Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Duplicate of #9636" ]

[]

[]

[ "Hi, thank you for opening an issue! Could you put the full stack-trace?\r\n\r\nI guess this comes from the tokenizer, rather than the model as we've already seen this error in [tokenizers](https://github.com/huggingface/tokenizers/issues/537).\r\n\r\nAs a means of debugging, can you let me know what happens if you change this line:\r\n```py\r\ntokenizer = AutoTokenizer.from_pretrained(model_name)\r\n```\r\nto\r\n```py\r\ntokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)\r\n```", "When use_fast is set to false I no longer see the borrow exception but randomly experience this one instead. This appears to be an issue in the qa pipeline code rather than tokenizer code though. I can test the tokenizer in isolation if that'd be helpful.\r\n\r\n```\r\n/usr/local/lib/python3.6/dist-packages/transformers/pipelines/question_answering.py in <listcomp>(.0)\r\n 360 ),\r\n 361 }\r\n--> 362 for s, e, score in zip(starts, ends, scores)\r\n 363 ]\r\n 364 else:\r\n\r\nKeyError: 132\r\n```", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hi! Do you want to open a PR to fix it? Thanks!" ]

[]

[]

[ "Solved by using `RagConfig` and initializing it with a `DPRQuestionEncoder` and the custom `BartForConditionalGeneration` generator configs. Passed the `RagConfig`, question encoder, generator, and retriever, to `RagModel` to initialize the model.", "Could you provide a complete sample code about how to do it? I'm stuck at how to do the config initialization for a DPR + customized BartForCondGen. Thanks!" ]

[]

[]

[]

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

[]

[ "In PyTorch, [`nn.CrossEntropyLoss`](https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html) has an optional `weight` parameter which you can specify. This should be a 1D Tensor assigning a weight to each of the classes.\r\n\r\nSo if you want `BertForTokenClassification` with a weighted cross entropy loss, you can simply replace [this line](https://github.com/huggingface/transformers/blob/c60e0e1ee45f4bf1017736b146c51729f120bb83/src/transformers/models/bert/modeling_bert.py#L1685) by a weighted loss. For example, you can define it as follows (I just copied the relevant code from `modeling_bert.py` and slightly adapted the cross entropy loss):\r\n\r\n```\r\nclass BertForTokenClassification(BertPreTrainedModel):\r\n\r\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\r\n\r\n def __init__(self, config):\r\n super().__init__(config)\r\n self.num_labels = config.num_labels\r\n\r\n self.bert = BertModel(config, add_pooling_layer=False)\r\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\r\n self.classifier = nn.Linear(config.hidden_size, config.num_labels)\r\n\r\n self.init_weights()\r\n\r\n @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format(\"batch_size, sequence_length\"))\r\n @add_code_sample_docstrings(\r\n tokenizer_class=_TOKENIZER_FOR_DOC,\r\n checkpoint=\"bert-base-uncased\",\r\n output_type=TokenClassifierOutput,\r\n config_class=_CONFIG_FOR_DOC,\r\n )\r\n def forward(\r\n self,\r\n input_ids=None,\r\n attention_mask=None,\r\n token_type_ids=None,\r\n position_ids=None,\r\n head_mask=None,\r\n inputs_embeds=None,\r\n labels=None,\r\n output_attentions=None,\r\n output_hidden_states=None,\r\n return_dict=None,\r\n ):\r\n r\"\"\"\r\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\r\n Labels for computing the token classification loss. Indices should be in ``[0, ..., config.num_labels -\r\n 1]``.\r\n \"\"\"\r\n return_dict = return_dict if return_dict is not None else self.config.use_return_dict\r\n\r\n outputs = self.bert(\r\n input_ids,\r\n attention_mask=attention_mask,\r\n token_type_ids=token_type_ids,\r\n position_ids=position_ids,\r\n head_mask=head_mask,\r\n inputs_embeds=inputs_embeds,\r\n output_attentions=output_attentions,\r\n output_hidden_states=output_hidden_states,\r\n return_dict=return_dict,\r\n )\r\n\r\n sequence_output = outputs[0]\r\n\r\n sequence_output = self.dropout(sequence_output)\r\n logits = self.classifier(sequence_output)\r\n\r\n loss = None\r\n if labels is not None:\r\n weights = torch.tensor([0.6, 0.3, 0.1])\r\n loss_fct = CrossEntropyLoss(weights=weights)\r\n # Only keep active parts of the loss\r\n if attention_mask is not None:\r\n active_loss = attention_mask.view(-1) == 1\r\n active_logits = logits.view(-1, self.num_labels)\r\n active_labels = torch.where(\r\n active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)\r\n )\r\n loss = loss_fct(active_logits, active_labels)\r\n else:\r\n loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))\r\n\r\n if not return_dict:\r\n output = (logits,) + outputs[2:]\r\n return ((loss,) + output) if loss is not None else output\r\n\r\n return TokenClassifierOutput(\r\n loss=loss,\r\n logits=logits,\r\n hidden_states=outputs.hidden_states,\r\n attentions=outputs.attentions,\r\n )\r\n```\r\n", "@NielsRogge \r\nYou are right. I had done exactly this in my local (huggingface) transformers codebase.\r\nWorked as expected.\r\n\r\nI think this would be a useful feature if huggingface models come with it.", "Oh ok so you want this to be an added feature. Not sure if this is possible. @LysandreJik what do you think?", "Hi, thanks for opening an issue! The losses in the models are not made to be completely customizable, but to be the most common loss used in most cases; we favor simplicity here.\r\n\r\nThis is because defining your custom loss in a PyTorch model is very simple: when you do not pass the labels to your model, then you retrieve the model logits. You can then define a loss (and customize it as you wish!) and compute its value using these logits and your labels.\r\n\r\nHowever, this is not the first time this feature has been requested, and we could probably come up with an implementation that wouldn't complexify the code-base too much. If we see more of this request we'll take a deeper look at how to implement it.\r\n\r\nHere's a past issue discussing the same/similar: https://github.com/huggingface/transformers/issues/7024\r\n\r\ncc @sgugger @patrickvonplaten", "I'm not sure whether it's a good idea to add such functionality to `modeling_bert.py` - there are too many possibilities. I think it could very well be added to the examples though.", "Yes. Unfortunately, there are too many of these possibilities. \r\nMost users who are familiar with PyTorch can anyway make necessary changes to their local codebase quite easily.\r\nThanks.", "Also see [this example in the documentation](https://huggingface.co/transformers/main_classes/trainer.html) (scroll a tiny bit down to the first example showing a subclass of `Trainer`) on how to change just the loss computation while using a model with `Trainer`.", "Hi everyone,\r\n\r\nI am a student and therefore not yet very familiar with the way issues report work on git, so I aplogize in advance if this is not the proper place to post this message.\r\nI've stumbled onto an error when using the aforementioned method for designing a custom loss function.\r\nMy code is the following\r\n```\r\nconfig = AutoConfig.from_pretrained(\"bert-base-cased\", num_labels=2, finetuning_task=\"SST-2\")\r\n\r\n# Test with modified trainer for weighted CrossEntropyLoss\r\n\r\nmodel = AutoModelForSequenceClassification.from_pretrained(\r\n \"dmis-lab/biobert-base-cased-v1.1\",\r\n from_tf=False,\r\n config=config)\r\n\r\nfrom torch import FloatTensor\r\n\r\nclassDistribution_raw = [97, 3]\r\nclassDistribution = [0.8, 0.2]\r\nnormedWeights = [1 - (x / sum(classDistribution)) for x in classDistribution]\r\nnormedWeights = FloatTensor(normedWeights).cuda()\r\n\r\nfrom torch.nn import CrossEntropyLoss\r\n\r\nclass MyTrainer(Trainer):\r\n def compute_loss(self, model, inputs, return_outputs=False):\r\n \r\n if \"labels\" in inputs:\r\n labels = inputs.pop(\"labels\")\r\n \r\n outputs = model(**inputs)\r\n logits = outputs.logits\r\n loss_function = CrossEntropyLoss(weight = normedWeights)\r\n\r\n if self.args.past_index >= 0:\r\n self._past = outputs[self.args.past_index]\r\n\r\n if labels is not None:\r\n loss = loss_function(logits, labels)\r\n else:\r\n # We don't use .loss here since the model may return tuples instead of ModelOutput.\r\n loss = outputs[\"loss\"] if isinstance(outputs, dict) else outputs[0]\r\n\r\n return (loss, outputs) if return_outputs else loss\r\n\r\ntrainer = MyTrainer(\r\n model=model,\r\n args=training_args,\r\n train_dataset=train_dataset,\r\n eval_dataset=eval_dataset,\r\n compute_metrics=compute_metrics_fn,\r\n tokenizer=tokenizer,\r\n )\r\n\r\n```\r\nAnd when I try to train the model using trainer.train(), i get the following error\r\n'NoneType' object has no attribute 'detach'\r\n\r\nThere is probably something wrong with the way I customized the loss function but I can't find where.\r\n\r\nBest regards,\r\nArthur\r\n", "> Hi, thanks for opening an issue! The losses in the models are not made to be completely customizable, but to be the most common loss used in most cases; we favor simplicity here.\r\n> \r\n> This is because defining your custom loss in a PyTorch model is very simple: when you do not pass the labels to your model, then you retrieve the model logits. You can then define a loss (and customize it as you wish!) and compute its value using these logits and your labels.\r\n> \r\n> However, this is not the first time this feature has been requested, and we could probably come up with an implementation that wouldn't complexify the code-base too much. If we see more of this request we'll take a deeper look at how to implement it.\r\n> \r\n> Here's a past issue discussing the same/similar: #7024\r\n> \r\n> cc @sgugger @patrickvonplaten\r\n\r\n@sgugger @LysandreJik @NielsRogge \r\n\r\nI want to put a +1 on this feature request. Datasets with imbalanced datasets would benefit a lot from custom loss functions. And this shouldn't be a complex add (should just be one more kwarg?)." ]

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[]

[]

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

[]

[ "@sgugger, I think the culprit for the 2nd error, when I uninstalled wandb is:\r\n```\r\ndef is_wandb_available():\r\n if os.getenv(\"WANDB_DISABLED\"):\r\n return False\r\n return importlib.util.find_spec(\"wandb\") is not None\r\n```\r\nas it returns `True`, when it shouldn't since:\r\n```\r\nls -l /home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb\r\nls: cannot access '/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb': No such file or directory\r\n```\r\n\r\nYou can see it with any ddp test, so you don't need to install deepspeed or fairscale to see it, e.g. this fails too:\r\n```\r\npytest -sv examples/seq2seq/test_finetune_trainer.py::TestFinetuneTrainer::test_finetune_trainer_ddp\r\n```\r\nBut a single unforked process test works just fine:\r\n```\r\npytest -sv examples/seq2seq/test_finetune_trainer.py::TestFinetuneTrainer::test_finetune_trainer_dp\r\n```\r\n\r\n-----------------\r\n\r\nand then there is another problem which occurs with `wandb` installed. See the first error in OP.\r\n", "But with `wandb` installed the 1st error I get with DDP too, w/o needing to fork a process in tests:\r\n\r\n```\r\npython -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path sshleifer/tiny-mbart --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size 4 --sortish_sampler --src_lang en_XX --task translation --tgt_lang ro_RO --val_max_target_length 128 --warmup_steps 500 --n_train 500\r\n[...]\r\n[INFO|integrations.py:521] 2021-01-16 20:47:40,853 >> Automatic Weights & Biases logging enabled, to disable set os.environ[\"WANDB_DISABLED\"] = \"true\"\r\nwandb: Currently logged in as: stason (use `wandb login --relogin` to force relogin)\r\n2021-01-16 20:47:42.440849: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0\r\nwandb: Tracking run with wandb version 0.10.14\r\nwandb: Syncing run output_dir\r\nwandb: ⭐️ View project at https://wandb.ai/stason/huggingface\r\nwandb: 🚀 View run at https://wandb.ai/stason/huggingface/runs/82q4zxt2\r\nwandb: Run data is saved locally in /mnt/nvme1/code/huggingface/transformers-master/examples/seq2seq/wandb/run-20210116_204741-82q4zxt2\r\nwandb: Run `wandb offline` to turn off syncing.\r\n 0%| | 0/63 [00:00<?, ?it/s]\r\n[...]\r\nTraining completed. Do not forget to share your model on huggingface.co/models =)\r\n\r\nTraceback (most recent call last):\r\n File \"./finetune_trainer.py\", line 367, in <module>\r\n main()\r\n File \"./finetune_trainer.py\", line 297, in main\r\n train_result = trainer.train(\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/trainer.py\", line 998, in train\r\n self.control = self.callback_handler.on_train_end(self.args, self.state, self.control)\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/trainer_callback.py\", line 342, in on_train_end\r\n return self.call_event(\"on_train_end\", args, state, control)\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/trainer_callback.py\", line 377, in call_event\r\n result = getattr(callback, event)(\r\n File \"/mnt/nvme1/code/huggingface/transformers-master/src/transformers/integrations.py\", line 565, in on_train_end\r\n self._wandb.log({})\r\n File \"/home/stas/anaconda3/envs/main-38/lib/python3.8/site-packages/wandb/sdk/lib/preinit.py\", line 38, in preinit_wrapper\r\n raise wandb.Error(\"You must call wandb.init() before {}()\".format(name))\r\nwandb.errors.error.Error: You must call wandb.init() before wandb.log()\r\n 2021-01-16 20:47:46 | INFO | wandb.sdk.internal.internal | Internal process exited\r\n``` \r\n", "I'm not sure I understand your first error. Could you give us more details? Are you saying that `importlib.from_spec` finds some weird \"wandb\" module but only in a distributed setting? I don't have wandb installed so I can't reproduce this at all.\r\n\r\nFor the last error, pinging @borisdayma ", "I had a similar issue recently with python 3.8 but it worked with 3.7. It was due to a function from \"importlib\" which changed name. Is it the same?", "@borisdayma, I have just installed python-3.7.9 and have the same issue there. Perhaps you had it working with python < 3.7.9?\r\nThe issue occurs with python-3.6.12 too.\r\n\r\n@sgugger yes, the problem occurs only when there is DDP. If I drop `-m torch.distributed.launch` the problem goes away so it has to do with forking/multi-processes. If you remember there was an Issue where someone also had the problem of using some transformers models because they were importing apex at load time and then it was crushing under `torch.mp` - this is definitely a totally different issue, but it's related that it has to do with multiproc.\r\n\r\nTo reproduce:\r\n\r\n```\r\npip install wandb\r\ncd examples/seq2seq\r\npython -m torch.distributed.launch --nproc_per_node=2 ./finetune_trainer.py --model_name_or_path sshleifer/tiny-mbart --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_train --freeze_embeds --label_smoothing 0.1 --learning_rate 3e-5 --logging_first_step --logging_steps 1000 --max_source_length 128 --max_target_length 128 --num_train_epochs 1 --overwrite_output_dir --per_device_train_batch_size 4 --sortish_sampler --src_lang en_XX --task translation --tgt_lang ro_RO --val_max_target_length 128 --warmup_steps 50 --n_train 50\r\n```\r\nwhich results in:\r\n```\r\nwandb.errors.error.Error: You must call wandb.init() before wandb.log()\r\n```\r\n\r\nIf you then remove wand:\r\n```\r\npip uninstall wandb -y\r\n```\r\nThe 2nd error happens:\r\n```\r\nAttributeError: module 'wandb' has no attribute 'ensure_configured'\r\n```\r\n\r\nThe full traces are in the OP.\r\n\r\nPlease let me know if you need any other info.\r\n", "I am running into the same issue with DDP @stas00 has https://github.com/huggingface/transformers/issues/9623#issuecomment-761731532\r\nI believe this might be due to the call to `on_train_end`, which calls `wandb.log({})` on all processes, and not just on world process 0, while [`wandb.init` was called only on world process 0](https://github.com/huggingface/transformers/blob/897a24c869e2ac2ed44f17956f1009fd8f055f5e/src/transformers/integrations.py#L541-L564): https://github.com/huggingface/transformers/blob/897a24c869e2ac2ed44f17956f1009fd8f055f5e/src/transformers/integrations.py#L586", "Interesting, can you check it solves the issue on your side @tristandeleu ?\r\nIf so I'll be happy to make a PR.", "It does work for me when I replace it with\r\n```python\r\nif state.is_world_process_zero:\r\n self._wandb.log({})\r\n```\r\n\r\nThere is also another thing I ran into at the same time: `_log_model` was not initialized on processes other than world 0, making the following check fail because it didn't know `self._log_model`. Adding `self._log_model = False` to `__init__` solved the issue.\r\n\r\nEDIT: This solves the issue with DDP though, I don't know if it also solves the original issue https://github.com/huggingface/transformers/issues/9623#issue-787077821", "Don't hesitate to suggest a PR with your fix @tristandeleu ", "> It does work for me when I replace it with\r\n> \r\n> ```python\r\n> if state.is_world_process_zero:\r\n> self._wandb.log({})\r\n\r\n\r\n> ```\r\n> \r\n> There is also another thing I ran into at the same time: `_log_model` was not initialized on processes other than world 0, making the following check fail because it didn't know `self._log_model`. Adding `self._log_model = False` to `__init__` solved the issue.\r\n> \r\n> EDIT: This solves the issue with DDP though, I don't know if it also solves the original issue [#9623 (comment)](https://github.com/huggingface/transformers/issues/9623#issue-787077821)\r\n\r\nI had the same problem. and I just use > if state.is_world_process_zero: self._wandb.log({}), forget self._log_model = False. Thanks !!!", "> It does work for me when I replace it with\r\n> \r\n> ```python\r\n> if state.is_world_process_zero:\r\n> self._wandb.log({})\r\n> ```\r\n> \r\n> There is also another thing I ran into at the same time: `_log_model` was not initialized on processes other than world 0, making the following check fail because it didn't know `self._log_model`. Adding `self._log_model = False` to `__init__` solved the issue.\r\n> \r\n> EDIT: This solves the issue with DDP though, I don't know if it also solves the original issue [#9623 (comment)](https://github.com/huggingface/transformers/issues/9623#issue-787077821)\r\n\r\nEven with revising these codes, the program(with TPU) doesn't seem to stop at the end", "@lkk12014402 can you confirm it still happens with latest HF master branch?\r\nIf so do you have a reproducible example you could share?", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[]

[]

[]

[]

[]

[]

[]

[ "@sgugger would know about this TODO; I think the fix has landed in `datasets`, right?", "Yes, this should be fixed directly from `datasets` now, will update the script this afternoon." ]

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

[]

[ "Roberta was train on a causal language model objective, therefore the `LineByLineDataset` is not adapted to train it: it considers one line for one text when the roberta objective is to have several lines concatenated and separated by the sep token until it reaches the block size, to avoid padding. \r\n\r\nYou need to use a different dataset for this. You should also check the new [`run_mlm` script](https://github.com/huggingface/transformers/blob/master/examples/language-modeling/run_mlm.py) that offers both options.", "Hi, I had the same issue, here are the workarounds I used\r\nPlatform: Ubuntu 18\r\nPython version: 3.7.9\r\nPyTorch version (GPU): 1.7.1 \r\ncuda11\r\n\r\n- save your tokenizer with `save_model()`, instead of `save()`, will save a `merges.json` and a` vocab.json`. \r\n\r\n` tokenizer.save_model('models/BPEtokenizer')`\r\n\r\n- you'll need `config.json`, a `tokenizer_config.json` and a `special_tokens_map.json` files in your tokenizer repo, you can get them from the base model you want to use your tokenizer with, i.e. just quickly run the `run_mlm` script with 2 batches to get them and add them in your tokenizer repo.\r\n\r\nI'm not sure the config.json is actually loaded, as it is the model config and not the tokenizer's, but the script wants is to accept your tokenizer path.\r\ntokenizer repo should contain:\r\n```\r\n|__config.json\r\n|__merges.txt\r\n|__special_tokens_map.json\r\n|__tokenizer_config.json\r\n|__vocab.json\r\n```\r\n\r\n- in `tokenizer_config.json`, change the ` name_or_path:\"roberta-base\"` to `model_type: \"roberta\"`\r\n\r\n- then train your model running the mlm script with your options and\r\n\r\n`-- tokenizer_name ./models/BPEtokenizer`\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Hitting this issue as well." ]

[]

[]

[ "I agree with Sylvain that while this is not tested, it's hard to recommend using it." ]

[]

[]

[ "Is it normal that this bias is missing?", "By answering your question I realized this could be upstreamed directly in the RoBERTa model, which I just did.\r\n\r\nYou can take a look at my answer this morning to a similar question: https://github.com/huggingface/transformers/issues/6193#issuecomment-760797867.\r\n\r\nXLM-R is an alias of the RoBERTa model, hence why they both need 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 and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Here the logs of a training with\r\n```\r\npython run_tf_glue.py --task_name mrpc --model_name_or_path bert-base-cased --output_dir model --num_train_epochs 4 --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --do_train --do_eval --do_predict --logging_steps 10 --overwrite_output_dir --gradient_accumulation_steps 2\r\n```\r\n\r\n```\r\n[INFO|trainer_tf.py:522] 2021-01-15 10:27:59,116 >> ***** Running training *****\r\n[INFO|trainer_tf.py:523] 2021-01-15 10:27:59,123 >> Num examples = 3668\r\n[INFO|trainer_tf.py:525] 2021-01-15 10:27:59,124 >> Num Epochs = 4\r\n[INFO|trainer_tf.py:526] 2021-01-15 10:27:59,124 >> Instantaneous batch size per device = 16\r\n[INFO|trainer_tf.py:527] 2021-01-15 10:27:59,124 >> Total train batch size (w. parallel, distributed & accumulation) = 32\r\n[INFO|trainer_tf.py:530] 2021-01-15 10:27:59,125 >> Gradient Accumulation steps = 2\r\n[INFO|trainer_tf.py:531] 2021-01-15 10:27:59,136 >> Steps per epoch = 115\r\n[INFO|trainer_tf.py:532] 2021-01-15 10:27:59,137 >> Total optimization steps = 460\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:28:50,371 >> {'loss': 0.6347228, 'learning_rate': 4.891304e-05, 'epoch': 0.08695652173913043, 'step': 10}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:28:56,791 >> {'loss': 0.604829, 'learning_rate': 4.7826084e-05, 'epoch': 0.17391304347826086, 'step': 20}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:03,192 >> {'loss': 0.62615454, 'learning_rate': 4.673913e-05, 'epoch': 0.2608695652173913, 'step': 30}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:09,614 >> {'loss': 0.61436784, 'learning_rate': 4.5652174e-05, 'epoch': 0.34782608695652173, 'step': 40}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:16,163 >> {'loss': 0.60542804, 'learning_rate': 4.456522e-05, 'epoch': 0.43478260869565216, 'step': 50}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:22,633 >> {'loss': 0.60221016, 'learning_rate': 4.347826e-05, 'epoch': 0.5217391304347826, 'step': 60}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:29,129 >> {'loss': 0.59315145, 'learning_rate': 4.2391304e-05, 'epoch': 0.6086956521739131, 'step': 70}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:35,655 >> {'loss': 0.5896678, 'learning_rate': 4.1304345e-05, 'epoch': 0.6956521739130435, 'step': 80}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:42,209 >> {'loss': 0.5796127, 'learning_rate': 4.0217386e-05, 'epoch': 0.782608695652174, 'step': 90}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:48,779 >> {'loss': 0.5678522, 'learning_rate': 3.9130435e-05, 'epoch': 0.8695652173913043, 'step': 100}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:29:55,365 >> {'loss': 0.55807614, 'learning_rate': 3.8043476e-05, 'epoch': 0.9565217391304348, 'step': 110}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:04,348 >> {'loss': 0.32373077, 'learning_rate': 3.695652e-05, 'epoch': 1.0434782608695652, 'step': 120}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:10,920 >> {'loss': 0.3261666, 'learning_rate': 3.5869565e-05, 'epoch': 1.1304347826086956, 'step': 130}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:17,516 >> {'loss': 0.34052417, 'learning_rate': 3.478261e-05, 'epoch': 1.2173913043478262, 'step': 140}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:24,125 >> {'loss': 0.35018474, 'learning_rate': 3.369565e-05, 'epoch': 1.3043478260869565, 'step': 150}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:30,714 >> {'loss': 0.35887596, 'learning_rate': 3.260869e-05, 'epoch': 1.391304347826087, 'step': 160}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:37,303 >> {'loss': 0.34891757, 'learning_rate': 3.1521737e-05, 'epoch': 1.4782608695652173, 'step': 170}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:43,900 >> {'loss': 0.33256933, 'learning_rate': 3.0434781e-05, 'epoch': 1.5652173913043477, 'step': 180}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:50,481 >> {'loss': 0.32668048, 'learning_rate': 2.934782e-05, 'epoch': 1.6521739130434783, 'step': 190}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:30:57,079 >> {'loss': 0.31888676, 'learning_rate': 2.8260865e-05, 'epoch': 1.7391304347826086, 'step': 200}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:03,688 >> {'loss': 0.31276095, 'learning_rate': 2.7173912e-05, 'epoch': 1.8260869565217392, 'step': 210}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:10,284 >> {'loss': 0.30366346, 'learning_rate': 2.6086956e-05, 'epoch': 1.9130434782608696, 'step': 220}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:16,885 >> {'loss': 0.2903903, 'learning_rate': 2.5e-05, 'epoch': 2.0, 'step': 230}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:26,095 >> {'loss': 0.15675393, 'learning_rate': 2.3913042e-05, 'epoch': 2.0869565217391304, 'step': 240}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:32,671 >> {'loss': 0.14483282, 'learning_rate': 2.2826087e-05, 'epoch': 2.1739130434782608, 'step': 250}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:39,275 >> {'loss': 0.14147088, 'learning_rate': 2.173913e-05, 'epoch': 2.260869565217391, 'step': 260}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:45,866 >> {'loss': 0.13758971, 'learning_rate': 2.0652174e-05, 'epoch': 2.3478260869565215, 'step': 270}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:52,464 >> {'loss': 0.13357341, 'learning_rate': 1.9565217e-05, 'epoch': 2.4347826086956523, 'step': 280}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:31:59,049 >> {'loss': 0.12877393, 'learning_rate': 1.8478258e-05, 'epoch': 2.5217391304347827, 'step': 290}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:05,682 >> {'loss': 0.13753517, 'learning_rate': 1.7391301e-05, 'epoch': 2.608695652173913, 'step': 300}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:12,281 >> {'loss': 0.1319594, 'learning_rate': 1.6304344e-05, 'epoch': 2.6956521739130435, 'step': 310}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:18,883 >> {'loss': 0.12644322, 'learning_rate': 1.5217389e-05, 'epoch': 2.782608695652174, 'step': 320}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:25,472 >> {'loss': 0.12481367, 'learning_rate': 1.41304345e-05, 'epoch': 2.869565217391304, 'step': 330}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:32,082 >> {'loss': 0.12073966, 'learning_rate': 1.3043478e-05, 'epoch': 2.9565217391304346, 'step': 340}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:41,403 >> {'loss': 0.10288413, 'learning_rate': 1.1956521e-05, 'epoch': 3.0434782608695654, 'step': 350}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:47,955 >> {'loss': 0.09858045, 'learning_rate': 1.0869565e-05, 'epoch': 3.130434782608696, 'step': 360}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:32:54,532 >> {'loss': 0.07963112, 'learning_rate': 9.782609e-06, 'epoch': 3.217391304347826, 'step': 370}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:01,104 >> {'loss': 0.08428383, 'learning_rate': 8.6956525e-06, 'epoch': 3.3043478260869565, 'step': 380}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:07,684 >> {'loss': 0.0844244, 'learning_rate': 7.6086967e-06, 'epoch': 3.391304347826087, 'step': 390}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:14,284 >> {'loss': 0.08690852, 'learning_rate': 6.5217405e-06, 'epoch': 3.4782608695652173, 'step': 400}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:20,877 >> {'loss': 0.0832295, 'learning_rate': 5.434781e-06, 'epoch': 3.5652173913043477, 'step': 410}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:27,494 >> {'loss': 0.078029804, 'learning_rate': 4.3478244e-06, 'epoch': 3.6521739130434785, 'step': 420}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:34,095 >> {'loss': 0.079320244, 'learning_rate': 3.2608687e-06, 'epoch': 3.7391304347826084, 'step': 430}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:40,709 >> {'loss': 0.076877564, 'learning_rate': 2.1739122e-06, 'epoch': 3.8260869565217392, 'step': 440}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:47,324 >> {'loss': 0.07551385, 'learning_rate': 1.0869561e-06, 'epoch': 3.9130434782608696, 'step': 450}\r\n[INFO|trainer_tf.py:398] 2021-01-15 10:33:53,941 >> {'loss': 0.07157838, 'learning_rate': 0.0, 'epoch': 4.0, 'step': 460}\r\n```\r\n\r\nNothing seems wrong in the loss computation.\r\n\r\n```\r\neval_acc = 0.8518518518518519\r\neval_f1 = 0.8954248366013072\r\neval_acc_and_f1 = 0.8736383442265796\r\n```", "@jplu \r\nYes, you are right, and I am wrong. \r\nMy dataset format was wrong (```labels``` in dataset for ```TFGPT2LMHead``` should be ```tensor```, but was ```dict``` yesterday). Sorry for the confusion.\r\n\r\nHowever, there is one problem though. ```training_loss``` is not properly calculated with successive training. Run ```run_tf_glue.py``` with ```save_steps```. I train 40 steps, and train again from that ckpt. Results are shown below, where loss increases and decreases.\r\n\r\n```\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:19:09,011 >> {'loss': 0.10616198, 'learning_rate': 4.456522e-05, 'epoch': 0.43478260869565216, 'step': 50}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:19:20,271 >> {'loss': 0.17419913, 'learning_rate': 4.347826e-05, 'epoch': 0.5217391304347826, 'step': 60}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:19:33,935 >> {'loss': 0.2174806, 'learning_rate': 4.2391304e-05, 'epoch': 0.6086956521739131, 'step': 70}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:19:47,342 >> {'loss': 0.25698015, 'learning_rate': 4.1304345e-05, 'epoch': 0.6956521739130435, 'step': 80}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:20:04,633 >> {'loss': 0.28348362, 'learning_rate': 4.0217386e-05, 'epoch': 0.782608695652174, 'step': 90}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:20:22,000 >> {'loss': 0.29971126, 'learning_rate': 3.9130435e-05, 'epoch': 0.8695652173913043, 'step': 100}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:20:35,515 >> {'loss': 0.30674613, 'learning_rate': 3.8043476e-05, 'epoch': 0.9565217391304348, 'step': 110}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:20:57,368 >> {'loss': 0.3384542, 'learning_rate': 3.695652e-05, 'epoch': 1.0434782608695652, 'step': 120}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:21:10,066 >> {'loss': 0.25661522, 'learning_rate': 3.5869565e-05, 'epoch': 1.1304347826086956, 'step': 130}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:21:23,626 >> {'loss': 0.2714903, 'learning_rate': 3.478261e-05, 'epoch': 1.2173913043478262, 'step': 140}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:21:42,122 >> {'loss': 0.262019, 'learning_rate': 3.369565e-05, 'epoch': 1.3043478260869565, 'step': 150}\r\n[INFO|trainer_tf.py:398] 2021-01-15 16:21:55,652 >> {'loss': 0.27134755, 'learning_rate': 3.260869e-05, 'epoch': 1.391304347826087, 'step': 160}\r\n```\r\n\r\nThis comes from ```steps_trained_in_current_epoch``` and ```training_loss = self.train_loss.result() / (step + 1)``` in ```TFTrainer```.\r\nFor 41 step (1 step after running from cpkt-40 in this case), only one loss is accumulated but step is 40.\r\n\r\nI simply revise it by initially saving ```steps_trained_in_current_epoch``` to another constant. \r\nThis may be treated in different PR.", "Humm looks to be an issue indeed. You can keep this PR open to fix this, feel free to ask questions if I can help :)" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1897896961, "node_id": "MDU6TGFiZWwxODk3ODk2OTYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Migration", "name": "Migration", "color": "e99695", "default": false, "description": "" } ]

[]

[ "'BertSelfAttention+BertSelfOutput' is tensorflow style\r\n'torch.nn.MultiheadAttention' is real pytorch style", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "The layers from the Pytorch will be significantly faster than using the two classes like in TensorFlow. Can't we make an exception for pytorch to use the optimized layers?" ]

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

[]

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

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[]

[]

[]

[]

[ "Hi, thank you for opening a PR!\r\n\r\nOur goal is to stay as close to the initial implementation as possible. The original implementation by OpenAI uses -1e4, so we will keep it this way.", "I find the initial implementation is `-1e10` in https://github.com/openai/gpt-2/blob/master/src/model.py#L88\r\n\r\n```py\r\nw = w*b - tf.cast(1e10, w.dtype)*(1-b)\r\n```\r\nrelated issue #9594\r\n\r\nI am not quite sure, I guess `1e-10` is not compatible with `fp16`, that may be the reason behind huggingface implementation.\r\n" ]

[]

[]

[ "Hello!\r\n\r\nYou have to use the `save_model` method of the trainer.", "ok. thanks!" ]

[]

[]

[ "Hi,\r\n\r\nI would like to report the same problem. I see this problem only with RoBERTa base or large and I am also using transformers4.2.2.\r\n\r\nAny suggestions or help would be appreciated. \r\nThanks.", "Hi, \r\nI had the same issue. I solved it by adding add_prefix_space=True to the tokenizer.\r\n\r\nBest", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hi,\r\n\r\nI am have the same issue. \r\nI am loading from json - \r\n\r\n`python $SCRATCH/transformers/examples/token-classification/run_ner.py \\\r\n --model_name_or_path roberta-base \\\r\n --train_file dict_structure/trivia_training.json \\\r\n --validation_file dict_structure/trivia_val.json \\\r\n --output_dir roberta_base_on_MITMovieNER/ \\\r\n --do_train \\\r\n --do_eval \\\r\n --per_device_train_batch_size 64 \\\r\n --per_device_eval_batch_size 20 \\\r\n --num_train_epochs 40 \\\r\n --overwrite_output_dir \\\r\n --evaluation_strategy steps \\\r\n --save_steps 1000 \\\r\n --eval_steps 500 \\\r\n --logging_first_step \\`\r\n\r\nSorry, not sure if this is an issue on my end. @stefan-it ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "This remains an issue using the official example and official task; it would be great to see this addressed." ]

[ { "id": 2648621985, "node_id": "MDU6TGFiZWwyNjQ4NjIxOTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Feature%20request", "name": "Feature request", "color": "FBCA04", "default": false, "description": "Request for a new feature" }, { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[]

[ "hi，we noticed Deepspeed transformer kernel is much faster than the original PyTorch version with less memory consumption. I would like to know if you have any future plan to integrate Deepspeed transformer kernel in huggingface.\r\nThanks!", "Personally my focus at the moment is to enable fitting big models on small hardware, because if we can do such training slowly it's better than not being able to do so.\r\n\r\nNext come the speed optimizations.\r\n\r\nI added `Deepspeed transformer kernel` to the list above. Thank you for the recommendation.\r\n\r\nBut if you'd like to do some experimentation and get some good results and submit a PR that would be fantastic. It doesn't have to be perfect, just good enough that it can be seen the speed up improvement the docs are alluding to.", "> Personally my focus at the moment is to enable fitting big models on small hardware, because if we can do such training slowly it's better than not being able to do so.\r\n> \r\n> Next come the speed optimizations.\r\n> \r\n> I added `Deepspeed transformer kernel` to the list above. Thank you for the recommendation.\r\n> \r\n> But if you'd like to do some experimentation and get some good results and submit a PR that would be fantastic. It doesn't have to be perfect, just good enough that it can be seen the speed up improvement the docs are alluding to.\r\n\r\nHi, I did a simple test with the bert-large model，The following are the test results\r\n\r\n\r\n", "Thank you for sharing the benchmarks, @gongjingcs\r\n\r\nThat's a nice speed up. \r\n\r\nI assume you also tested deepspeed w/o \"Deepspeed transformer kernel\" as a baseline, to know that it's that feature that gave the speed up and not DeepSpeed's other features.\r\n\r\nI encourage you to try to make a PR to integrate this aspect of Deepspeed if you are inspired to do so.", "Hi @stas00, \r\n\r\nThank you for sharing those awesome topics. Are the features still requested/up-to-date ? I would like to follow the point made by @gongjingcs about the Deepspeed Transformer Kernel. ", "Hi Simon,\r\n\r\nre: up-to-date I'm sure Deepspeed came up with new advancements since this was last updated, if that's what you asking about. And the list in the OP is still outstanding.\r\n\r\nSo wrt Deepspeed Transformer Kernel. How would you envision us integrating it - i.e. which components of HF transformers do you want? HF models have a lot of features inside the transformer layers, so swapping in a different Transformer block won't work easily. pytorch too has a Transformer block in its arsenal.\r\n\r\nIn other words I'm seeing to understand how you see those replacements to be used?\r\n\r\nAdditionally are you after inference or training? For inference we will soon have fast fused kernels via:\r\nhttps://github.com/huggingface/transformers/pull/14426 and @hyunwoongko has just announced https://github.com/tunib-ai/oslo https://github.com/huggingface/transformers/issues/13690#issuecomment-998492192 which does kernel fusion, though we haven't done any benchmarking yet, but check it out.\r\n\r\nThank you!", "Thank you for your answer @stas00 \r\n\r\n> re: up-to-date I'm sure Deepspeed came up with new advancements since this was last updated, if that's what you asking about. And the list in the OP is still outstanding.\r\n\r\nI was looking at the features you provided in the list and wondered if they were still requested or if anyone was already working on it.\r\n\r\n> So wrt Deepspeed Transformer Kernel. How would you envision us integrating it - i.e. which components of HF transformers do you want? HF models have a lot of features inside the transformer layers, so swapping in a different Transformer block won't work easily. pytorch too has a Transformer block in its arsenal.\r\n> \r\n> In other words I'm seeing to understand how you see those replacements to be used?\r\n\r\nI just finished to benchmark the Transformer Kernel with the models provide in the DeepSpeedExamples repo. So I don't have a clear plan on how to do this. I was wondering if we could first do an in-place operation to swap out the Transformer layer in the Trainer s.t we can keep the HF components code unchanged while taking advantage of the throughput speed-up and the batch size improvement provided. But I don't know if it will impact other features.\r\n\r\n> Additionally are you after inference or training? For inference we will soon have fast fused kernels via:\r\n> #14426 and @hyunwoongko has just announced https://github.com/tunib-ai/oslo #13690 (comment) which does kernel fusion, though we haven't done any benchmarking yet, but check it out.\r\n\r\nI have been focusing on training: pre-training and fine-tuning. I haven't look at the deepspeed pre-training yet. OSLO seems really nice, do you think it's still worth looking at the deepspeed Transformer Kernel ?\r\n\r\nThank you ", "The problem is that the weight names will be different and any custom features that HF Transformers model expects will not be provided by an external implementation. You can try to import the \"normal\" model and then monkeypatching the transformers layer to the deepspeed version and see if you get anywhere with it. \r\n\r\nAnd which architecture are you trying to speed up?\r\n\r\nI'm yet to try OSLO myself, so can't give any first hand experience, but since it suggests that it can fuse the model, perhaps it can do much better already than the plain pytorch version. I'd make a request at https://github.com/tunib-ai/oslo to support the arch you want and compare the performance. That would probably be the low hanging fruit.\r\n\r\nThen you can also try to compile the model into ONNX as described here https://huggingface.co/docs/transformers/serialization and use one of the optimized runtimes. But I don't yet have an experience with that tech yet, hoping to fill the gap in the new year.\r\n", "OSLO only fuses certain parts, just like Megatron-LM. (scale+mask+softmax, bias+gelu, bias+dropout) Therefore, it is slower than the fully fusable kernels like DeepSpeed. I also reviewed DeepSpeed's transformer kernel (not the inference kernel), but I gave up because it is a structure that is difficult to apply to various architectures and cannot do tensor model parallelization.", "On the other hand, DeepSpeed inference is a much more scalable structure. It can also perform tensor model parallelization. However, no backward kernel is provided. It would be nice if @RezaYazdaniAminabadi could provide a backward kernels. (If the backward kernels are available, I will also add them to OSLO)", "Note that there are also lightseq kernels by bytedance which improve DeepSpeed transformer kernels.\r\nhttps://github.com/bytedance/lightseq The speed of the kernels is similar, but various kernels have been added (embedding, cross-entropy, etc...) and It provides a little more flexible Pybind API.", "Hi, @stas00 , could you please confirm that [DeepSpeed Activation Checkpointing] is working properly?\r\nI was seeing some issues with activation partitioning feature (I need it to reduce activation memory usage)\r\nAlso, where are the code changes located for this feature?\r\nThanks!", "we currently don't use Deepspeed's Activation Checkpointing as it'd be very difficult to integrate into transformers (it'd require massively changing all models). The normal pytorch activation available in most models works just fine. To activate it use this API:\r\nhttps://huggingface.co/docs/transformers/main_classes/model#transformers.PreTrainedModel.gradient_checkpointing_enable\r\n\r\nDeepspeed's Activation Checkpointing however has additional features that pytorch implementation lacks.\r\n\r\n" ]

[]

[]

[ "Could we discuss the naming of this script and others?\r\n\r\nThe description goes:\r\n\r\n> Fine-tuning the library models for sequence to sequence.\r\n\r\n`run_seq2seq.py` is much less descriptive or intuitive than `finetune_trainer.py `- why not go back to `finetune.py` to replace the script that was PL-based and moved to experiments? \r\n\r\n1. If we are cleaning up the naming, just as well we could drop any `run_` prefices that we now have in many `examples/*/run_*.py` - they are all scripts, they all get to **run**. The names are great when they are focused on their purpose and not how they are executed. \r\n\r\n2. This script is already inside `seq2seq` - So in `examples/seq2seq/run_seq2seq.py` - how does it help to repeat it twice? I can see where you'd want to uniquely identify each script if they are taken out of context of their `examples/*/` subdir - perhaps this is the intention? perhaps if you open them all in the editor and end up with 10 `finetune.py`? If that's the case, then I can see your point of repeating the \"domain\" in the name of the script.\r\n\r\nIf the 2nd item is trying to solve the uniqueness issue, then repetition works just fine, but I strongly recommend replacing `run_` with `finetune_` to at the very least have some mnemonics about what it does.\r\n\r\n", "Also, since you will need to update README.md to show users how to run the new script - could we have some of it in this PR? Even just the basic command lines - that would help testing this PR and not needing to figure out the new args? \r\n\r\nIf possible that is?\r\n\r\nThank you!", "The scripts are all named `run_xxx` precisely for reason 2, the same way we didn't rename `modeling_xxx` files to just `modeling.py` when restructuring the repo. I have no strong objection to changing `run` to `finetune` but it will break lots of links in the documentation and may confuse users, so not sure if it's worth it. I'll let @LysandreJik and @patrickvonplaten chime in on that subject.\r\n\r\nI'll add command examples in the README (this PR is not quite ready to be merged yet, there is also the small test to add), first I wanted to grab comments on the actual script before finishing :-) I don't expect it to work fully (though if it's magically the case I'll be happy :-) ) which is why this PR does not delete the old script, so we can make some tests and make sure there is no regression, then progressively fix this new script as needed.", "> The scripts are all named run_xxx precisely for reason 2 [...]\r\n\r\nI understand. Thank you for clarifying that part. Easy editing is a strong pro for sure.\r\n\r\nThinking more about it perhaps `finetune` isn't the right name either because it does finetuning plus prediction, so perhaps `run` actually is somewhat of a better choice, as it's less committing to anything ;)\r\n\r\nI think what I'm experiencing here is the pain of pattern breaking. First I was using \"finetune.py\", then I switched to \"finetune_trainer.py\" and now \"run_seq2seq.py\" - say, what? :)\r\n\r\n> I'll add command examples in the README [...]\r\n\r\nI'd like to contribute with the review, but I need context to do such things and there is neither diff nor a way to run it, I'm just not sure how to approach such type of review. So perhaps I will be able to do that at a later stage when I can test the new script, or if you'd like me to look at a particular part of it I'm game too.", "I am merging as a first step. @stas00 I know it's missing examples of use and that there still is the memory regression, I plan to address those in follow-up PRs (also anyone should feel free to suggest improvements to the new scripts).", "It's a good plan, @sgugger! I know you won't forget these. Thank you for considering my concerns.\r\n\r\nThe only thing I am not sure about is that nobody commented on the new script's naming." ]

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

[]

[ "Indeed! Do you want to open a PR with a doc fix?", "I would like to, but I couldn't find where that exact doc in the repo is", "Also, a built-in tool for pointing out mistakes in the docs would be usable (like those when you highlight an error and press Ctrl+Enter). I notice a few mistakes and typos from time to time.\r\nI am speaking not about the docs themselves, but about the guides and tutorials which are more community-oriented", "> Indeed! Do you want to open a PR with a doc fix?\r\n\r\n@LysandreJik, could you please point out the doc I could fix?", "Here it is: https://github.com/huggingface/transformers/blob/master/docs/source/task_summary.rst", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

[]

[ "You are using a pytorch lightning callback instead of a Hugging Face `TrainerCallback`, I'm unsure of why you would think this will work. If you want to use pytorch ligthning, you will have to user their `Trainer` as well.", "thanks @sgugger . The reason I used pytorch lightning callback is because I couldn't find in transformers something that saves only the best checkpoint. Is there something like that? (which I can use instead of using pytorch-lightings- ModelCheckpoint)", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread.", "> You are using a pytorch lightning callback instead of a Hugging Face `TrainerCallback`, I'm unsure of why you would think this will work. If you want to use pytorch ligthning, you will have to user their `Trainer` as well.\r\n\r\nSuper Helpful" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Please avoid spamming the repository with multiple duplicate issues.\r\nAlso, those questions should go in the [forums](https://discuss.huggingface.co/), the issues are kept for bugs and feature requests only.", "sorry...I created two streams by mistake.\r\n" ]

[]

[]

[ "Please avoid spamming the repository with multiple duplicate issues.\r\nAlso, those questions should go in the [forums](https://discuss.huggingface.co/), the issues are kept for bugs and feature requests only." ]

[]

[]

[ "_Note:_ that'd be a better question for the forums at discuss.huggingface.co\r\n\r\nThe `optimizer.pt` is a snapshot of the optimizer's internal state, for inference you can delete it and only keep your `model.bin` (= the weights)", "Thank you, that is very helpful and for directing me to the forums, I did not even know it existed :)\r\n\r\nAs regards to my other question, is it possible to do the fine tuning in tensorflow? or even export to a tensorflow model? \r\n\r\n or would this discussion be better suited to the forum?\r\n\r\nThanks" ]

[]

[]

[ "CircleCI error messages says as below.\r\n\r\nIn `run_tests_torch`:\r\n``` \r\n-- Docs: https://docs.pytest.org/en/stable/warnings.html\r\n=========================== short test summary info ============================\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_sample_generate\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_search_generate\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_search_generate_dict_outputs_use_cache\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_gpt2_gradient_checkpointing\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_group_beam_search_generate\r\n==== 5 failed, 4202 passed, 1775 skipped, 744 warnings in 216.47s (0:03:36) ====\r\n\r\nExited with code exit status 1\r\nCircleCI received exit code 1\r\n```\r\n\r\nIn `run_tests_flax`:\r\n```\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_sample_generate\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_search_generate\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_beam_search_generate_dict_outputs_use_cache\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_gpt2_gradient_checkpointing\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_group_beam_search_generate\r\n==== 5 failed, 4172 passed, 1805 skipped, 751 warnings in 282.27s (0:04:42) ====\r\n\r\nExited with code exit status 1\r\nCircleCI received exit code 1\r\n```\r\n", "Is there a difference between `past_key_value` and `layer_past`?　I understand that they both represent the contents of `past_key_values`, the past of each layer, but are they different?\r\n\r\nI first thought it might be a difference between the Causal language model and the Seq2Seq language model, but it seems that both `past_key_value` and `layer_past` are used in `modeling_bart.py`.\r\n\r\nAnd as for the contents of `layer_past`, should it be named `past_state`, as the following part of `modeling_bart.py` shows?\r\n\r\nhttps://github.com/huggingface/transformers/blob/236cc365aff2512ef773c6b1786555dab6fb182f/src/transformers/models/bart/modeling_bart.py#L1236-L1244", "I've updated `generation_utils.py`, and it seems `mems` in transfo_xl and xlnet causes a new error.\r\n\r\n```\r\n=========================== short test summary info ============================\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_gpt2_gradient_checkpointing\r\nFAILED tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_beam_sample_generate\r\nFAILED tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_beam_sample_generate_dict_output\r\nFAILED tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_beam_search_generate\r\nFAILED tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_beam_search_generate_dict_output\r\nFAILED tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_group_beam_search_generate\r\nFAILED tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_group_beam_search_generate_dict_output\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_beam_sample_generate\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_beam_sample_generate_dict_output\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_beam_search_generate\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_beam_search_generate_dict_output\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_group_beam_search_generate\r\nFAILED tests/test_modeling_xlnet.py::XLNetModelTest::test_group_beam_search_generate_dict_output\r\n=== 13 failed, 4194 passed, 1775 skipped, 743 warnings in 205.38s (0:03:25) ====\r\n\r\nExited with code exit status 1\r\nCircleCI received exit code 1\r\n```\r\n\r\nhttps://github.com/huggingface/transformers/blob/236cc365aff2512ef773c6b1786555dab6fb182f/src/transformers/models/xlnet/modeling_xlnet.py#L581-L607\r\n\r\nIt seems `mems` is something similar to `past_key_values`. \r\nIs there any difference between these two elements with different names?\r\nAlso, is it safe to change `mems` from `List[torch.Tensor]` to `Tuple[Tuple[torch.Tensor]]`?", "Hey @forest1988, \r\n\r\nYou're PR looks very nice! Yes, it is expected that `XLNet` and `TransfoXL` fail actually since they also have been using the \"default\" `_reorder_cache` function of `modeling_utils.py`. Could you do the following changes to correct this:\r\n\r\n1) Copy that old `_reorder_cache` (the one before you did your changes) function that was in `generation_utils.py` to both `modeling_xlnet.py` and `modeling_transfo_xl.py` file so that those have the same function as before? \r\n2) Copy the current `_reorder_cache` function of `generation_utils.py` into `modeling_gpt2.py`?\r\n3) Add a default `_reorder_cache` function to `generation_utils.py` that looks as follows:\r\n\r\n```python\r\ndef _reorder_cache(self, past, beam_idx):\r\n raise NotImplementedError(...)\r\n```", "I've just updated `torch.utils.checkpoint.checkpoint` check in `modeling_gpt2.py`, referring to `modeling_bart.py`.", "This way it's much cleaner and correct :-) The reason I'm proposing this change is that the `_reorder_cache` function is so different for each model that there should be **no** default function. A default function could confuse people that want to add a new model in a way that they think it works out of the box, but in most cases it just doesn't. A clear error message such as:\r\n\r\n\r\n```python\r\ndef _reorder_cache(self, past, beam_idx):\r\n raise NotImplementedError(f\"Make sure that a `_reorder_cache` function is correctly implemented in {self.__class__.__module__} to enable beam search for {self.__class__}\")\r\n```\r\n```\r\n", "I think this should solve the problems, let me know if you need more help :-) ", "Thank you for your advice! I'll update `_reorder_cache` soon and commit it.", "Hi @patrickvonplaten,\r\n\r\nThanks to your kind advice, I could solve the problem of `_reorder_cache` in `GPT-2`, `XLNet`, `TransfoXL` (, and `CTRL`).\r\nReferring to `modeling_bart.py`, in which `_reorder_cache` is placed in `ConditionalGeneration` Model, I added `_reoder_cache` in `LMHead` Models in each Causal Language Models.\r\n\r\nThe last one remaining bug is:\r\n```\r\nFAILED tests/test_modeling_gpt2.py::GPT2ModelTest::test_gpt2_gradient_checkpointing\r\n```\r\n\r\nI think I should modify `test_gpt2_gradient_checkpointing` so that it has `use_cache=False`, or reconsider my previous update and re-modify the usage of `checkpoint` in modeling_gpt2.\r\n> I've just updated torch.utils.checkpoint.checkpoint check in modeling_gpt2.py, referring to modeling_bart.py.\r\n>\r\n", "All checks have passed!\r\nI appreciate all your help.\r\n\r\nHowever, in the documentation of `_reorder_cache`, there are references to both `past_key_values` and `mems` regardless of which object is used.\r\nI think we can fix that and only mention the one we use, or we can leave the reference to both to show that the aim of the function is the same.\r\nIf there is a need to modify it, please let me know.\r\n", "Hi @patrickvonplaten,\r\n\r\n> I hope it's fine for you that I went into the PR to do some final fixes. Thanks a lot for cleaning this up :-)\r\n\r\nOf course! Thank you for adding fixes to make this PR more valuable!", "Awesome, merging - great job @forest1988 !", "Thank you for your advice and encouraging comments!\r\nIt’s my pleasure to have opened this PR!" ]

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

[]

[ "Hi! Order doesn’t matter in a dictionary. \r\n\r\nIt only matters if you use the arguments as positional arguments, which is not recommended.", "@LysandreJik \r\n\r\nSo, order does matter when using lists? What is now the right order?\r\n", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

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

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

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

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

[]

[ "For the `WordPiece` decoder, which is used in `BertTokenizerFast`, It seems that `cleanup` cannot be changed after initialization.\r\nhttps://github.com/huggingface/tokenizers/blob/python-v0.10.0/tokenizers/src/tokenizer/mod.rs#L762\r\nhttps://github.com/huggingface/tokenizers/blob/python-v0.10.0/tokenizers/src/decoders/wordpiece.rs#L35\r\nhttps://github.com/huggingface/tokenizers/blob/python-v0.10.0/bindings/python/py_src/tokenizers/decoders/__init__.pyi#L113\r\nhttps://github.com/huggingface/transformers/blob/v4.2.0/src/transformers/convert_slow_tokenizer.py#L106\r\n\r\nI confirmed that a tokenizer and the fast tokenizer return the same string when they are based on SentencePiece because it treats whitespace as a symbol and can reconstruct the original sentence.\r\nSo when specifying `clean_up_tokenization_spaces=False`, spaces before punctuation depend on `ids`, but there are no differences in the decoded string between a tokenizer (e.g. `T5Tokenizer`) and the fast tokenizer (e.g. `T5TokenizerFast`).", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "You can change the logging level:\r\n\r\n```py\r\nfrom transformers import logging as hf_logging\r\n\r\nhf_logging.set_verbosity_error()\r\n```" ]

[]

[]

[ "You can change the logging level:\r\n\r\n```\r\nfrom transformers import logging as hf_logging\r\n\r\nhf_logging.set_verbosity_error()\r\n```" ]

[]

[ { "login": "jplu", "id": 959590, "node_id": "MDQ6VXNlcjk1OTU5MA==", "avatar_url": "https://avatars.githubusercontent.com/u/959590?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jplu", "html_url": "https://github.com/jplu", "followers_url": "https://api.github.com/users/jplu/followers", "following_url": "https://api.github.com/users/jplu/following{/other_user}", "gists_url": "https://api.github.com/users/jplu/gists{/gist_id}", "starred_url": "https://api.github.com/users/jplu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jplu/subscriptions", "organizations_url": "https://api.github.com/users/jplu/orgs", "repos_url": "https://api.github.com/users/jplu/repos", "events_url": "https://api.github.com/users/jplu/events{/privacy}", "received_events_url": "https://api.github.com/users/jplu/received_events", "type": "User", "site_admin": false } ]

[ "Hello !\r\n\r\nYou can safely ignore those warnings, no worries.", "@jplu Thanks. Might the results of training a model may different to new version (like because of the new kind of tokenizer)?\r\n\r\nBut \"why\" do I get the same output as in v3, since in the doc is stated that the ouput structure somehow chnaged and you cannot do unpacking like \r\n\r\n`a, b, c = model(input)\r\n`\r\n\r\nBut which stills works.\r\n\r\nHow can I ignore (disable) this messages? I tried a lot, but nothing worked!\r\n\r\n", "In graph mode you cannot get tuples anymore, the dict output is forced and you cannot disable this message for now. This will be possible in a future release, as it will be displayed only when you set yourself the `output_attentions`, `output_hidden_states` or `return_dict` yourself in the method call while running your model in graph mode.", "@jplu Thanky. What do you mean by graph mode? As stated above, I still get the tuples as output?", "You are not getting tuples, by doing:\r\n```\r\na, b, c = model(input)\r\n```\r\nYou are getting the keys of the dict.\r\n\r\nBy graph mode, I mean TensorFlow graph mode, and not eager mode.", "@jplu Ah ok, that is why I still get the tuple, because by default in tf2 eager mode is activated, right?", "Yes eager mode is activated by default, and no you don't get tuples, you get a dict because `return_dict` is set to `True` is all the configs by default.", "Ôk, but then I am confused why above unpacked worked for, although return dict is true?", "Because you can unpack a dict, and you get the keys of the dict.", "This issue has been stale for 1 month." ]

[]

[]

[]

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

[]

[ "Comparing codebase of version `3.0.2` and `4.2.0` I have noticed that `forward` function differs. I have added deleted lines right at the beginning of the function:\r\n```python\r\n def forward(\r\n self,\r\n hidden_states,\r\n attention_mask=None,\r\n is_index_masked=None,\r\n is_index_global_attn=None,\r\n is_global_attn=None,\r\n output_attentions=False,\r\n ):\r\n \"\"\"\r\n :class:`LongformerSelfAttention` expects `len(hidden_states)` to be multiple of `attention_window`. Padding to\r\n `attention_window` happens in :meth:`LongformerModel.forward` to avoid redoing the padding on each layer.\r\n\r\n The `attention_mask` is changed in :meth:`LongformerModel.forward` from 0, 1, 2 to:\r\n\r\n * -10000: no attention\r\n * 0: local attention\r\n * +10000: global attention\r\n \"\"\"\r\n attention_mask = attention_mask.squeeze(dim=2).squeeze(dim=1)\r\n\r\n # is index masked or global attention\r\n is_index_masked = attention_mask < 0\r\n is_index_global_attn = attention_mask > 0\r\n is_global_attn = any(is_index_global_attn.flatten())\r\n```\r\n\r\nand now model seems to be working, but returns:\r\n```bash\r\n{'eval_loss': nan, 'eval_runtime': 20.6319, 'eval_samples_per_second': 1.939}\r\n```\r\nBelow You can find results of consecutive steps in `forward` function. Can You see something wrong here?\r\n```bash\r\ndiagonal_mask: tensor([[[[-inf, -inf, -inf, ..., 0., 0., 0.]],\r\n\r\n [[-inf, -inf, -inf, ..., 0., 0., 0.]],\r\n\r\n [[-inf, -inf, -inf, ..., 0., 0., 0.]],\r\n\r\n ...,\r\n\r\n [[0., 0., 0., ..., -inf, -inf, -inf]],\r\n\r\n [[0., 0., 0., ..., -inf, -inf, -inf]],\r\n\r\n [[0., 0., 0., ..., -inf, -inf, -inf]]],\r\n\r\n\r\n [[[-inf, -inf, -inf, ..., 0., 0., 0.]],\r\n\r\n [[-inf, -inf, -inf, ..., 0., 0., 0.]],\r\n\r\n [[-inf, -inf, -inf, ..., 0., 0., 0.]],\r\n\r\n ...,\r\n\r\n [[0., 0., 0., ..., -inf, -inf, -inf]],\r\n\r\n [[0., 0., 0., ..., -inf, -inf, -inf]],\r\n\r\n [[0., 0., 0., ..., -inf, -inf, -inf]]]], device='cuda:0',\r\n dtype=torch.float16)\r\nattn_scores: tensor([[[[ -inf, -inf, -inf, ..., 0.5771, 0.2065, -1.0449],\r\n [ -inf, -inf, -inf, ..., -1.3174, -1.5547, -0.6240],\r\n [ -inf, -inf, -inf, ..., -1.3691, -1.3555, -0.3799],\r\n ...,\r\n [ -inf, -inf, -inf, ..., 1.7402, 1.6152, 0.8242],\r\n [ -inf, -inf, -inf, ..., 0.5122, 1.0342, 0.2091],\r\n [ -inf, -inf, -inf, ..., 1.7568, -0.1534, 0.7505]],\r\n\r\n [[ -inf, -inf, -inf, ..., -0.8066, -1.7480, -2.5527],\r\n [ -inf, -inf, -inf, ..., -3.3652, 0.1046, -0.5811],\r\n [ -inf, -inf, -inf, ..., -0.0958, -1.0957, -0.2377],\r\n ...,\r\n [ -inf, -inf, -inf, ..., -0.4148, -0.9497, -0.1229],\r\n [ -inf, -inf, -inf, ..., -1.9443, -1.3467, -1.5342],\r\n [ -inf, -inf, -inf, ..., 0.1263, -0.4407, 0.1486]],\r\n\r\n [[ -inf, -inf, -inf, ..., -0.9077, -0.1603, -0.5762],\r\n [ -inf, -inf, -inf, ..., -0.2454, 0.1932, -0.5034],\r\n [ -inf, -inf, -inf, ..., -1.4375, -1.2793, -1.0488],\r\n ...,\r\n [ -inf, -inf, -inf, ..., -0.3452, 0.1405, 1.3643],\r\n [ -inf, -inf, -inf, ..., -0.2168, -1.0000, -0.9956],\r\n [ -inf, -inf, -inf, ..., -1.7451, 0.1410, -0.6221]],\r\n\r\n ...,\r\n\r\n [[-1.3965, 0.7798, 0.4707, ..., -inf, -inf, -inf],\r\n [ 0.6260, -0.4146, 0.9180, ..., -inf, -inf, -inf],\r\n [ 0.4807, -1.0742, 1.2803, ..., -inf, -inf, -inf],\r\n ...,\r\n [ 0.0909, 0.8022, -0.4170, ..., -inf, -inf, -inf],\r\n [-2.6035, -1.2988, 0.5586, ..., -inf, -inf, -inf],\r\n [-0.6953, -0.8232, 0.0436, ..., -inf, -inf, -inf]],\r\n\r\n [[ 1.0889, -0.2776, -0.0632, ..., -inf, -inf, -inf],\r\n [-0.4128, 0.4834, -0.3848, ..., -inf, -inf, -inf],\r\n [-0.8794, 0.9150, -1.5107, ..., -inf, -inf, -inf],\r\n ...,\r\n [ 0.8867, -0.4731, 0.3389, ..., -inf, -inf, -inf],\r\n [-0.1365, 0.4905, -2.0000, ..., -inf, -inf, -inf],\r\n [-0.0205, -0.5464, -0.6851, ..., -inf, -inf, -inf]],\r\n\r\n [[ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n ...,\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf]]],\r\n\r\n\r\n [[[ -inf, -inf, -inf, ..., -4.0469, -2.6270, -5.4805],\r\n [ -inf, -inf, -inf, ..., -0.9312, -0.6743, -1.9688],\r\n [ -inf, -inf, -inf, ..., -0.0593, -0.9507, -0.6392],\r\n ...,\r\n [ -inf, -inf, -inf, ..., 0.3105, 2.3926, 1.0664],\r\n [ -inf, -inf, -inf, ..., -0.0166, 2.2754, 1.0449],\r\n [ -inf, -inf, -inf, ..., -0.4224, 1.7686, -0.2603]],\r\n\r\n [[ -inf, -inf, -inf, ..., -0.5088, -1.2666, -0.4363],\r\n [ -inf, -inf, -inf, ..., -0.3823, -1.7998, -0.4504],\r\n [ -inf, -inf, -inf, ..., -0.1525, 0.1614, -0.0267],\r\n ...,\r\n [ -inf, -inf, -inf, ..., 0.0225, -0.5737, 0.2318],\r\n [ -inf, -inf, -inf, ..., 0.7139, 0.6099, 0.3767],\r\n [ -inf, -inf, -inf, ..., 0.2008, -0.6714, 0.5869]],\r\n\r\n [[ -inf, -inf, -inf, ..., -0.9302, -1.5303, -2.7637],\r\n [ -inf, -inf, -inf, ..., -0.1124, -0.5850, 0.0818],\r\n [ -inf, -inf, -inf, ..., -1.5176, -1.7822, -0.9111],\r\n ...,\r\n [ -inf, -inf, -inf, ..., -0.3618, 0.3486, 0.4368],\r\n [ -inf, -inf, -inf, ..., -0.4158, -1.1660, -0.9106],\r\n [ -inf, -inf, -inf, ..., -0.4636, -0.7012, -0.9570]],\r\n\r\n ...,\r\n\r\n [[-1.0137, -1.2324, -0.2091, ..., -inf, -inf, -inf],\r\n [ 0.0793, 0.1862, -0.6162, ..., -inf, -inf, -inf],\r\n [ 0.2406, 0.1237, -1.0420, ..., -inf, -inf, -inf],\r\n ...,\r\n [ 0.5308, 0.3862, 0.9731, ..., -inf, -inf, -inf],\r\n [-0.5752, -0.8174, 0.4766, ..., -inf, -inf, -inf],\r\n [-0.4299, -0.7031, -0.6240, ..., -inf, -inf, -inf]],\r\n\r\n [[-2.9512, -1.0410, 0.9194, ..., -inf, -inf, -inf],\r\n [-0.0306, -0.8579, 0.1930, ..., -inf, -inf, -inf],\r\n [ 0.2927, -1.4600, -1.6787, ..., -inf, -inf, -inf],\r\n ...,\r\n [ 0.6128, -0.8921, 1.2861, ..., -inf, -inf, -inf],\r\n [-0.7778, -0.8564, 2.3457, ..., -inf, -inf, -inf],\r\n [-0.8877, -1.4834, 0.7783, ..., -inf, -inf, -inf]],\r\n\r\n [[ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n ...,\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf],\r\n [ nan, nan, nan, ..., -inf, -inf, -inf]]]],\r\n device='cuda:0', dtype=torch.float16)\r\n``` ", "Hey @adamwawrzynski,\r\n\r\nsadly we cannot maintain `convert_model_to_longformer.py` as I think it's not in the core transformers library `src/transformers/...`. Feel free to ask your question on the forum: https://discuss.huggingface.co/ though - maybe someone from the community wants to help", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Hi! The `BertJapaneseTokenizer` you mention was created specifically for Japanese, so it should encode the Japanese language well.\r\n\r\nYou can find the list of models that have Japanese checkpoints [here](https://huggingface.co/models?filter=ja).", "Tokenizers can be decoupled from their models, so you can indeed use a BERT tokenizer with a BART model; however, this requires the tokenizer and model to be trained together.", "Thanks for the quick reply! \r\nI don't know the exact procedure to train tokenizer and model **together**. Could you explain it in detail?" ]

[]

[]

[ "You can avoid removing that column by setting `remove_unused_columns=False` in your `TrainingArguments`.", "I will try it,thank you so much " ]

[]

[]

[ "Sorry we have to revert this PR as `labels` is not a dict but a tensor and makes fails all our examples.", "Thanks a lot for having spotted this cas, a more adapted fix will be available here #9616 very sorry for the inconvenience.", "Never mind. Thanks again for the fix." ]

[]

[]

[]

[]

[]

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

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[]

[ "Thank you for your awesome suggestions and tweaks - all done." ]

[]

[]

[]

[]

[]

[ "Hi, thanks for raising an issue!\r\nIndeed, this is problematic. We're going to do a patch release this morning (v4.2.1) with a fix for this.", "Ths fix is here: https://github.com/huggingface/transformers/pull/9584\r\n\r\nIt should be merged in a couple of hours, after which we'll release a patch." ]

[]

[]

[ "Could you put the full error in the description of the issue rather than in the title? We don't know which weights are not initialized.", "> Could you put the full error in the description of the issue rather than in the title? We don't know which weights are not initialized.\r\nCode:\r\nmodel = XLMRobertaForMaskedLM.from_pretrained('xlm-roberta-base')\r\nWarning:\r\nSome weights of XLMRobertaForMaskedLM were not initialized from the model checkpoint at xlm-roberta-base and are newly initialized: ['lm_head.decoder.bias']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.", "You don't need to worry about that message, it lets you know that the bias in the LM head is not initialized - it will be initialized to all zeros.\r\n\r\nI'm removing this warning in #9615." ]

[]

[]

[]

[]

[]

[ "The `self.args._n_gpu = 1` is to avoid parallelizing the data so it has nothing to do with your problem (and it is right, we can't set `self.args.n_gpu` which is a property but that's a whole different story!)\r\n\r\nHow is your model parallelized? Without that piece of code we can't reproduce the bug and help you.", "Thanks @sgugger.\r\n\r\nIn my test, I'm using some code originally derived from the run_clm.py example. I'm trying to fine-tune a GPT2 model I've trained from scratch. The model was parallelized with the following lines, and this exact fine-tuning script ran successfully yesterday in 4.1.1, using the `--model_parallel` training arg. \r\n\r\n```\r\n device_map = {0: range(0, 15),\r\n 1: range(15, 32)}\r\n model.parallelize(device_map)\r\n```\r\n\r\nThe error I'm getting now looks a lot like what would happen if I left out the `--model_parallel` flag in 4.1.1.\r\n", "> RuntimeError: module must have its parameters and buffers on device cuda:0 (device_ids[0]) but found one of them on device: cuda:1\r\n\r\nPlease post the full trace.\r\n\r\nI have only experimented with t5 and bart MP so far, but gpt2 is supposed to be very similar.\r\n\r\nMost likely the outputs aren't being copied back to the 0th gpu on return, so this won't have anything to do with the trainer. Most likely the issue you encountered has to do with evaluation and not training.\r\n\r\nI had to fix t5-MP to do that, but the PR with the fix hasn't been merged. \r\n\r\nhttps://github.com/huggingface/transformers/blob/58d047a596a97fbb815acb3e657102bf1960b06a/src/transformers/models/t5/modeling_t5.py#L1263-L1266\r\n\r\nI won't be surprised if gpt2 is missing that too.\r\n\r\n`model_parallel_inputs_to_specific_device` is a new function that isn't in master, but part of these 2 PRs: https://github.com/huggingface/transformers/pull/9323 and https://github.com/huggingface/transformers/pull/9384 - it relies on another function - the full new file is here: https://github.com/huggingface/transformers/blob/fe21c43745fcf3f7958c17c2ac461bd784094205/src/transformers/utils/model_parallel_utils.py\r\n\r\nThe current MP implementations are very limited and at the moment I highly recommend you look at DeepSpeed instead, see:\r\nhttps://github.com/huggingface/transformers/issues/8771#issuecomment-759176685 and\r\nhttps://github.com/huggingface/transformers/issues/8771#issuecomment-759248400\r\nYou will need master for that as it was just merged 2 days ago.\r\n\r\nWe also removed `--model_parallel` in trainer master as it wasn't fully baked in first place.", "@stas00 This is linked to how `TrainingArguments.n_gpu` was computed. Could reproduce and test the fix in #9578 removes the bug.", "That's easy then. The error though very much reminded me of the issue I described in my comment above.", "Thanks both!\r\n\r\n@stas00 Definitely excited to check out DeepSpeed – that's the reason I started testing my code in 4.2.0" ]

[]

[]

[ "The [documentation](https://huggingface.co/transformers/main_classes/pipelines.html#transformers.pipeline) of the pipeline function clearly shows the `truncation` argument is not accepted, so i'm not sure why you are filing this as a bug.\r\n\r\nThe `__call__` method of a class is not what is used when you create it but when you... well, call it. So `results = nlp(narratives, **kwargs)` will probably work better.", "@sgugger , you're right. Thanks for the quick response. Sorry, while I looked at https://github.com/huggingface/transformers/pull/9432, I didn't look close enough at https://github.com/huggingface/transformers/blob/master/tests/test_pipelines_summarization.py#L78 or the updated docs. It works now! Thanks @Narsil for adding this feature." ]

[]

[]

[ "The `BertForPreTraining` contains two heads: the NSP and the MLM heads. Therefore, by using the tf1 conversion script, you're already porting the entire model!\r\n\r\nFor the configuration, it should align pretty seamlessly to Google's configurations, but you can check the expected field here: https://github.com/huggingface/transformers/blob/master/src/transformers/models/bert/configuration_bert.py#L120-L138", "Great, thank you! For anyone who's curious, I ended up updating my config file with a few extra fields from here to make `Auto*` detection work: https://huggingface.co/bert-base-multilingual-cased/blob/main/config.json" ]

[]

[]

[]

[]

[]

[ "Alright! I found the solution. For the tokenizer, XLMRobertaTokenizer should be used instead of AutoTokenizer. ", "Thanks for reporting. Now that we have model versioning, the author(s) of [`\"microsoft/Multilingual-MiniLM-L12-H384\"`](https://huggingface.co/microsoft/Multilingual-MiniLM-L12-H384) could update the model's config.json to specify a `tokenizer_class` so that AutoTokenizer works out of the box.\r\n\r\n@JetRunner @LysandreJik do you remember who the model author(s) are?", "I think I did the uploading. I'll update the config tomorrow!", "I believe the config is not yet updated because the error is still there", "@sersoage Re-uploading it now. Thanks for the note!", "Done", "@JetRunner Thank you!" ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nThanks!", "yes,. I asked but forum is inactive, please do not close it I really need\nhelp on this.\n\nOn Thu, Jan 14, 2021 at 9:35 AM Lysandre Debut <[email protected]>\nwrote:\n\n> Hello, thanks for opening an issue! We try to keep the github issues for\n> bugs/feature requests.\n> Could you ask your question on the forum <https://discusss.huggingface.co>\n> instead?\n>\n> Thanks!\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/9572#issuecomment-760020852>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AM3GZM3PY5THOZ4TKXEDGB3SZ2UFXANCNFSM4WBHELJA>\n> .\n>\n", "please do not close it I really need help on this and forrum is inactive", "Our github issue policy is detailed in the related [ISSUE.md](https://github.com/huggingface/transformers/blob/master/ISSUES.md) and very clearly indicate we reserve the issue tracker for bugs and features request which this issue is not, as well as several other issues you have recently opened.\r\n\r\nFeel free to post here a link to the thread you should open on the forum if you want to be visible in both location (though this should stay very exceptional).\r\n\r\nThe forum is NOT inactive, I see that you already have several answers to your [related post](https://discuss.huggingface.co/t/training-models-for-smaller-epochs-and-then-continue-trianing/3153). This is the place for discussion. NOT here in the issues.\r\n\r\nOverall, please note that if you persist in not following the guidelines and open-source collaboration policies that we have defined and shared with the community on the repository in the [CODE_OF_CONDUCT](https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md), the [CONTRIBUTING](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) and [ISSUES](https://github.com/huggingface/transformers/blob/master/ISSUES.md) documents, we reserve the right to take the moderation actions advised by GitHub in the [Community guidelines](https://docs.github.com/en/free-pro-team@latest/github/site-policy/github-community-guidelines).", "Hi Thomas,\nI did received some response, but those response were not helpful, this\nneeds a response from someone developed the codes, to know the small\ndetails which can help,\nplease assist me with the issue, is there a way I could get a better help\nin forume? so someone with more knowledge see the question?\nthanks\n\nOn Thu, Jan 14, 2021 at 1:54 PM Thomas Wolf <[email protected]>\nwrote:\n\n> Our github issue policy is detailed in the related ISSUE.md\n> <https://github.com/huggingface/transformers/blob/master/ISSUES.md> and\n> very clearly indicate we reserve the issue tracker for bugs and features\n> request which this issue is not, as well as several other issues you have\n> recently opened.\n>\n> Feel free to post here a link to the thread you should open on the forum\n> if you want to be visible in both location (though this should stay very\n> exceptional).\n>\n> The forum is NOT inactive, I see that you already have several answers to\n> your related post\n> <https://discuss.huggingface.co/t/training-models-for-smaller-epochs-and-then-continue-trianing/3153>.\n> This is the place for discussion. NOT here in the issues.\n>\n> Overall, please note that if you persist in not following the guidelines\n> and open-source collaboration policies that we have defined and shared with\n> the community on the repository in the CODE_OF_CONDUCT\n> <https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md>,\n> the CONTRIBUTING\n> <https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md>\n> and ISSUES\n> <https://github.com/huggingface/transformers/blob/master/ISSUES.md>\n> documents, we reserve the right to take the moderation actions advised by\n> GitHub in the Community guidelines\n> <https://docs.github.com/en/free-pro-team@latest/github/site-policy/github-community-guidelines>\n> .\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/9572#issuecomment-760178500>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AM3GZM7FRFC7SZF7DBIGIZLSZ3SPHANCNFSM4WBHELJA>\n> .\n>\n", "To me this question can safely be marked as a bug, since if one trains a\nfinetune_trainer.py till some epochs and start retraining from the saved\ncheckpoints, this does not get the same accuracy as full training, can I\nfile a bug for this issue? thanks\n\nOn Sat, Jan 16, 2021 at 12:13 PM julia hane <[email protected]> wrote:\n\n> Hi Thomas,\n> I did received some response, but those response were not helpful, this\n> needs a response from someone developed the codes, to know the small\n> details which can help,\n> please assist me with the issue, is there a way I could get a better help\n> in forume? so someone with more knowledge see the question?\n> thanks\n>\n> On Thu, Jan 14, 2021 at 1:54 PM Thomas Wolf <[email protected]>\n> wrote:\n>\n>> Our github issue policy is detailed in the related ISSUE.md\n>> <https://github.com/huggingface/transformers/blob/master/ISSUES.md> and\n>> very clearly indicate we reserve the issue tracker for bugs and features\n>> request which this issue is not, as well as several other issues you have\n>> recently opened.\n>>\n>> Feel free to post here a link to the thread you should open on the forum\n>> if you want to be visible in both location (though this should stay very\n>> exceptional).\n>>\n>> The forum is NOT inactive, I see that you already have several answers to\n>> your related post\n>> <https://discuss.huggingface.co/t/training-models-for-smaller-epochs-and-then-continue-trianing/3153>.\n>> This is the place for discussion. NOT here in the issues.\n>>\n>> Overall, please note that if you persist in not following the guidelines\n>> and open-source collaboration policies that we have defined and shared with\n>> the community on the repository in the CODE_OF_CONDUCT\n>> <https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md>,\n>> the CONTRIBUTING\n>> <https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md>\n>> and ISSUES\n>> <https://github.com/huggingface/transformers/blob/master/ISSUES.md>\n>> documents, we reserve the right to take the moderation actions advised by\n>> GitHub in the Community guidelines\n>> <https://docs.github.com/en/free-pro-team@latest/github/site-policy/github-community-guidelines>\n>> .\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/9572#issuecomment-760178500>,\n>> or unsubscribe\n>> <https://github.com/notifications/unsubscribe-auth/AM3GZM7FRFC7SZF7DBIGIZLSZ3SPHANCNFSM4WBHELJA>\n>> .\n>>\n>\n", "The maintainers have very clearly labelled it as NOT a bug ([here](https://github.com/huggingface/transformers/issues/9572#issuecomment-760020852)).\r\n\r\nWhat you are asking here is a free consultation which unfortunately we don't provide at HuggingFace so here is the next step I think are the most adapted in the present case:\r\n- you've probably reached the limit of the amount of consulting the community could provide for free, the best would be now to hire a consultant to build a solution for you\r\n- regarding the current issue and your usage of the open-source repository, my mission is now to step up and protect the maintainers from the reprository to be able to conduct their mission of maintaining the repository and not diverging from their task to conduct free consulting missions. As such this is now the second warning I send you to follow the guidelines and open-source collaboration policies that we have defined and shared with the community on the repository in the CODE_OF_CONDUCT, the CONTRIBUTING and ISSUES documents (see my message [here](https://github.com/huggingface/transformers/issues/9572#issuecomment-760178500)).\r\n- Which lead us to the last point: if I have to spend more time and send you a third warning to use the repository as it was designed for the community I will have to limit your ability to open issues on our repositories following the GitHub Community guidelines as I explained to you in my message [here](https://github.com/huggingface/transformers/issues/9572#issuecomment-760178500). That would be the first time I have to do that with a community of thousand of people so please just use our forum tools like we designed them according to our guidelines stated [here](https://github.com/huggingface/transformers/issues/9572#issuecomment-760178500)." ]

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

[]

[ "Hello!\r\n\r\nUnfortunately the TF models are not yet compliant with the \"float16\" mixed precision. This is our main goal for the next release (the one after 4.2.X) as we are actively working on this.\r\n\r\nSorry the the inconvenience. I will update this post once done. ", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[ "Ok, just restored the previous version checking." ]

[]

[]

[ "Thanks for opening a new PR. Let me know if you need a review (It's also ok if I go into the PR and fix some things if your stuck :-) )", "@patrickvonplaten I hope this PR is again ready for review. The only thing remaining to resolve is that issue in `test_headmasking` described above. Currently, I've been trying to fix this one, but I'll be grateful for sure if you can have a look at that too :)", "Hey @patrickvonplaten. I would like to inform you I fixed `test_headmasking` for BART-based. The problem was that code inside\r\n```\r\nself.assertNotEqual(attentions[1][..., 0, :, :].flatten().sum().item(), 0.0)\r\n```\r\npointed to the last layer of encoder/decoder (encoder-decoder models have only 2 layers in each module while BERT has 5 layers during testing). At the end of the day, this condition was invalid for BART-based models considering the `head_mask` to be\r\n```\r\nhead_mask = torch.ones(\r\n self.model_tester.num_hidden_layers,\r\n self.model_tester.num_attention_heads,\r\n device=torch_device,\r\n)\r\nhead_mask[0, 0] = 0\r\nhead_mask[-1, :-1] = 0\r\n```\r\n\r\nI hope this PR is then ready for review." ]

[]

[]

[ "Hey @cheop-byeon, \r\n\r\nwe no longer actively maintain the `research_projects` folder ourselves. To solve your problem you can however just add \r\n`model.config.decoder_start_token_id` as the third argument to the function.", "Note that we recommend that you use the research project with it's proposed version being `pip install transformers==4.1.0`. We won't do actively maintain the code in `research_projects` anymore." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Okey for me" ]