n> \n> \n> Try replacing `self.roberta = RobertaModel(config)`?\n\nThank you very much for replying, i did as you said and replaced `self.cls = RobertaLMHead(config)` with `self.lm_head = RobertaLMHead(config)`, it worked well, Thanks!", "Hi @LysandreJik, sorry for reviving this old thread, but could you point to me where can I find this info in the docs? I'm interested to know what is the identifier used for different models." ]

[ { "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" } ]

[]

[]

[ { "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" } ]

[]

[]

[]

[]

[ "Hey @lessenko - the model is useless for any real application as it's just randomly initialized. It's only used for testing purposes.", "Hi @patrickvonplaten,\r\nThanks. " ]

[]

[]

[ "I managed to install `nvidia-apex` binary via conda:\r\n```\r\nconda install nvidia-apex -c conda-forge\r\n``` \r\nSo now I was able to validate that with pytorch-1.5 + nvdia-apex `--fp16` consumes less memory, than w/o `--fp16`. \r\n\r\nI was able to squeeze bs=20 (!) onto a 8gb card.\r\n\r\nSo the problem has to do with pytorch's 1.6 fp16\r\n", "Found another report of memory increase with fp16:\r\nhttps://discuss.pytorch.org/t/fp16-training-with-feedforward-network-slower-time-and-no-memory-reduction/95560/\r\n", "Michael Carilli suggested I add this to the top of the script to find the problem:\r\n```\r\nimport torch\r\ntorch.cuda.amp.autocast = \"hfdj\"\r\n```\r\nrerunning gives:\r\n```\r\nValidation sanity check: 0it [00:00, ?it/s]Traceback (most recent call last):\r\n File \"finetune.py\", line 449, in <module>\r\n main(args)\r\n File \"finetune.py\", line 416, in main\r\n trainer: pl.Trainer = generic_train(\r\n File \"/mnt/nvme1/code/huggingface/transformers-watchdog/examples/lightning_base.py\", line 395, in generic_train\r\n trainer.fit(model)\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/trainer/trainer.py\", line 445, in fit\r\n results = self.accelerator_backend.train()\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/accelerators/gpu_accelerator.py\", line 54, in train\r\n results = self.train_or_test()\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/accelerators/accelerator.py\", line 74, in train_or_test\r\n results = self.trainer.train()\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/trainer/trainer.py\", line 467, in train\r\n self.run_sanity_check(self.get_model())\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/trainer/trainer.py\", line 671, in run_sanity_check\r\n _, eval_results = self.run_evaluation(test_mode=False, max_batches=self.num_sanity_val_batches)\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/trainer/trainer.py\", line 591, in run_evaluation\r\n output = self.evaluation_loop.evaluation_step(test_mode, batch, batch_idx, dataloader_idx)\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/trainer/evaluation_loop.py\", line 176, in evaluation_step\r\n output = self.trainer.accelerator_backend.validation_step(args)\r\n File \"/mnt/nvme1/code/github/00pytorch/pytorch-lightning/pytorch_lightning/accelerators/gpu_accelerator.py\", line 75, in validation_step\r\n with torch.cuda.amp.autocast():\r\nTypeError: 'str' object is not callable\r\n```\r\n\r\nI am on the master of PL (pytorch-lightning), but tried bisecting on earlier versions (3 months back) with no change in behavior.", "The same issue, when I run the train_distilbart_xsum.sh, in sanity check step, meet OOM in 12G card", "Based on some initial debugging Lightning is calling into `trainer,model.validation_step`, which is calling `BartForConditionalGeneration` with fixed tensor sizes of:\r\n```python\r\ntorch.Size([48, 1, 1024]) torch.Size([50264, 1024]) torch.Size([1, 50264])\r\n```\r\nIn each iteration more tensors are allocated and never freed, which yields the OOM in the `sanity_check`.\r\n\r\nAMP run before OOM:\r\n```python\r\n|===========================================================================|\r\n| PyTorch CUDA memory summary, device ID 0 |\r\n|---------------------------------------------------------------------------|\r\n| CUDA OOMs: 0 | cudaMalloc retries: 0 |\r\n|===========================================================================|\r\n| Metric | Cur Usage | Peak Usage | Tot Alloc | Tot Freed |\r\n|---------------------------------------------------------------------------|\r\n| Allocated memory | 13553 MB | 13555 MB | 139840 MB | 126286 MB |\r\n| from large pool | 13551 MB | 13551 MB | 137780 MB | 124228 MB |\r\n| from small pool | 2 MB | 25 MB | 2059 MB | 2057 MB |\r\n|---------------------------------------------------------------------------|\r\n| Active memory | 13553 MB | 13555 MB | 139840 MB | 126286 MB |\r\n| from large pool | 13551 MB | 13551 MB | 137780 MB | 124228 MB |\r\n| from small pool | 2 MB | 25 MB | 2059 MB | 2057 MB |\r\n|---------------------------------------------------------------------------|\r\n| GPU reserved memory | 13888 MB | 13888 MB | 13888 MB | 0 B |\r\n| from large pool | 13858 MB | 13858 MB | 13858 MB | 0 B |\r\n| from small pool | 30 MB | 30 MB | 30 MB | 0 B |\r\n|---------------------------------------------------------------------------|\r\n| Non-releasable memory | 315554 KB | 913 MB | 109380 MB | 109072 MB |\r\n| from large pool | 313557 KB | 912 MB | 107043 MB | 106737 MB |\r\n| from small pool | 1997 KB | 8 MB | 2336 MB | 2334 MB |\r\n|---------------------------------------------------------------------------|\r\n| Allocations | 3403 | 3410 | 29885 | 26482 |\r\n| from large pool | 3161 | 3161 | 7265 | 4104 |\r\n| from small pool | 242 | 267 | 22620 | 22378 |\r\n|---------------------------------------------------------------------------|\r\n| Active allocs | 3403 | 3410 | 29885 | 26482 |\r\n| from large pool | 3161 | 3161 | 7265 | 4104 |\r\n| from small pool | 242 | 267 | 22620 | 22378 |\r\n|---------------------------------------------------------------------------|\r\n| GPU reserved segments | 368 | 368 | 368 | 0 |\r\n| from large pool | 353 | 353 | 353 | 0 |\r\n| from small pool | 15 | 15 | 15 | 0 |\r\n|---------------------------------------------------------------------------|\r\n| Non-releasable allocs | 204 | 210 | 14948 | 14744 |\r\n| from large pool | 200 | 205 | 3772 | 3572 |\r\n| from small pool | 4 | 20 | 11176 | 11172 |\r\n|===========================================================================|\r\n```\r\n\r\nFP32 run for the same step:\r\n```python\r\n|===========================================================================|\r\n| PyTorch CUDA memory summary, device ID 0 |\r\n|---------------------------------------------------------------------------|\r\n| CUDA OOMs: 0 | cudaMalloc retries: 0 |\r\n|===========================================================================|\r\n| Metric | Cur Usage | Peak Usage | Tot Alloc | Tot Freed |\r\n|---------------------------------------------------------------------------|\r\n| Allocated memory | 3948 MB | 4000 MB | 203229 MB | 199281 MB |\r\n| from large pool | 3946 MB | 3998 MB | 201302 MB | 197356 MB |\r\n| from small pool | 2 MB | 24 MB | 1927 MB | 1925 MB |\r\n|---------------------------------------------------------------------------|\r\n| Active memory | 3948 MB | 4000 MB | 203229 MB | 199281 MB |\r\n| from large pool | 3946 MB | 3998 MB | 201302 MB | 197356 MB |\r\n| from small pool | 2 MB | 24 MB | 1927 MB | 1925 MB |\r\n|---------------------------------------------------------------------------|\r\n| GPU reserved memory | 4226 MB | 4226 MB | 4226 MB | 0 B |\r\n| from large pool | 4198 MB | 4198 MB | 4198 MB | 0 B |\r\n| from small pool | 28 MB | 28 MB | 28 MB | 0 B |\r\n|---------------------------------------------------------------------------|\r\n| Non-releasable memory | 259725 KB | 778 MB | 118090 MB | 117836 MB |\r\n| from large pool | 257727 KB | 777 MB | 115922 MB | 115670 MB |\r\n| from small pool | 1997 KB | 8 MB | 2168 MB | 2166 MB |\r\n|---------------------------------------------------------------------------|\r\n| Allocations | 415 | 422 | 23263 | 22848 |\r\n| from large pool | 173 | 175 | 3796 | 3623 |\r\n| from small pool | 242 | 267 | 19467 | 19225 |\r\n|---------------------------------------------------------------------------|\r\n| Active allocs | 415 | 422 | 23263 | 22848 |\r\n| from large pool | 173 | 175 | 3796 | 3623 |\r\n| from small pool | 242 | 267 | 19467 | 19225 |\r\n|---------------------------------------------------------------------------|\r\n| GPU reserved segments | 85 | 85 | 85 | 0 |\r\n| from large pool | 71 | 71 | 71 | 0 |\r\n| from small pool | 14 | 14 | 14 | 0 |\r\n|---------------------------------------------------------------------------|\r\n| Non-releasable allocs | 17 | 26 | 13706 | 13689 |\r\n| from large pool | 13 | 14 | 2264 | 2251 |\r\n| from small pool | 4 | 21 | 11442 | 11438 |\r\n|===========================================================================|\r\n```\r\nThe allocations increase using `AMP` by ~50 in each iteration, while FP32 increases them sometimes by ~1.\r\nSo far I wasn't able to isolate the memory increase. ", "@ptrblck, and if the same is done with pt15/nvidia-apex - why doesn't the same happen there? How are the two different (native vs apex)", "I don't know, why an older PyTorch version with apex/amp works, as I wasn't able to isolate the issue yet.\r\nThe native amp implementation differs in various ways from `apex/amp`.\r\nBtw. do you only see the OOM, if you are using Lightning or also using the standalone model + amp?", "@ptrblck, I have tried the hugging face trainer and the problem doesn't seem to happen there. I don't think `--fp16` makes any difference in that trainer but it doesn't increase memory requirements. So is it possible this is specifically a PL issue?\r\n\r\nFor testing I used:\r\n\r\nsetup:\r\n```\r\ncd examples/seq2seq\r\nwget https://cdn-datasets.huggingface.co/translation/wmt_en_ro.tar.gz\r\ntar -xzvf wmt_en_ro.tar.gz\r\n```\r\nand then run:\r\n```\r\n\r\nbs=11; rm -rf tmpdir; PYTHONPATH=\"../../src\" python ./finetune_trainer.py \\\r\n--model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --data_dir wmt_en_ro --output_dir tmpdir \\\r\n--overwrite_output_dir --max_source_length 128 --max_target_length 128 --val_max_target_length 128 \\\r\n--do_train --do_eval --do_predict --num_train_epochs 10 --per_device_train_batch_size $bs \\\r\n--per_device_eval_batch_size $bs --learning_rate 3e-4 --warmup_steps 2 --evaluate_during_training \\\r\n--predict_with_generate --logging_steps 0 --save_steps 2 --eval_steps 2 --sortish_sampler \\\r\n--label_smoothing 0.1 --adafactor --task translation --tgt_lang ro_RO --src_lang en_XX --n_train 100 \\\r\n--n_val 50 --fp16\r\n```\r\n(same memory consumption w/o `--fp16`)\r\n\r\nThis command also uses `BartForConditionalGeneration`.\r\n\r\n`bs=11` is the biggest batch size I could fit onto the 8GB card, `bs=12` OOMs\r\n", "@ptrblck How did you make such a nice table?\r\n\r\n@stas00 I will check your #s on my card. cc @patil-suraj \r\n\r\n", "I replicated the OOM and fixed by passing `--amp_backend='apex'` in my torch 1.6 environment on a 24GB card. Would still be good to see if there is any easy way to get native amp working well.", "Thank you for validating that, @sshleifer. So it really has something to do with the native amp in pytorch.\r\n\r\nHere is the summary of my experiments:\r\n\r\n- pt15 + confa-force apex w/ `--fp16` works at the start - reduces memory consumption\r\n- pt16 + conda-forge apex w/ `--fp16 --amp_backend='apex'` works at the start too!\r\n\r\nbut both fail at the end with:\r\n```\r\n File \"python3.8/site-packages/apex/amp/_amp_state.py\", line 32, in warn_or_err\r\n raise RuntimeError(msg)\r\nRuntimeError: Found param model.model.shared.weight with type torch.cuda.HalfTensor, expected torch.cuda.FloatTensor.\r\nWhen using amp.initialize, you do not need to call .half() on your model\r\n```\r\nIf I use `--fp16_opt_level O1`, the failure changes to:\r\n```\r\n File \"python3.8/site-packages/torch/optim/lr_scheduler.py\", line 56, in with_counter\r\n instance_ref = weakref.ref(method.__self__)\r\nAttributeError: 'function' object has no attribute '__self__'\r\n```\r\nboth envs use the binary apex from `conda-forge`\r\n\r\n- apex doesn't support cuda11 at the moment\r\n- rtx-30* cards don't support cuda<11 (and really cuda<11.1).\r\n- ubuntu-20.4 doesn't support cuda<11, since it dropped gcc7, so can't build apex from source even for cuda-10\r\n\r\nBottom line, apex is not a great option at the moment, but may work for some short term - need to sort out native amp.\r\n\r\nI will poke at it with debugger today.", "I'm running in parallel `pt15+apex` and `pt16+native` with debugger:\r\n\r\nFound the first issue. \r\n\r\nAt this point in stack (we are in PL domain) - both have about 1.7GB allocated on GPU:\r\n```\r\nrestore_weights, checkpoint_connector.py:64\r\nsetup_training, training_loop.py:174\r\ntrain, gpu_accelerator.py:51\r\nfit, trainer.py:444\r\ngeneric_train, lightning_base.py:398\r\nmain, finetune.py:413\r\n<module>, finetune.py:446\r\n```\r\n\r\nnext step: `torch.cuda.empty_cache()` frees about 0.6GB on pt15, but 0 on pt16.\r\n\r\nCould `GradScaler` be holding onto objects and not letting them go? I commented out its init and there is no change. \r\n\r\nPerhaps there are some circular references preventing the objects from being cleared out.\r\n", "@sshleifer I used [`torch.cuda.memory_summary()`](https://pytorch.org/docs/stable/cuda.html#torch.cuda.memory_summary) and added it for each iteration in the `BartForConditionalGeneration` model.\r\n\r\nAfter some more debugging it seems that the `autocast` cache is blowing up.\r\nAs a workaround you can add `torch.clear_autocast_cache()` in [BartForConditionalGeneration.forward](https://github.com/huggingface/transformers/blob/eb3bd73ce35bfef56eeb722d697f2d39a06a8f8d/src/transformers/modeling_bart.py#L1036), which might slow down your code but should at least work.\r\n\r\nBased on @stas00 's debugging it seems that PL is interacting with native AMP in a way that the cache is increasing.\r\nCC @mcarilli", "The autocast cache is [cleared automatically](https://github.com/pytorch/pytorch/blob/88ec72e1c2a4b1e2a15cbe4703b9567bf9369a09/torch/cuda/amp/autocast_mode.py#L127) every time you exit an autocast context, which is one reason autocast should wrap the forward pass then exit.\r\n\r\n@ptrblck where in `BartConditionalGeneration.forward` did you call `clear_autocast_cache()` to resolve the memory blowup?\r\n\r\nIt's also helpful to write in the following sequence:\r\n```\r\ntorch.autocast_increment_nesting()\r\nprint(torch.autocast_decrement_nesting())\r\n```\r\nto see how deeply we're nested in autocast contexts at that point in forward. It should print `1`.", "@ptrblck, if I use your suggestion to add `torch.clear_autocast_cache()` torch blows up:\r\n\r\n```\r\nEpoch 0: 50%|████████████████Traceback (most recent call last): | 1/2 [00:02<00:02, 2.37s/it, loss=2.934, v_num=156]\r\n File \"finetune.py\", line 447, in <module>\r\n main(args)\r\n File \"finetune.py\", line 414, in main\r\n trainer: pl.Trainer = generic_train(\r\n File \"/mnt/nvme1/code/huggingface/transformers-watchdog/examples/lightning_base.py\", line 403, in generic_train\r\n trainer.fit(model)\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/trainer/trainer.py\", line 444, in fit\r\n results = self.accelerator_backend.train()\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py\", line 63, in train\r\n results = self.train_or_test()\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/accelerators/accelerator.py\", line 74, in train_or_test\r\n results = self.trainer.train()\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/trainer/trainer.py\", line 493, in train\r\n self.train_loop.run_training_epoch()\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/trainer/training_loop.py\", line 589, in run_training_epoch\r\n self.trainer.run_evaluation(test_mode=False)\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/trainer/trainer.py\", line 578, in run_evaluation\r\n output = self.evaluation_loop.evaluation_step(test_mode, batch, batch_idx, dataloader_idx)\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/trainer/evaluation_loop.py\", line 171, in evaluation_step\r\n output = self.trainer.accelerator_backend.validation_step(args)\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py\", line 85, in validation_step\r\n output = self.__validation_step(args)\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py\", line 95, in __validation_step\r\n output = self.trainer.model.validation_step(*args)\r\n File \"finetune.py\", line 183, in validation_step\r\n return self._generative_step(batch)\r\n File \"finetune.py\", line 216, in _generative_step\r\n generated_ids = self.model.generate(\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/torch/autograd/grad_mode.py\", line 26, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/mnt/nvme1/code/huggingface/transformers-watchdog/src/transformers/generation_utils.py\", line 553, in generate\r\n return self.beam_search(\r\n File \"/mnt/nvme1/code/huggingface/transformers-watchdog/src/transformers/generation_utils.py\", line 950, in beam_search\r\n beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)\r\nUnboundLocalError: local variable 'torch' referenced before assignment\r\nException ignored in: <function tqdm.__del__ at 0x7f1ec3812ee0>\r\nTraceback (most recent call last):\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/tqdm/std.py\", line 1122, in __del__\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/tqdm/std.py\", line 1335, in close\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/tqdm/std.py\", line 1514, in display\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/tqdm/std.py\", line 1125, in __repr__\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/tqdm/std.py\", line 1475, in format_dict\r\nTypeError: cannot unpack non-iterable NoneType object\r\n```\r\n\r\nthis is on pt-nightly, the only pt version I can use rtx-3090 24G card.\r\n\r\n**correction**: on pt-nightly it blows up with this error w/ or w/o cache clearing - it just doesn't work. I checked `torch` is imported at the point it says it's not defined.\r\n\r\n**edit2**: some problem related to `--warmup 500` - I set it to `1` and the above failure is gone - something to deal with separately.\r\n\r\nLet me see if I can figure out how to test this on pt16 - the problem is the 8GB card still can't fit even a single sample.", "> It's also helpful to write in the following sequence:\r\n> \r\n> ```\r\n> torch.autocast_increment_nesting()\r\n> print(torch.autocast_decrement_nesting())\r\n> ```\r\n> \r\n> to see how deeply we're nested in autocast contexts at that point in forward. It should print `1`.\r\n\r\nI validated - it prints `1`.", "for calling `torch.clear_autocast_cache()` how does a forward know it's in the autocast context?\r\n```\r\nif torch.autocast_enabled:\r\n torch.clear_autocast_cache()\r\n```\r\nI don't see a public API to check that [here](https://pytorch.org/docs/stable/_modules/torch/cuda/amp/autocast_mode.html#autocast)\r\n", "`torch.is_autocast_enabled()`", "Thank you, I see it now in the source code - should it be in the docs too? https://pytorch.org/docs/stable/amp.html", "I'm not sure, I want people to use autocast through the context manager interface. I guess it's useful for debugging.", "Oh, for sure. I don't think any of the `transformers` models should be made autocast-aware - this is the job of the trainer.", "Hmm, now I'm able to use the 24GB card so it's much easier to debug as I don't hit OOM all the time. Though I can't compare with apex, as I can't build it for cuda-11 - but I hope it should still be OK.\r\n\r\nWhat it appears to be is the beam search (size=4) consumes some 20GB just to search for 1 sample. It calls forward about 100 times each time allocating about 200MB - never releasing memory. I suppose because with apex the model is much more lean it requires much less memory.\r\n\r\nIf I add @ptrblck's suggestion to clear the autocast cache in forward it now consumes only 2.5GB - that's 1/8th of the same with the cache.\r\n\r\nW/o fp16 it consumes 5GB for the same beam_seach operation.\r\n\r\nLet's summarize:\r\n\r\ntype | memory\r\n----------|-------\r\nw/o fp16 | 5GB\r\nw/ fp16 | 20GB\r\nw/ fp16 + cache flush | 2.5GB\r\n\r\nSo definitely this is at least a huge part of the culprit. I think we are making progress. Much appreciation for your suggestion, @ptrblck~\r\n\r\nSo how do you recommend to proceed? Should PL have an option to clear the autocast cache?\r\n\r\nShould `autocast` cache be made smarter and flush automatically if gpu ram is 90% full and be called to check when this happens ala `gc.collect()`-timing and not just in the context of its use? Since now it created an additional cache in to add to to `cuda.cache`.", "If you follow along this is the command I use now: `PYTHONPATH=\"../../src\" CUDA_VISIBLE_DEVICES=0 python finetune.py --learning_rate 3e-5 --gpus 1 --do_train --val_check_interval 1 --num_train_epochs 1 --freeze_encoder --freeze_embeds --data_dir cnn_dm --max_target_length 142 --val_max_target_length 142 --train_batch_size 1 --eval_batch_size 1 --gradient_accumulation_steps 1 --model_name_or_path sshleifer/student_cnn_12_6 --tokenizer_name facebook/bart-large --warmup_steps 1 --output_dir distilbart-cnn-12-6 --overwrite_output_dir --num_sanity_val_steps 0 --n_train 1 --n_val 1 --fp16`", "Don't think this is a trainer issue, I've been able to replicate this OOM crash putting autocast directly into the forward call, wrapping the code found here with `torch.cuda.amp.autocast`:\r\n\r\nhttps://github.com/huggingface/transformers/blob/master/src/transformers/modeling_bart.py#L419\r\n\r\nand turning off `--fp16`. Haven't been able to investigate further into why memory isn't being freed in this block!", "Thank you for looking into it, @SeanNaren!\r\n\r\n> Don't think this is a trainer issue\r\n\r\nIf it's an issue of `autocast` cache then it is by definition a PL issue - since it's the one managing it - but let's see what @ptrblck and @mcarilli say about what's the correct approach of taking advantage of native amp w/ incurring unreasonable increased memory demands. \r\n\r\nSurely there must be a way to manage it efficiently, and if caching is bad for whatever reason (application specific?) - there must be a way to turn it off in first place, rather than wasting resources/time copying/deleting it repeatedly.\r\n\r\n> Haven't been able to investigate further into why memory isn't being freed in this block!\r\n\r\n@ptrblck and I already did - it's the `autocast` cache (at least a huge part of it). See https://github.com/huggingface/transformers/issues/8403#issuecomment-725562117", "Thanks @stas00! Any idea why within that block of code I posteted above that the autocast cache is not being freed? Btw if you wrap that particular section with `with torch.cuda.amp.autocast(enabled=False)`, the code runs (which I assume just turns off autocast functionality for that region of code). ", "@SeanNaren, if PL still invokes the `autocast` context then this sub-context it won't free the cache on its exit. It will only free it when the most outer context will exit. You will have to remove the `autocast` call in PL for this to work.", "Updates so far:\r\n\r\nIt appears that `autocast` is not designed to handle a massive number of `forward` calls in a single context - it caches them all! If I understand it correctly, it has to be called as close as possible to the first `forward` call that actually needs the casting. Currently, in `finetune.py` we end up with PL calling `autocast` on `SummarizationModule` which has nothing to do with pytorch math (i.e. needs no casting), which then goes through a massive logic of generate/beam_search which again has nothing to do with math, and only when it hits `BartForConditionalGeneration`'s `forward` we need the `autocast`.\r\n\r\nThe problem is that `generate` (which already runs under `autocast` via PL - so caching is on) ends up calling `BartForConditionalGeneration`'s `forward` 100s of times, and every such call in the debug example was adding 200MB - so in 100 calls for beam search of size 4 it accumulated 20GB - we have a big problem.\r\n\r\nSo a workaround suggested by @ptrblck is to call \r\n```\r\nif torch.autocast_enabled:\r\n torch.clear_autocast_cache()\r\n```\r\ninside `BartForConditionalGeneration.forward` - but ideally it should be called in a more generic way in `generation_utils` `while` loop ([here](https://github.com/huggingface/transformers/blob/121c24efa4453e4e726b5f0b2cf7095b14b7e74e/src/transformers/generation_utils.py#L954)), so that it works for any `transformers` model:\r\n\r\n```\r\n while cur_len < max_length:\r\n model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)\r\n if torch.is_autocast_enabled():\r\n torch.clear_autocast_cache()\r\n outputs = self(**model_inputs, return_dict=True)\r\n```\r\nbut this is clearly far from optimal as a lot of resources will be wasted on filling the cache and immediately emptying it.\r\n\r\n(Perhaps there is a way to disable the cache completely, but I don't know about its performance implications).\r\n\r\nSo a more efficient solution would be to `autocast` here instead:\r\n```\r\n while cur_len < max_length:\r\n model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)\r\n if somehow_we_know_autocast_should_be_used():\r\n with autocast(): \r\n outputs = self(**model_inputs, return_dict=True)\r\n else:\r\n outputs = self(**model_inputs, return_dict=True) \r\n```\r\n\r\nbut we have two issues here:\r\n\r\n1. we have a problem if someone invoked `autocast()` sooner - e.g. as PL or HF trainer do it now. As this will defeat the purpose and the cache will blow up again. There is no cache-per-context, but only a single cache, regardless of whether the contexts are stacked. The outer call defines the scope for the cache and it'll clear only on the exit of that call. So the `autocast` call above for all means and purposes is a no-op if `autocast` has already been called in earlier frames.\r\n\r\n2. now we are mixing trainer logic with the middle-layer (`generate` is neither a trainer nor a model - it's in between, and `SummarizationModule` in `finetune.py` is definitely not a model, but more of a trainer). How would `generation_utils` know about `somehow_we_know_autocast_should_be_used()`?\r\n\r\nAt this point since the problem is better understood I invite @LysandreJik, @patrickvonplaten, @sgugger and others to chime in and suggest how to move forward.", "I found at least part of the culprit or trigger of the leak - it's `@torch.no_grad()` used for `generate` https://github.com/huggingface/transformers/blob/eb3bd73ce35bfef56eeb722d697f2d39a06a8f8d/src/transformers/generation_utils.py#L281-L282\r\n\r\nHere is a short script that reproduces the leakage. It removes all the generate/search logic and feeds the same random input_ids to `BartForConditionalGeneration` pre-trained model.\r\n\r\nPlease first run:\r\n```\r\npip install ipyexperiments \r\n```\r\nto get the memory tracing, but feel free to disable it if for some reason it's not working for you. (it should)\r\n\r\n```\r\n#!/usr/bin/env python\r\n\r\nimport os\r\nimport sys\r\nimport torch\r\nos.environ[\"USE_TF\"] = \"0\"\r\nsys.path.insert(1, \"src\")\r\n\r\n# !pip install ipyexperiments \r\nfrom ipyexperiments.utils.mem import gpu_mem_get_used_mbs, gpu_mem_get_used_no_cache_mbs\r\n\r\nfrom transformers import BartForConditionalGeneration\r\n\r\ndevice = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')\r\n\r\nmodel = BartForConditionalGeneration.from_pretrained('sshleifer/student_cnn_12_6').to(device)\r\nmodel.eval()\r\n\r\nvocab_size = 50264 # model.config.vocab_size\r\nlength = 10\r\n\r\nAUTOCAST = False if \"-f\" in sys.argv else True\r\nprint(f\"autocast: {AUTOCAST}\")\r\n\r\nclass MemReport():\r\n def __init__(self, gc_collect=True):\r\n self.get_mem = gpu_mem_get_used_no_cache_mbs if gc_collect else gpu_mem_get_used_mbs\r\n self.cur = self.get_mem()\r\n def delta(self, id):\r\n peak = torch.cuda.memory_stats()[\"allocated_bytes.all.peak\"]\r\n print(f\"{id}: {gpu_mem_get_used_mbs()-self.cur}MB (peak {peak>>20}MB)\")\r\n self.cur = self.get_mem()\r\n \r\nmr = MemReport(gc_collect=False)\r\n\r\n### reproducible code starts here ###\r\n\r\[email protected]_grad()\r\ndef logic():\r\n input_ids = torch.randint(vocab_size, (1,length)).to(device)\r\n mr.delta(0)\r\n for i in range(1,10):\r\n outputs = model(input_ids)\r\n mr.delta(i)\r\n\r\nif AUTOCAST:\r\n with torch.cuda.amp.autocast():\r\n logic()\r\nelse:\r\n logic()\r\n```\r\n\r\nSo if I run it with `-f` which disables `autocast`, I get:\r\n\r\n```\r\n./reproduce.py -f\r\nautocast: False\r\n0: 0MB (peak 1165MB)\r\n1: 12MB (peak 1167MB)\r\n2: 0MB (peak 1169MB)\r\n3: 0MB (peak 1169MB)\r\n4: 0MB (peak 1169MB)\r\n5: 0MB (peak 1169MB)\r\n6: 0MB (peak 1169MB)\r\n7: 0MB (peak 1169MB)\r\n8: 0MB (peak 1169MB)\r\n9: 0MB (peak 1169MB)\r\n```\r\nno leak.\r\n\r\nIf however I remove `-f` and `autocast` gets enabled, we get:\r\n```\r\n./reproduce.py -f\r\nautocast: True\r\n0: 0MB (peak 1165MB)\r\n1: 592MB (peak 1744MB)\r\n2: 580MB (peak 2324MB)\r\n3: 580MB (peak 2902MB)\r\n4: 580MB (peak 3480MB)\r\n5: 580MB (peak 4058MB)\r\n6: 580MB (peak 4636MB)\r\n7: 580MB (peak 5214MB)\r\n8: 580MB (peak 5793MB)\r\n9: 580MB (peak 6371MB)\r\n```\r\nthe memory logger prints the delta for each `forward` call in the loop and the peak memory.\r\n\r\nYou can see that we are leaking 600Mb per forward call here.\r\n\r\nIf I comment out `@torch.no_grad()`, the total memory usage doubles but there is no leak:\r\n\r\n```\r\nautocast: True\r\n0: 0MB (peak 1165MB)\r\n1: 602MB (peak 1754MB)\r\n2: 590MB (peak 2343MB)\r\n3: 0MB (peak 2343MB)\r\n4: 0MB (peak 2343MB)\r\n5: 0MB (peak 2343MB)\r\n6: 0MB (peak 2343MB)\r\n7: 0MB (peak 2343MB)\r\n8: 0MB (peak 2343MB)\r\n9: 0MB (peak 2343MB)\r\n```\r\n\r\nI was using pycharm to debug this and to write a small script and boy it got me so delayed as it leaks gpu ram on its own, since it has to save all those variables on cuda, but I wasn't aware of it. Well, now I know not to do that. Luckily I had https://github.com/stas00/ipyexperiments handy to give me easy memory tracing.\r\n\r\nNote I'm importing two gpu mem tracking functions - one of them clears cuda cache - but here it appears it's better not use that version. ", "One other issue to look into is that what happens under `autocast` to weights that are deterministic (such as positional weights [SinusoidalPositionalEmbedding](https://github.com/huggingface/transformers/blob/24184e73c441397edd51e9068e0f49c0418d25ab/src/transformers/modeling_bart.py#L1340)) as these are set with `requires_grad = False`.\r\n\r\nSeeing how the caching logic [works](https://github.com/pytorch/pytorch/blob/21f447ee2c6ebbd72b6c3608c4df17c74edd4784/aten/src/ATen/autocast_mode.cpp#L69-L71):\r\n```\r\n bool can_try_cache = (to_type == at::kHalf && arg.scalar_type() == at::kFloat && arg.requires_grad() && arg.is_leaf());\r\n```\r\nthe conversion of these to fp16 will not be cached if I read the code correctly. This probably belongs to a separate issue though." ]

[ { "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" } ]

[]

[ "The Inference API is failing to load your model so there might be a configuration issue:\r\n\r\n<img width=\"770\" alt=\"Screenshot 2020-11-08 at 11 02 01\" src=\"https://user-images.githubusercontent.com/326577/98462120-a5bb2280-217f-11eb-8777-8456534b99ef.png\">\r\n\r\nI suspect this line is not working: `tokenizer = AutoTokenizer.from_pretrained(\"mymusise/gpt2-medium-chinese\")`\r\n\r\nWhat kind of tokenizer are you using?", "(cc @Narsil @JetRunner)", "Well many models actually use BERT Chinese's tokenizer so maybe that's the case. If so, how can we combine the BERT tokenizer with GPT model? (I think we've handled similar situations before)", "> The Inference API is failing to load your model so there might be a configuration issue:\r\n> \r\n> <img alt=\"Screenshot 2020-11-08 at 11 02 01\" width=\"770\" src=\"https://user-images.githubusercontent.com/326577/98462120-a5bb2280-217f-11eb-8777-8456534b99ef.png\">\r\n> \r\n> I suspect this line is not working: `tokenizer = AutoTokenizer.from_pretrained(\"mymusise/gpt2-medium-chinese\")`\r\n> \r\n> What kind of tokenizer are you using?\r\n\r\nYes, I use BertTokenizer, but how can i change the Tokenizer and Model class of the default example?\r\n```\r\nfrom transformers import AutoTokenizer, TFAutoModelWithLMHead\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(\"mymusise/gpt2-medium-chinese\")\r\n\r\nmodel = TFAutoModelWithLMHead.from_pretrained(\"mymusise/gpt2-medium-chinese\")\r\n```\r\nI specify `architectures=[\"TFGPT2LMHeadModel\"]` on my configs, but it didn't show in the model card.\r\n\r\nIs there any document I can reference?", "> Well many models actually use BERT Chinese's tokenizer so maybe that's the case. If so, how can we combine the BERT tokenizer with GPT model? (I think we've handled similar situations before)\r\n\r\nYes, I have tried to use `GPT2Tokenizer` by specific `vocab_file` and `merges_file`, but it didn't work.\r\n```\r\ntokenizer = GPT2Tokenizer(vocab_file=my_vocab_file, merges_file=my_merges_file)\r\n```\r\nIs there any examples about how to custom GPT2Tokenizer that I can reference? Many thanks.", "> > Well many models actually use BERT Chinese's tokenizer so maybe that's the case. If so, how can we combine the BERT tokenizer with GPT model? (I think we've handled similar situations before)\n> \n> \n> \n> Yes, I have tried to use `GPT2Tokenizer` by specific `vocab_file` and `merges_file`, but it didn't work.\n> \n> ```\n> \n> tokenizer = GPT2Tokenizer(vocab_file=my_vocab_file, merges_file=my_merges_file)\n> \n> ```\n> \n> Is there any examples about how to custom GPT2Tokenizer that I can reference? Many thanks.\n\ncc @julien-c ", "Yes, see the doc here: https://huggingface.co/transformers/master/main_classes/configuration.html – look for `tokenizer_class`\r\n\r\nIn config.json:\r\n\r\n> The name of the associated tokenizer class to use (if none is set, will use the tokenizer associated to the model by default).", "> Yes, see the doc here: https://huggingface.co/transformers/master/main_classes/configuration.html – look for `tokenizer_class`\r\n> \r\n> In config.json:\r\n> \r\n> > The name of the associated tokenizer class to use (if none is set, will use the tokenizer associated to the model by default).\r\n\r\n\r\nThanks julien, I have added `tokenizer_class=\"BertTokenizer\"` to [configs](https://s3.amazonaws.com/models.huggingface.co/bert/mymusise/gpt2-medium-chinese/config.json) and update it. But the web page still using `AutoTokenizer` as example. Cache?\r\n", "The webpage will still display `AutoTokenizer` – but it will now work.", "As you can see the Inference widget now works: https://huggingface.co/mymusise/gpt2-medium-chinese?text=%E6%88%91%E5%8F%AB%E6%9C%B1%E5%88%A9%E5%AE%89%EF%BC%8C%E6%88%91%E5%96%9C%E6%AC%A2", "@julien-c I think many Chinese models in our hub have the same problem. I'll have them fixed." ]

[]

[]

[ "> For example, if someone were to use `self.get_auto_remove_tmp_dir()` but didn't like where the tmp directory was set, they would specify the directory. They then re-run their code, but the behavior now has changed, even though they only updated the path. That seems unexpected and unwanted to me.\r\n\r\nYou're making a valid point, @LysandreJik and that's how I coded it in first place, except it lead to needing to type out many arguments. So I'm trying to make it smarter based on the experience of using it.\r\n\r\n1. I'd ask you - what do you mean specifically you don't like where it's located. Since by default it gets wiped out anyway what difference does its location make? The only time you'd want to set it to a desired location is when you want to see its contents - therefore, you will not want it wiped out ;) Does my thinking process make sense?\r\n\r\n If however the temp mechanism is wrong (say it fails to create a tmp dir under /tmp/ - dir is full), then it needs to be dealt with on the system/shell level and redefine where `/tmp` is located or empty it since this is a system-level problem. It'd be wrong to try to solve it in this helper function.\r\n\r\n2. The intention here is to make the debugging process as painless and quick as possible, so if the function behaves slightly differently than you'd normally expect from a function, but you use it all the time and you're well aware of its quirks and it saves you a lot of time and mistakes, then it serves the purpose, IMHO.\r\n\r\n3. I'm totally open to other solutions as long as they lead to easier debug.\r\n\r\nI don't know whether other devs have been using it and have some experiential input to share.\r\n", "Great, thanks for explaining!\r\n\r\n> I'd ask you - what do you mean specifically you don't like where it's located. Since by default it gets wiped out anyway what difference does its location make?\r\n\r\nYou're definitely right about this. Thanks for humoring me! Good to merge for me.", "this one is ready to merge when you get a chance - thank you!", "it should be safe to merge - the CI failures are unrelated" ]

[]

[]

[ "Agreed that more than 8 records is the way to go. Feel free to address that in a future PR!" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi @Backpackerice did you solve the issue? I also encounter the same issue while training with `Trainer` API.", "I have found the issue. I was giving the model that I am initialising before `Trainer` object. So, it wasn't initialised with a seed. I solved the issue by setting the seed before initialising the model. It can be also solved by providing `model_init` arg for `Trainer`. " ]

[ { "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 good now!\r\n<img width=\"645\" alt=\"Screenshot 2020-11-08 at 10 25 27\" src=\"https://user-images.githubusercontent.com/326577/98461482-748c2380-217a-11eb-9fe4-38d6dc555233.png\">\r\n" ]

[]

[]

[ "Wait, the example for `run_mlm_wwm.py` doesn't seem to run:\r\n```\r\nTraceback (most recent call last):\r\n File \"run_mlm_wwm.py\", line 334, in <module>\r\n main()\r\n File \"run_mlm_wwm.py\", line 161, in main\r\n model_args, data_args, training_args = parser.parse_args_into_dataclasses()\r\n File \".../transformers/src/transformers/hf_argparser.py\", line 144, in parse_args_into_dataclasses\r\n raise ValueError(f\"Some specified arguments are not used by the HfArgumentParser: {remaining_args}\")\r\nValueError: Some specified arguments are not used by the HfArgumentParser: ['--dataset_name', 'wikitext', '--dataset_config_name', 'wikitext-2-raw-v1']\r\n```\r\n\r\nI will try to fix it too.", "Thanks for the fixes. I'm unsure the script can run on datasets hosted on the hub because of the need for reference files, that's why I hadn't added the `dataset_name` and `dataset_config_name` arguments in this one.", "You can still ignore the `dataset_name` arg and use the old behavior, so adding `dataset_name` should be fine.\r\nI want to mention that I did not run the example with ref file.", "Ok, then we should add a check in the post_init to make sure no reference file is passed if we're using `dataset_name`, otherwise it won't 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." ]

[]

[]

[ "You might also test that the config is identical to `BartConfig`.\r\n", "Thank you @sshleifer for your review, I added some additional integration tests.", "Hi @sgugger, thank you for your review, I added all the proposed changes.\r\n\r\nAs for the tokenizer, the reason for having a vocab file, is that mBarthez uses the same sentencepiece tokenizer as mBart, while discarding tokens with non-Latin characters from the embedding layers. So basically the token-ids mapping has changed. I am not sure if it is possible to change the sentencepiece model itself. Anyway I think we can keep it like that for the moment. \r\n\r\nPlease let me know if you would like to recommend any other changes. ", "Hi @LysandreJik, thank you for the review. Yes you're right, actually I was hesitating whether to redefine barthez models or not. Anyway I modified the code as requested. Hope it's ok now. ", "@LysandreJik @julien-c Please let me know if any other modifications are required. Thank you in advance :) ", "Hi @moussaKam! Sorry about getting back to you so late. The issue with this PR is with the implementation of the tokenizer and its fast tokenizer counterpart. Right now there is no `BarthezTokenizerFast`, which we would really need as SentencePiece is not installed by default anymore.\r\n\r\nIt seems that the `BarthezTokenizer` is very similar to the `XLMRobertaTokenizer`, so I'm trying to see if we can't do a similar conversion between your tokenizer and the resulting tokenizer. It seems completely feasible.\r\n\r\nDo you think you could take a look at the [`convert_slow_tokenizers_checkpoints_to_fast.py` module](https://github.com/huggingface/transformers/blob/master/src/transformers/convert_slow_tokenizers_checkpoints_to_fast.py), and at the [`XLMRobertaConverter` object available in the `convert_slow_tokenizer.py` module](https://github.com/huggingface/transformers/blob/master/src/transformers/convert_slow_tokenizer.py#L466) to see if such a conversion if possible? Thank you.", "Hi @LysandreJik, I added the fast tokenizer, thank you for the tip! Please let me know if we're good now! :)", "Hi @LysandreJik, do you think we still need any changes?", "I think we can merge it as it is right now, and use the `legacy_format=False` when saving the slow tokenizer. We're thinking of a way to enable this by default for fast tokenizers, but this isn't blocking for this PR. Thanks!" ]

[]

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

[ "> The training script should work for all datasets in huggingface datasets.\r\n\r\nTo clarify our intent with the examples we provide, this is untrue. We can't support every use case of every user. The script can easily be modified and has plenty of comments to help the user tune them to their needs.\r\n\r\nIn this particular instance however, I think your proposed fix doesn't break anything and allows us to support more datasets with little change, so I think we can add it. Don't hesitate to suggest a PR and tag me on it.", "Thanks for the reply~ Since this is my first time to do make a pr, I followed on the [guide](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) and run `make test` as suggested. However, most of the test failed.\r\n`947 failed, 188 passed, 124 skipped, 25 warnings in 37.09s`\r\nThose test failed before I even makes any changes to the code. Sorry if this is a newbie question but should we expect all the test to pass?", "Yes, the guide is what is also used by the CI to run all tests, so they should all pass. Make sure you do have the installation from source with all the dev dependencies installed. Apart from that, there is little I can do to help without seeing any error message.", "I finally found out the reason I fail most of the test. It's because I have another training on my gpu which interferes with testing for some reason. \r\n\r\nNow I have all test passed except for 8 of them. Here are those 8 test: \r\n\r\n================================ short test summary info ================================\r\nFAILED tests/test_modeling_ctrl.py::CTRLModelTest::test_multi_gpu_data_parallel_forward\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_multi_gpu_data_parallel_forward\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_multi_gpu_data_parallel_forward\r\nFAILED tests/test_trainer.py::TrainerIntegrationTest::test_dynamic_shapes - IndexError...\r\nFAILED tests/test_trainer.py::TrainerIntegrationTest::test_evaluate - AssertionError: ...\r\nFAILED tests/test_trainer.py::TrainerIntegrationTest::test_load_best_model_at_end - As...\r\nFAILED tests/test_trainer.py::TrainerIntegrationTest::test_number_of_steps_in_training\r\nFAILED tests/test_trainer_distributed.py::TestTrainerDistributed::test_trainer - Runti...\r\n======== 8 failed, 4035 passed, 688 skipped, 608 warnings in 2855.87s (0:47:35) =========\r\n\r\nAgain, those test are run after I follows the guide by installing from the source with the dev dependencies **but before I made any change to the source code**. So I think just installing from the source with dev dependencies and run \"make test\" will be enough to reproduce the error.\r\n\r\nHere's the traceback error message for 4 of those erorr:\r\n\r\n**FAILED tests/test_modeling_ctrl.py::CTRLModelTest::test_multi_gpu_data_parallel_forward**\r\n```\r\n\r\n ...[-1.1481e-01, -8.0696e-02, 7.7946e-02, ..., -6.4783e-02,\r\n 1.2656e-01, 7.6060e-02]]]]], device='cuda:2'))\r\ndim = 0, destination = 0\r\n\r\n def gather(tensors, dim=0, destination=None, *, out=None):\r\n r\"\"\"Gathers tensors from multiple GPU devices.\r\n \r\n Arguments:\r\n tensors (Iterable[Tensor]): an iterable of tensors to gather.\r\n Tensor sizes in all dimensions other than :attr:`dim` have to match.\r\n dim (int, optional): a dimension along which the tensors will be\r\n concatenated. Default: ``0``.\r\n destination (torch.device, str, or int, optional): the output device.\r\n Can be CPU or CUDA. Default: the current CUDA device.\r\n out (Tensor, optional, keyword-only): the tensor to store gather result.\r\n Its sizes must match those of :attr:`tensors`, except for :attr:`dim`,\r\n where the size must equal ``sum(tensor.size(dim) for tensor in tensors)``.\r\n Can be on CPU or CUDA.\r\n \r\n .. note::\r\n :attr:`destination` must not be specified when :attr:`out` is specified.\r\n \r\n Returns:\r\n - If :attr:`destination` is specified,\r\n a tensor located on :attr:`destination` device, that is a result of\r\n concatenating :attr:`tensors` along :attr:`dim`.\r\n - If :attr:`out` is specified,\r\n the :attr:`out` tensor, now containing results of concatenating\r\n :attr:`tensors` along :attr:`dim`.\r\n \"\"\"\r\n if out is None:\r\n if destination == -1:\r\n warnings.warn(\r\n 'Using -1 to represent CPU tensor is deprecated. Please use a '\r\n 'device object or string instead, e.g., \"cpu\".')\r\n destination = _get_device_index(destination, allow_cpu=True, optional=True)\r\n> return torch._C._gather(tensors, dim, destination)\r\nE RuntimeError: Input tensor at index 2 has invalid shape [2, 4, 4, 7, 8], but expected [2, 5, 4, 7, 8]\r\n\r\n../../test_env_6/lib/python3.8/site-packages/torch/nn/parallel/comm.py:230: RuntimeError\r\n\r\n```\r\n\r\n\r\n\r\n**FAILED tests/test_trainer.py::TrainerIntegrationTest::test_dynamic_shapes - IndexError: list index out of range**\r\n```\r\n\r\ntests/test_trainer.py:366: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _\r\nsrc/transformers/trainer.py:1308: in evaluate\r\n output = self.prediction_loop(\r\nsrc/transformers/trainer.py:1416: in prediction_loop\r\n for step, inputs in enumerate(dataloader):\r\n../../test_env_6/lib/python3.8/site-packages/torch/utils/data/dataloader.py:435: in __next__\r\n data = self._next_data()\r\n../../test_env_6/lib/python3.8/site-packages/torch/utils/data/dataloader.py:475: in _next_data\r\n data = self._dataset_fetcher.fetch(index) # may raise StopIteration\r\n../../test_env_6/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py:44: in fetch\r\n data = [self.dataset[idx] for idx in possibly_batched_index]\r\n../../test_env_6/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py:44: in <listcomp>\r\n data = [self.dataset[idx] for idx in possibly_batched_index]\r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _\r\n\r\nself = <tests.test_trainer.DynamicShapesDataset object at 0x7f547e1b0250>, i = 48\r\n\r\n def __getitem__(self, i):\r\n> return {\"input_x\": self.xs[i], \"labels\": self.ys[i]}\r\nE IndexError: list index out of range\r\n```\r\n\r\n\r\n**FAILED tests/test_trainer.py::TrainerIntegrationTest::test_evaluate - AssertionError: ...**\r\n```\r\n\r\n_______________________________________________________________________ TrainerIntegrationTest.test_evaluate ________________________________________________________________________\r\n\r\nself = <tests.test_trainer.TrainerIntegrationTest testMethod=test_evaluate>\r\n\r\n def test_evaluate(self):\r\n trainer = get_regression_trainer(a=1.5, b=2.5, compute_metrics=AlmostAccuracy())\r\n results = trainer.evaluate()\r\n \r\n x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]\r\n pred = 1.5 * x + 2.5\r\n expected_loss = ((pred - y) ** 2).mean()\r\n> self.assertAlmostEqual(results[\"eval_loss\"], expected_loss)\r\nE AssertionError: 0.37259840965270996 != 0.3790205 within 7 places (0.006422102451324463 difference)\r\n\r\ntests/test_trainer.py:312: AssertionError\r\n```\r\n\r\n\r\n**FAILED tests/test_trainer.py::TrainerIntegrationTest::test_load_best_model_at_end - As...**\r\n\r\n```\r\n> self.check_saved_checkpoints(tmpdir, 64 // self.batch_size, total)\r\n\r\ntests/test_trainer.py:565: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _\r\ntests/test_trainer.py:211: in check_saved_checkpoints\r\n self.assertTrue(os.path.isdir(checkpoint))\r\nE AssertionError: False is not true\r\n```", "Again, this is exactly what the CI does, so those failures are linked to your particular environment. Since you didn't tell us what it is we can't reproduce and fix potential issues.\r\nOne way to make sure you don't use your GPU if it's busy is to run `CUDA_VISIBLE_DEVICES='' make tests`.", "Thanks for the reply. Sorry I forgot to list my environment. But as you suggest, `CUDA_VISIBLE_DEVICES=''`, all the test pass when I run them without GPU using. You are right that the test failures are linked to my environment. It's irrelevant to this issue so I think I'll make another post about it.", "Hi newbie questions are better suited for the forum at https://discuss.huggingface.co\r\n\r\nWe try to keep the issues for bug reports and features/model requests.", "> Hi newbie questions are better suited for the forum at https://discuss.huggingface.co\r\n> \r\n> We try to keep the issues for bug reports and features/model requests.\r\n\r\nYeah, you are right. I will make another post in the forum instead. " ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "The problem can be fixed by changing `e[\"input_ids\"].tolist()` to `e[\"input_ids\"]` " ]

[]

[]

[ "I don't know if this PR is in its final state and ready for review now, but here are a few pointers:\r\n\r\nThe PR looks good, but there are a few things to do before merging.\r\n\r\n- You should take a look at the failing tests and fix them if they're related to your PR (they should be)\r\n- You should implement the tests related to the added classes\r\n- You should add these classes to the auto models and init.py", "Hi @LysandreJik. Thanks for the comments. Apologies though, this PR is still a WIP so no need for review atm. I plan on adding my classes to auto models and init.py and running the appropriate tests before committing my final version.", "Great, thanks for letting me know. Will wait for your ping before reviewing again!", "HI @LysandreJik & @patrickvonplaten, I'm hoping to get your help with the only remaining test that's failing for this PR. The `tests/test_modeling_distilbert.py` script fails due to an error in the `tests/test_modeling_common.py` in line 966. I've detailed the error message below but in short what I think is happening is that check_equivalence fails on the following check:\r\n\r\n```\r\ncheck_equivalence(\r\n model, tuple_inputs, dict_inputs, {\"output_hidden_states\": True, \"output_attentions\": True}\r\n)\r\n```\r\n\r\nWhat's strange is that all the other equivalence checks pass and I haven't changed the hidden states or attention layers.\r\n\r\nError Message:\r\n\r\n```\r\ntests/test_modeling_common.py:1037: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _\r\ntests/test_modeling_common.py:1001: in check_equivalence\r\n recursive_check(tuple_output, dict_output)\r\ntests/test_modeling_common.py:990: in recursive_check\r\n recursive_check(tuple_iterable_value, dict_iterable_value)\r\ntests/test_modeling_common.py:990: in recursive_check\r\n recursive_check(tuple_iterable_value, dict_iterable_value)\r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _\r\n\r\ntuple_object = tensor([[[[0.4205, 0.0000, 0.2311, ..., 0.0000, 0.0000, 0.3484],\r\n [0.4074, 0.0000, 0.2666, ..., 0.0000, 0.0...71, 0.0000, 0.1483, ..., 0.1858, 0.1648, 0.1354],\r\n [0.1422, 0.0000, 0.2250, ..., 0.1523, 0.1632, 0.1669]]]])\r\ndict_object = tensor([[[ 1.3640e+00, 2.6618e-01, 5.9669e-01, ..., -7.1322e-01,\r\n -3.5144e-02, -1.3287e+00],\r\n [ 1....1.0870e-01],\r\n [-9.9694e-01, -7.1065e-01, -5.7949e-01, ..., 1.4234e+00,\r\n -2.1929e+00, 1.2390e-01]]])\r\n\r\n def recursive_check(tuple_object, dict_object):\r\n if isinstance(tuple_object, (List, Tuple)):\r\n for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):\r\n recursive_check(tuple_iterable_value, dict_iterable_value)\r\n elif tuple_object is None:\r\n return\r\n else:\r\n self.assertTrue(\r\n torch.allclose(\r\n> set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), atol=1e-5\r\n ),\r\n msg=f\"Tuple and dict output are not equal. Difference: {torch.max(torch.abs(tuple_object - dict_object))}. Tuple has `nan`: {torch.isnan(tuple_object).any()} and `inf`: {torch.isinf(tuple_object)}. Dict has `nan`: {torch.isnan(dict_object).any()} and `inf`: {torch.isinf(dict_object)}.\",\r\n )\r\nE RuntimeError: The size of tensor a (7) must match the size of tensor b (768) at non-singleton dimension 3\r\n\r\ntests/test_modeling_common.py:996: RuntimeError\r\n```", "Hey @KMFODA, \r\n\r\nThanks for all the hard work already! We'll definitely help you get this merged :-) @patil-suraj, could you maybe help here since you recently added the caching mechanism for `BertForCausalLM`? Otherwise I'll try to take a look by the end of this week!", "Thanks @patrickvonplaten. I'm working through Lysandre's comment which was very helpful and hoping this will fix the remaining error. Will await any further comments / advice on how to merge this from you / suraj as well in case I've missed something else out.", "Hi @KMFODA \r\n\r\nThanks for all the work! \r\n\r\n`DistillBertForCausalLM` can be used as a decoder in `EncoderDecoder` models. To use it as a decoder we should now implement `cross attention` (or `encoder-decoder` attention) so that it can attend to `encoder_hidden_states` and caching the `past_key_values` to speed-up inference.\r\n\r\nWe should implement this feature analogous to how it is implemented in `Bert`. This means that we should\r\n\r\n- Add cross-attention and caching mechanism `MultiHeadSelfAttention` as shown here in Bert\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/src/transformers/models/bert/modeling_bert.py#L259\r\n\r\n- this means we need to pass `encoder_hidden_states`, `encoder_attention_mask`, `past_key_values` through the layers. So introduce these parameters in `forward` of `DistilBertForCausalLM`, `DistilBertModel`, `Transformer`, `TransformerBlock` and `MultiHeadSelfAttention`.\r\n\r\n- compute cross attention in `TransformerBlock` as shown here in `BertLayer`\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/src/transformers/models/bert/modeling_bert.py#L478-L480\r\n\r\n- Adjust the position embedding and attention masks accordingly as done in `BertEmbeddings` and `BertModel`\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/src/transformers/models/bert/modeling_bert.py#L194\r\nand\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/src/transformers/models/bert/modeling_bert.py#L194\r\n\r\nThe important thing to note here is that, when `past_key_values` are enabled, the actual seq length is `cur_seq_length + pas_key_values_length`\r\n\r\n- Add tests for `DistillBertForCausalLM` that checks that the decoder and caching mechanism works as expected:\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/tests/test_modeling_bert.py#L170\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/tests/test_modeling_bert.py#L203\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/tests/test_modeling_bert.py#L230\r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/tests/test_modeling_bert.py#L263\r\n\r\n\r\nAll of this should be pretty similar to how it's implemented in Bert, so you could use that as a reference. Let me know if something is not clear. Happy to help you here :)\r\n\r\n", "Apolgoies, this PR got too messy when I tried to rebase to master and it closed automatically when I tried to revert back to an old commit. I'm working on @patil-suraj's very useful recommendations and will post a new PR when it passes all tests. Thanks!", "Hey @KMFODA \r\n\r\nGlad to know you're still working on this :) Don't hesitate to ping me if you're stuck or need some help." ]

[]

[]

[ "It seems you have a mismatch between several `transformers` versions.\r\n\r\nWe generally recommend using virtual environments not relying on `conda`, as full `transformers` support is not yet available on `conda`. Can you reproduce using a pip virtual env?\r\n\r\n```py\r\npython -m venv .env\r\nsource .env/bin/activate\r\npip install transformers\r\n```", "that works in the virtual env, thank you!" ]

[]

[]

[ "I understood this is implemented in Seq2SeqDataCollator with adding padding=longest. " ]

[]

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

[ "Great catch, you are correct." ]

[]

[]

[ "My gpu instances does not have access to download the files, could you please assist me? could it be not hardcoded in the codes? so user could download the files and put them in the directory?", "could you tell me what is this trying to download in the beginning? ", "apparently the code downloads some stuff in datasets/utils/file_utils.py ? could you tell me how to resolve this issue, I cannot really get the code to run on a gpu without access to download from amazon? thanks ", "my gpu instance cannot download \"\"https://s3.amazonaws.com/datasets.huggingface.co/datasets/datasets\" inside this file." ]

[]

[]

[ "You need to run `make fixup`" ]

[]

[]

[]

[]

[]

[ "Hi, these questions are better suited for the forum as we try to keep the github issues for bugs/feature requests only. Could you open an issue on the [forum](https://discuss.huggingface.co) instead?", "Hi, the link is broken. thanks\n\nOn Mon, Nov 9, 2020 at 3:50 PM Lysandre Debut <[email protected]>\nwrote:\n\n> Hi, these questions are better suited for the forum as we try to keep the\n> github issues for bugs/feature requests only. Could you open an issue on\n> the forum <https://hdiscuss.huggingface.co> instead?\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/8380#issuecomment-724060649>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ARPXHHZTAWRY5UBXSGMBOVTSO76R7ANCNFSM4TNT3AXA>\n> .\n>\n", "fixed!" ]

[]

[]

[ "Oh I had missed that this one was a subclass of `DataCollatorForLanguageModeling`, thanks for fixing. The last one isn't a subclass, but we can certainly change it to use the same function too!", "Merging the PR to have the fix as quickly as possible, but if you want to add a test of this data collator, please go ahead! That will make the library more robust." ]

[]

[]

[ "@RedneckedCrake Hey please checkout my PR and let me know if it works!", "Should be fixed by #8379, thanks for flagging this!" ]

[]

[]

[ "+1 It would be good to document supported/tested versions, but I think it would be useful to try to keep up to date (to the extent possible.) I think most people choose whether to upload to hf **after** they have trained a model, and simply won't upload if their version if it is not supported. (Rather than re-train with a supported `fairseq` version.) ", "@stas00 You could also write an integration test that runs only if fairseq is installed (probably in examples), that tests conversion. This wouldn't get run by CI, but might make trouble-shooting easier.", "Well, in this case, they haven't changed the output for the model - they changed their hub tools instead (i.e. how the model is loaded) - so no re-training is necessary. So requiring a fixed fairseq model would probably be sufficient.\r\n\r\nThe problem is that they stopped making releases last Dec. https://github.com/pytorch/fairseq/tags How can we tell a developer to install version X if it doesn't exist (or check for it)?\r\n\r\nI would leave it as is for now, since so far there is barely any interest in fsmt and should it start getting more traction I'll be all over it making it super-resilient. ", "Makes sense, feel free to merge @LysandreJik " ]

[]

[]

[ "For your test, you could check the code runs correctly with `dict(teacher=BART_TINY_RANDOM, student=T5_TINY_RANDOM)`", "I've pushed some fixes to get the tests to run successfully! However, I realized that an issue I was wondering about earlier is indeed important, and that is, dealing with the student and teacher models having different tokenizers, which would make the `input_ids` different in the two cases. However since the (language) models we are looking at all have the same output space (set of tokens) as the input space, in order for the logits from the output layers to be comparable, they have to have the same output space, and hence the same set of tokens. Does this make sense? And if so, it is a fair assumption that the teacher and student models should have the same tokenizer (specified via the existing `tokenizer_name` parameter), correct? In this case, unit testing distillation from `BART_TINY` to `T5_TINY` is not valid (and I do run into errors). I did try `T5_TINY` to `T5_TINY` and that works fine but of course it's not very satisfying. Can I add a test case for `T5_SMALL` to `T5_TINY`? Worried about slowing down the unit tests. Please let me know your suggestion. Sorry for the long comment!", "For testing,\r\nI just ran\r\n```bash\r\npython make_student.py patrickvonplaten/t5-tiny-random t5-tinier-random -e 1 -d 1\r\ntransformers-cli upload t5-tinier-random\r\n```\r\n\r\nSo now you can test distilling `patrickvonplaten/t5-tiny-random` (2/2 layers) to `sshleifer/t5-tinier-random` (1/1 layer).\r\n\r\n\r\n\r\n", "You also need to run `make fixup`", "> You also need to run `make fixup`\r\n\r\nI get the following message:\r\n```\r\n/bin/sh: 3: black: not found\r\n/bin/sh: 4: isort: not found\r\n/bin/sh: 5: flake8: not found\r\nMakefile:7: recipe for target 'modified_only_fixup' failed\r\nmake: *** [modified_only_fixup] Error 127\r\n```\r\n\r\nAny idea?", "> > You also need to run `make fixup`\r\n> \r\n> I get the following message:\r\n> \r\n> ```\r\n> /bin/sh: 3: black: not found\r\n> /bin/sh: 4: isort: not found\r\n> /bin/sh: 5: flake8: not found\r\n> Makefile:7: recipe for target 'modified_only_fixup' failed\r\n> make: *** [modified_only_fixup] Error 127\r\n> ```\r\n> \r\n> Any idea?\r\n\r\nAh. I wasn't familiar with these pre-commit libraries. Looks like I need to `pip install` those...", "I've added a unit test to distill from `T5_TINIER` to `T5_TINY` and it passes. Also ran `make fixup` and fixed formatting issues. Could you please review again?\r\nAll CircleCI tests above pass except the first one, `check_code_quality`. I checked the details on that one and they seem to be unrelated:\r\n\r\n```\r\nRun python -c \"import torch; print(torch.hub.list('huggingface/transformers:$BRANCH'))\"\r\nDownloading: \"https://github.com/huggingface/transformers/archive/add_student_base_model.zip\" to /home/runner/.cache/torch/hub/add_student_base_model.zip\r\nTraceback (most recent call last):\r\n File \"<string>\", line 1, in <module>\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/site-packages/torch/hub.py\", line 269, in list\r\n repo_dir = _get_cache_or_reload(github, force_reload, True)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/site-packages/torch/hub.py\", line 141, in _get_cache_or_reload\r\n download_url_to_file(url, cached_file, progress=False)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/site-packages/torch/hub.py\", line 425, in download_url_to_file\r\n u = urlopen(req)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 222, in urlopen\r\n return opener.open(url, data, timeout)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 531, in open\r\n response = meth(req, response)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 641, in http_response\r\n 'http', request, response, code, msg, hdrs)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 563, in error\r\n result = self._call_chain(*args)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 503, in _call_chain\r\n result = func(*args)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 755, in http_error_302\r\n return self.parent.open(new, timeout=req.timeout)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 531, in open\r\n response = meth(req, response)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 641, in http_response\r\n 'http', request, response, code, msg, hdrs)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 569, in error\r\n return self._call_chain(*args)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 503, in _call_chain\r\n result = func(*args)\r\n File \"/opt/hostedtoolcache/Python/3.7.9/x64/lib/python3.7/urllib/request.py\", line 649, in http_error_default\r\n raise HTTPError(req.full_url, code, msg, hdrs, fp)\r\nurllib.error.HTTPError: HTTP Error 404: Not Found\r\nError: Process completed with exit code 1.\r\n```", "thank you for the style fix! are you going to merge?", "@danyaljj FYI", "CircleCI `check_code_quality` is failing. I see--\r\n```\r\n#!/bin/bash -eo pipefail\r\nblack --check examples tests src utils\r\nwould reformat /home/circleci/transformers/examples/seq2seq/distillation.py\r\nOh no! 💥 💔 💥\r\n1 file would be reformatted, 562 files would be left unchanged.\r\n\r\nExited with code exit status 1\r\nCircleCI received exit code 1\r\n```\r\n\r\nBut not sure how to format it. Tried `black --code examples/seq2seq/distillation.py` but that doesn't work cause it expects code as a string, not a filename. How do you format a file?", "pip install -e .[dev]\nmake fixup \n", "> make fixup\r\n\r\nI did run `make fixup` before committing. But just ran both your above steps again. I don't get any errors from `distillation.py`.", "On a high-level, do you have a fine-tuned t5-large that you want to distill?\r\nThis makes the code a lot more complex so I am trying to understand the use-case better.", "Yes. We wish to create a small model (under 500 MB), distilled from a large or 11B T5 model. For this we want to start with a small T5 model and train using the larger T5 as the teacher. Conceptually distillation can be done from any large model to any small model as long they share the same output space, so this enables that.", "> For this we want to start with a small T5 model and train using the larger T5 as the teacher. \r\n\r\nMy point was that this only really makes sense if the large t5 is fine-tuned.", "ah, yes. we are using a fine-tuned large model.", "Cleaned up some things and now ready to merge. Thanks for the contribution @sbhaktha !", "Thanks Sam!" ]

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

[]

[ "Maybe related to https://github.com/huggingface/transformers/issues/6962", "Both tokenizers are very unrelated, so I don't think they can be directly compared like this.", "@LysandreJik can you elaborate on this unrelatedness? \r\nI think both are implemented in Rust and I assume the tokenizers use same algorithm for tokenization part? Maybe I am missing a detail here. ", "BERT's tokenizer uses WordPiece while RoBERTa's tokenizer is a byte-level BPE. They're not the same algorithm. You can have more information [here](https://huggingface.co/docs/tokenizers/python/latest/components.html#models).", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Thanks @LysandreJik for the explanation. I was not aware of this difference. " ]

[]

[]

[ "You were right, it was an EN to EL model so the src-tgt were inversed, however I changed the args and I tried again. \r\nIt gave me an error: \r\n```\r\nRuntimeError: Error(s) in loading state_dict for FSMTForConditionalGeneration:\r\n size mismatch for model.decoder.embed_tokens.weight: copying a param with shape torch.Size([12896, 512]) from checkpoint, the shape in current model is torch.Size([9936, 512]).\r\n```\r\nwhich I fixed by changing one of the ignore keys due to a typo probably, in line 251 (see comment below):\r\n```\r\n # remove unneeded keys\r\n ignore_keys = [\r\n \"model.model\",\r\n \"model.encoder.version\",\r\n \"model.decoder.version\",\r\n \"model.encoder_embed_tokens.weight\", \r\n \"model.decoder.embed_tokens.weight\",#here, the original was: model.decoder_embed_tokens.weight\r\n \"model.encoder.embed_positions._float_tensor\",\r\n \"model.decoder.embed_positions._float_tensor\",\r\n ]\r\n for k in ignore_keys:\r\n model_state_dict.pop(k, None)\r\n```\r\nAfter that, the conversion script concludes succesfully: \r\n```\r\nGenerating data/wmt16-el-en-dist/vocab-src.json\r\nGenerating data/wmt16-el-en-dist/vocab-tgt.json\r\nGenerating data/wmt16-el-en-dist/merges.txt\r\nGenerating data/wmt16-el-en-dist/config.json\r\nGenerating data/wmt16-el-en-dist/tokenizer_config.json\r\nGenerating data/wmt16-el-en-dist/pytorch_model.bin\r\nConversion is done!\r\n\r\nLast step is to upload the files to s3\r\ncd data\r\ntransformers-cli upload wmt16-el-en-dist\r\n\r\n```\r\nI guess the next thing is to try to load this model locally to test that it's working.", "Fantastic! Yes, load it up and let us know whether it works.\r\n\r\nIf you want ready scripts to adapt from, perhaps try https://github.com/stas00/porting/blob/master/transformers/fairseq-wmt19/scripts/fsmt-translate.py\r\nand if you convert the reversed model (hint: just swap src and tgt languages in the conversion script) you can even do a paraphrase:\r\nhttps://github.com/stas00/porting/blob/master/transformers/fairseq-wmt19/scripts/fsmt-paraphrase.py\r\n", "> which I fixed by changing one of the ignore keys due to a typo probably, in line 251 (see comment below):\r\n> remove unneeded keys\r\n> ignore_keys = [\r\n> \"model.decoder.embed_tokens.weight\",#here, the original was: model.decoder_embed_tokens.weight\r\n\r\nI don't think that this solution works. While you made the conversion script complete, your ported model now has random weights for `model.decoder.embed_tokens.weight` - you probably are going to see garbage output.\r\n\r\nThe error is:\r\n```\r\nsize mismatch for model.decoder.embed_tokens.weight: copying a param with shape torch.Size([9936, 512]) from checkpoint, the shape in current model is torch.Size([12896, 512]).\r\n```\r\n* source here is 12892 (el)\r\n* and target is 9932 (en) \r\n\r\nFor some reason it's creating a decoder with the size of the encoder dict.\r\n\r\nI think there might have been a bug introduced since I wrote and used this script. I will debug and get back to you.", "Hmm, it looks like `fairseq` has introduced some breaking changes - that's why the script wasn't working out of the box. The `args` in the checkpoint appears to be mostly empty, so none of the wmt19 models can be converted either. Will investigate.\r\n\r\nThis is the breaking change: https://github.com/pytorch/fairseq/commit/3b27ed7996b0315f471c795cf9b7dfcc18467cbe\r\n\r\nThey did away with the `args` object", "OK, I updated the conversion script to support the latest fairseq and it now converts your model out of the box w/o needing any changes, please use the version in this PR if it hasn't been merged yet. https://github.com/huggingface/transformers/pull/8377\r\n\r\nPlease let me know whether the results are satisfactory and you get a good translation out of it - note it uses some default hparams (see the script) so you can adjust those to your liking. \r\n\r\nOnce you're happy you can upload the model to s3 as explained here: https://huggingface.co/transformers/model_sharing.html", "Resolved in https://github.com/huggingface/transformers/pull/8377", "This is great @stas00, I just tried it locally and it works as intended. Many thanks!\r\n\r\nOne quick question, during training with fairseq, the tokenization also converted all letters to lower-case (to reduce the vocab I assume) so now in order to get correct translations the input text needs to be lowercase only. I can of course add a line to my script to do that automatically, but I was wondering how I can force the uploaded model to do that (so that anyone wanting to test it doesn't have to download it locally and add that additional line). Probably with a config argument...?\r\n\r\nI am uploading this to s3 soon, your help has been invaluable :+1: ", "Excellent. I'm glad to hear we sorted it out.\r\n\r\nSo have you validated that the translation works and the bleu score evals are satisfactory? You can do it easily with `transformers` using the examples here: https://github.com/huggingface/transformers/tree/master/scripts/fsmt (scripts starting with `eval_`).\r\n\r\n> One quick question, during training with fairseq, the tokenization also converted all letters to lower-case (to reduce the vocab I assume) so now in order to get correct translations the input text needs to be lowercase only. I can of course add a line to my script to do that automatically, but I was wondering how I can force the uploaded model to do that (so that anyone wanting to test it doesn't have to download it locally and add that additional line). Probably with a config argument...?\r\n\r\nIt's currently not supported as all the models I worked with didn't have that restriction. I will add this functionality to the `transformers` implementation of FSMT now that I know that this is still needed. I will let you know when this is done. Perhaps hold off on making the release to ensure that your model works out of box. You will also need to update your ported model's config when this is done.\r\n\r\nOut of curiosity, if you don't mind sharing, is there a special reason why you chose to train an older more restricted architecture and not one of the newer ones? Surely, losing the normal casing would be a hurdle for practical use.", "\r\n> So have you validated that the translation works and the bleu score evals are satisfactory? You can do it easily with `transformers` using the examples here: https://github.com/huggingface/transformers/tree/master/scripts/fsmt (scripts starting with `eval_`).\r\n\r\nI validated the bleu and chrF scores on the fairseq equivalent of the model (before converting it to huggingface) on the Tatoeba testset, but now that there are additional evaluation scripts I will try these as well!\r\n\r\n> It's currently not supported as all the models I worked with didn't have that restriction. I will add this functionality to the `transformers` implementation of FSMT now that I know that this is still needed. I will let you know when this is done. Perhaps hold off on making the release to ensure that your model works out of box. You will also need to update your ported model's config when this is done.\r\n\r\nThanks for this, sure I can wait if there's the option of adding that feature too!\r\n\r\n> Out of curiosity, if you don't mind sharing, is there a special reason why you chose to train an older more restricted architecture and not one of the newer ones? Surely, losing the normal casing would be a hurdle for practical use.\r\n\r\nTbh, I was just following a fairseq [guide](https://github.com/pytorch/fairseq/tree/master/examples/translation#iwslt14-german-to-english-transformer) that was suggesting this arch over the following possible choices:\r\n`Possible choices: transformer, transformer_iwslt_de_en, transformer_wmt_en_de, transformer_vaswani_wmt_en_de_big, transformer_vaswani_wmt_en_fr_big, transformer_wmt_en_de_big, transformer_wmt_en_de_big_t2t, multilingual_transformer, multilingual_transformer_iwslt_de_en, fconv, fconv_iwslt_de_en, fconv_wmt_en_ro, fconv_wmt_en_de, fconv_wmt_en_fr, nonautoregressive_transformer, nonautoregressive_transformer_wmt_en_de, nacrf_transformer, iterative_nonautoregressive_transformer, iterative_nonautoregressive_transformer_wmt_en_de, cmlm_transformer, cmlm_transformer_wmt_en_de, levenshtein_transformer, levenshtein_transformer_wmt_en_de, levenshtein_transformer_vaswani_wmt_en_de_big, levenshtein_transformer_wmt_en_de_big, insertion_transformer, bart_large, bart_base, mbart_large, mbart_base, mbart_base_wmt20, lstm, lstm_wiseman_iwslt_de_en, lstm_luong_wmt_en_de, transformer_lm, transformer_lm_big, transformer_lm_baevski_wiki103, transformer_lm_wiki103, transformer_lm_baevski_gbw, transformer_lm_gbw, transformer_lm_gpt, transformer_lm_gpt2_small, transformer_lm_gpt2_medium, transformer_lm_gpt2_big, transformer_align, transformer_wmt_en_de_big_align, hf_gpt2, hf_gpt2_medium, hf_gpt2_large, hf_gpt2_xl, transformer_from_pretrained_xlm, lightconv, lightconv_iwslt_de_en, lightconv_wmt_en_de, lightconv_wmt_en_de_big, lightconv_wmt_en_fr_big, lightconv_wmt_zh_en_big, lightconv_lm, lightconv_lm_gbw, fconv_self_att, fconv_self_att_wp, fconv_lm, fconv_lm_dauphin_wikitext103, fconv_lm_dauphin_gbw, lstm_lm, roberta, roberta_base, roberta_large, xlm, masked_lm, bert_base, bert_large, xlm_base, s2t_berard, s2t_berard_256_3_3, s2t_berard_512_3_2, s2t_berard_512_5_3, s2t_transformer, s2t_transformer_s, s2t_transformer_sp, s2t_transformer_m, s2t_transformer_mp, s2t_transformer_l, s2t_transformer_lp, wav2vec, wav2vec2, wav2vec_ctc, wav2vec_seq2seq, dummy_model, transformer_lm_megatron, transformer_lm_megatron_11b, transformer_iwslt_de_en_pipeline_parallel, transformer_wmt_en_de_big_pipeline_parallel, model_parallel_roberta, model_parallel_roberta_base, model_parallel_roberta_large`\r\n\r\nIf you could indicate a more recent architecture that has about the same number of parameters (I have a constraint on complexity as I am using a single GTX2080SUPER) I would be happy to re-train!\r\n\r\n", "> > So have you validated that the translation works and the bleu score evals are satisfactory? You can do it easily with `transformers` using the examples here: https://github.com/huggingface/transformers/tree/master/scripts/fsmt (scripts starting with `eval_`).\r\n> \r\n> I validated the bleu and chrF scores on the fairseq equivalent of the model (before converting it to huggingface) on the Tatoeba testset, but now that there are additional evaluation scripts I will try these as well!\r\n\r\nMy only concern here is that the forced lower-case which won't be the case with the references bleu scores are evaled against.\r\n\r\n> > Out of curiosity, if you don't mind sharing, is there a special reason why you chose to train an older more restricted architecture and not one of the newer ones? Surely, losing the normal casing would be a hurdle for practical use.\r\n> \r\n> Tbh, I was just following a fairseq [guide](https://github.com/pytorch/fairseq/tree/master/examples/translation#iwslt14-german-to-english-transformer) that was suggesting this arch over the following possible choices:\r\n> `Possible choices: transformer, transformer_iwslt_de_en, transformer_wmt_en_de, transformer_vaswani_wmt_en_de_big, [...]\r\n> \r\n> If you could indicate a more recent architecture that has about the same number of parameters (I have a constraint on complexity as I am using a single GTX2080SUPER) I would be happy to re-train!\r\n\r\nI re-read the guide and I'm not sure what you mean when you said: \"was suggesting this arch over the following possible choices\" - I can't find any recommendations to use this particular model over the dozens of the ones you listed. e.g. how did you know that it's a smaller model than some others?\r\n\r\nI'm gradually getting to know the fairseq models and have only dealt with wmt-variations of `transformer`. I suppose you can see all the variations defined here and below: https://github.com/pytorch/fairseq/blob/master/fairseq/models/transformer.py#L985\r\nSo primarily these appear to differ in the size and shape of the model.\r\n\r\nWhen you did the training, was there an option not to force lowercase input or did it come automatic with the `transformer_iwslt_de_en`? I don't see an option to toggle this on/off in `fairseq-train` command. And looking around the code I don't quite see a configurable option to do so.", "True, the forced lower-case may give slightly higher BLEU score.\r\n\r\n> I re-read the guide and I'm not sure what you mean when you said: \"was suggesting this arch over the following possible choices\" - I can't find any recommendations to use this particular model over the dozens of the ones you listed. e.g. how did you know that it's a smaller model than some others?\r\n\r\nBy \"suggesting\" I mean that I just used the pre-defined arch on the available script (see below):\r\n```\r\nCUDA_VISIBLE_DEVICES=0 fairseq-train \\\r\n data-bin/iwslt14.tokenized.de-en \\\r\n --arch transformer_iwslt_de_en --share-decoder-input-output-embed \\\r\n --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \\\r\n --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \\\r\n --dropout 0.3 --weight-decay 0.0001 \\\r\n --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \\\r\n --max-tokens 4096 \\\r\n --eval-bleu \\\r\n --eval-bleu-args '{\"beam\": 5, \"max_len_a\": 1.2, \"max_len_b\": 10}' \\\r\n --eval-bleu-detok moses \\\r\n --eval-bleu-remove-bpe \\\r\n --eval-bleu-print-samples \\\r\n --best-checkpoint-metric bleu --maximize-best-checkpoint-metric\r\n```\r\nI didn't know it would lead to a smaller model before digging it up a bit and discovering that e.g. there's a difference to the ffd hidden layer dimension. I experimented with some of them (not all) and since I am not an expert ofc I ended up on that one based on a.the fact that it actually worked, b. that the perplexity I was getting was getting lower quicker that other cases, and c.more importantly based on whether I would get an OOM after a while (most of the cases) :P. \r\nDo you think that there would be a significant gain from trying a newer architecture that you have in mind?\r\n\r\n> When you did the training, was there an option not to force lowercase input or did it come automatic with the transformer_iwslt_de_en? I don't see an option to toggle this on/off in fairseq-train command. And looking around the code I don't quite see a configurable option to do so.\r\n\r\nThe lowercase command comes at the data preparation script provided in the guide: https://github.com/pytorch/fairseq/blob/master/examples/translation/prepare-iwslt14.sh\r\n(line 11). It was not part of the fairseq training/preprocessing.\r\n ", "> If you could indicate a more recent architecture that has about the same number of parameters (I have a constraint on complexity as I am using a single GTX2080SUPER) I would be happy to re-train!\r\n\r\nI think most of the recent ones are quite much bigger, but one that we ported that may warrant your attention is the distilled variation: https://github.com/jungokasai/deep-shallow/\r\n(ported scripts https://github.com/huggingface/transformers/tree/master/scripts/fsmt - start with `convert-allenai-`, \r\nthe wmt model cards are at the end of the list here https://huggingface.co/allenai)", "> I didn't know it would lead to a smaller model before digging it up a bit and discovering that e.g. there's a difference to the ffd hidden layer dimension. I experimented with some of them (not all) and since I am not an expert ofc I ended up on that one based on a.the fact that it actually worked, b. that the perplexity I was getting was getting lower quicker that other cases, and c.more importantly based on whether I would get an OOM after a while (most of the cases) :P.\r\n> Do you think that there would be a significant gain from trying a newer architecture that you have in mind?\r\n\r\nI'm relatively new to this myself, so I haven't tried enough variations yet to make such recommendations. Perhaps asking at the forums stating your limitations would lead to some excellent recommendations - or perhaps what you have done is just fine - it all depends on your needs. Do check out the distilled approach I mentioned in the comment above.\r\n\r\n> The lowercase command comes at the data preparation script provided in the guide: https://github.com/pytorch/fairseq/blob/master/examples/translation/prepare-iwslt14.sh\r\n> (line 11). It was not part of the fairseq training/preprocessing.\r\n\r\nOh, I see, this is totally circumstantial - you just trained on lower-cased input so this is the world it knows. This makes total sense. Thank you for helping me understand this nuance.", "OK, I implemented the lowercase config, and wanted to automate the discovery of when this option should be pre-set, but the latter didn't work - my detector code discovered up-case letters - I looked at both vocabs you supplied and both have upcase letters in them - a lot of them in the el one and some in en-one (merges/code too).\r\n\r\nI tried this very simple heuristic:\r\n```\r\n # detect whether this is a do_lower_case situation, which can be derived by checking whether we\r\n # have at least one upcase letter in the source vocab\r\n do_lower_case = True\r\n for k in src_vocab.keys():\r\n if not k.islower():\r\n do_lower_case = False\r\n break\r\n```\r\n\r\nI suppose this has to be set manually then because you know you trained mainly on lowercase - but perhaps there is a bug somewhere on the fairseq side and should not have any upcase letters in either of the 3 files if it were to be properly lower-cased?", "The PR that adds lower-case support is here https://github.com/huggingface/transformers/pull/8389, but for the converter to work with the recent fairseq https://github.com/huggingface/transformers/pull/8377 is needed to be merged first, or you can do just this over 8389:\r\n```\r\ndiff --git a/src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py b/src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py\r\nindex 2cc42718..61ef9010 100755\r\n--- a/src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py\r\n+++ b/src/transformers/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py\r\n@@ -113,7 +113,7 @@ def convert_fsmt_checkpoint_to_pytorch(fsmt_checkpoint_path, pytorch_dump_folder\r\n fsmt_folder_path, checkpoint_file, data_name_or_path, archive_map=models, **kwargs\r\n )\r\n\r\n- args = dict(vars(chkpt[\"args\"]))\r\n+ args = vars(chkpt[\"args\"][\"model\"])\r\n```\r\nOr if you don't want to mess with these, we can wait until both are merged.\r\n\r\nIn either case before uploading to s3 you will need to manually set `\"do_lower_case\": true` in `tokenizer_config.json` of the converted model - since as I mentioned in the comment above there is no way of automatically detecting the need to preset `\"do_lower_case\": true` during conversion as all vocabs in your model have uppercase letters in them.\r\n", "> I think most of the recent ones are quite much bigger, but one that we ported that may warrant your attention is the distilled variation: https://github.com/jungokasai/deep-shallow/\r\n(ported scripts https://github.com/huggingface/transformers/tree/master/scripts/fsmt - start with convert-allenai-,\r\nthe wmt model cards are at the end of the list here https://huggingface.co/allenai)\r\n\r\nI will try this for sure, although I remember that the `--arch transformer` used in the script led to OOM in my machine.\r\n\r\n> Or if you don't want to mess with these, we can wait until both are merged.\r\n> In either case before uploading to s3 you will need to manually set \"do_lower_case\": true in tokenizer_config.json of the converted model - since as I mentioned in the comment above there is no way of automatically detecting the need to preset \"do_lower_case\": true during conversion as all vocabs in your model have uppercase letters in them.\r\n\r\nWell I couldn't wait, so I tried following your steps and it works perfectly! :D Kudos once again!\r\nNow, my only question is should I upload the EN2EL and EL2EN models or wait until the PR is merged? I guess that the transformers version that is currently loading all s3-uploaded models is not up-to-date yet, so it will actually miss the capital letters.\r\n", "> I will try this for sure, although I remember that the `--arch transformer` used in the script led to OOM in my machine.\r\n\r\nI'm not sure whether fairseq has some doc that compares the different arch configs, but this piece of code seems to be pretty clear on the differences: https://github.com/pytorch/fairseq/blob/master/fairseq/models/transformer.py#L985\r\n\r\nThe default `transformer` uses pretty big layers - so it requires a lot of gpu memory.\r\n\r\n> Well I couldn't wait, so I tried following your steps and it works perfectly! :D Kudos once again!\r\n\r\nFantastic!\r\n\r\nI suppose you're not concerned with the upcase letters in the dict/merge files of your pre-trained model. I'd have thought that fairseq pre-processor would lowercased all inputs. But if you think it's no problem, then all is good.\r\n\r\n> Now, my only question is should I upload the EN2EL and EL2EN models or wait until the PR is merged? I guess that the transformers version that is currently loading all s3-uploaded models is not up-to-date yet, so it will actually miss the capital letters.\r\n\r\nYou have to first wait till the lowercasing-PR merged - probably Mon or early next week, and then AFAIK [the online version](https://huggingface.co/models) doesn't get updated automatically - the models' code doesn't change often - so we will have to ask for this to happen. And once you see the demo working on the site, then it's in the clear to share with others.", "Thank you for the code showcasing the differences between models. I couldn't find a doc with that info.\r\n\r\n> I suppose you're not concerned with the upcase letters in the dict/merge files of your pre-trained model. I'd have thought that fairseq pre-processor would lowercased all inputs. But if you think it's no problem, then all is good.\r\n\r\nProbably the perl command included in the fairseq preparation script didn't catch all cases; I can't think of another explanation. In any case I will be modifying this script to re-train without lower-casing and with a bigger number of BPE tokens, just to see if I get a more convenient model that doesn't need the lower-case argument (and hopefully without losing much BLEU).\r\n\r\n> You have to first wait till the lowercasing-PR merged - probably Mon or early next week, and then AFAIK the online version doesn't get updated automatically - the models' code doesn't change often - so we will have to ask for this to happen. And once you see the demo working on the site, then it's in the clear to share with others.\r\n\r\nOK so I'm waiting for the merge, then upload and probably come back in this thread to request an update on the online version if possible. Thanks for the help @stas00 ! \r\n ", "FYI, the lower-casing PR has been merged, so please let me know whether you're waiting to re-train with mixed-casing or whether you want to upload the lower-case model and I will then ask to update the code on the models server.", "I already started the mixed-casing training and I was thinking I can upload all 4 of them (lower EN2EL, lower EL2EN, mixed EN2EL, mixed EL2EN) together. The mixed ones also have a bigger vocabulary (almost double) and BLEU scores are very similar to the older lower case ones.", "Great!\r\n\r\nI made the request to update the server - I will update when this is done.\r\n\r\nIt's good to know that there is not much difference with the bigger vocab. I'm curious to how the scores would have been different if your original model was truly lower-case (since it's not at the moment if you check the vocab). (this is just for my learning should you ever run this test)", "Hi @stas00, just pinging to check if the code on the model hub is updated.\r\nI also trained cased models, and I uploaded one to the hub already: https://huggingface.co/lighteternal/SSE-TUC-mt-en-el-cased\r\n\r\nBut it returns an error when used online from the link above: \r\n```\r\nUnrecognized configuration class for this kind of AutoModel: AutoModelForCausalLM. Model type should be one of CamembertConfig, XLMRobertaConfig, RobertaConfig, BertConfig, OpenAIGPTConfig, GPT2Config, TransfoXLConfig, XLNetConfig, XLMConfig, CTRLConfig, ReformerConfig, BertGenerationConfig, XLMProphetNetConfig, ProphetNetConfig.\r\n```\r\nI also noticed that the tag directly above the field box is falsely assigned as \"text-generation\".\r\n", "Hi, it seems your model card is defined in markdown format not in Yaml: https://huggingface.co/lighteternal/SSE-TUC-mt-en-el-cased leading to incorrect pipeline detection (hence the error you are seeing). Can you try setting the pieline correctly ? \r\n\r\nhttps://huggingface.co/docs#how-are-model-tags-determined\r\n\r\nLet us know if it works better", "Thanks @Narsil I updated the model card, but it doesn't seem to have made any difference yet. Is it possible that it takes some time to change pipeline?", "It appears that editing the README from the browser, doesn't work; after pulling, editing and pushing again it worked! Many thanks! :) ", "> Thanks @Narsil I updated the model card, but it doesn't seem to have made any difference yet. Is it possible that it takes some time to change pipeline?\r\n\r\nMaybe we forgot to hook a refresh on commits from the website, @Pierrci?", "Indeed, I pushed a fix that will be deployed soon, next time you edit a readme or another file on the website the changes will reflect instantly @lighteternal, thanks for reporting this!" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi,\r\n\r\nYou can move the path of the \"**_runs/_**\" folder by adding `logging_dir` to your `TrainingArguments` such like this:\r\n\r\n```python\r\ntraining_args = TrainingArguments(\r\n output_dir=OUTPUT_PATH,\r\n overwrite_output_dir=True,\r\n logging_dir=OUTPUT_PATH + \"/logs\",\r\n ...\r\n)\r\n```" ]

[]

[]

[ "This is for **extractive** QA which means the answer needs to be in a continuous sequence in the context. The way the model works is by predicting the start and end indices of the answer. ", "> This is for **extractive** QA which means the answer needs to be in a continuous sequence in the context. The way the model works is by predicting the start and end indices of the answer.\r\n\r\nthanks for help \r\nI found what I needed" ]

[]

[]

[ "Nice!! Now I have with `make fixup`:\r\n\r\n```\r\n-n was unexpected.\r\nmake: *** [modified_only_fixup] Error 255\r\n```", "OK, so it sounds that perhaps I was trying to fix the wrong thing in first place. Let's see if we can find a different way to check whether the variable is empty that also works on windows.", "@jplu, please try now - I used `test -n` instead.", "same error", "ok, can you perhaps tell me which test would work on windows? \r\n\r\nWe are doing a very simple thing\r\n```\r\nif len(foo):\r\n do a\r\nelse:\r\n do b\r\n```\r\nwhat's a good test for us to use for `len(foo)` on windows?\r\n\r\nalso what's your `test -h` shows - perhaps no `-n` option there?\r\n\r\non unix:\r\n```\r\n\r\n -n STRING\r\n the length of STRING is nonzero\r\n\r\n STRING equivalent to -n STRING\r\n```\r\n\r\ndoes it work if you just remove \"-n\", i.e. `test -n` and just `test` are supposed to do the same.\r\n", "Doesn't matter which command I do with `test` it doesn't works. I don't think we will succeed to run any shell code, only the commands are available not the shell interpreter. Would it be possible to turn that into a Python subprocess that runs itself the `black`, `isort` and `flake8` command?", "It's possible, but surely there must be a way to do such a basic thing on windows.\r\n\r\nHow about this?\r\n\r\n```\r\n\t@if [ \"$(modified_py_files)\" != \"\" ]; then \\\r\n\t\techo \"Checking/fixing $(modified_py_files)\"; \\\r\n\telse \\\r\n\t\techo \"No library .py files were modified\"; \\\r\n\tfi\r\n```\r\n\r\nIf that doesn't work either, perhaps it doesn't like the spacing? remove space before `\\`?", "I get:\r\n```\r\n\"utils/check_repo.py\" was unexpected.\r\nmake: *** [modified_only_fixup] Erreur 255\r\n```\r\n\r\nWith:\r\n\r\n```\r\n$(eval modified_py_files := $(shell python utils/get_modified_files.py $(check_dirs)))\r\n@if [ \"$(modified_py_files)\" != \"\" ]; then \\\r\n\techo \"Checking/fixing $(modified_py_files)\"; \\\r\nelse \\\r\n\techo \"No library .py files were modified\"; \\\r\nfi\r\n```\r\n\r\n\"utils/check_repo.py\" is the value of `$(modified_py_files)`", "Thank you for checking that and overall your patience so far. \r\n\r\nThis back-n-force doesn't seem to lead to a productive way of sorting it out. I need to have direct access to experiment on.\r\n\r\nIs there a way I could get access to a windows box with a similar setup to yours - I won't suppose there is google colab like environment but on windows?\r\n\r\n@sgugger, is there someone on the hf team who groks windows shell/make and could help us out here? We need to be able to do such basic things in `Makefile` as checking if a variable is not \"\".", "If it can help, PowerShell is an opensource project https://github.com/PowerShell/PowerShell so you can install it (on Mac or Linux) to have a Windows interpreter like mine. But I don't know if you will have access to your tools from it, never tried this can of thing on my Linux.\r\n", "I don't think there is anyone that can help, maybe @mfuntowicz ? I personally only use WSL on Windows on my laptop, so the make commands work without any problem", "OK, so since the problematic on windows code is now hidden in the optional target, we can go ahead and merge this PR - as it no longer interferes with the normal functioning of Makefile for @jplu and other users on the same setup. And meanwhile we will look for a solution for this basic check. How does that sound, @jplu? \r\n\r\nOne possible solution would be to require windows users to use one of the approved set of shells, since as @sgugger says he doesn't have this issue on his windows box. Does it make sense?", "> If it can help, PowerShell is an opensource project https://github.com/PowerShell/PowerShell so you can install it (on Mac or Linux) to have a Windows interpreter like mine. But I don't know if you will have access to your tools from it, never tried this can of thing on my Linux.\r\n\r\nThank you, @jplu. I will give it a quick try. And if it takes too long I will defer to someone who is a windows user to sort this out, but it doesn't need to keep us from merging this fix.", "@jplu, fwiw I installed PowerShell and everything works there just fine:\r\n```\r\n/hf/transformers-fixup-python> pwsh\r\nPS /hf/transformers-fixup-python> make fixup\r\nChecking/fixing src/transformers/testing_utils.py utils/get_modified_files.py\r\nAll done! ✨ 🍰 ✨\r\n2 files left unchanged.\r\npython utils/check_copies.py\r\npython utils/check_dummies.py\r\npython utils/check_repo.py\r\n2020-11-06 21:02:02.519250: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0\r\nChecking all models are properly tested.\r\nChecking all models are properly documented.\r\npython utils/style_doc.py src/transformers docs/source --max_len 119\r\n```\r\n\r\nEither of these works:\r\n```\r\n @if test -n \"$(modified_py_files)\"; then \\\r\n @if [ \"$(modified_py_files)\" != \"\" ]; then \\\r\n```\r\n", "I confirm it is still not working on my side, there should be some differences between both version. I think in your case it is biased because you have access to a shell code interpreter. I checked a bit how to do a `if` in usual PowerShell and apparently it is not the same way https://docs.microsoft.com/en-gb/powershell/scripting/learn/deep-dives/everything-about-if?view=powershell-7 they don't if the `if` and `fi`. So on your side I think that `make` is still using your usual shell and not powershell.\r\n\r\n", "So we need a not shell-specific way.\r\n\r\nI don't know anything about the Windows world - but is the shell you use better/preferable to the one that @sgugger uses, which he says doesn't have this problem?\r\n\r\nUnrelated, could you please confirm that with this PR the rest of the `Makefile` works for you - that is you have no trouble running all targets, except `fixup`?\r\n\r\nAs I mentioned `fixup` is an optional target, as long as `style` and `quality` targets work. The former is much faster, but it's not required for the normal operation. And I trust that eventually we will find a way to sort it out for all setups.", "I'm using as well the one that uses @sgugger which is not a simple shell, but a real Linux distribution installed on Windows (mine for example is Ubuntu 20.04), this is called WSL, and this is mostly what I'm using. The advantage of this is that you can test your code on Windows + Linux in same time. But I don't think we can oblige the Windows users to install this.\r\n\r\nI can confirm that the \"problematic\" part is only the `if` now.\r\n\r\nThis issue is not blocking as you said, I'm not against dealing with it right now that the previous issue (the long egrep command) is solved with the new Python script.", "Thank you for confirming that this PR can be merged, @jplu and following up with the details on where it works.\r\n\r\nWould you kindly open a new issue and describe the specifics of this hurdle we couldn't overcome here, ideally listing all the ways we have tried and failed and most importantly how one could reproduce this situation by having the exact setup you discovered this problem on.\r\n\r\nThen we will seek out someone who knows how to solve this. \r\n\r\nI suppose this could be as simple as posting a tiny Makefile that's unrelated to `transformers` to SO, which demonstrates the issue and surely someone will have the answer.\r\n\r\nThank you!\r\n", "I can take care of it. I will open a PR to try to solve this in a pure Python way. It should not be too difficult.", "While you can do that - this is not the best way if we want to continue using `Makefile`. Otherwise just as well we do away with `Makefile` completely and write everything in python.\r\n\r\nIf we use `Makefile` we need to be able to use it properly, and because we have stumbled upon a tiny problem on some variation of Windows setup and don't have an immediate expertise to solve it is not a strong enough reason to warrant such change.", "Then what about using something like?\r\n```\r\nOS = `uname -a`\r\nifeq ($(OS), Win32)\r\n doing some Windows here\r\nelse\r\n doing some Linux here\r\n```", "Absolutely! I suggested that very early on - please make a version that works for your setup and then we can integrate it with the main version. I thought you didn't know how to make it work, hence I was seeking out variations that may work for you. Your suggestion makes total sense.\r\n\r\nPerhaps let's do it in a separate issue/PR so that we can move this one along in case some windows user is stuck needing this - but if you guys feel that there is no urgency then I will slap the [WIP] back on this PR and then we will go at it until it's perfect. Your call. I just don't want to waste time of reviewers who are probably very confused by now what to do here ;)", "Let's merge this one, and I will rebase on mine to take care of this 👍 ", "Looks to be a know bug in the make version of Windows http://gnu-make.2324884.n4.nabble.com/process-begin-CreateProcess-NULL-quot-quot-failed-td13819.html#a13820\r\n\r\nWe have to recompile Make to make it works, so I think we won't be able to fix this issue with a pure Makefile solution.", "Ok, I did install the last version of Make (4.3) and now it works perfectly. Problem solved 👍 ", "Amazing! Thank you for investigating this and getting to the root of it, @jplu! ", "I don't know whether it's worth documenting in the CONTRIBUTING.md as a possible caveat and giving the solution?", "Good idea! I will do a PR to add a section for Windows.", "Added here https://github.com/huggingface/transformers/pull/8436" ]

[]

[]

[ "Hi! `ids_to_token` is legacy code that should have its visibility set to private. We recommend using the `convert_ids_to_tokens` method, or the `get_vocab()` if you wish to take a look at the whole vocabulary. Both should be available on both the `BertTokenizer` and `BertTokenizerFast`." ]

[ { "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" } ]

[]

[ "Should I add tags like`question answering`and so on? @julien-c ", "If it's intended to be consumed (through the inference API) as a QA model, you can add `pipeline_tag: question-answering`.\r\n\r\ncc @Narsil " ]

[]

[]

[]

[]

[]

[ "Hi! Questions like this are welcome on the forum, we try to keep the github issues for bugs/feature requests only :) Thanks!\r\n\r\nForum link: https://discuss.huggingface.co" ]

[]

[]

[]

[]

[]

[ "Indeed. It has been fixed on master and will be in the next release (early next week). In the meantime you can install from source to have the fix.", "Ah ok, thanks for the quick response 👍 " ]

[]

[]

[ "Merging now to ensure the bug is patched in v3.5.0. Will address your comments in a future PR @Narsil." ]

[ { "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" } ]

[]

[]

[ { "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" } ]

[]

[]

[ { "id": 1862634478, "node_id": "MDU6TGFiZWwxODYyNjM0NDc4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Should%20Fix", "name": "Should Fix", "color": "FF0000", "default": false, "description": "This has been identified as a bug and should be fixed." } ]

[ { "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 } ]

[ "Hey @sebastianGehrmann,\r\n\r\nThanks for the super in-detail error description! You're 100% correct. The PR attached below does an rather ugly fix for now (just cut the `encoder_outputs` to the last hidden states) since the `attentions` and `hidden_states` cannot be output anyways at the moment. \r\n\r\nWith the PT generate refactor: https://github.com/huggingface/transformers/pull/6949 it will now be pretty easy to add functionality to output `attentions` and `hidden_states` for generate in PT - so we should do this soon for PT. \r\n\r\nFor TF we first need to do the same `generate()` refactor and then we can implement this functionality as well. \r\n\r\nSorry that it takes such a long time :-/", "Thank you for the insanely quick fix @patrickvonplaten! No worries about the time, given that no one else has been bumping into this suggests that is is quite a niche requirement. \r\n\r\n", "Would this be resolved at any point? Thanks," ]

[ { "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 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 seems that the tokenizer in that directory is not a Byte level BPE, or at least not in a format the library can understand. A byte-level BPE like the RoBERTa tokenizer should have a merges files as well.\r\n\r\nCould you try to load it in a BERT tokenizer? The BERT tokenizer saves its vocabulary as `vocab.txt` so it is possible this is the tokenizer that was used.", "> It seems that the tokenizer in that directory is not a Byte level BPE, or at least not in a format the library can understand. A byte-level BPE like the RoBERTa tokenizer should have a merges files as well.\r\n> \r\n> Could you try to load it in a BERT tokenizer? The BERT tokenizer saves its vocabulary as `vocab.txt` so it is possible this is the tokenizer that was used.\r\n\r\nThank you! It works." ]

[]

[]

[ "@patil-suraj ", "I am guessing maybe this issue is making it slower on tpus? on en-ro which it does not happen it is faster", "@rabeehkarimimahabadi did you solve the 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." ]