Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Interested in this as well. Do not find a solution from blog [\"How to monitor both train and validation metrics at the same step?\"](https://discuss.huggingface.co/t/how-to-monitor-both-train-and-validation-metrics-at-the-same-step/1301).", "> Hi\r\n> I want to find the best model per evaluation score. Could you please give me more info, how I can checkpoint all evaluation scores in each step of training to find the best model? thanks\r\n\r\nI think I figure it out:\r\n```diff\r\ntraining_args = TrainingArguments(\r\n output_dir='./results', # output directory\r\n num_train_epochs=3, # total number of training epochs\r\n per_device_train_batch_size=16, # batch size per device during training\r\n per_device_eval_batch_size=64, # batch size for evaluation\r\n warmup_steps=500, # number of warmup steps for learning rate scheduler\r\n weight_decay=0.01, # strength of weight decay\r\n logging_dir='./logs', # directory for storing logs\r\n logging_steps=100,\r\n++ evaluation_strategy='steps',\r\n)\r\n```", "> Interested in this as well. Do not find a solution from blog [\"How to monitor both train and validation metrics at the same step?\"](https://discuss.huggingface.co/t/how-to-monitor-both-train-and-validation-metrics-at-the-same-step/1301).\r\n\r\n\r\n\r\n> Hi\r\n> I want to find the best model per evaluation score. Could you please give me more info, how I can checkpoint all evaluation scores in each step of training to find the best model? thanks\r\n\r\nCorrespondingly, I put \r\n```diff\r\n# https://huggingface.co/transformers/training.html\r\n#metric = load_metric('glue', 'mrpc')\r\ndef compute_metrics(p):#: EvalPrediction\r\n# def compute_metrics(p):\r\n preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions\r\n# preds = np.squeeze(preds) if is_regression else np.argmax(preds, axis=1)\r\n preds = np.argmax(preds, axis=1)\r\n# if data_args.task_name is not None:\r\n# result = metric.compute(predictions=preds, references=p.label_ids)\r\n# if len(result) > 1:\r\n# result[\"combined_score\"] = np.mean(list(result.values())).item()\r\n# return result\r\n# elif is_regression:\r\n# return {\"mse\": ((preds - p.label_ids) ** 2).mean().item()}\r\n# else:\r\n return {\"accuracy\": (preds == p.label_ids).astype(np.float32).mean().item()}\r\n\r\ntrainer = Trainer(\r\n model=model, # the instantiated 🤗 Transformers model to be trained\r\n args=training_args, # training arguments, defined above\r\n train_dataset=train_dataset, # training dataset\r\n eval_dataset=val_dataset, # evaluation dataset\r\n++ compute_metrics = compute_metrics\r\n)\r\n\r\ntrainer.train()\r\n```\r\nand result in output:\r\n\r\n\r\n", "\r\n- However, in terms of `Accuracy`, Not for sure it is on training dataset or validation dataset.", "`Accuracy` here is for validation dataset.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hey @hjzhang1018,\r\n\r\nThanks for your bug report. Could you try to run your command again - I think it should be fixed now: https://huggingface.co/seyonec/SMILES_tokenized_PubChem_shard00_160k/commit/7ef67531cfe96d0e2aa3ae913352c8e9a8c1df4f", "@seyonechithrananda I added some files to your repo here: https://huggingface.co/seyonec/SMILES_tokenized_PubChem_shard00_160k/commit/7ef67531cfe96d0e2aa3ae913352c8e9a8c1df4f - feel free to take a look and see whether this is OK for you. It might be possible that other models require this change as well. ", "@hjzhang1018 @patrickvonplaten Hi all, thanks for the fix! I believe the reason this may be happening is because the tokenizer we use is custom (a subclass of BertTokenizer) and thus we run into the issue of fitting directly with AutoTokenizer. We have a PR for a tutorial in the DeepChem [library](https://github.com/deepchem/deepchem/pull/2302), which demonstrates how to call our subclass instead of using AutoTokenizer. If you refer to Part 23 hopefully that is of use!\r\n\r\nDocs: https://deepchem.readthedocs.io/en/latest/api_reference/tokenizers.html", "Link to SmilesTokenizer class which these models utilize: https://github.com/deepchem/deepchem/blob/master/deepchem/feat/smiles_tokenizer.py#L39-L282", "Just read the changes, it looks like @patrickvonplaten directly converted the vocab.txt file sufficient for BertTokenizer into the vocab.json format necessary for RoBERTa tokenizers, which should run smoothly. I will try this out once I get more time but this fix should work. Thanks a lot for the quick fix!", "Is there a way to add this change to the other models with the 'SmilesTokenizer', @patrickvonplaten? Thanks again for the support.", "@patrickvonplaten Thank you so much for your help! Now this worked!\r\n@seyonechithrananda Thank you for the explanation. I'll read the tutorials carefully. Very useful!", "> Is there a way to add this change to the other models with the 'SmilesTokenizer', @patrickvonplaten? Thanks again for the support.\r\n\r\nYes on simply has to create an empty 'merges.txt' file and create `vocab.json` from vocab.txt", "Hi, I have the same issues here, I have a custom roberta model and I am using https://github.com/UKPLab/sentence-transformers. Here is the full detail of my problem: https://github.com/UKPLab/sentence-transformers/issues/658\r\n\r\nThe output after training from this sentence transformers yield files that doesn't contain vocab.json or vocab.txt. But I have a file called `unigram.json` and it looks something like this:\r\n\r\n\r\n\r\n\r\n```\r\n{\r\n \"unk_id\": 0,\r\n \"vocab\": [\r\n [\r\n \"<unk>\",\r\n 0.0\r\n ],\r\n [\r\n \"<sep>\",\r\n 0.0\r\n ],\r\n [\r\n \"<pad>\",\r\n 0.0\r\n ],\r\n ....\r\n ]\r\n}\r\n```\r\n\r\nI also faced this TypeError, the same as the title of this issue, when trying to use AutoTokenizer", "In short every Roberta-like Tokenizer requires two files:\r\n\r\n1) One merges.txt file. This file describes the BPE algorithm (which letters are merged in which order)\r\n2) One vocab.json file. This file describes the vocabulary.\r\n\r\nTo get an idea of how the format of these files should be, I'd recommend taking a look at some the files in `roberta-base` here:https://huggingface.co/roberta-base/tree/main\r\n\r\n@seyonechithrananda, I think in your case, you don't need a merges.txt file because of the small vocabulary and because there are no words, just tokens\r\n@hjzhang1018 If the other library: UKPLab/sentence-transformers uses the same format for loading/saving files that we do, then file should be renamed to be called `vocab.json` and should have a different format (check out the format here: https://huggingface.co/roberta-base/blob/main/vocab.json).", "So after I call `sentence-transformers` save which gives me back 6 files:\r\n1. I will need to rename `unigram.json` to `vocab.json`\r\n2. Change the format of `unigram.json` to follows `vocab.json` structure\r\n3. Create an empty `merge.txt` file\r\n\r\nAnd currently my `unigram.json` contains a word weight:\r\n```\r\n{\r\n \"unk_id\": 0,\r\n \"vocab\": [\r\n [\r\n \"<unk>\",\r\n 0.0\r\n ],\r\n [\r\n \"<sep>\",\r\n 0.0\r\n ],\r\n [\r\n \"<pad>\",\r\n 0.0\r\n ],\r\n [\r\n \"<cls>\",\r\n 0.0\r\n ],\r\n [\r\n \"<mask>\",\r\n 0.0\r\n ],\r\n [\r\n \",\",\r\n -3.1215689182281494\r\n ],\r\n [\r\n \".\",\r\n -3.642984628677368\r\n ],\r\n [\r\n \"a\",\r\n -4.921720027923584\r\n ],\r\n ......\r\n ]\r\n}\r\n```\r\n\r\nDo I just ignore all the weights and created a new file `vocab.json` with this format ?\r\n```\r\n{\r\n \"<unk>\": 0,\r\n \"<sep>\": 1,\r\n \"<pad>\": 2,\r\n \"<cls>\": 3,\r\n \"<mask>\": 4,\r\n \",\": 5,\r\n \".\": 6,\r\n \"a\": 7,\r\n .......\r\n}\r\n```", "I started getting this error only about an hour ago without any changes on my side in my old Colab notebook. There must be some version change in Colab env that triggers this error. Any ideas what it might be?\r\n", "@lenyabloko could you open a new issue with the issue you're facing and how to reproduce it? Thank you.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hey @hjzhang1018,\r\n\r\nThis does not seem to be a bug in Transformers, but rather in `seyonechithrananda/simpletransformers.git` so I'm not sure here is the correct place to post the issue. I think one has to change the line `from transformers.modeling_albert import ....` to `from transformers.models.albert.modeling_albert import ...` in the respective repo." ]

[]

[]

[ "Hey @PremalMatalia, \r\n\r\nCould you please provide a copy/paste ready code-snippet that can be used to reproduce the error. By copy/past ready code snippet I mean something like:\r\n\r\n```python\r\nfrom transformers import DistilBertTokenizerFast\r\ntokenizer = DistilBertTokenizerFast.from_pretrained('distilbert-base-uncased')\r\n\r\n# ... add all the necessary code here to be able to reproduce your error\r\n```\r\n\r\n. Thanks! ", "Hello Patrick,\r\nPlease find entire code starting from SQuAD 2.0 training data download to encoding to adding start and end position as below:\r\n\r\n```python\r\n!pip install wget\r\n!pip install transformers==4.0.1\r\n\r\nimport wget\r\nimport json\r\nfrom pathlib import Path\r\nimport os\r\nimport json\r\nfrom transformers import DistilBertTokenizerFast,TFDistilBertForQuestionAnswering\r\nimport tensorflow as tf\r\n\r\n# Import training data\r\n!mkdir squad\r\ntrain_source = 'https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json'\r\ntrain_dest = 'squad/train-v2.0.json'\r\nwget.download(train_source,train_dest)\r\n\r\n## Function to extract context, questions and answers\r\ndef read_squad(path,dataset='train'):\r\n \r\n contexts = []\r\n questions = []\r\n question_ids = []\r\n answers = []\r\n blank_counter = 0\r\n append_flag = False\r\n\r\n with open(path) as f:\r\n data = json.load(f)\r\n \r\n ## Loop over entire dataset\r\n for article_id in range(len(data['data'])):\r\n paragraphs = data['data'][article_id]['paragraphs']\r\n ## Loop over all the paragraphs\r\n for paragraph in paragraphs:\r\n context = paragraph['context']\r\n qas = paragraph['qas']\r\n ## Loop over Questions and Answers\r\n for qa in qas:\r\n append_flag=False\r\n question = qa['question']\r\n question_id = qa['id']\r\n ## Select 1st answer if answers are available \r\n if qa['answers']:\r\n answer = qa['answers'][0]\r\n append_flag = True\r\n ## Append contexts and questions in a list and answers in a list as dictionary\r\n contexts.append(context)\r\n questions.append(question)\r\n question_ids.append(question_id)\r\n answers.append(answer) \r\n \r\n return contexts, questions, question_ids, answers\r\n\r\ntrain_contexts, train_questions,_, train_answers = read_squad('squad/train-v2.0.json')\r\n\r\n\r\n## Function to update answer_start and answer_end \r\ndef add_end_idx(answers, contexts):\r\n '''\r\n Description: This function is to find out character position at which the answer ends in the passage. \r\n Also corrects answer start and end position if the SQuAD answers are off by one or two characters\r\n Input: List of all answers, List of all contexts\r\n Output: Updated list with answer end position\r\n '''\r\n for answer, context in zip(answers, contexts):\r\n # Your code here\r\n if answer['answer_start'] is None:\r\n answer['answer_end'] = None\r\n else:\r\n answer_text = answer['text']\r\n answer_start = answer['answer_start']\r\n answer_end = len(answer_text) + answer_start\r\n\r\n #Sometimes answers are off by a character or two\r\n if context[answer_start:answer_end] == answer['text']:\r\n answer['answer_end'] = answer_end\r\n # If the answer text is off by 1 character\r\n elif context[answer_start-1:answer_end-1] == answer_text:\r\n answer['answer_start'] = answer_start - 1\r\n answer['answer_end'] = answer_end - 1 \r\n # If the answer text is off by 2 characters\r\n elif context[answer_start-2:answer_end-2] == answer_text:\r\n answer['answer_start'] = answer_start - 2\r\n answer['answer_end'] = answer_end - 2 \r\n\r\nadd_end_idx(train_answers, train_contexts)\r\n\r\n## Tokenize training data\r\ntokenizer = DistilBertTokenizerFast.from_pretrained('distilbert-base-uncased')\r\ntrain_encodings = tokenizer(train_contexts, train_questions, truncation=True, padding=True,\r\n return_offsets_mapping=True,\r\n return_overflowing_tokens=True)\r\n\r\n## Find out start_position and end_position in encoded dataset\r\ndef add_token_positions(encodings, answers):\r\n start_positions = []\r\n end_positions = []\r\n\r\n for idx in range(len(answers)):\r\n start_positions.append(encodings.char_to_token(idx, answers[idx]['answer_start']))\r\n if answers[idx]['answer_end'] is None:\r\n end_positions.append(encodings.char_to_token(idx, answers[idx]['answer_end']))\r\n else:\r\n end_positions.append(encodings.char_to_token(idx, answers[idx]['answer_end'] - 1))\r\n \r\n #if None, the answer passage has been truncated due to words > 512 so setting last position as 511\r\n if start_positions[-1] is None:\r\n start_positions[-1] = tokenizer.model_max_length-1\r\n if end_positions[-1] is None:\r\n end_positions[-1] = tokenizer.model_max_length-1\r\n \r\n encodings.update({'start_positions': start_positions, 'end_positions': end_positions})\r\n\r\nadd_token_positions(train_encodings, train_answers)\r\n\r\n## Validate answers based on start_position and end_position with actual answer for some random index\r\nidx=8\r\nprint(f'Actual context: {train_contexts[idx]}')\r\nprint(f'Actual question: {train_questions[idx]}')\r\nprint(f\"Actual answer: {train_answers[idx]['text']}\")\r\n\r\nstart_position=train_encodings.char_to_token(idx,train_answers[idx]['answer_start'])\r\nend_position =train_encodings.char_to_token(idx,train_answers[idx]['answer_end'])\r\n\r\n## ******This shows how start_position and end_position derived by using char_to_token() function is not correct******\r\nprint(f\"Answer after tokenization: {tokenizer.convert_ids_to_tokens(train_encodings['input_ids'][idx][start_position:end_position])}\")\r\n```", "Completely agree with @PremalMatalia. \r\nThe problem is with -\r\nstart_positions.append(encodings.char_to_token(i, answers[i]['answer_start']))\r\n\r\nWe are getting None where as we should have got start token position. \r\nchar_to_token is not able to convert from string position to token position.", "Thanks @PremalMatalia - I think I can reproduce.\r\n\r\nThe PR attached below should fix the problem. Can you check it again with the proposed fix?", "Thanks @patrickvonplaten for quick action.\r\nIf I understand correctly, fix has been merged to original char_to_token() function? If yes, we can directly use the same function without any changes in code from myside. Is that correct?", "No the `char_to_token()` function was always correct (It's actually a rust function from tokenizers that is used with python bindings). The function was simply used incorrectly, so I updated the docs.", "> Hello Patrick,\r\n> Please find entire code starting from SQuAD 2.0 training data download to encoding to adding start and end position as below:\r\n> \r\n> ```python\r\n> !pip install wget\r\n> !pip install transformers==4.0.1\r\n> \r\n> import wget\r\n> import json\r\n> from pathlib import Path\r\n> import os\r\n> import json\r\n> from transformers import DistilBertTokenizerFast,TFDistilBertForQuestionAnswering\r\n> import tensorflow as tf\r\n> \r\n> # Import training data\r\n> !mkdir squad\r\n> train_source = 'https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json'\r\n> train_dest = 'squad/train-v2.0.json'\r\n> wget.download(train_source,train_dest)\r\n> \r\n> ## Function to extract context, questions and answers\r\n> def read_squad(path,dataset='train'):\r\n> \r\n> contexts = []\r\n> questions = []\r\n> question_ids = []\r\n> answers = []\r\n> blank_counter = 0\r\n> append_flag = False\r\n> \r\n> with open(path) as f:\r\n> data = json.load(f)\r\n> \r\n> ## Loop over entire dataset\r\n> for article_id in range(len(data['data'])):\r\n> paragraphs = data['data'][article_id]['paragraphs']\r\n> ## Loop over all the paragraphs\r\n> for paragraph in paragraphs:\r\n> context = paragraph['context']\r\n> qas = paragraph['qas']\r\n> ## Loop over Questions and Answers\r\n> for qa in qas:\r\n> append_flag=False\r\n> question = qa['question']\r\n> question_id = qa['id']\r\n> ## Select 1st answer if answers are available \r\n> if qa['answers']:\r\n> answer = qa['answers'][0]\r\n> append_flag = True\r\n> ## Append contexts and questions in a list and answers in a list as dictionary\r\n> contexts.append(context)\r\n> questions.append(question)\r\n> question_ids.append(question_id)\r\n> answers.append(answer) \r\n> \r\n> return contexts, questions, question_ids, answers\r\n> \r\n> train_contexts, train_questions,_, train_answers = read_squad('squad/train-v2.0.json')\r\n> \r\n> \r\n> ## Function to update answer_start and answer_end \r\n> def add_end_idx(answers, contexts):\r\n> '''\r\n> Description: This function is to find out character position at which the answer ends in the passage. \r\n> Also corrects answer start and end position if the SQuAD answers are off by one or two characters\r\n> Input: List of all answers, List of all contexts\r\n> Output: Updated list with answer end position\r\n> '''\r\n> for answer, context in zip(answers, contexts):\r\n> # Your code here\r\n> if answer['answer_start'] is None:\r\n> answer['answer_end'] = None\r\n> else:\r\n> answer_text = answer['text']\r\n> answer_start = answer['answer_start']\r\n> answer_end = len(answer_text) + answer_start\r\n> \r\n> #Sometimes answers are off by a character or two\r\n> if context[answer_start:answer_end] == answer['text']:\r\n> answer['answer_end'] = answer_end\r\n> # If the answer text is off by 1 character\r\n> elif context[answer_start-1:answer_end-1] == answer_text:\r\n> answer['answer_start'] = answer_start - 1\r\n> answer['answer_end'] = answer_end - 1 \r\n> # If the answer text is off by 2 characters\r\n> elif context[answer_start-2:answer_end-2] == answer_text:\r\n> answer['answer_start'] = answer_start - 2\r\n> answer['answer_end'] = answer_end - 2 \r\n> \r\n> add_end_idx(train_answers, train_contexts)\r\n> \r\n> ## Tokenize training data\r\n> tokenizer = DistilBertTokenizerFast.from_pretrained('distilbert-base-uncased')\r\n> train_encodings = tokenizer(train_contexts, train_questions, truncation=True, padding=True,\r\n> return_offsets_mapping=True,\r\n> return_overflowing_tokens=True)\r\n> \r\n> ## Find out start_position and end_position in encoded dataset\r\n> def add_token_positions(encodings, answers):\r\n> start_positions = []\r\n> end_positions = []\r\n> \r\n> for idx in range(len(answers)):\r\n> start_positions.append(encodings.char_to_token(idx, answers[idx]['answer_start']))\r\n> if answers[idx]['answer_end'] is None:\r\n> end_positions.append(encodings.char_to_token(idx, answers[idx]['answer_end']))\r\n> else:\r\n> end_positions.append(encodings.char_to_token(idx, answers[idx]['answer_end'] - 1))\r\n> \r\n> #if None, the answer passage has been truncated due to words > 512 so setting last position as 511\r\n> if start_positions[-1] is None:\r\n> start_positions[-1] = tokenizer.model_max_length-1\r\n> if end_positions[-1] is None:\r\n> end_positions[-1] = tokenizer.model_max_length-1\r\n> \r\n> encodings.update({'start_positions': start_positions, 'end_positions': end_positions})\r\n> \r\n> add_token_positions(train_encodings, train_answers)\r\n> \r\n> ## Validate answers based on start_position and end_position with actual answer for some random index\r\n> idx=8\r\n> print(f'Actual context: {train_contexts[idx]}')\r\n> print(f'Actual question: {train_questions[idx]}')\r\n> print(f\"Actual answer: {train_answers[idx]['text']}\")\r\n> \r\n> start_position=train_encodings.char_to_token(idx,train_answers[idx]['answer_start'])\r\n> end_position =train_encodings.char_to_token(idx,train_answers[idx]['answer_end'])\r\n> \r\n> ## ******This shows how start_position and end_position derived by using char_to_token() function is not correct******\r\n> print(f\"Answer after tokenization: {tokenizer.convert_ids_to_tokens(train_encodings['input_ids'][idx][start_position:end_position])}\")\r\n> ```\r\n\r\nSince Pytorch removed SAVE_STATE_WARNING now it will pop up an error if install transformers==4.0.1. I use transformers>=4.5 and it works" ]

[]

[]

[ "@norabelrose Great job! Thanks for letting me know about this :)", "@norabelrose - you've done an amazing job here! Having Performer in PyTorch is a huge contribution.\r\n\r\nI completely understand that you've already invested a lot of time in making this PR and we're happy to complete your PR!", "also pinging @TevenLeScao here. \r\n\r\nIn terms of next steps, I think we should do the following (We're happy to take over those tasks @norabelrose :-)) : \r\n\r\n- Check that the added PyTorch and Tensorflow Performer self-attention yields identical results as the flax version: Compare Bert model to this Performer BertFlax model: https://github.com/huggingface/transformers/pull/8358\r\n- Fine-tune some pre-trained weighs to be compatible with Performer attention (ideally Bert or DistilBert)", "Now, the painful discussion on how to integrate Performer.\r\n\r\n**Context**:\r\n\r\nPerformer attention is special in the sense that it is fully compatible with a pre-existing model architecture and does not require any weights to be different from normal attention. This means that a `bert-base-cased` model does not require any changes in its architecture to use Performer's attention. The only change will be how the weights are used to compute the self-attention layer output. \r\nOne can easily see this on the original code base: https://github.com/google-research/google-research/tree/master/performer/fast_attention/jax#jax-variant-of-favor where only the attention function has to be changed to `make_fast_softmax_attention` with no changes required to the parameters dict.\r\n\r\nThis is a huge argument for simply making Performer's self-attention available to all models by changing their respective `modeling_....py` file. \r\n\r\n**My opinion on the integration into Transformers**\r\n\r\nNevertheless, I'm in favor of implementing Performer **only** in a stand-alone file (at least at first), a.k.a. `PerformerModel` or maye in this case `PerformerBertModel`, which is **different** from the current version of the PR. I've the following arguments:\r\n\r\n- It's the standard in Transformers to add a new model for a new attention function. We've done the same for Longformer even though the Longformer attention could have been added for each model. It's easier for users to navigate between models, *e.g.* Performer will have its own model page vs. some docstring in utils.\r\n- It's actually not that easy to convert an existing BERT-like model to a Performer-BERT model. E.g. le'ts say we integrate Performer attention into `modeling_bert.py`. If one wants to convert the model to performer attention, the user would have to manually copy the positional embeddings (which are limited to 512 in Bert) to be as long as 64K+. We could write a convert function for this, but this convert function would probably be different for each model.\r\n- Performer cannot support all the functionalities of Bert. This means if we integrate Performer into Bert, a ```BertModel.from_pretrained(...., is_performer=True)``` model will not have all the functionalities that a Bert model will have, such as `output_attentions=True`, `is_decoder=True`, `is_encoder_decoder=True` -> Performer never creates the complete attention_mask so the `ouput_attentions=True` functionality gets lost, Performer does not support Encoder-Decoder out-of-the-box without requiring more if-else clauses. This will necessarily lead to many issues and some `if self.is_performer` code in BERT which I don't want to do.\r\n- Performer is still a very novel feature that is still somewhat experimental IMO. If Performer really takes off, we can always integrate the Performer attention more deeply into the library as proposed in this PR. The `modeling_bert.py` code is now used by 100K+ people, so I want to be very very careful with changes to this code especially. It's just a safer option to have a standalone Performer model in the beginning IMO. \r\n- I don't really think that users are interested to be able to use Performer Attention for all models. I think the models of interest will be `Bert` (`DistilBERT`), `GPT2`, `T5`, and `Bart`. Some models will never be used with Performer, such as Reformer, XLNet, Transfo-XL, Longformer, ConvBert, Routing Transformer, LED. \r\n\r\n\r\nI'd be thrilled you hear your opinions here @norabelrose, @sgugger, @LysandreJik, and @thomwolf", "Thanks for this amazing contribution! \r\n\r\nI think long document classification and summarisation tasks are an important use case for this so having performer attention for some representative models in those scenarios would be fantastic. Personally I am looking forward to using performer attention with Roberta sequence and token classification models, but I understand not every model can get performer support right away so it would be great to also have a few examples on how we could add performer attention to other models ourselves, if possible.\r\n\r\nReally excited about this, thanks so much!", "@patrickvonplaten Thank you for the thoughtful feedback. I understand your concerns about building Performer attention right into existing models like BERT. On the other hand, as @onclue mentioned, having only one model that supports Performer attention would really restrict the usefulness of the feature.\r\n\r\nIt seems like there should be some \"compromise\" option here. What if we just added simple Performer-supporting subclasses to a few different models, something like this:\r\n```\r\n@supports_performer_attention\r\nclass PerformerBertConfig(BertConfig):\r\n\tpass\r\n```\r\n\r\nand have `PerformerBertModel` be a subclass of `BertModel` that uses the following module for its attention mechanism:\r\n\r\n```\r\nclass PerformerBertAttention(BertAttention):\r\n\t@init_performer_attention_bertlike(BertSelfAttention)\r\n\tdef __init__(self, config):\r\n\t\tsuper().__init__(self, config)\r\n```\r\n\r\nAnd then the same process could be done for RoBERTa, DistilBERT, GPT-2, etc. I recognize that trying to add Performer support to \"all\" models is sort of silly and wouldn't work, but there are quite a few models that would benefit from it. It would also be nice if `PerformerAttention` and `PerformerAttentionConfig` remained public APIs, as they are in this PR, so that users could just take the attention mechanism and drop it into whatever custom model they want.", "Thanks for your answer @norabelrose! I understand your point. The goal should definitely to support all the \"highly\" used models: DistilBERT, BERT, RoBERTa, T5, Bart\r\n\r\nI think we need to dive a bit deeper into the PR and play around with Performer to see how to best integrate the model, but in general the philosophy of the library has always been:\r\n\r\n- Model files should be kept as independent from each other as possible\r\n- Readability is more important than the drawback of duplicated code -> so we don't mind it too much if we duplicate Performer code across 5 or so model files\r\n- We try to minimize \"magic\" internal functionalities that are hard to understand when first seeing the code to a minimum meaning that we're not a huge fan of function decorators for important functionalities in general. \r\n\r\nBut we'll have to dive deeper into the PR to get a better understanding here - sorry for being so slow here!\r\n\r\nAlso, is there already a model that has successfully been fine-tuned to \"long\" inputs?", "@norabelrose Where do the keys, queries & values come from when calling the Performer attention? \r\n\r\ndef call(self, query, key, value, mask=None, head_mask=None, output_attentions=False):\r\n\r\nThe call method of the Bert attention its replacing (i guess) takes the hidden states instead and then calculates q,k,v; \r\n\r\nHow does it work from a coding perspective that it can have different call inputs? \r\n& Should i just feed in the hidden states 3x for q, k, v when using performer attention? \r\n\r\n\r\n", "To get the TFPerformerAttention working I had to had to apply three fixes:\r\n- Swap all shape calls for shape_list\r\n- Add mask = tf.reshape(mask, shape=(shape_list(k_prime)[0], shape_list(k_prime)[2])) in compute_attention_with_projected_queries_and_keys due to problems with the extended attn mask\r\n- Remove the reshapes in _finalize_attention_output, as we need the shape to stay in [..., num_heads, dim_per_head]-like shape\r\n\r\nperhaps it helps sb else // @norabelrose can correct me if im doing sth wrong", "Hello,\r\n\r\nAmazing work @norabelrose! I have been trying your performer implementation. I have copied your attention implementation\r\n```PerformerAttention``` and have replaced that attention with the normal self-attention in Mobilebert. I have tracked some metrics with respect to other implementations. I have seen that the memory consumption on 512 tokens long it consume the same memory that the normal self attention. And it is also the same fast.\r\n\r\nI have logged the metrics with Wandb:\r\nhttps://wandb.ai/gaceladri/new_berts/reports/Memory-and-speed-comparison--Vmlldzo0NDA4MTI \r\n\r\nDoes that makes sense? I have seen in Long Range Arena https://arxiv.org/abs/2011.04006 that it is 1.2x faster with 1k tokens but I have not tried with that long. The point where I am confused is with the memory consumption. At shorter values, the attention mechanism, being linear with respect to sequence length, not should be consuming less memory?", "@norabelrose I tried the `TFPerformerAttention` with some minor adaptions and it works fine during training. I must say it is a very nice implementation 👍 \r\nHowever, when I train my model, I save the weights at each checkpoint, and I quantize it into `model.pb` as well as into a `TF-Lite` model. When loading all the models again (from saved weights, quantized and tensorflow lite), the output of the model with loaded weights differ from the rest. Any idea why this is the case? ", "@gcuder Would you mind sharing your code? I have been getting speed & memory improvements but the TFPerformer doesn't really converge... \r\n\r\n", "What are the plans for this MR @patrickvonplaten ?", "Based on @norabelrose great work, I set up a fork with the performer as a separate model at [https://github.com/Muennighoff/transformers](https://github.com/Muennighoff/transformers). I removed the decorators but kept the separate performer attention config. For now I only added Distilbert, the question being whether we should add new performer_xyz folders for each model or fit them in one performer folder.\r\nIt can just be used as \r\n`from transformers import DistilBertPerformerModel, DistilBertPerformerConfig`\r\n`configuration = DistilBertPerformerConfig()`\r\n`model = DistilBertPerformerModel(configuration)`\r\n\r\nHere's an example notebook comparing the distilbert perf/trans performance on seq. classification:\r\n[https://colab.research.google.com/drive/1o8ioYUIvvIol7PXrDguftQtHyCpyDSJL#scrollTo=JGxH15LIN66M](https://colab.research.google.com/drive/1o8ioYUIvvIol7PXrDguftQtHyCpyDSJL#scrollTo=JGxH15LIN66M)\r\n\r\nperhaps it can help us bring this forward? @marrrcin @patrickvonplaten \r\n\r\n", "Hey guys, sorry I don't really have the bandwidth to take a closer look here at the moment, but it's definitely on my ToDo List! One thing that would be extremely useful would be to have a script that shows how a pretrained model such as `distilbert` can be extended to a its performer version and subsequently be fine-tuned for long-range sequence modeling. @TevenLeScao ran some initial experiments and didn't find the fine-tuning to be that easy...", "Can it be made compatible with T5? As far as I know Performer and relative attention together is an open research question.", "How about this? If I understand correctly Performer calculates `Q' * (K' * V)` instead of `softmax(Q * K) * V` (Q: queries, K: keys, V: values, *: matmul). T5 calculates `softmax(Q * K + B) * V` (B: relative positional biases). A new kind of model that initializes most of its weights from T5 could calculate `(softmax(Q * K) + B') * V = (Q' * K' + B') * V = Q' * (K' * V) + B' * V`. This way at least the first term can be calculated with FAVOR+ and the second term is much smaller/faster to calculate even if its complexity is quadratic. `B'` could be initialized in a way that on average the activations in the training set remain unchanged. We loose backward compatibility so more finetuning is necessary.", "> How about this? If I understand correctly Performer calculates `Q' * (K' * V)` instead of `softmax(Q * K) * V` (Q: queries, K: keys, V: values, *: matmul). T5 calculates `softmax(Q * K + B) * V` (B: relative positional biases). I could calculate `(Q' * K' + B') * V` but then I would not gain much from using FAVOR+. But if I calculate `Q' * (K' * V) + B' * V` then at least the first term can be calculated with FAVOR+ and the second term is much smaller/faster to calculate even if it's quadratic complexity. `B'` can be initialized with `B` and finetuned.\r\n\r\nthat's an interesting idea! I will try to add T5 to the repo I set up so we can experiment with that; Currently for some reason only distilbert converges, while bert doesn't (https://colab.research.google.com/drive/1o8ioYUIvvIol7PXrDguftQtHyCpyDSJL#scrollTo=9pom5i196Bwg), so i need to figure that out first;; if anybody got bert to work let me know!", "I refactored a lot of code & now works like a charm for me \r\nHere's a large notebook with TF & Torch comparisons for Perf/Trans on BERT, DBERT, T5 on short sequences (SST2 Dataset): https://colab.research.google.com/drive/1A9reiUZbA7DELuJ8keTo73sIQ4dJJVoT?usp=sharing\r\n\r\nTo add a new model all one needs to do is: \r\n- Copy over the config, tf/torch modeling file of the model you want favor attention for to its own folder [here](https://github.com/Muennighoff/transformers/tree/master/src/transformers/models/performer)\r\n- Add a modelXPerformer Config (should be about the same as the current modelXperformer configs)\r\n- Init favor attention directly in the self attention module (i.e. one level lower than in the implementation of this PR) - this is preferrable as it scales better to models with different attention modules / linear layers\r\n- feed q,k,v,mask after their linear layers through the favor attention to get back the final attention output\r\n- Remove all the softmax business😃 \r\n- Last thing to adapt is the (extended) attention mask -- We want the shape to be (bs, 1, seq_len, 1) instead of (bs, 1, 1, seq_len) & we don't want to fill it with the -infs, i.e. just leave it as 1's & 0's, as we multiply it not add it\r\n- If you want to import it with `from transformers import ModelXPerformer` like the current performer models, rename the models & add them to the `__init__.py` in the performer folder & transformer parent folder\r\n\r\nReg. T5: \r\n- Based on @marton-avrios proposal, I added T5 - it got a bit more complex due to the attn mask so I compute:\r\n\t- `Q' @ ((K' * M) @ V) + (B * M) @ V` & it converges🎉\r\n- However, B is a matrix of (bs, n_heads, L, L) where L is the seq len so it scales quadratically with seq len, the exact problem performers try to solve ;_; \r\n- Removing the Rel Pos Encoding entirely surprisingly has about the same performance and is much faster (i got a 30% speedup for 1000 seq_len) - Still need to test it for EncDec model; Another option is just using bert's abs. pos. embeddings.; The pretrained model shouldnt be affected much \r\n\r\nDecoders:\r\n- Only worked with T5Encoder so far; The code for doing causal perf. attention should be there, so adding EncDec / GPT-2 like models should be pretty simple; If somebody wants to try it let me know!; We just need to be careful with the attn mask when we implement it\r\n\r\nA couple pointers if you dont get the desired performance:\r\n- The approximation error propagates through the layers, i.e. the more layers the worse it may get (A 6-layer bert performer gives me about as good a performance as a 12-layer one)\r\n- Try increasing the random features & the feature drawing interval - The more random features the better the softmax approximation, though it also gets more expensive\r\n- Make sure the masking is correct!", "I'm sorry I haven't responded to mentions on this PR recently— I've been quite busy with an unrelated project.\r\n\r\nThank you @Muennighoff for all your hard work extending/refining the PR! I just merged your changes.", "wow, great work @Muennighoff !\r\n\r\nRegarding T5:\r\n - despite still being quadratic complexity have you measured any speed/memory improvements compared to vanilla T5? Or significantly worse (better?) performance? In vanilla T5 there are 2 computations of quadratic complexity: `QK` and `B` but the calculation of `QK` plays a much bigger role in the overall speed of T5. Also it is calculated (and stored) in every layer while `B` is only calculated (and stored) in the first layer.\r\n - when you mention 30% speedup and same performance by removing relative positional attention is it a PerformerT5 compared to a PerformerT5 without it or a vanilla T5 compared to a vanilla T5 without it? Because if it is a PerformerT5 comparison then I think it means that it cannot learn meaningful weights for `B` anyway.", "> wow, great work @Muennighoff !\r\n> \r\n> Regarding T5:\r\n> \r\n> * despite still being quadratic complexity have you measured any speed/memory improvements compared to vanilla T5? Or significantly worse (better?) performance? In vanilla T5 there are 2 computations of quadratic complexity: `QK` and `B` but the calculation of `QK` plays a much bigger role in the overall speed of T5. Also it is calculated (and stored) in every layer while `B` is only calculated (and stored) in the first layer.\r\n> * when you mention 30% speedup and same performance by removing relative positional attention is it a PerformerT5 compared to a PerformerT5 without it or a vanilla T5 compared to a vanilla T5 without it? Because if it is a PerformerT5 comparison then I think it means that it cannot learn meaningful weights for `B` anyway.\r\n\r\nCheck out the T5 Tensorflow experiments here: https://colab.research.google.com/drive/1A9reiUZbA7DELuJ8keTo73sIQ4dJJVoT?usp=sharing\r\nFor each configuration (performer/transformer /// raw/pretrained) i ran it w/ & w/o pos bias, but only on short seq task of sst-2. \r\n\r\nFor the 1000 seq len task i mentioned, i ran only performer encoders (w/ & w/o pos bias) on byte-text level classification from the LRA paper & they had the same performance within +- 1% accuracy.\r\n\r\ni'm not yet sure what to make of it; I think we could confirm that they are of no use after training a full t5 enc-dec model in performer mode & benchmarking that", "I realized a mistake in my formulation which would explain why PerformerT5 could not make use of `B'`.\r\n\r\nVanilla T5 calculates this: `inverse(D) * exp(Q * t(K) + B) * V`\r\n- ...which is equivalent to `softmax(Q * t(K) + B) * V`\r\n- ...where `D = diag(exp(Q * t(K) + B) * 1L)`, `t()` is the transpose function and `1L` is the all 1 vector of length L.\r\n\r\nI propose to calculate this: `inverse(D) * (exp(Q * t(K)) + B') * V`\r\n- ...which is equivalent to `inverse(D) * exp(Q * t(K)) * V + inverse(D) * B' * V`\r\n- ...where `D = diag(exp(Q * t(K)) * 1L + B' * 1L)`\r\n- ...and when finetuning `B'` should not be initialized from `B` but randomly instead.\r\n\r\nI propose to approximate it with: `inverse(D') * Q' * t(K') * V + inverse(D') * B' * V`\r\n- ...where `D' = diag(Q' * t(K') * 1L + B' * 1L)`.\r\n\r\nMy previous (incorrect) approximation was: `inverse(D') * Q' * t(K') * V + B' * V`\r\n- ...which approximates `inverse(D) * exp(Q * t(K)) * V + B' * V`\r\n- ...and NOT `inverse(D) * exp(Q * t(K)) * V + inverse(D) * B' * V`.", "> * inverse\r\n\r\n\r\n\r\n> I realized a mistake in my formulation which would explain why PerformerT5 could not make use of `B'`.\r\n> \r\n> Vanilla T5 calculates this: `inverse(D) * exp(Q * t(K) + B) * V`\r\n> \r\n> * ...which is equivalent to `softmax(Q * t(K) + B) * V`\r\n> * ...where `D = diag(exp(Q * t(K) + B) * 1L)`, `t()` is the transpose function and `1L` is the all 1 vector of length L.\r\n> \r\n> I propose to calculate this: `inverse(D) * (exp(Q * t(K)) + B') * V`\r\n> \r\n> * ...which is equivalent to `inverse(D) * exp(Q * t(K)) * V + inverse(D) * B' * V`\r\n> * ...where `D = diag(exp(Q * t(K)) * 1L + B' * 1L)`\r\n> * ...and when finetuning `B'` should not be initialized from `B` but randomly instead.\r\n> \r\n> I propose to approximate it with: `inverse(D') * Q' * t(K') * V + inverse(D') * B' * V`\r\n> \r\n> * ...where `D' = diag(Q' * t(K') * 1L + B' * 1L)`.\r\n> \r\n> My previous (incorrect) approximation was: `inverse(D') * Q' * t(K') * V + B' * V`\r\n> \r\n> * ...which approximates `inverse(D) * exp(Q * t(K)) * V + B' * V`\r\n> * ...and NOT `inverse(D) * exp(Q * t(K)) * V + inverse(D) * B' * V`.\r\n\r\nYeah you're right the previous approximation wasn't correct; I also forgot to include it in D when doing the code experiments;\r\nWe could try the approximation you propose. \r\n\r\nAnother angle could be:\r\nSince \r\n`exp(Q @ t(K) + B) = exp(Q @ t(K)) * exp(B)`\r\nand \r\n`exp(Q @ t(K)) ~ ϕ(Q) @ t(ϕ(K))` \r\ni think we can do\r\n`exp(Q @ t(K) + B) ~ ϕ(Q) @ t(ϕ(K)) * exp(B) `\r\nbut rearranging \r\n`(ϕ(Q) @ t(ϕ(K)) * exp(B)) @ V `\r\nto avoid calculating Q @ K first is a pain ", "Is this close? My teammates and I want to use performers in T5", "I was just looking through the code and this is stuff of legends! Great work.\r\n\r\nIn the T5 implementation, I noticed that performer attention forward method is called with position bias, yet it is not currently a valid parameter. Is that residual from the conversation about the above position bias conversations?\r\n\r\nEDIT: Ignore the above, I was looking at the wrong implementation of `PerformerAttention`", "> I was just looking through the code and this is stuff of legends! Great work.\r\n> \r\n> In the T5 implementation, I noticed that performer attention forward method is called with position bias, yet it is not currently a valid parameter. Is that residual from the conversation about the above position bias conversations?\r\n> \r\n> EDIT: Ignore the above, I was looking at the wrong implementation of `PerformerAttention`\r\n\r\nI removed the position bias temporarily, as not using it at all worked best. I havn't tried @marton-avrios most recent idea though, so perhaps somebody might want to try it and report back. \r\n\r\nIf you only need an Encoder T5, you should be able to work with what's there -- For Encoder-Decoder, The causal decoder is currently still prohibitively expensive due to the for loop & cumsum operation (@mymusise and me are working on it [here](https://github.com/mymusise/gpt2-quickly/issues/5)). Let us know if you get the decoder to perform! ", "I'm thinking about the position bias, and it doesn't seem like there's a good way to include it. What's been mentioned above seems correct, that the mathematical starting point is \r\n\r\n(Q'K'^T * B')V, where B' = e^B (elementwise)\r\n\r\nBut, this can't be computed without computing Q'K'^T first, defeating the purpose.\r\n\r\nThe alternative is to add some position encoding into each of Q' and K' (a la 'Attention Is All You Need'). I think this is the only / best way to do position bias in this context. That said, it would be getting kind of wonky / outside the spirit of the performers paper, so I'm not sure position bias should even be allowed in this PR.\r\n\r\nDo you all agree with this?", "Hi, I have been trying to run finetuning with `T5PerformerForConditionalGeneration` using this pull request branch, and I have got few minor issues or questions I wanted to ask about.\r\n\r\n1. Merge conflict comments were still left under `/src/transformers/__init__.py`, which is not a serious issue.\r\n2. After getting the attention output from `PerformerAttention`, I had to add `unshape` call it to concat the head attentions and multiply to `W0` in `forward()` of `T5Attention`. I found original call to unshape was commented out since it included matmul of `V`.\r\n3. Both in encoder and decoder, I was getting matrix multiplication exception by wrong dimension on the line when multiplying(in `PerformerAttention`) `mask` to `k_prime`. Was this the reason why @norabelrose mentioned T5 Decoders is not fully working yet? I am trying to fix this attention mask issue for the decoder, but for encoder case, is transposing the attention mask the right way to fix?\r\n> Decoders:\r\n> \r\n> Only worked with T5Encoder so far; The code for doing causal perf. attention should be there, so adding EncDec / GPT-2 like models should be pretty simple; If somebody wants to try it let me know!; We just need to be careful with the attn mask when we implement it", "> Hi, I have been trying to run finetuning with `T5PerformerForConditionalGeneration` using this pull request branch, and I have got few minor issues or questions I wanted to ask about.\r\n> \r\n> 1. Merge conflict comments were still left under `/src/transformers/__init__.py`, which is not a serious issue.\r\n> 2. After getting the attention output from `PerformerAttention`, I had to add `unshape` call it to concat the head attentions and multiply to `W0` in `forward()` of `T5Attention`. I found original call to unshape was commented out since it included matmul of `V`.\r\n> 3. Both in encoder and decoder, I was getting matrix multiplication exception by wrong dimension on the line when multiplying(in `PerformerAttention`) `mask` to `k_prime`. Was this the reason why @norabelrose mentioned T5 Decoders is not fully working yet? I am trying to fix this attention mask issue for the decoder, but for encoder case, is transposing the attention mask the right way to fix?\r\n> \r\n> > Decoders:\r\n> > Only worked with T5Encoder so far; The code for doing causal perf. attention should be there, so adding EncDec / GPT-2 like models should be pretty simple; If somebody wants to try it let me know!; We just need to be careful with the attn mask when we implement it\r\n\r\nI think 1,2 & 3 are all fixed here: https://github.com/Muennighoff/transformers ; \r\nThe masking for the decoder however may not yet work\r\n" ]

[]

[]

[ "Hey @asigalov61,\r\n\r\nI think applying `Transformers` to Music is a super cool idea! Regarding the best model to use for music composition, IMO it depends strongly on:\r\n\r\n- What is the input to the model? Do you input tokens or float vectors?\r\n- How long is the input? *e.g.* how many float vectors or tokens? GPT2 is limited to 1024 tokens / float vectors -> is this too short? \r\n- For generation (composition), I think only our `autoregressive models` make sense: https://huggingface.co/transformers/model_summary.html#autoregressive-models so mostly GPT2. For \"classification\" it would mostly be BERT. If you need very long inputs, it would be interesting to check-out ReformerLM: https://huggingface.co/google/reformer-enwik8\r\n\r\nI bet people would be very interested in Transformers + Music. We've created a new examples folder structure for such projects, so feel free to open a PR to add a dir \"music_transformers\" here: https://github.com/huggingface/transformers/tree/master/examples/research_projects", "Hey Patrick,\r\n\r\nThank you for your help/guidance and for the welcome 🙂\r\n\r\nI think I have created the proper PR for the new dir as you have suggested. Please check it and let me know if it is ok. I am new to PRs so it's still kinda difficult to do it right sometimes 🙂\r\n\r\nWhat can I add there? Can I add my GPT2 implementation there? I should put it in a separate dirs there? Right?\r\n\r\nRegarding your questions for me:\r\n\r\n1) I am sorta working with what is available so atm I just use existing implementations. So I usually use implementation's way of feeding the model. I.e. for my GPT2 implementation, I use the minGPT char-based approach which is painfully slow and inefficient. minGPT does not have BPE yet so I can't really improve it and do it properly as it is very complex for me and difficult. This is why I was very interested in your work cuz you guys provide a standardized and easy way to do it.\r\n\r\nSo basically in my GPT2 implementation I simply feed it the text char tokens. I have attached the example of input/output in my original post. Check it out if you can, please.\r\n\r\n2) I figured that GPT2 is most capable (OpenAI did the same thing with MuseNet). So I was wondering if you guys have a nice GPT2 version that is tuned to the limit. This would really help.\r\nAlso I need better tokenizer but I do not know how to do it. So if you can help/give me specific pointers, I will really appreciate it.\r\n\r\n3) I most certainly heard about the Reformer. And it would be super cool to try it. But again, I have no idea how to make it compatible with the text input/text tokens I use. So if you can help, this also will be much appreciated.\r\n\r\nAgain, thank you for your advice.\r\n\r\nMost sincerely,\r\n\r\nAlex\r\n\r\n\r\n________________________________\r\nFrom: Patrick von Platen <[email protected]>\r\nSent: Monday, December 28, 2020 4:35 AM\r\nTo: huggingface/transformers <[email protected]>\r\nCc: Alex <[email protected]>; Mention <[email protected]>\r\nSubject: Re: [huggingface/transformers] Music Transformers (#9324)\r\n\r\n\r\nHey @asigalov61<https://github.com/asigalov61>,\r\n\r\nI think applying Transformers to Music is a super cool idea! Regarding the best model to use for music composition, IMO it depends strongly on:\r\n\r\n * What is the input to the model? Do you input tokens or float vectors?\r\n * How long is the input? e.g. how many float vectors or tokens? GPT2 is limited to 1024 tokens / float vectors -> is this too short?\r\n * For generation (composition), I think only our autoregressive models make sense: https://huggingface.co/transformers/model_summary.html#autoregressive-models so mostly GPT2. For \"classification\" it would mostly be BERT. If you need very long inputs, it would be interesting to check-out ReformerLM: https://huggingface.co/google/reformer-enwik8\r\n\r\nI bet people would be very interested in Transformers + Music. We've created a new examples folder structure for such projects, so feel free to open a PR to add a dir \"music_transformers\" here: https://github.com/huggingface/transformers/tree/master/examples/research_projects\r\n\r\n—\r\nYou are receiving this because you were mentioned.\r\nReply to this email directly, view it on GitHub<https://github.com/huggingface/transformers/issues/9324#issuecomment-751698309>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ANNXLI3QSD2OFJT5JQ4FTGLSXB3PVANCNFSM4VLQPGAA>.\r\n", "Hello! I’d be willing to contribute work in this space if anyone would like to collaborate. In my previous life I was a professional audio engineer, now I’m an enterprise AI systems architect. https://www.paulprae.com/", "@praeducer Hey Paul!\r\n\r\nThank you for responding to this thread. I would love to collab and create something based on hugginface implementations so if you can help, I would really appreciate it.\r\n\r\nBasically, huggingface docs are very convoluted and unclear to me atm so if you can create a working collab with GPT2 hugginface implementation, I can take it from there and add music parts to it.\r\n\r\nI need something similar to my own GPT2 implementation but based on huggingface so that we can add it here and contribute to their repo/library.\r\n\r\nThis is what I have and this is what I need:\r\nhttps://github.com/asigalov61/Optimus-VIRTUOSO\r\n\r\nAnd my attempt to use huggingface implementation is posted above in the thread so check it out also.\r\n\r\nThanks a lot. Looking forward to working together with like-minded people.\r\n\r\nAlex.\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[ { "id": 2627272588, "node_id": "MDU6TGFiZWwyNjI3MjcyNTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Model%20Parallel", "name": "Model Parallel", "color": "8B66A5", "default": false, "description": "Model Parallelilsm Implementations" }, { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "I don't have an in-depth knowledge of our model parallelism features, so it would be great if @LysandreJik can take a look here as well. \r\n\r\nI think in general, I'm in favor of this PR. However, I'm not sure if a function decorator is better than just having two lines of \r\n```\r\nif self.is_parallel: \r\n # call map to device function\r\n```\r\n\r\nin the respective forward function. We've decided against using function decorators in Pytorch at multiple points (gradient checkpointing e.g.), so I'm not convinced it's the better option to do it here. Function decorators do reduce code readability quite a lot IMO.", "I'm not sure how your suggestion would work since it needs to be generic, and once inside `forward` the function args are no longer generic. Remember, I'm trying to build a generic functionality that can work out of the box in any `transformers` model and not specific to t5.\r\n\r\nThe other approach that doesn't need a decorator is to override `self.__call__` via `self.parallelize` to set to a variation of this wrapper.\r\n\r\n```\r\n def parallelize(self, device_map=None):\r\n $self.__call__ = model_parallel__call__\r\n [...]\r\n\r\n def deparallelize(self):\r\n $self.__call__ = nn.Module.__call__\r\n [...]\r\n```\r\nand:\r\n```\r\ndef model_parallel__call__(self, *input, **kwargs):\r\n \r\n # get device of any of the params of this layer\r\n try:\r\n device = next(self.parameters(recurse=True)).device\r\n except StopIteration:\r\n device = None\r\n\r\n if device is not None:\r\n\r\n input = list(input)\r\n for i, v in enumerate(input):\r\n if v is not None:\r\n input[i] = v.to(device)\r\n input = tuple(input)\r\n\r\n for k in kwargs.keys():\r\n if kwargs[k] is not None and torch.is_tensor(kwargs[k]):\r\n kwargs[k] = kwargs[k].to(device)\r\n\r\n return nn.Module.__call__(self, *input, **kwargs)\r\n```\r\n(or could save the original `self.__call__` to be more flexible and to allow for others to override this too)\r\n\r\nthis in fact is even better since it will have 0 impact on non-MP functionality as this wrapper will be called only under MP.", "This is great progress, @stas00! From my perspective, to create a general way of doing model parallelism, we need four things:\r\n* a format for `device_map` that can be used on any model\r\n* `device_map` and `model_parallel` need to be attributes on all models, probably by assigning them to `PreTrainedModel`\r\n* `parallelize()` and `deparallelize()` should be on all models, again probably by assigning them to `PreTrainedModel`\r\n* changes to the forward methods need to be abstracted if at all possible (this is by far the most challenging)\r\n\r\nThis PR makes a lot of progress, the strongest of which is a potential abstraction/simplification of the changes to the forward method. Not sure if a decorator is the solution. @LysandreJik will have that insight when he's back. I like the suggestion by @patrickvonplaten that it's instead a two line implementation `if self.model_parallel` instead of a decorator. But the BIG thing is if most or all of the code in the forward method can be replaced with with something like:\r\n```\r\nif self.model_parallel:\r\n hidden_states, attention_mask, position_bias, encoder_hidden_states, encoder_attention_mask, encoder_decoder_position_bias, head_mask, past_key_value = _call__mp(hidden_states, attention_mask, position_bias, encoder_hidden_states, encoder_attention_mask, encoder_decoder_position_bias, head_mask, past_key_value)\r\n```\r\nIf we can get that right, it might turn model parallelism from a day or weekend project per model into something that takes a few minutes. Much more scalable and sustainable. \r\n\r\nSupporting non-sequential GPUs could be more trouble than its worth -- not entirely sure on this, it's just my instincts. With the billion + parameter models that we're dealing with -- and all indications are that it's only getting bigger going forward -- it's pretty fair to say that most workflows in enterprise and research will be: \r\n\r\n1. develop locally on a machine with one or maybe two GPUs on a small sized version of a model, and \r\n\r\n2. train a final model on a cloud instance or cluster with multiple identical GPUs. \r\n\r\nSequential hand-offs between GPUs will be the norm in cases like that, which I think are going to be most of them. \r\n\r\nThe other thing I worry about is a challenge with PyTorch 1.5 and 1.6 model parallelism behavior. The seemingly redundant clauses and `set_device` statements are there to prevent PyTorch's inferential logic from moving modules or inputs around after `.to()` assignments have been called. It's very annoying. I don't know if it's fixed in 1.7. You'll notice that output layers like the `lm_head` are always on the first device instead of the last device. A more logical workflow would have the embedding layers on the first device and the output layers on the last device. \r\n\r\nI got that to work just fine in forward passes, but I must've tried 10 different ways to get it to behave in backprop before conceding that for whatever reason PyTorch's quantum device superposition just wouldn't allow it. So the output layers are on the same device as the embedding layers. You'd think that matters for load balance between GPUs, and it does -- for gpt2-xl. But since we're practically limited in most situations to PyTorch's 8 GPU per machine preference (inherited from CUDA), by the time you're at 3 billion parameters the embedding and `lm_head` layers are so small in comparison to the attention blocks that it doesn't matter that they're both on the first device, and a custom `device_map` solves the problem for cases where that matters. The implementation implies that there is an extra hand-off or two of a large tensor between GPUs, but I don't think having a perfectly optimal setup will save even 10% on training time. Happy to be proven wrong on this though. What will save a TON of time and $$$ though is deepspeed integration.\r\n\r\nI got t5-11b with 1024 tokens to train quickly on the new p4 instance AWS released last month with its epic 320 GB of GPU memory so I was like \"ok fine whatever... that's pretty good\". ", "That's awesome, @alexorona. Do continue to share your insights from the frontier!\r\n\r\nLet's wait for @LysandreJik to come back to plan ahead and meanwhile I will experiment with Bart.\r\n\r\n> The seemingly redundant clauses and set_device statements are there to prevent PyTorch's inferential logic from moving modules or inputs around after .to() assignments have been called. It's very annoying. I don't know if it's fixed in 1.7. \r\n\r\nOh, so glad you flagged that. Would it be enough to run the existing parallel tests with pt-1.5, pt-1.6 to detect these failures?\r\n\r\nI'm developing on pt-nightly since rtx-30* work only there (well 1.7.1 should be usable too, but mainly waiting for cuda-11.2 support in pytorch, which is again pt-nightly - won't be in 1.7.x). But it means I can't use it with older pt versions.\r\n\r\nBut since we have to support pt-1.4, I will then put `set_device` back as you had them originally. But this time let's add specific comments why there are there, otherwise someone like myself will think they are some left-overs from earlier experimentation and swipe them away.\r\n\r\nActually, I think we should have a design document where we explain why this or that is done. Rather than make a lot of noise in the model files. A developer-oriented doc. \r\n\r\nThe `set_device` was just one thing, right? Or have I naively nuked any other essentials?\r\n\r\nThanks again!", "@alexorona, one more question. If pt-1.7+ removes the need for jumping through hoops, as you're suggesting older versions have all kinds of issues, perhaps it'd be a reasonable approach to make MP in `transformers` require pt-1.7?\r\n\r\nIf and when you get some time could you please test if what wasn't working in pt < 1.7 works in pt-1.7? And if not - perhaps we need to file some Issues with pytorch if there are bugs to be solved.\r\n\r\nThank you.", "@stas00 Will try to do so, but in the middle of moving so I don't think I'll get to this until the end of January at the soonest. The team would have to make the call about only support model parallelism for PyTorch >= 1.7.0 if it won't work on earlier versions. I would be very tempted to support that idea, but don't have enough usage information to know what the impact would be.", "I guess once everybody is back next week we can start having some discussion with the HF team.\r\n\r\nHave an easy move!", "@stas00 Yeah, should be able to get some input when everyone is back. In the meantime, I'm still not sure on the final form of the `device_map`. There are two issues left to work out: \r\n\r\n1. Some models don't have decoder architectures\r\n2. No ability to map embeddings and output layers (always on first device), which _might_ be just fine. I think most output layers and embeddings are going to be comparatively small to attention blocks going forward, but we should confirm that. We are allowing people to create custom a `device_map` that should enable them to get around any potential situations where the first device is becoming overloaded.\r\n\r\nTo confirm, this looks good for decoder architectures:\r\n```\r\ndevice_map = {\r\n\tencoder: {\r\n\t\t\t0: [0, 1, 2, 3, 4, 5],\r\n\t\t\t1: [6, 7, 8, 9, 10, 11]\r\n\t\t\t},\r\n\tdecoder: {\r\n\t\t\t2: [0, 1, 2],\r\n\t\t\t3: [3, 4, 5]\r\n\t\t\t}\r\n}\r\n```\r\n\r\nMaybe we use the keys to map to the attribute? In gpt2, `self.h` contains the attention blocks, so:\r\n```\r\ndevice_map = {\r\n\th: {\r\n\t\t0: [0, 1, 2, 3, 4, 5],\r\n\t\t1: [6, 7, 8, 9, 10, 11]\r\n\t\t}\r\n}\r\n```\r\nIn trying to generalize `parallelize()`, we still need access to list of all modules. For example, in `GPT2LMHeadModel`, we would need to know: `self.lm_head`, `self.transformer.h`, `self.transformer.wte`, `self.transformer.wpe` and `self.transformer.ln_f`. ", "I haven't looked into gpt2, yet. t5 and bart are very similar structure-wise. We probably need to map out all the different archs `transformers` has and then generalize.\r\n\r\nWhat is emerging so far is that the device map might have various keys, none required, and each model architecture will have:\r\n\r\n1. its required keys\r\n2. its own default map generator - so that the user doesn't have to provide one and overtime it can be improved to have smarts to create a balanced map based on the \"insider\" information.\r\n\r\nSo if some architectures need to explicitly manage the mapping of non-block/layers, rather than just assigning them by default on the \"`main_device`\", because they are significantly big, they could do that too. Otherwise, leave the `main_device` to all the \"smallish-fish\" and use the other devices for \"the big fish\" if that makes sense. The main advantage of this \"lazy\" approach is that there is less device-hopping and less code needed to match the hopping.", "Yes, that's right. \r\n\r\nSo it turns out the `self._modules` attribute has all of the modules. To move `parallelize()` to `PreTrainedModel`, I think all we need is a per-model `module_map` object to map between the `device_map` and the model placements. With a little work, we might be able to reduce making a model parallel to:\r\n\r\n1. Adding a few lines of code in the forward method per your work\r\n2. Modifying the validation function to check for errors in a custom `device_map`\r\n3. Creating a `module_map` dictionary for that model and adding it to the `get_module_map()` function \r\n\r\nWe can embed special placement rules where non-attention block modules need to be on the same device as another module by creating a tuple in `module_map['dependent_modules']`:\r\n\r\n```\r\n# Device map for GPT2LMHead. T5 would have 'encoder', 'decoder' as keys instead of 'h' and validate_device_map would\r\n# check to see if the device_map has the right keys.\r\ndevice_map = {\r\n\t'h': {\r\n\t\t\t0: [0, 1, 3, 4],\r\n\t\t\t1: [5, 6, 7, 8],\r\n\t\t\t2: [9, 10, 11, 12]\r\n\t}\r\n}\r\n\r\n\r\nclass PreTrainedModel():\r\n...\r\n\t# Probably use get_model_map(), but just to make it simple:\r\n\tself.module_map = {\r\n\t\t'h': self._modules['transformer'].h,\r\n\t\t'embeddings': [\r\n\t\t\t\t\tself._modules['transformer'].wte,\r\n\t\t\t\t\tself._modules['transformer'].wpe\r\n\t\t\t\t\t],\r\n\t\t'dependent_modules': [\r\n\t\t\t\t\t(\r\n\t\t\t\t\t\tself._modules['transformer'].ln_f], \r\n\t\t\t\t\t\tmodel._modules['transformer'].h[-1],\r\n\t\t\t\t\t),\r\n\t\t\t\t\t(\r\n\t\t\t\t\t\tself._modules['lm_head'],\r\n\t\t\t\t\t\tself._modules['transformer'].wte\r\n\t\t\t\t\t)\r\n\t\t\t\t]\r\n\t}\r\n\r\ndef parallelize(self, device_map = None):\r\n\t\r\n\tself.device_map = device_map\r\n\r\n\t# validate_device_map extended to check for valid keys for model\r\n\t\r\n\t...\r\n\r\n\t# Set all embeddings to first device\r\n\tif 'embeddings' in self.module_map:\r\n\t\tfor layer in self.module_map['embeddings'].items():\r\n\t\t\tlayer.to(self.first_device)\r\n\r\n\r\n\t# Assign attention blocks to the appropriate device.\r\n\tfor module_group, group_map in self.device_map.items():\r\n\t for device, layers in group_map.items():\r\n\t for layer in layers:\r\n\t self.module_map[module_group][layer].block_parallelize(f\"cuda:{device}\")\r\n\r\n\r\n\t# Some modules should always be on the same device as another module. We can express \r\n\t# this as a tuple pair where tuple[0] needs to be on tuple[1]\r\n\tif 'dependent_modules' in self.module_map:\r\n\t\t for i in self.module_map['dependent_modules']:\r\n\t\t \ti[0].to(i[1].device)\r\n```\r\n", "All, awesome suggestions that should be looked at next once the current work has been merged.\r\n\r\nI'm going to wait implementing anything new, since there are already too many partial PRs that need to be carefully merged and rebased and once that is done we can do another round of generalization integrating your suggestions.", "So if there is no objection, I will merge this one, and then start integrating with https://github.com/huggingface/transformers/pull/9384, which is ahead functionality-wise - so I want to sync the two, switching t5 to the improved version of MP backend. I will implement the suggestions in that new PR.", "As we have discovered the original PR didn't make t5 work with trainer. I have just fixed that in the last commit here, bringing some goodies over from the Bart MP PR.\r\n\r\nSo this now works:\r\n\r\n```\r\nexport BS=20; rm -r output_dir; CUDA_VISIBLE_DEVICES=0,1 PYTHONPATH=../../src USE_TF=0 ./finetune_trainer.py --model_name_or_path t5-small --output_dir output_dir --adam_eps 1e-06 --data_dir wmt_en_ro --do_eval --do_predict --do_train --evaluation_strategy=steps --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_eval_batch_size $BS --per_device_train_batch_size $BS --predict_with_generate --eval_steps 25000 --sortish_sampler --task translation_en_to_ro --test_max_target_length 128 --val_max_target_length 128 --warmup_steps 500 --n_train 200 --n_val 200 --n_test 200 --fp16 --save_steps 2 --model_parallel\r\n```\r\n\r\nBut! while it's fine in the training stage, it's 10x slower on eval than w/o `--model_parallel`", "Hello @stas00 kudos to all the hard work you do, especially around continuing the ambitious work around supporting parallelism. \r\n\r\nInterested in doing some inference with the t5-11b model variant.\r\nCan you provide some insights on how many gpus would be needed to achieve that?\r\n\r\nI tried this branch with 8xV100 (16gb) on GCE. \r\nAll good while I created the model and called parallelize, but got a out of memory error on inference step when moving inputs to the first gpu device.\r\n\r\nLet me know if I have a wrong mental model about achieving this. Thanks again!\r\n\r\n\r\n", "Thank you for the kind words, @kznmft!\r\n\r\nPlease have a look at https://github.com/huggingface/transformers/pull/9765\r\nwhich implements a very inefficient in my opinion but nevertheless working pipeline parallelism on t5, which should be superior to this naive implementation, speed-wise but it's not quite there yet. Please read the first post carefully for all the details. and you can see the follow up comments with the experiments that have been done. So 4x40gb A100s gpus weren't enough for t5-11b in initial experiments. But 5-6 of those probably should be enough.\r\n\r\nI finally got access to a machine with 4 gpus just now, so I'm going to start looking at implementing 2D parallelism - using Pipeline with DeepSpeed ZeRO-DP, so I will post news once I get something working.\r\n\r\nSubscribe to watch https://github.com/huggingface/transformers/pull/9765 and I most likely will update that PR with new info or a link to a new PR once I have something working.\r\n\r\n-----\r\n\r\n> I tried this branch with 8xV100 (16gb) on GCE.\r\n> All good while I created the model and called parallelize, but got a out of memory error on inference step when moving inputs to the first gpu device.\r\n\r\nBut you're not telling me the device map you were using. You need to spread out the layers over the 8 gpus, have you done it? unless you were relying on the default map which should spread things out.\r\n\r\nThe problem is that it doesn't take into an account that gpu 0 is always overtaxed, so I'd always try a few layers less on the first gpu 0. And then watch nvidia-smi (and later we will have better tools) to see that you get each GPU getting a somewhat equal memory allocation.\r\n\r\nBut if 4x40 couldn't fit it, I doubt that 8x16 will.\r\n\r\nRemember in t5-11b you have 45GB of params, plus optimizer states plus gradients.\r\n\r\nAlso probably need to try to use a more lean optimizer, say Adam instead of AdamW which needs more memory.\r\n\r\n\r\n\r\n", "too long. closing." ]

[]

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

[ "Hey @ZOUG,\r\n\r\nThanks for the issue. @LysandreJik is on holiday at the moment, but I'm sure he's more than happy to take a look when he's back :-) ", "Hello! We have started officially maintaining the anaconda packages in version v4.0.0. Installing a version anterior to that one would result in you using the `transformers` version from another channel (such as `conda-forge`), which we do not maintain.\r\n\r\nDo you get the same error when installing `transformers` from our channel (on a more recent version)?", "> Hello! We have started officially maintaining the anaconda packages in version v4.0.0. Installing a version anterior to that one would result in you using the `transformers` version from another channel (such as `conda-forge`), which we do not maintain.\r\n> \r\n> Do you get the same error when installing `transformers` from our channel (on a more recent version)?\r\n\r\nNo, the error does not occur on the most recent version. The problem is that packages dependent on `transformers` may not be compatible with v4.x at the moment so that the error will still arise. It might be better to provide a pip installation package for v3.5.1 that is compatible with the `conda-forge` dependencies.\r\n\r\nIn my case, I got lucky that the package that I need just released a new version today that is compatible with transformers v4.x.", "I believe this is still the case in Docker-based environments (ex. Kaggle). I removed existing transformers and tokenizers, installed new ones (transformers 4.2.1 and tokenizers 0.9.4). \r\nIn the code, it goes back to conda and complains about tokenizers being 0.9.3\r\n```\r\n/opt/conda/lib/python3.7/site-packages/transformers/__init__.py in <module>\r\n 41 \r\n 42 # Check the dependencies satisfy the minimal versions required.\r\n---> 43 from . import dependency_versions_check\r\n 44 from .file_utils import (\r\n 45 _BaseLazyModule,\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/dependency_versions_check.py in <module>\r\n 39 continue # not required, check version only if installed\r\n 40 \r\n---> 41 require_version_core(deps[pkg])\r\n 42 else:\r\n 43 raise ValueError(f\"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py\")\r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/utils/versions.py in require_version_core(requirement)\r\n 92 \"\"\" require_version wrapper which emits a core-specific hint on failure \"\"\"\r\n 93 hint = \"Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git master\"\r\n---> 94 return require_version(requirement, hint)\r\n 95 \r\n 96 \r\n\r\n/opt/conda/lib/python3.7/site-packages/transformers/utils/versions.py in require_version(requirement, hint)\r\n 85 if want_ver is not None and not ops[op](version.parse(got_ver), version.parse(want_ver)):\r\n 86 raise pkg_resources.VersionConflict(\r\n---> 87 f\"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}\"\r\n 88 )\r\n 89 \r\n\r\nVersionConflict: tokenizers==0.9.4 is required for a normal functioning of this module, but found tokenizers==0.9.3.\r\n``` \r\nEdit: found a work around to re-import modules:\r\n```\r\nimport importlib, pkg_resources, tokenizers\r\nimportlib.reload(pkg_resources)\r\nimportlib.reload(tokenizers)\r\n``` \r\ntqdm may also complain if 4.50 or later.", "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.", "Issue still persists\r\n\r\n---------------------------------------------------------------------------\r\nVersionConflict Traceback (most recent call last)\r\n<ipython-input-24-3b738e6ed358> in <module>\r\n----> 1 from transformers import PreTrainedTokenizerFast\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _find_and_load(name, import_)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _find_and_load_unlocked(name, import_)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _load_unlocked(spec)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _load_backward_compatible(spec)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/site-packages/transformers-4.2.2-py3.8.egg/transformers/__init__.py in <module>\r\n 41 \r\n 42 # Check the dependencies satisfy the minimal versions required.\r\n---> 43 from . import dependency_versions_check\r\n 44 from .file_utils import (\r\n 45 _BaseLazyModule,\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _find_and_load(name, import_)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _find_and_load_unlocked(name, import_)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _load_unlocked(spec)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/importlib/_bootstrap.py in _load_backward_compatible(spec)\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/site-packages/transformers-4.2.2-py3.8.egg/transformers/dependency_versions_check.py in <module>\r\n 39 continue # not required, check version only if installed\r\n 40 \r\n---> 41 require_version_core(deps[pkg])\r\n 42 else:\r\n 43 raise ValueError(f\"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py\")\r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/site-packages/transformers-4.2.2-py3.8.egg/transformers/utils/versions.py in require_version_core(requirement)\r\n 92 \"\"\" require_version wrapper which emits a core-specific hint on failure \"\"\"\r\n 93 hint = \"Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git master\"\r\n---> 94 return require_version(requirement, hint)\r\n 95 \r\n 96 \r\n\r\n/app1/anaconda3/envs/praveen_tfu/lib/python3.6/site-packages/transformers-4.2.2-py3.8.egg/transformers/utils/versions.py in require_version(requirement, hint)\r\n 85 if want_ver is not None and not ops[op](version.parse(got_ver), version.parse(want_ver)):\r\n 86 raise pkg_resources.VersionConflict(\r\n---> 87 f\"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}\"\r\n 88 )\r\n 89 \r\n\r\nVersionConflict: tokenizers==0.9.4 is required for a normal functioning of this module, but found tokenizers==0.11.6.\r\nTry: pip install transformers -U or pip install -e '.[dev]' if you're working with git master\r\n" ]

[]

[]

[ "I think this notebook could help you: https://github.com/huggingface/notebooks/blob/master/examples/language_modeling.ipynb you should check out the `def tokenizer(...)`, `def group_texts(...)` functions. I think they should help at what you want to achieve. ", "Thank you for the fast response @patrickvonplaten.\r\nI reviewed your link only to find out it was an input problem on my side that I did not see before. Sorry to bother you for that.\r\nJust in case anyone comes across a similar issue here is the solution I found to be working for me.\r\n\r\n```\r\nclass german_bert_sentiment: \r\n\t\"\"\"\r\n\tSentiment analyzer module based on a range of sources including twitter, facebook, product reviews\r\n\thttps://huggingface.co/oliverguhr/german-sentiment-bert?text = Du+Arsch%21\r\n\t\"\"\"\r\n\r\n\tdef __init__(self, truncate=False):\r\n\t\tself.tokenizer = AutoTokenizer.from_pretrained(\"oliverguhr/german-sentiment-bert\")\r\n\t\tself.model = AutoModelForSequenceClassification.from_pretrained(\"oliverguhr/german-sentiment-bert\")\r\n\t\tself.truncate=truncate\r\n\t\tself.max_length=512\r\n\r\n\tdef analyze(self, text):\r\n\t\taverages=[]\r\n\t\terrors=[]\r\n\t\tinputs = self.tokenizer(text, return_tensors = \"pt\")#, max_length=512, stride=0, return_overflowing_tokens=True, truncation=True, padding=True)\r\n\t\tlength=len(inputs['input_ids'][0])\r\n\t\twhile length>0:\r\n\t\t\tif length>self.max_length:\r\n\t\t\t\tnext_inputs={k: (i[0][self.max_length:]).reshape(1,len(i[0][self.max_length:])) for k, i in inputs.items()}\r\n\t\t\t\tinputs={k: (i[0][:self.max_length]).reshape(1,len(i[0][:self.max_length])) for k, i in inputs.items()}\r\n\t\t\telse:\r\n\t\t\t\tnext_inputs=False\r\n\t\t\tproOrCon = self.model(**inputs)\r\n\t\t\tweights = proOrCon[0].detach().numpy()[0]\r\n\t\t\tweights[2], weights[1] = weights[1], weights[2]\r\n\t\t\tweights = softmax(weights)\r\n\t\t\taverage=np.average(np.linspace(1, -1, 3), weights = weights)\r\n\t\t\taverages.append(average)\r\n\t\t\terrors.append(\r\n\t\t\t\tnp.sqrt(np.average(np.array(np.linspace(1, -1, 3)-average)**2, weights = weights))\r\n\t\t\t\t)\r\n\t\t\t#from IPython import embed; embed()\r\n\t\t\tif self.truncate:\r\n\t\t\t\tbreak\r\n\t\t\tif next_inputs:\r\n\t\t\t\tinputs=next_inputs\r\n\t\t\telse:\r\n\t\t\t\tbreak\r\n\t\t\tlength=len(inputs['input_ids'][0])\r\n\t\taverage = np.average(averages, weights = 1./np.array(errors)**2)\r\n\t\terror = np.sqrt(1./np.sum(1./np.array(errors)**2))\r\n\t\treturn [average, error]\r\n```\r\n " ]

[]

[]

[]

[]

[]

[ "I'm not sure we have a completely fine-tuned SOP classification model. My best advice is to try out different models and see which model not randomly allocates the weights for those layers." ]

[]

[]

[ "That example is unfortunately unmaintained. Have you tried playing around with LXMERT, which is also a multi-modal model? There is a demo available [here](https://github.com/huggingface/transformers/tree/master/examples/research_projects/lxmert).", "Oh, I didn't know there is one with LXMERT. I will try that. Thanks.", "You can make it work with a little inference call modification. \r\nAdd **\"return_dict\": False** to **inputs** dict.\r\nLike this:\r\n```\r\ninputs = {\r\n \"input_ids\": batch[0],\r\n \"input_modal\": batch[2],\r\n \"attention_mask\": batch[1],\r\n \"modal_start_tokens\": batch[3],\r\n \"modal_end_tokens\": batch[4],\r\n \"return_dict\": False\r\n }\r\noutputs = model(**inputs)\r\n```" ]

[]

[]

[ "Issue resolved with adding this line in evaluate function:\r\n\r\n self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, results)\r\n", "reopened, since this is not solved after this line, indeed this looks like a bug, could you have a look please?", "Hello, could you please put all of your environment information as asked in the template, as well as the command you used to launch the script? We need this in order to help you. Thank you.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "In my case, I found that trainer does not do the evaluation event at the end of each epoch even I set the trainer.args eval_stategy = epoch. So maybe we find the log info such as {eval_loss: xxx, eval_acc: xxx, **epoch: 1.98**}. And finally I prefer to use callback which is defined by myself to log metric." ]

[ { "id": 2627272588, "node_id": "MDU6TGFiZWwyNjI3MjcyNTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Model%20Parallel", "name": "Model Parallel", "color": "8B66A5", "default": false, "description": "Model Parallelilsm Implementations" }, { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "Somehow I'm feeling that this approach of having special logic to remap inputs ahead of time is over-complicated. I haven't tried it yet, but won't it be much much simpler to remap inputs once the model layer is visible and just before they are used by that layer - i.e. at the point where one gets:\r\n```\r\nRuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:1 and cuda:0!\r\n```\r\nThen we just do `input_foo = input_foo.to(next(self.parameters()).device)` and we are done? No logic required other than just put the inputs on the same device as this layer.\r\n\r\nWe might be able to even remove all those `if self.model_parallel` in most places in `forward`, and have the same code for w/ or w/o MP. Perhaps with some wrapper that will be noop when not under MP. It could also handle `None` to avoid a gazillion of `if not None` checks. I'd also make it an in-place operation, just like `nn.Module.to` does.\r\n\r\n\r\n**edit** I branched off from this PR and implemented this - works amazingly simple: https://github.com/huggingface/transformers/pull/9323", "I also think #9323 is the way to go here", "too long. closing." ]

[]

[]

[ "cc'ing @gary149 and @beurkinger ", "Hmm, I just discovered I had 2 related issues opened some months back:\r\n- https://github.com/huggingface/transformers/issues/8531\r\n- https://github.com/huggingface/transformers/issues/7206" ]

[]

[]

[ "I updated all 5 models to include fr/ro/de language tags." ]

[]

[]

[ "This bug forced me to disable the corresponding test in the new `TFLed` model for now, see: https://github.com/huggingface/transformers/pull/9278/files#r549042909", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hey @XinyuHua,\r\n\r\nregarding 1) I'm not really sure either. Maybe the author can give a better answer here (hope it's ok to ping you here @ola13) For 2) Yes it should be correct. You can see from the paper that at every generation step `i` the marginal probability over all tokens `z` is calculated (compare to the equation of RAG-Token Model in 2.1). In the equation, we sum over all `z` which corresponds to summing overall `doc_logprobs` above -> so this looks correct to me. Just the fact that the `marginalize` function (your code above) is executed at every forward pass shows that this has to correspond to `RagToken`.", "Thanks for the explanation!\r\n\r\n> the marginal probability over all tokens `z` is calculated\r\n\r\n`z` is the document instead of token, right? And the paper says \"we can draw a different latent document for each\r\ntarget token and marginalize accordingly\", so `p(z|x)` should be a different score for different `y_i`, but the code looks like `p(z|x)` is unchanged for all `y_i`s, whose log form is `doc_logprobs`. So effectively this is how `RagSequence` model is framed. Does that mean the `RagToken` model is actually trained the same way as `RagSequence`, but just the generation is different? \r\n\r\nAnother related question, I checked the pre-trained `RagConfig` for `RagToken`, and the `do_marginalize` is actually set to False, so the marginalize method is never called during forward?", "`z` is the tensor of containing the logprob of all docs -> this is why it's called `doc_logprobs`. If you check out the dimensions of this tensor you should see that one dimension exactly corresponds to `n_docs`. \r\n\r\n`do_marginalize` is called at every forward pass because it's set to `True` in the function argument here: https://github.com/huggingface/transformers/blob/8e74eca7f2b3235f8d5340d66361ea656c67bac7/src/transformers/models/rag/modeling_rag.py#L1099", "Hi @XinyuHua, thanks for the questions! Regarding your first point:\r\n\r\n> The marginalization for RagSequenceForGeneration seems a bit strange. From line 998 to line 1001 (link), only the second tokens in seq_logprobs are getting scored by doc_logprobs\r\n\r\nSince we're operating in the log-space, multiplications from the formulas in section 2.1. of the paper become additions. So instead of multiplying the `p(z|x) * p(y| x,z)`, we sum: `log p(z|x) + log p(y | z,x)` (or `doc_logprobs` + `seq_logprobs` using our variables names from the code), where x is the input sequence, y is the output sequence and z is the retrieved document. Note that in the logspace, `log p(y | z,x)` decomposes into the sum of logprobs of each token in y. In the part of the code you linked we perform this summation - we only want to add `doc_logprobs` once per sequence - we don't need to add it to each token of the sequence.\r\n\r\nNow the reason we add `doc_logprobs` to the second token is that we want to avoid adding it to the BOS token, in case the target sequence doesn't contain one or in case the `exclude_bos_score` argument is used - otherwise we would effectively do no marginalization at all in these cases.\r\n\r\nI hope this helps, but let us know if anything's still unclear!", "Hi @ola13 and @patrickvonplaten , thanks for the detailed explanations! " ]

[]

[]

[ "Here are things you may try (they are unrelated to each other, so you can try in any order that resonates):\r\n\r\n1. turn off `--fp16` or keep it but switch to [pytorch-nightly](https://pytorch.org/get-started/locally/) - there was a large memory leak fixed a few weeks ago related to autocast (fp16) - if this is not related to `autocast`/ftp16 this won't help then. `--fp16` was triggering the leak. Switching to apex amp is another option to try if you're hitting this memory leak in pytorch.\r\n\r\n2. If you are using huggingface trainer (I assume `finetune_trainer.py` is from examples/seq2seq then you're good) and if you can use `transformers` master, I'd suggest using the just added `--sharded_ddp` option. In my few experiments I was able to get 2-3 times bigger batches. It's documented in this PR https://github.com/huggingface/transformers/pull/9208 (we are just waiting for a new fairscale release to merge it). But you can just use it w/o needing to understand if you are short on time. So if you want to try it, install both transformers and [fairscale](https://github.com/facebookresearch/fairscale/) from master and then that new option will be available. \r\n\r\nAnd please edit your Issue to show the command line you use, so we can see what cl args and/or hyper parameters you're using.", "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." ]

[]

[]

[ "Sorry, wrong repo" ]

[]

[]

[ "Hi @yezhengli-Mr9 not sure what you are asking here,\r\nby `Trainer` demo do you mean an example showing how to fine-tune `LXMERT`?\r\n\r\nIf you are looking for how to use `LXMERT` then this [demo notebook](https://colab.research.google.com/drive/18TyuMfZYlgQ_nXo-tr8LCnzUaoX0KS-h?usp=sharing) show's how to use LXMERT for visual QA ", "> Hi @yezhengli-Mr9 not sure what you are asking here,\r\n> by `Trainer` demo do you mean an example showing how to fine-tune `LXMERT`?\r\n> \r\n> If you are looking for how to use `LXMERT` then this [demo notebook](https://colab.research.google.com/drive/18TyuMfZYlgQ_nXo-tr8LCnzUaoX0KS-h?usp=sharing) show's how to use LXMERT for visual QA\r\n\r\nHi @patil-suraj @patrickvonplaten , thanks a lot but [`examples/lxmert/`](https://github.com/huggingface/transformers/blob/master/examples/lxmert/) no longer exists although I am reconstructing some functionality since the [demo notebook](https://colab.research.google.com/drive/18TyuMfZYlgQ_nXo-tr8LCnzUaoX0KS-h?usp=sharing) seems quite instructive." ]

[]

[]

[]

[]

[]

[ "In the documentation, this is written how from_pretrained works for untouched models, but I cannot see how this works when one modifies the model. ", "Hey @juliahane,\r\n\r\ncould you please provide a code snippet showcasing the unintended behavior? Thanks!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[ { "login": "lhoestq", "id": 42851186, "node_id": "MDQ6VXNlcjQyODUxMTg2", "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhoestq", "html_url": "https://github.com/lhoestq", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "repos_url": "https://api.github.com/users/lhoestq/repos", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "type": "User", "site_admin": false } ]

[ "Both comments are right.\r\n\r\n```python\r\nself.assertEqual(doc_dicts[0][\"id\"][0], \"1\") # max inner product is reached with second doc\r\n```\r\nmakes sure the first retrieved document of the first query is the second document of the corpus (it's the one that maximizes the inner product)\r\n\r\nwhile\r\n```python\r\nself.assertEqual(doc_dicts[1][\"id\"][0], \"0\") # max inner product is reached with first doc\r\n```\r\nmakes sure that the first retrieved document of the second query is the first document of the corpus (it's the one that maximizes the inner product)\r\n\r\nTo be clearer the indices in those statements could be written as \r\n```python\r\nquery_idx = 0\r\nretrieved_document_idx = 0\r\nexpected_id = \"0\"\r\nself.assertEqual(doc_dicts[query_idx][\"id\"][retrieved_document_idx], expected_id)\r\n```", "Thanks for your reply!" ]

[]

[]

[ "You're right that the current design is sub-optimal, especially for the tokenizers since we have introduced tokenizers decoupled from models.\r\n\r\n- Having this approach would imply modifying most configuration files on the hubs, given that you use the approach\r\n\r\n ```\r\n (config.class, config.tokenizer_class) -> ...\r\n ``` \r\n as most models configurations have no tokenizer class defined.\r\n\r\n- The `isinstance` should be replaced by `type` imo, which would prevent having such a test\r\n\r\nOverall I'm definitely not against refactoring this part to ensure better compatibility, but let's try to find a way of making sure we don't have to update 1000s of configurations on the hub. Maybe adding a `tokenizer_class = XXXTokenizer` field in the configurations would prevent this. ", "> You're right that the current design is sub-optimal, especially for the tokenizers since we have introduced tokenizers decoupled from models.\r\n> \r\n> * Having this approach would imply modifying most configuration files on the hubs, given that you use the approach\r\n> ```\r\n> (config.class, config.tokenizer_class) -> ...\r\n> ```\r\n> \r\n> \r\n> as most models configurations have no tokenizer class defined.\r\n> * The `isinstance` should be replaced by `type` imo, which would prevent having such a test\r\n> \r\n> Overall I'm definitely not against refactoring this part to ensure better compatibility, but let's try to find a way of making sure we don't have to update 1000s of configurations on the hub. Maybe adding a `tokenizer_class = XXXTokenizer` field in the configurations would prevent this.\r\n\r\nSorry, I think my explanation wasn't very clear above - I modified the description. I didn't mean to force configs to have a `tokenizer_class` attribute. The idea was just that the `TOKENIZER_MAPPING` should expose a function that allows one to get the correct tokenizer not only by the config but also by the tokenizer_class as a string. So the idea is that we could replace the current `TOKENIZER_MAPPING` with a class like as (now) shown above, but then this class can be used in whatever way is best by `AutoTokenizer`, *e.g.* the AutoTokenizer's `from_pretrained(...)` method could then call the `TOKENIZER_MAPPING` class above as follows:\r\n\r\n```python\r\nif hasattr(config, tokenizer_class):\r\n tokenizer = TOKENIZER_MAPPING[(config, config.tokenizer_class)]\r\nelse\r\n tokenizer = TOKENIZER_MAPPING[config]\r\n```\r\n ", "I'd need to see a PoC to be sure, but this looks like an interesting idea to me. There are certainly big limitations in the way those AUTO variables are currently structured.", "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." ]

[]

[]

[ "Sorry, I don't really understand the question here. Could you clarify a bit? In case this is a question about the behavior of `DataCollatorForLanguageModeling`, it would be awesome if you could use the forum: https://discuss.huggingface.co/ .\r\nOtherwise it would be great if you provide a code snippet showcasing the unexpected behavior. Thanks!", "in run_mlm.py\r\n\r\nfirst use:\r\n```python\r\ndef tokenize_function(examples):\r\n # Remove empty lines\r\n examples[\"text\"] = [line for line in examples[\"text\"] if len(line) > 0 and not line.isspace()]\r\n return tokenizer(\r\n examples[\"text\"],\r\n padding=padding,\r\n truncation=True,\r\n max_length=data_args.max_seq_length,\r\n # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it\r\n # receives the `special_tokens_mask`.\r\n return_special_tokens_mask=True,\r\n )\r\n```\r\n\r\nif i set padding=\"max_length\", the inputs will be padded. that means the inputs will already be padded to [1, 4, 5, 6, 2, 0, 0..] and attention_mask will be set to [1, 1, 1, 1, 1, 0, 0...]. so when use DataCollatorForLanguageModeling, the inputs will be padded again (tokenizer.pad). the inputs will not change but the attention_mask will be [1, 1, 1, 1, 1, 1, 1..].\r\nif i set padding=\"false\", tokenizer.pad in DataCollatorForLanguageModeling will not pad the inputs to max_length. ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hey @xieyuchen13 - I don't think the `attention_mask` will change from [1,1,1,...0,0,0] to [1,1,1,....1,1,1] => could you show me an example that proves otherwise? ", "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, @vasudevgupta7 thanks for adding this! \r\nCould you link this notebook in the community notebooks table [here](https://github.com/huggingface/transformers/tree/master/notebooks#community-notebooks) instead of adding it to `/notebooks` ", "done.", "Thanks!\r\n\r\nI re-worded the description a bit, hope you don't mind ;)" ]

[]

[]

[]

[]

[]

[ "Already run of course 😉 and I can tell you that they all pass!" ]

[]

[]

[ "@LysandreJik feel free to merge if it looks ok for you and if @sgugger approves the last fix on `pooled_hidden`." ]

[]

[]

[]

[]

[]

[ "Hey @forest1988,\r\n\r\nThanks a lot for your issue - you're 100% correct, the docs need to be updated here - I'll open a PR for this and tag you.\r\nSo to give some context on why this function doesn't exist anymore.\r\n\r\n- We are trying to align the API of all models which makes it easier for users to switch from one model to the other. No other model had such a function.\r\n- The function was only called internally, so you don't really to care about the change as long as you only call the public API of Bart being at the moment the `forward pass` of `BartModel` and `BartForConditionalGeneration`. All possible behaviors of pubic API functions should have stayed 1-to-1 the same for Bart.\r\n- In the bullet point in question, it says that the model will create the `decoder_input_ids` if they are not passed. This is a very Bart-specific feature and is only really be used in two cases:\r\n 1) You want to do <mask-filling> for bart as shown in the example in this section: https://huggingface.co/transformers/model_doc/bart.html#transformers.BartForConditionalGeneration . In this case you only have to pass the `input_ids` and Bart will correctly output something. All non-Bart seq2seq model yield an error here because they expect the `decoder_input_ids` to be passed as well (which should be the default case IMO). Bart is able to do this task thanks to its rather specific pre-training objective.\r\n2) (This is the same for all Seq2Seq models) you pass `labels` and `input_ids` => this is used fro training and in this case Bart (and all other seq2seq models except EncoderDecoderModel) shift the labels to the right to create the `decoder_input_ids`.", "So in short, the function doesn't exist anymore in the new code. If you adapted \"old\" bart code to your specific needs and are now stuck to \"port\" it to the new Bart code, I'd suggest to write an integration test using your \"old\" bart code and then closely look at what was done in #8900 to adapt your code analogously. If you're completely stuck, feel free to post an issue here and tag me - I'll help you then :-) ", "@patrickvonplaten \r\nThank you for your quick comments and for solving the issue! I've checked PR #9299 and would like to say thank you for tagging me there.\r\n\r\nSome time ago, we tried to use various Seq2SeqLMs but had trouble using them in a unified way. The update of aligning APIs is very helpful!\r\n\r\nI read carefully your comments and linked documents, will write an integration test, and closely look #8900 to adapt my code analogously. \r\nIf I’ll be still completely stuck, I would like to take your word and ask for your help.\r\n\r\nThank you again!" ]

[]

[]

[]

[]

[]

[]

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

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false }, { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ " here's what I found\r\n\r\n`t5-small` is the only T5 model that works in fp16 at the moment. The rest of the models produce `nan` loss/logits.\r\n\r\n for all the models and versions (v1, v1.1, mT5) at some point we get `inf` values in `hidden_states` after applying the final linear layer (`wo`) in `T5DenseReluDense` and `T5DenseGatedGeluDense`.\r\nhttps://github.com/huggingface/transformers/blob/02e05fb0a532e572b56ba75dad6ba3db625bbdeb/src/transformers/models/t5/modeling_t5.py#L248-L278\r\n\r\nwhich results in `nan` values in `T5LayerNorm`.\r\n\r\nAlso for `t5-large`, `t5-v1_1-base`, `t5-v1_1-large`, there are `inf` values in the output of `T5LayerSelfAttention` and `T5LayerCrossAttention`, specifically where we add the attn output with the `hidden_states`\r\n\r\nhttps://github.com/huggingface/transformers/blob/02e05fb0a532e572b56ba75dad6ba3db625bbdeb/src/transformers/models/t5/modeling_t5.py#L548\r\n\r\nhttps://github.com/huggingface/transformers/blob/02e05fb0a532e572b56ba75dad6ba3db625bbdeb/src/transformers/models/t5/modeling_t5.py#L584\r\n\r\nThis happens during both training and inference, to reproduce \r\n\r\n```python\r\nmodel = T5ForConditionalGeneration.from_pretrained(\"t5-base\").to(\"cuda:0\").eval()\r\nmodel.half()\r\n\r\ntokenizer = T5Tokenizer.from_pretrained(\"t5-base\")\r\n\r\nARTICLE = \"\"\"summarize: Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that \"so far no videos were used in the crash investigation.\" He added, \"A person who has such a video needs to immediately give it to the investigators.\" Robin's comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. \"One can hear cries of 'My God' in several languages,\" Paris Match reported. \"Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object. Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing.\" \"It is a very disturbing scene,\" said Julian Reichelt, editor-in-chief of Bild online. An official with France's accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were \"completely wrong\" and \"unwarranted.\" Cell phones have been collected at the site, he said, but that they \"hadn't been exploited yet.\" Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical \"no.\" Reichelt told \"Erin Burnett: Outfront\" that he had watched the video and stood by the report, saying Bild and Paris Match are \"very confident\" that the clip is real. He noted that investigators only revealed they'd recovered cell phones from the crash site after Bild and Paris Match published their reports. \"That is something we did not know before. ... Overall we can say many things of the investigation weren't revealed by the investigation at the beginning,\" he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he's accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a \"previous episode of severe depression,\" the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz's battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a \"swift and seamless clarification\" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz's correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz's possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and \"held all the licenses required.\" Earlier, a spokesman for the prosecutor's office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot's license. Kumpa emphasized there's no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot's license, a European government official briefed on the investigation told CNN on Tuesday. While flying was \"a big part of his life,\" the source said, it's only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz's girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there's more to the story, said Brian Russell, a forensic psychologist. \"Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren't going to keep doing their job and they're upset about that and so they're suicidal,\" he said. \"But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person's problems.\" Germanwings crash compensation: What we know. Who was the captain of Germanwings Flight 9525? CNN's Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN's Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel, and Anna-Maja Rappard contributed to this report.\"\"\"\r\n\r\ninputs = tokenizer(ARTICLE, max_length=512, truncation=True, return_tensors=\"pt\").to(\"cuda:0\")\r\nout = model(**inputs, decoder_input_ids=torch.tensor([[tokenizer.pad_token_id]]).to(\"cuda:0\"))\r\ntorch.isnan(out.logits).any()\r\n# => True\r\n```\r\n\r\n## Proposed fix\r\n\r\nTo avoid `inf` values we could clamp the `hidden_states` to the max values for the current data type if there are `inf` in it. i.e\r\n```python\r\nif torch.isinf(hidden_states).any():\r\n clamp_value = torch.finfo(hidden_states.dtype).max - 1000\r\n hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)\r\n```\r\n\r\nwe need to add this after self attn, cross-attn, and the feed-forward layer which is where the `inf` values occur. This works for both `apex` and `amp`\r\n\r\nTo verify this fix, I trained `t5-base`, `t5-v1_1-base` and `t5-v1_1-small` on `cnn/dm` for 10k steps (1.11 epochs)\r\nHere's the training command, to run this clone [this fork](https://github.com/patil-suraj/transformers) and check out the `fix-t5-fp16` branch. navigate to `examples/seq2seq` dir, follow the instructions in the readme to download `cnn_dm` and dataset, and then run the following command\r\n\r\n```bash\r\nexport M=google/t5-v1_1-base\r\nexport OUT_DIR=t5-v1_1-base-cnn-fp16\r\nexport DATA_DIR=cnn_dm\r\n\r\npython finetune_trainer.py \\\r\n --model_name_or_path $M \\\r\n --data_dir $DATA_DIR \\\r\n --output_dir $OUT_DIR --overwrite_output_dir \\\r\n --max_steps=10000 \\\r\n --gradient_accumulation_steps=8 \\\r\n --learning_rate=1e-4 \\\r\n --per_device_train_batch_size=4 \\\r\n --n_val 500 \\\r\n --max_target_length=56 --val_max_target_length=128 \\\r\n --fp16 --fp16_backend apex \\\r\n --do_train --do_eval --evaluation_strategy steps \\\r\n --logging_steps=100 --logging_first_step --eval_steps=2500 --save_steps=2500 --save_total_limit=2 \\\r\n --sortish_sampler \\\r\n```\r\n\r\nfor evaluation \r\n```bash\r\npython run_eval.py \\\r\n t5-v1_1-base-cnn-fp16 cnn_dm/test.source hypothesis.txt \\\r\n --reference_path cnn_dm/test.target \\\r\n --score_path metrics.json \\\r\n --device cuda:0 \\\r\n --prefix summarize: \\\r\n --bs 16 \\\r\n --fp16 \\\r\n```\r\n\r\nand got the following metrics (ROUGE2)\r\n1. for `t5-base`: 19.2804\r\n2. for `t5-v1.1-base`: 18.4316\r\n(note that the score for `t5-base` is more because it's already pre-trained on `cnn/dm`)\r\n\r\nTo compare this, evaluated the pre-trained `t5-base` in both `fp32` and `fp16`, which gave the following results\r\n1. `fp16`: 18.3681\r\n2. `fp32`: 18.394\r\n\r\nSo the results are close enough.\r\n\r\nTo verify the fix for `t5-large`, I evaluated the pre-trained `t5-large` in `fp32` and `fp16` (use the same command above to evaluate `t5-large`) and got the following results\r\n1. `fp16`: 19.2734\r\n2. `fp32`: 19.2342\r\n\r\nSurprisingly, rouge2 is slightly better in `fp16`.\r\n\r\nSo with the above fix, the following model types now work in `fp16` (opt level `01`), and give descent speed-up :)\r\n- **T5v1**: `t5-small`, `t5-base`, `t5-large`\r\n- **T5v1_1**: `google/t5-v1_1-small`, `google/t5-v1_1-base`\r\n- **MT5**: `google/mt5-small`, `google/mt5-base`\r\n\r\n`google/t5-v1_1-large` and `google/mt5-large` should also work, will confirm after running few experiments.\r\n\r\n\r\nOne interesting observation,\r\nFor inference, the `t5-base` fine-tuned with `fp16` and evaluated in `fp32` is faster than pre-trained `t5-base` evaluated in `fp16`. See this [colab](https://colab.research.google.com/drive/1UaMBsWp3e1Qf-fYKxXmtulsvPXViKa72?usp=sharing) \r\n\r\n\r\n**Update**:\r\n`google/t5-v1_1-large` still gives `nan` loss after about 200 steps\r\n", "Great work! We should also share those results on the forum: https://discuss.huggingface.co/ :-) ", "Hi @exelents\r\n\r\nTo answer your question,\r\nas mentioned above these changes will enable fp16 for all small and base version with `apex` `01` and native `amp`. \r\nFor large models, I only tested it for inference, and it works. Right now I'm training large models and will report the results here.\r\n\r\nDeepSpeed handles it's own fp16 and I don't know all the details about it, so won't be able to help there at the moment. @stas00 might have some ideas as he's working with deepspeed.\r\n\r\nTo sum up, this fix works with `apex 01` and `native amp` with `Seq2SeqTrainer` for training and with `.half` for inference.", "> DeepSpeed handles it's own fp16 and I don't know all the details about it, so won't be able to help there at the moment. @stas00 might have some ideas as he's working with deepspeed.\r\n\r\nI would like the DeepSpeed integration to be merged and then anybody can start experimenting and seeing what else might be needed to be tweaked. To start with I've been primarily focusing on training/eval just working. The next stage would be using and tuning up.", "Hi @patil-suraj \r\nIt seems like huggingface still hasn't repaired the FP16 problem in MT5-large or MT5-xl, do you or anynoe else have any plans on it?\r\n", "Hey @mxa4646,\r\n\r\nT5 was never made to be fully compatible with FP16, it was trained using bfloat16, which has a different range than PyTorch's fp16. There is a good chance though that training T5 with deepspeed and fp16 will work!", "Hi \r\nI am training mt5-small with deepspeed, with fp16, and I am getting always nan, so far could not managed to make it work, do you mind to share how you set parameters to make it work? I am having a hard time with this and kindly appreciate your help @patrickvonplaten ", "> T5 was never made to be fully compatible with FP16, it was trained using bfloat16,\r\n\r\nThank you for this insight, @patrickvonplaten - I didn't know that!\r\n\r\nI was reading up on bfloat16 for a related issue https://github.com/huggingface/transformers/issues/10816 and it looks like the main issue is that whenever one does an aggregate operation on big numbers in bfloat16 or fp16 - the accumulate needs to be in fp32. So for example the fix applied here: https://github.com/huggingface/transformers/pull/10815 - so perhaps it is possible to identify such operations and change them to `some_torch_operator(..., , dtype=torch.float32)` so most of the math will still be fp16, but there will be no overflow. And it won't impact the normal fp32 logic, as it'd already be of the same type. And this operation doesn't take much extra memory (other than doubling of the resulting variable size).\r\n\r\nBut here it sounds like the problem is different and it's that bfloat16 may not convert to the same value in fp16. I wonder if someone tried to convert the weights and compare the difference. \r\n\r\nPerhaps it's enough to take the models and finetune them on the same data but in mixed precision and perhaps it'd rectify its level of precision.\r\n\r\n", "I tried the T5-large in fp16 and it is slower which is really strange. For everything else the same for the same test data i get 5.62 sec with Fp32 and 6.95 sec for Fp16. However fp16 uses almost 50% less memory. ", "Has this model been implemented for PyTorch yet?" ]

[]

[]

[ "I commented on the corresponding issue, I don't fully understand what's going on there in the error", "This is on purpose because all the methods of a Keras Model allow to have `np.ndarray` as input. You can check for example `fit`, `predict` or `evaluate` here https://www.tensorflow.org/api_docs/python/tf/keras/Model. They all take a numpy array as possible input", "Okey, I see. I still don't think we should provide this feature, just because keras has some automatic conversion internally. Is there a use case where one cannot forward a TF tensor and has to forward a `nd.array`? The general philosophy of the lib is to \"not add too many magic functions\" and allowing `nd.arrays` as inputs for TF seems like opening the door for lots of future issues to me. Let's see what @sgugger @LysandreJik think about it", "On my side I would prefer to keep as much compliancy as possible with TF. But if everyone are not confident because of this Keras magic, I'm ok to do not provide it :)", "The TF models do not accept numpy arrays inputs, so this would allow bad inputs to be passed to TF models. I think we should stick with inputs accepted by TF models only.", "The TF models allow to have numpy arrays as inputs (if we had the `np.ndarray` type among the allowed ones). As example:\r\n```python\r\nfrom transformers import TFBertForSequenceClassification, BertTokenizer\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-cased\")\r\ninputs = tokenizer(\"Hello\", return_tensors=\"np\")\r\nmodel = TFBertForSequenceClassification.from_pretrained(\"bert-base-cased\")\r\nmodel(inputs)\r\n```\r\nGives:\r\n```\r\nTFSequenceClassifierOutput(loss=None, logits=<tf.Tensor: shape=(1, 2), dtype=float32, numpy=array([[0.08065815, 0.58226204]], dtype=float32)>, hidden_states=None, attentions=None)\r\n```\r\n\r\nKeras layers/models are by default compliant with numpy arrays.", "Ah my bad, I tried but forgot to checkout the PR before :facepalm:\r\nIf TF models do accept those inputs then, I have no strong objection.", "@LysandreJik you mean adding a test in `test_modeling_tf_common` with numpy array as input for each model?", "Yes, for example !", "There is a now a new test to be sure that the models can be properly executed with numpy inputs.", "Does it looks ok to be merged?" ]

[]

[]

[ "Thanks!" ]

[]

[]

[]

[]

[]

[]

[ { "id": 1834056635, "node_id": "MDU6TGFiZWwxODM0MDU2NjM1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Core:%20Tokenization", "name": "Core: Tokenization", "color": "FF4446", "default": false, "description": "Internals of the library; Tokenization." }, { "id": 1920687293, "node_id": "MDU6TGFiZWwxOTIwNjg3Mjkz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Fast%20Tokenizers", "name": "Fast Tokenizers", "color": "b60205", "default": false, "description": "" } ]

[]

[ "That looks like a tough one...we might need help from @n1t0 here", "The files for this model can be found [here](https://huggingface.co/pdelobelle/robbert-v2-dutch-base/tree/main).\r\n\r\nIn `merges.txt` is the merge rule: `ĉ Ċ`\r\nThat means that the symbol `Ċ` should be present in `vocab.json`\r\nIt is not.\r\nSo the error does kind of make sense.\r\n\r\n @iPieter are these the correct files?\r\nIn the paper you mention:\r\n> We limited the vocabulary to 40k words,\r\n> which is 10k words less than RobBERT v1, due to\r\n> additional tokens including non-negligible num-\r\n> ber of Unicode tokens that are not used in Dutch.\r\n\r\nThere are 39982 words in the vocabulary. \r\nIs it possible that some of the symbols/tokens are missing?\r\n\r\n\r\n", "@schelv Thanks for looking into the issue. Those files should be correct, there are indeed 39982 tokens.\r\n\r\nI am using the same files internally without any issues on the old tokenizer (i.e. correct behaviour and sensible predictions), for which the files were specifically created. I also looked into this error before, but cannot find an easy fix. There are a few other merge rules that are also conflicting. \r\n\r\nMy suspicion is that this stems from the translation of the tokenizer's files from Fairseq to HF. This was done a year ago, so the details might be a bit fuzzy. The problem was that the Fairseq library, where we trained RobBERT, used the vocab and merges file, but also generated an additional file (`dict.txt`) that was used to count the number of occurrences for each token (which is ok, not the issue) and also orders the tokens. \r\n\r\nThese new, ordered positions were then used in Fairseq, while HF uses the id's from the vocab.json file. So this means there is an additional dictionary lookup. To fix this behaviour in HF transformers, I created a script to merge the vocab.sjon with the dict.txt. Otherwise, the token id's from unrelated tokens would be used for the embedding layer and the MLM task, giving a garbage output. \r\n\r\nI will investigate this translation step again, but I'm confused by the fact that the behaviour is correct with the slow tokenizer.\r\n\r\n", "Thanks for the quick answer!\r\n\r\nJust checking if I understand you correctly:\r\nFairseq uses token id's that are based on the token occurrence count. This information is stored in dict.txt\r\nThe token id's of the original vocab.json were updated with the information from dict.txt? did this create the current vocab.json that is loaded by the transformers library?\r\n\r\nIf you upload the original files somewhere I can also take a look at them.🙂", "Yes, that's exactly right. The conversion script is [here](https://github.com/iPieter/RobBERT/blob/340bd9d87ef362462fccf8f44e7740c7dfd1d865/src/convert_roberta_dict.py) and [this is the unit test](https://github.com/iPieter/RobBERT/blob/340bd9d87ef362462fccf8f44e7740c7dfd1d865/tests/test_convert_roberta_dict.py) that missed this case. The original files are also downloadable from our github release [here](https://github.com/iPieter/RobBERT/releases/tag/v2.0), but you have to download the entire model. \r\n\r\nHowever, when writing the previous response, I had an insight in what the issue might be. So'll try to debug and report the results here.\r\n\r\n_Long report ahead, TLDR: it now works_ 😃 \r\n\r\nThere might be tokens in the `vocab.json` that was generated by the HuggingFace tokenizer library that are not found by the Fairseq tokenizer, thus they don't occur in the `dict.txt`. After some debugging, I found the these tokens (the special tokens are handled later): \r\n\r\n```python\r\n00:'Á'\r\n01:'÷'\r\n02:'À'\r\n03:'þ'\r\n04:'ø'\r\n05:'ÿ'\r\n06:'ú'\r\n07:'ö'\r\n08:'ĉĊ'\r\n09:'</s>'\r\n10:'č'\r\n11:'ĠTheCompany'\r\n12:'û'\r\n13:'ü'\r\n14:'<unk>'\r\n15:'ý'\r\n16:'õ'\r\n17:'<pad>'\r\n18:'Ċ'\r\n19:'<s>'\r\n20:'<mask>'\r\n21:'ù'\r\n```\r\n\r\nSo these tokens are what is causing the fast tokenizer to complain, since they appear in the `vocab.json` set and not in the `dict.txt` set. Ignoring the special tokens (`<unk>`, `<s>`, `</s>` and `<pad>`), this brings the latest vocab id to 39996, not yet 40k. So there is a second bug in my conversion script. \r\n\r\nThe second bug has to do with the fact that Fairseq adds 2 custom tokens by default that I didn't remove. That's not a big deal, but they do affect the vocab length, so let's be totally correct and add those two tokens and the mask token (since `robbert-v2-dutch-base` is and MLM model) as well:\r\n\r\n<img width=\"1019\" alt=\"image\" src=\"https://user-images.githubusercontent.com/6965756/109134694-9a9ff880-7756-11eb-9734-f678d6dc8845.png\">\r\n\r\nTime for a sanity check:\r\n```python\r\n[{'sequence': 'Er staat een boom in mijn tuin.', 'score': 0.16917602717876434, 'token': 2600, 'token_str': ' boom'},\r\n {'sequence': 'Er staat een bankje in mijn tuin.', 'score': 0.08176644891500473, 'token': 21620, 'token_str': ' bankje'}, \r\n{'sequence': 'Er staat een schutting in mijn tuin.', 'score': 0.0384209081530571, 'token': 15000, 'token_str': ' schutting'}, \r\n{'sequence': 'Er staat een vijver in mijn tuin.', 'score': 0.038086555898189545, 'token': 8217, 'token_str': ' vijver'}, \r\n{'sequence': 'Er staat een plant in mijn tuin.', 'score': 0.03249552100896835, 'token': 2721, 'token_str': ' plant'}]\r\n````\r\n\r\nThe vocab.json in `robbert-v2-dutch-base` is updated, so this issue can be closed.\r\n\r\n", "Great, thanks a lot for the investigation @iPieter! Now I can use the fast tokenizers in all their glory.", "Thanks from me as well! Looking forward to using it! " ]

[]

[]

[]

[]

[]

[ "I really like the idea of adding such a document. Two general things:\r\n\r\n1) I think we should make this document a bit more positive and not a \"mandatory read\" before posting an issue. IMO, in general we are (or should be) happy about all issues. Even if the issue is very badly formulated, it gives us a good signal on how the users work with the library and which features are more and which are less used.\r\n\r\n2) I'd make a clearer distinction with what should be in the forum and what should be an issue. I've probably already answered ~30 times on issues that the user should please redirect the question to the forum. I'd like to make a clearer distinction here:\r\n\r\nThe issues should ideally only be used for bug reports.\r\n\r\n**_Forum:_** All \"please explain\" questions or objectively very user-specific feature requests should land in the forum. IMO those should never land in the issues. What I mean by that are *e.g.* the following kinds of issues:\r\n\r\ni. \"I would like to use a BertModel within a RL-Agent for a customer support service. How can I use a `BertForMaskedLM` in my `ChatBotModel`?\"\r\n\r\nii. \"Could you please explain why T5 has no positional embedding matrix under `T5Model`?\"\r\n\r\niii. \"How should I set my generation parameters for translation?\"\r\n\r\niiii. \"How to train T5 on De->En translation?\" \r\n\r\n=> all these kinds of questions do not belong to the issues IMO. None of those issues hint at a bug in Transformers and have definitely a better place in the forum IMO. But, again, we **do** want people to ask exactly these questions and I'm more than happy to answer all of them (maybe i. a bit less). Not sure what @thomwolf @LysandreJik @sgugger think here.\r\n\r\n**_Issues:_** Everything which hints at a bug should be opened as an issue in Transformers. Here again, I'd like to encourage people to open issues as I'm much happier with users posting an objectively badly written issue than with users discovering an issue, but being afraid to post it in Transformers. Having said this, I really like the points you've written down so far @stas00! One thing, I like to add (as one of the first points) is that users should google (or whatever SE) their issue before opening the exact same one (yes using a SE with \"your issue here\" + \"transformers\" + \"huggingface\" often gives better results than searching on github itself). I often just link a new issue to an already answered one (which is also not that bad since it shows us again which parts of the transformer are heavily used). And I think in some cases it is fine if the user posts a link to a colab if it's bug absolutely requires a big data set.\r\n\r\nIn general, I think such a document can save us a lot of time because we can just link to it on issues that are badly written and usually are as a consequence just ignored by us. Think the author of a \"badly\" written issue can then better understand why we sometimes stop answering.\r\n", "Great suggestions, @patrickvonplaten - Thank you - I hope I integrated them well.\r\n\r\nI think at this point we are in the gathering stage - so bring on the ideas and points you feel are important.\r\n\r\nWhen this stage is done we will do a final edit so that the document feels most welcoming to the users.", "All comments/suggestions you have kindly offered so far have been addressed. A vous!", "> After applying Patrick's suggestion, I think this document is in excellent shape. One nit I would have is that it's written in perfect English. We have a lot of users that are not native speakers, and possibly won't understand some parts of it. Then again, I don't think we should reduce the quality of this document, but let's think about what we can do in general to be friendly to non-English speaking users/contributors.\r\n\r\nOh, thank you for the complement as I'm one very non-native English speaker ;) But I totally hear what you mean.\r\n\r\nPerhaps the simplest thing to do is to add a note that if someone struggles with understanding anything in the document, they can ask us to make it more easy to understand? That way it doesn't have to be imperfectly perfect from the get-going.", "Merging, thanks a lot @stas00!", "Copied this into [this forum post](https://discuss.huggingface.co/t/how-to-request-support/3128) and added a link to it in the document in [this commit](https://github.com/huggingface/transformers/commit/6009668c631aa5773c66aa30c6bfd9c191e2a6be)." ]

[]

[]

[ "Hi @ziqi-zhang,\r\nthe distillation code is written using the `pytorch-lightning` framework and `SummarizationModule` is a lightning module. You should go through the lightning docs to see how training works for lightning modules and how to customize it.\r\n\r\n`Trainer` is transformers training helper which is different from lightning and there is no connection between the two.", "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." ]

[]

[]

[ "Hmm, I don't really know how to best answer your question here...maybe @JetRunner ?", "using lowercase=False preserves the information about casing and this information maybe helpful in the context of the some work. such as for sentiment analysis i think casing is not important there but for the task like NER it maybe useful. Again but for many task where we deal with multiple languages is it recommended to use cased model because i think every language has its own grammar and syntax and maybe casing helps in some or the other way?\r\n\r\nCould this be the reason why Bert-chinese uses lowercase=False by default?", "I think your explanation makes sense. \nBut there are no capital letters in the Chinese vocab.txt, so all words contain capitals will be regarded as [unk].\n\nSent from my iPhone\n\n> On Dec 25, 2020, at 4:57 AM, Shubham kumar <[email protected]> wrote:\n> \n> \n> using lowercase=False preserves the information about casing and this information maybe helpful in the context of the some work. such as for sentiment analysis i think casing is not important there but for the task like NER it maybe useful. Again but for many task where we deal with multiple languages is it recommended to use cased model because i think every language has its own grammar and syntax and maybe casing helps in some or the other way?\n> \n> Could this be the reason why Bert-chinese uses lowercase=False by default?\n> \n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n", "Well I would say the design of Chinese BERT is not necessarily the best. It makes sense to use only lower cases to resolve the data sparsity since there are not many English sentences in Chinese Wikipedia.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Hello!\r\n\r\nThese two messages are just a warning, you can ignore them if you are not concerned. Basically, these messages will always be displayed everytime the graph node is executed, and only in graph mode.", "Is there a way to suppress these warnings? They overwhelm the logs with useless messages...", "@jplu the issue still persists, what is the purpose of these warnings? Why are they displayed to the user of the library?", "@jklaise @Gilthans The logging happens through a TF Logger (See [this line](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_tf_utils.py#L50)) . \r\n\r\nYou can suppress them by using something like `tf.get_logger().setLevel('ERROR')`" ]

[]

[]

[]

[]

[]

[ "@patrickvonplaten , I reworked a bit the approach. Now `TFDPREncoder` and `TFDPRSpanPredictor` are still models and keep their features from `TFPreTrainedModel` while all the DPR models benefit of the serving.\r\n\r\nAll the slow/quick tests are still passing.", "All @sgugger comments have been addressed." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hello @rabeehk,\r\n\r\nWe sadly cannot fix issues of different repos, such as `rabeehk/seq2seq.git` - this is too time-consuming and not really our responsibility. We're happy to assist if you could provide a **short, precise, and complete** code snippet that is based on Transformers `Seq2SeqTrainer` only.", "@rabeehk, I think you may have not considered that open source projects are not a help desk. If you are going to continue in the same fashion you will not get any answers at all.\r\n\r\nMany people ask for help but you need to think how to ask for help so that it's easy for the developers to quickly understand what is going on, reproduce the problem and solve what needs to be solved. But if you dump 1000 line logs and say help me to fix this without investigating it first yourself you will not get anywhere here.\r\n\r\nFor example, in your 1000 line log dump in OP if you look closely you will see that the error is on your side since it tells you:\r\n```\r\noutputs/mixture1/meta-adapters-task-projector-new_sampler-num-gpus-4/mp-lr-3e-02-r-8-l-true/config.json not found\r\n```\r\ni.e. your setup is broken.\r\n\r\nSo you didn't really study the problem and yet want us to do this for you.\r\n\r\nThat's said I personally will not do it again, so please don't tag me unless it's related to what I'm working on and you found a bug in the code I wrote or maintain.\r\n\r\nAlso tagging multiple developers out of context is frowned upon - you tagged me on this issue:\r\n\r\n> Who can help\r\n> FSMT: @stas00\r\n\r\nwhat does it have to do with FSMT? \r\n\r\nThe tagging info is to help users to direct their questions to the right developers who are maintainers of certain domains. They can then decide at their own discretion to tag other developers if they feel it'd help move the issue forward. If you tag multiple people out of context you will gain no support. \r\n\r\nIf you are not willing to invest energy and time into investigating the problems you encounter and forming quality questions, please consider hiring someone who will be willing to answer the multitude of your questions and sort things out for you. Perhaps ask at the forums if someone is willing to work with you professionally where you pay them for the services provided.\r\n\r\nI hope this comment has been useful and trust you will find a way to receive the support you need.\r\n", "Yes @stas00 is completely right. @rabeehk your comments are borderline spammy. \r\n\r\nWe try to help as much as possible but you also need to put in the work so the community can actually help you efficiently.", "Hi Stephan, Hi Julien\nplease find my responses below:\n\nOn Wed, Dec 23, 2020 at 6:29 PM Stas Bekman <[email protected]>\nwrote:\n\n> rabeehk, I think you may have not considered that open source projects are\n> not a help desk. If you are going to continue in the same fashion you will\n> not get any answers at all.\n>\nSorry if this looks like a spam to you, but I really still think this was a\nbug, if you try to load the model twice inside the finetune_trainer.py in\nevaluation part, which is something the user might well need when one wants\nto apply more chnages to the trained model before evaluation, you would see\nthis is not multi-process safe resulting in the bug I reported.\n\n> Many people ask for help but you need to think how to ask for help so that\n> it's easy for the developers to quickly understand what is going on,\n> reproduce the problem and solve what needs to be solved. But if you dump\n> 1000 line logs and say help me to fix this without investigating it first\n> yourself you will not get anywhere here.\n>\nSorry I thought providing full logs help, if this is not sure, I would not\nprovide the full logs, I still included the one line error message above\nthese logs, I did investigate the issue myself, and I realized this is not\nmulti-process safe as I mentioned.\n\n> For example, in your 1000 line log dump in OP if you look closely you will\n> see that the error is on your side since it tells you:\n>\n> outputs/mixture1/meta-adapters-task-projector-new_sampler-num-gpus-4/mp-lr-3e-02-r-8-l-true/config.json not found\n>\n> no the setup is not broken, the files are there as I said, please read my\nbug report carefully, this is the result of the bug I said.\n\n> i.e. your setup is broken.\n>\n> So you didn't really study the problem and yet want us to do this for you.\n>\n> not really, I investigated it for hours and hours in fact.\n\n> That's said I personally will not do it again, so please don't tag me\n> unless it's related to what I'm working on and you found a bug in the code\n> I wrote or maintain.\n>\n> Sure, I was thinking you are working on seq2seq from various updates on\nthis, sorry for the mistake,\n\n> Also tagging multiple developers out of context is frowned upon - you\n> tagged me on this issue:\n>\n> Who can help\n> FSMT: @stas00 <https://github.com/stas00>\n>\n> what does it have to do with FSMT?\n>\nsorry for the mistake, I explained this above, I was really thinking you\nare working on seq2seq and though this is relevant.\n\n> The tagging info is to help users to direct their questions to the right\n> developers who are maintainers of certain domains. They can then decide to\n> tag other developers if they feel it'd help the issue along. If you tag\n> multiple people out of context you will gain no support.\n>\n> I was really mistaken thinking this is relevant.\n\n> If you are not willing to invest energy and time into investigating the\n> problems you encounter and forming quality questions, please consider\n> hiring someone who will be willing to answer the multitude of your\n> questions and sort things out for you. Perhaps ask at the forums if someone\n> is willing to work with you professionally where you pay them for the\n> services provided.\n>\nNo, this is not correct, I did spent hours and hours on this, to me this is\nstill a bug.\n\n> I hope this comment has been useful and trust you will find a way to\n> receive the support you need.\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/9280#issuecomment-750399462>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABP4ZCBMPNG4IN2NE5DGRPDSWISH7ANCNFSM4VG3SIFA>\n> .\n>\n", "I also should say putting comments like your last paragraph @stas00 is inappropripate. No matter I beleive this is a bug you think this is a spam, no matter in which position you are, no matter I mistakenly thought providing full logs helps, please behave people with respect. I still believe this is a bug.\r\n\r\nOn Wed, Dec 23, 2020, 8:43 PM Rabeeh Karimi <[email protected]> wrote:\r\n\r\n> Hi Stephan, Hi Julien\r\n> please find my responses below:\r\n>\r\n> On Wed, Dec 23, 2020 at 6:29 PM Stas Bekman <[email protected]>\r\n> wrote:\r\n>\r\n>> rabeehk, I think you may have not considered that open source projects\r\n>> are not a help desk. If you are going to continue in the same fashion you\r\n>> will not get any answers at all.\r\n>>\r\n> Sorry if this looks like a spam to you, but I really still think this was\r\n> a bug, if you try to load the model twice inside the finetune_trainer.py in\r\n> evaluation part, which is something the user might well need when one wants\r\n> to apply more chnages to the trained model before evaluation, you would see\r\n> this is not multi-process safe resulting in the bug I reported.\r\n>\r\n>> Many people ask for help but you need to think how to ask for help so\r\n>> that it's easy for the developers to quickly understand what is going on,\r\n>> reproduce the problem and solve what needs to be solved. But if you dump\r\n>> 1000 line logs and say help me to fix this without investigating it first\r\n>> yourself you will not get anywhere here.\r\n>>\r\n> Sorry I thought providing full logs help, if this is not sure, I would not\r\n> provide the full logs, I still included the one line error message above\r\n> these logs, I did investigate the issue myself, and I realized this is not\r\n> multi-process safe as I mentioned.\r\n>\r\n>> For example, in your 1000 line log dump in OP if you look closely you\r\n>> will see that the error is on your side since it tells you:\r\n>>\r\n>> outputs/mixture1/meta-adapters-task-projector-new_sampler-num-gpus-4/mp-lr-3e-02-r-8-l-true/config.json not found\r\n>>\r\n>> no the setup is not broken, the files are there as I said, please read my\r\n> bug report carefully, this is the result of the bug I said.\r\n>\r\n>> i.e. your setup is broken.\r\n>>\r\n>> So you didn't really study the problem and yet want us to do this for you.\r\n>>\r\n>> not really, I investigated it for hours and hours in fact.\r\n>\r\n>> That's said I personally will not do it again, so please don't tag me\r\n>> unless it's related to what I'm working on and you found a bug in the code\r\n>> I wrote or maintain.\r\n>>\r\n>> Sure, I was thinking you are working on seq2seq from various updates on\r\n> this, sorry for the mistake,\r\n>\r\n>> Also tagging multiple developers out of context is frowned upon - you\r\n>> tagged me on this issue:\r\n>>\r\n>> Who can help\r\n>> FSMT: @stas00 <https://github.com/stas00>\r\n>>\r\n>> what does it have to do with FSMT?\r\n>>\r\n> sorry for the mistake, I explained this above, I was really thinking you\r\n> are working on seq2seq and though this is relevant.\r\n>\r\n>> The tagging info is to help users to direct their questions to the right\r\n>> developers who are maintainers of certain domains. They can then decide to\r\n>> tag other developers if they feel it'd help the issue along. If you tag\r\n>> multiple people out of context you will gain no support.\r\n>>\r\n>> I was really mistaken thinking this is relevant.\r\n>\r\n>> If you are not willing to invest energy and time into investigating the\r\n>> problems you encounter and forming quality questions, please consider\r\n>> hiring someone who will be willing to answer the multitude of your\r\n>> questions and sort things out for you. Perhaps ask at the forums if someone\r\n>> is willing to work with you professionally where you pay them for the\r\n>> services provided.\r\n>>\r\n> No, this is not correct, I did spent hours and hours on this, to me this\r\n> is still a bug.\r\n>\r\n>> I hope this comment has been useful and trust you will find a way to\r\n>> receive the support you need.\r\n>>\r\n>> —\r\n>> You are receiving this because you were mentioned.\r\n>> Reply to this email directly, view it on GitHub\r\n>> <https://github.com/huggingface/transformers/issues/9280#issuecomment-750399462>,\r\n>> or unsubscribe\r\n>> <https://github.com/notifications/unsubscribe-auth/ABP4ZCBMPNG4IN2NE5DGRPDSWISH7ANCNFSM4VG3SIFA>\r\n>> .\r\n>>\r\n>\r\n", "The point we are trying to communicate is that you need to review how you communicate, @rabeehk. Your communications come across as too much and too indiscriminate.\r\n\r\nI'm totally accepting that you might be unaware of what is expected in good communications and perhaps HuggingFace needs to have a guidelines document at how users can ask for help in the most efficient way for all involved parties.\r\n\r\nAs of this moment I'd happy to invest a bit of my free time to support you to find a way for you to become an asset to this community and not an annoyance. If you are willing to listen and take action:\r\n\r\n1. Anybody looking at your first post will have an urge to flee - it's scary in its length and most people will not even try to understand what could be a very valid issue.\r\n\r\n So you need to edit the first post to remove any information that's not pertaining to the issue at hand. e.g. all those Download xx% logs are totally useless. You're saying you are attaching the traceback, but you're attaching the full log. I accept that you might have not known that.\r\n\r\n Attaching a full log can be helpful if it's done as an attachment, a link to a paste.bin or at the very least if you enclosed it in:\r\n\r\n```\r\n<details>\r\n<summary>Full log</summary>\r\n<pre>\r\n\r\nmany\r\nlines \r\ngo\r\nhere\r\n\r\n</pre>\r\n</details>\r\n```\r\n Here is an example of the outcome:\r\n\r\n<details>\r\n<summary>Full log</summary>\r\n<pre>\r\nmany\r\nlines \r\ngo\r\nhere\r\n</pre>\r\n</details>\r\n\r\n2. As @patrickvonplaten replied to you, you can't ask someone to go into your repository and figure out what you may have done. The code is already very complex and unless there is an easy way to do a diff and it's a small diff, nobody has the time to investigate. So you need to spend time to find a way to reproduce the problem in a minimal example which should introduce no more than a few lines of code change-wise (of course, there are exceptions, but this is more of norm).\r\n\r\n Usually the best way is to just show the relevant backtrace (in DDP just one of them, as each process will dump a copy), the command line and then ask if there is anything else that you could supply to help the developer reproduce the problem.\r\n\r\n3. Try to use the latest official version. We have no resources to go and debug older revisions, which could easily have bugs that have been fixed in the latest released version.\r\n\r\n I understand that this is not always possible. But this is the best way if it fits.\r\n\r\n4. Most of the time you can't ask to test with your data, since we don't have your data. So either you should use some existing dataset supported by HF datasets or you need to have the code that generates a small sample on the fly.\r\n\r\n5. Do not tag multiple people on the issue unless you know this is expected, either because you asked them and they gave you an explicit permission or the Issue template instructs you to do so. Having someone help you like I'm doing now is not an invitation to tag that person in the future on all your issues.\r\n\r\n I can see why you chose to tag me by looking at seq2seq commits, and while I made a few small changes in seq2seq recently, it just happened to be so because I was working on something totally unrelated and there were some changes that were required for me to proceed. But I'm not in charge of that domain.\r\n\r\n Remember that every time you tag someone, they get a notification and you're taking their time w/o their permission. Please be sensitive to that.\r\n\r\n6. Use the edit button. Delete and merge multiple comments into one if nobody followed up yet. As you merge them edit them to be coherent. Use bullets and items if it makes sense.\r\n\r\n I know my first comment version almost always comes out with typos and can be incoherent or too verbose. If you look at my comments' history I often make a ton of re-edits, since I want to make sure my communication is as clear as possible (and I know I myself can be too verbose, mea culpa). \r\n \r\n---------------\r\n\r\nThe key message of this comment is that when you ask for support you are given a tiny sliver of developer's time and you need to quickly communicate the essentials of the issue at hand. You're not expected to be born with that knowledge. You're not expected to be perfect at it. If I may recommend - learn from issues posted by other people - see which issues get responses and which are ignored - learn what the posters who did get responses did right. It's a simple pattern matching with some patience.\r\n\r\nThere is no harm in asking: \"look, I have all these questions and I don't know how to ask them in the best possible way. Can someone help?\"\r\n\r\nPerhaps, you need to find a mentor in the community at the forums, by asking if someone can support you to help you find a way to file good issues.\r\n\r\nWhen you tune up your communications then the developers of this fabulous project will be more than happy to address and resolve the issues you raise. This skill, of course, will help you at any other open source project.\r\n\r\nPlease let me know if you found this helpful. And perhaps if you'd like to continue this discussion because you need further clarifications let's go to the forums https://discuss.huggingface.co/ and leave the Issues section alone for now. You have my permission to tag me (just @stas) in the forums for this particular discussion if you think it'd be helpful to you.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "I think I fixed all of them. \r\n\r\nWhen moving everything around I felt more comfortable switching temporarily to absolute imports forgot to switch back.\r\n" ]

[ { "id": 2669577093, "node_id": "MDU6TGFiZWwyNjY5NTc3MDkz", "url": "https://api.github.com/repos/huggingface/transformers/labels/PR%20for%20Model%20Addition", "name": "PR for Model Addition", "color": "5319e7", "default": false, "description": "" } ]

[]

[ "@patrickvonplaten when you have time, can you fix the conflicts and apply the same updates merged in Longformer to LED. Thanks!" ]

[]

[]

[]

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

[ { "login": "czabo", "id": 75574105, "node_id": "MDQ6VXNlcjc1NTc0MTA1", "avatar_url": "https://avatars.githubusercontent.com/u/75574105?v=4", "gravatar_id": "", "url": "https://api.github.com/users/czabo", "html_url": "https://github.com/czabo", "followers_url": "https://api.github.com/users/czabo/followers", "following_url": "https://api.github.com/users/czabo/following{/other_user}", "gists_url": "https://api.github.com/users/czabo/gists{/gist_id}", "starred_url": "https://api.github.com/users/czabo/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/czabo/subscriptions", "organizations_url": "https://api.github.com/users/czabo/orgs", "repos_url": "https://api.github.com/users/czabo/repos", "events_url": "https://api.github.com/users/czabo/events{/privacy}", "received_events_url": "https://api.github.com/users/czabo/received_events", "type": "User", "site_admin": false } ]

[ "This was implemented in https://github.com/huggingface/transformers/pull/10950" ]

[]

[]

[ "If you set `disable_tqdm=False` in your `TrainingArguments`, you shouldn't have any progress bar from the library.", "Well, I think you meant `disable_tqdm=True`. By the way, the following worked:\r\n\r\n`args = TrainingArguments(disable_tqdm=True, output_dir=\"tmp_trainer\")`\r\n\r\nI am still getting progress bars for the `dataset.map()` though. Is there something like `verbose=False`?", "`dataset.map` comes from the Datasets library, not Transformers. So you should open an issue there for this part :-)" ]

[]

[]

[ "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 should help: https://github.com/huggingface/transformers/issues/1413#issuecomment-538083512", "Tanks a lot." ]

[]

[]

[ "You'll have it soon 😉 , once #9150 is merged ", "> You'll have it soon 😉 , once #9150 is merged\r\n\r\nThat's awesome. Thanks!", "@patil-suraj great, it got merged but how does translates now to your question_generation repo? How do I get the output probability/confidence score to the the predicted answers?", "I only found this comment from you https://discuss.huggingface.co/t/text-generation-pipeline-output-scores-parameter/3294/2: \r\n\r\n*the text-generation pipeline doesn’t return scores, however you could the generate method directly, to get the scores, this should help*\r\n\r\nWould be great if you could elaborate on that.\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.", "@patil-suraj Any chance that `generate_tf_utils` will get the same functionality? ", "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." ]

[]

[]

[ "LayoutLM only has a PyTorch implementation available. If you remove the `from_tf=True` statement, it will work.\r\n", "I am trying to load the model weights from [here](https://huggingface.co/microsoft/layoutlm-base-uncased/tree/main#) but `from_tf=False` doesn't work either. Traceback is below.\r\n```\r\nfile_share_pre_train_model_path = \"layoutlm-base-uncased\"\r\n... config = LayoutLMConfig.from_pretrained(\r\n... os.path.join(file_share_pre_train_model_path, \"config.json\"), num_labels=len(tag_labels), cache_dir=None\r\n... )\r\n... model = LayoutLMForTokenClassification.from_pretrained(\r\n... file_share_pre_train_model_path,\r\n... from_tf=False,\r\n... config=config,\r\n... cache_dir=None,\r\n... )\r\nTraceback (most recent call last):\r\n File \"/Users/hsk/Company/environments/ner_layoutlm/lib/python3.8/site-packages/transformers/modeling_utils.py\", line 1035, in from_pretrained\r\n state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")\r\n File \"/Users/hsk/Company/environments/ner_layoutlm/lib/python3.8/site-packages/torch/serialization.py\", line 527, in load\r\n with _open_zipfile_reader(f) as opened_zipfile:\r\n File \"/Users/hsk/Company/environments/ner_layoutlm/lib/python3.8/site-packages/torch/serialization.py\", line 224, in __init__\r\n super(_open_zipfile_reader, self).__init__(torch._C.PyTorchFileReader(name_or_buffer))\r\nRuntimeError: version_ <= kMaxSupportedFileFormatVersion INTERNAL ASSERT FAILED at ../caffe2/serialize/inline_container.cc:132, please report a bug to PyTorch. Attempted to read a PyTorch file with version 3, but the maximum supported version for reading is 2. Your PyTorch installation may be too old. (init at ../caffe2/serialize/inline_container.cc:132)\r\nframe #0: c10::Error::Error(c10::SourceLocation, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&) + 135 (0x111e90787 in libc10.dylib)\r\nframe #1: caffe2::serialize::PyTorchStreamReader::init() + 2350 (0x119f5e14e in libtorch.dylib)\r\nframe #2: caffe2::serialize::PyTorchStreamReader::PyTorchStreamReader(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&) + 143 (0x119f5d79f in libtorch.dylib)\r\nframe #3: void pybind11::cpp_function::initialize<void pybind11::detail::initimpl::constructor<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > >::execute<pybind11::class_<caffe2::serialize::PyTorchStreamReader>, 0>(pybind11::class_<caffe2::serialize::PyTorchStreamReader>&)::'lambda'(pybind11::detail::value_and_holder&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >), void, pybind11::detail::value_and_holder&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >, pybind11::name, pybind11::is_method, pybind11::sibling, pybind11::detail::is_new_style_constructor>(pybind11::class_<caffe2::serialize::PyTorchStreamReader>&&, (*)(0...), void pybind11::detail::initimpl::constructor<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > >::execute<pybind11::class_<caffe2::serialize::PyTorchStreamReader>, 0>(pybind11::class_<caffe2::serialize::PyTorchStreamReader>&)::'lambda'(pybind11::detail::value_and_holder&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> >) const&...)::'lambda'(pybind11::detail::function_call&)::operator()(pybind11::detail::function_call&) const + 147 (0x1113f57c3 in libtorch_python.dylib)\r\nframe #4: pybind11::cpp_function::dispatcher(_object*, _object*, _object*) + 3382 (0x110dede66 in libtorch_python.dylib)\r\nframe #5: cfunction_call_varargs + 120 (0x109b07518 in Python)\r\nframe #6: _PyObject_MakeTpCall + 373 (0x109b06f85 in Python)\r\nframe #7: method_vectorcall + 449 (0x109b0a0b1 in Python)\r\nframe #8: PyVectorcall_Call + 109 (0x109b072ad in Python)\r\nframe #9: slot_tp_init + 201 (0x109b5e619 in Python)\r\nframe #10: type_call + 297 (0x109b59a29 in Python)\r\nframe #11: _PyObject_MakeTpCall + 373 (0x109b06f85 in Python)\r\nframe #12: call_function + 533 (0x109bd5945 in Python)\r\nframe #13: _PyEval_EvalFrameDefault + 25678 (0x109bd274e in Python)\r\nframe #14: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #15: _PyFunction_Vectorcall + 270 (0x109b07a6e in Python)\r\nframe #16: _PyObject_FastCallDict + 247 (0x109b06dd7 in Python)\r\nframe #17: _PyObject_Call_Prepend + 143 (0x109b083df in Python)\r\nframe #18: slot_tp_init + 145 (0x109b5e5e1 in Python)\r\nframe #19: type_call + 297 (0x109b59a29 in Python)\r\nframe #20: _PyObject_MakeTpCall + 373 (0x109b06f85 in Python)\r\nframe #21: call_function + 533 (0x109bd5945 in Python)\r\nframe #22: _PyEval_EvalFrameDefault + 25829 (0x109bd27e5 in Python)\r\nframe #23: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #24: _PyFunction_Vectorcall + 270 (0x109b07a6e in Python)\r\nframe #25: call_function + 444 (0x109bd58ec in Python)\r\nframe #26: _PyEval_EvalFrameDefault + 25976 (0x109bd2878 in Python)\r\nframe #27: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #28: _PyFunction_Vectorcall + 270 (0x109b07a6e in Python)\r\nframe #29: method_vectorcall + 170 (0x109b09f9a in Python)\r\nframe #30: call_function + 444 (0x109bd58ec in Python)\r\nframe #31: _PyEval_EvalFrameDefault + 25976 (0x109bd2878 in Python)\r\nframe #32: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #33: PyEval_EvalCode + 100 (0x109bcc224 in Python)\r\nframe #34: builtin_exec + 626 (0x109bc9612 in Python)\r\nframe #35: cfunction_vectorcall_FASTCALL + 175 (0x109b438bf in Python)\r\nframe #36: call_function + 444 (0x109bd58ec in Python)\r\nframe #37: _PyEval_EvalFrameDefault + 25829 (0x109bd27e5 in Python)\r\nframe #38: function_code_fastcall + 128 (0x109b078d0 in Python)\r\nframe #39: call_function + 444 (0x109bd58ec in Python)\r\nframe #40: _PyEval_EvalFrameDefault + 25642 (0x109bd272a in Python)\r\nframe #41: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #42: _PyFunction_Vectorcall + 270 (0x109b07a6e in Python)\r\nframe #43: call_function + 444 (0x109bd58ec in Python)\r\nframe #44: _PyEval_EvalFrameDefault + 25642 (0x109bd272a in Python)\r\nframe #45: function_code_fastcall + 128 (0x109b078d0 in Python)\r\nframe #46: call_function + 444 (0x109bd58ec in Python)\r\nframe #47: _PyEval_EvalFrameDefault + 25642 (0x109bd272a in Python)\r\nframe #48: function_code_fastcall + 128 (0x109b078d0 in Python)\r\nframe #49: call_function + 444 (0x109bd58ec in Python)\r\nframe #50: _PyEval_EvalFrameDefault + 25642 (0x109bd272a in Python)\r\nframe #51: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #52: _PyFunction_Vectorcall + 270 (0x109b07a6e in Python)\r\nframe #53: call_function + 444 (0x109bd58ec in Python)\r\nframe #54: _PyEval_EvalFrameDefault + 25642 (0x109bd272a in Python)\r\nframe #55: function_code_fastcall + 128 (0x109b078d0 in Python)\r\nframe #56: call_function + 444 (0x109bd58ec in Python)\r\nframe #57: _PyEval_EvalFrameDefault + 25829 (0x109bd27e5 in Python)\r\nframe #58: function_code_fastcall + 128 (0x109b078d0 in Python)\r\nframe #59: call_function + 444 (0x109bd58ec in Python)\r\nframe #60: _PyEval_EvalFrameDefault + 25678 (0x109bd274e in Python)\r\nframe #61: _PyEval_EvalCodeWithName + 2804 (0x109bd6734 in Python)\r\nframe #62: PyEval_EvalCode + 100 (0x109bcc224 in Python)\r\nframe #63: PyRun_FileExFlags + 336 (0x109c1bed0 in Python)\r\nDuring handling of the above exception, another exception occurred:\r\nTraceback (most recent call last):\r\n File \"<input>\", line 5, in <module>\r\n File \"/Users/hsk/environments/ner_layoutlm/lib/python3.8/site-packages/transformers/modeling_utils.py\", line 1037, in from_pretrained\r\n raise OSError(\r\nOSError: Unable to load weights from pytorch checkpoint file for 'layoutlm-base-uncased' at 'layoutlm-base-uncased/pytorch_model.bin'If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. \r\n\r\n```", "Maybe try updating to Transformers 4.1.1 (I just ran the following in a notebook and it works):\r\n\r\n```\r\n!pip install transformers\r\n\r\nfrom transformers import LayoutLMForTokenClassification\r\nmodel = LayoutLMForTokenClassification.from_pretrained('microsoft/layoutlm-base-uncased')\r\n```", "Thanks @NielsRogge It was a pytorch version issue. I solved it after seeing [this](https://github.com/huggingface/transformers/issues/7739#issuecomment-707214148)", "Closing then, thanks for your help @NielsRogge !" ]

[]

[]

[ "This looks like it's actually an intended behavior of `ArgumentParser`: see [here](https://docs.python.org/3/library/argparse.html#argparse.ArgumentParser.parse_known_args) and in general of `argparse` (see [here](https://docs.python.org/3/library/argparse.html#prefix-matching)). So I don't think it categorizes as a bug, even if it should be documented.", "Oh, fantastic - thank you for finding that out, @sgugger \r\n\r\nThis feature surely bit us yesterday." ]

[]

[]

[ "Looks like you need to run `make style` on your branch to fix the formatting of the doc files. Let me know if you run into any trouble doing that.", "Thanks @sgugger 😄 I was able to run `make style` successfully (and update `preprocessing.rst` with the changes), but it looks like the `check_code_quality` check ran out of memory this time 😅 \r\n\r\n\r\n\r\nI tried rerunning it, but it seems like I don't have permission. Could you try rerunning it?\r\n\r\nI'm also happy to bump the `resource_class` for the check from `medium` to `medium+` if that would be helpful 🙂 ", "Yes it was just a spurious failure. Thanks!" ]

[]

[]

[ "After giving this some thought, given the fact there are two different lengths here (input and targets) I would propose keeping `--max_source_length` and `--max_target_length` to avoid any confusion for the user. In the same vein, `run_qa.py` contains `max_seq_length` and `max_answer_length` to clearly differentiate the two.\r\n\r\nAs for val/test I don't have any strong opinion, apart from the fact they are not used properly in the prediction at the end (only for the preprocessing), so they should be collapsed IMO", "This works for me!\r\n\r\nI especially would like to see all examples use same cl args for the same functionality.", "Works for me too,\r\n\r\nAnd for val/test targets lengths, IMO we can collapse it into one single `max_generate_length` or `eval_max_length` since most of the users (and the example scripts as well) use the same value for both args", "Having only `max_source_length` and `max_target_length` works for me!", "regarding the `val_max_target_length` and `test_max_target_length` args\r\n\r\nThe reason we (I and Sam) decided to add that\r\n\r\n- in general, it’s okay to have a bit smaller max target length for training/validation because some documents could be too long than avg length, it’s okay during training if these get truncated\r\n- for the test, we should set the max target length to be as long as the longest text in the test set so it won’t get truncated. The reason is if the text in the test set is truncated then the calculated metrics won’t be accurate.\r\n\r\nAlso, we should mention in the readme that it's best to use `run_eval.py` for calculating metrics. As there is an issue when calculating BLUE score this way as outlined in #9161", "> * for the test, we should set the max target length to be as long as the longest text in the test set so it won’t get truncated. The reason is if the text in the test set is truncated then the calculated metrics won’t be accurate.\r\n\r\nWhy not compute this in the script then? That would avoid having an argument that is half-used.", "What I'm hearing is that perhaps there was a concrete situation where the training needed a shorter max length than eval/test. So @sgugger's suggestion will solve your concern that the scoring is done on the full length, but not if for some reason the training stage should use shorter sequences. \r\n\r\nSo we have 2 related situations and I'm not sure @sgugger's solution covers the 2nd one. I just don't know whether it's a real use case or a may be.\r\n\r\nPlease let me know if I haven't explained myself clearly.\r\n", "Thinking more about it, @patil-suraj, won't fixing up `generate`'s `max_length` to match the longest max length of the test dataset lead to better scores than what they will be otherwise? And thus provide misleading results? \r\n\r\nFor example, let's take translation, the best model would do the most correct translation regardless of whether it is allowed to generate much longer sequences. So if we calculate any such max length dynamically to be fair I think there needs to be added some extra length beyond the longest test sequence. Say `max(len(tokenize(test_inputs)))*1.1`? Does it make sense?\r\n\r\nSo perhaps this is how it should work. \r\n- `max_target_length` is for training\r\n- for eval/train we derive max length from the input data plus some margin?", "What I meant here is that the `test_max_target_length` is also passed to the dataset, and the dataset then truncates the reference targets (translations, summaries) longer than that. So later the generated targets are compared with (possibly) truncated references which will result in incorrect metrics ", "You're absolutely correct - that doesn't sound right.\r\n\r\nSo before we can discuss the flags we then need to first discuss the algorithm, otherwise we won't get anywhere.\r\n\r\nIf in order to get the correct metrics we must not truncate the val/test datasets then why are we doing that in the current code?\r\n\r\nPerhaps what I suggested at the end of https://github.com/huggingface/transformers/issues/9265#issuecomment-750423347 is a better way to approach it?\r\n\r\nAlso please don't forget that max_length is used to deal with OOM limitations", "This has been resolved." ]

[]

[]

[ "If I understand correctly you raise 2 unrelated issues:\r\n\r\n1. there is only one place where `compute_metrics` is set and perhaps it needs to be changed through the life of the trainer object?\r\n\r\nSince you can always override it with:\r\n```\r\ntrainer.compute_metrics = new_compute_metrics\r\n```\r\nwhen you need to switch it to another version, so in a pinch you can do that. But clearly this is not a public API at the moment and can change at any time.\r\n\r\nPerhaps all is needed is a setable accessor for the `compute_metrics` attribute, so that a user can use it to swap in a new function at will, rather than adding new arguments?\r\n\r\n2. You're saying that users may need to pass more args to `compute_metrics`, but it's not possible.\r\n\r\nYou can do that via a closure mechanism, e.g. how it's done here:\r\nhttps://github.com/huggingface/transformers/blob/cbe63949d76efd153a1f389f38fe9ce1287e06b0/examples/seq2seq/utils.py#L80\r\nso you build your `compute_metrics` on the fly, getting whatever data you need into the closure function and then it'll have access to whatever other data you may need at run time.\r\n\r\nHere is a silly example:\r\n```\r\ndef make_compute_metrics():\r\n extra_input = 1\r\n def compute_metrics(pred):\r\n print(f\"Look ma, I can pass my own args: {extra_input}\")\r\n return compute_metrics\r\n\r\ntrainer.compute_metrics = make_compute_metrics()\r\n\r\n# and then some time later in `prediction_loop`:\r\nself.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))\r\n# calls your created on the fly function with whatever other data you want to be seen from it.\r\n```\r\n\r\nSo your custom `compute_metrics_fn` now can access whatever other data you want besides the `EvalPrediction` object.\r\n\r\n\r\n\r\n", "Hi there\nthank you for the response, yes, I agree this is possible, I solved this\nwith `functools.partial`, but still I think the better design would be to\nallow the user add extra parameters. so this was more feature request.\nPlease feel free to ignore if this does not make sense.\nthanks\nBest\nRabeeh\n\n\nOn Tue, Dec 22, 2020 at 11:12 PM Stas Bekman <[email protected]>\nwrote:\n\n> If I understand correctly you raise 2 unrelated issues:\n>\n> 1. there is only one place where compute_metrics is set and perhaps it\n> needs to be changed through the life of the trainer object?\n>\n> Since you can always override it with:\n>\n> trainer.compute_metrics = new_compute_metrics\n>\n> when you need to switch it to another version, so in a pinch you can do\n> that.\n>\n> Perhaps all is needed is an setable accessor for the compute_metrics\n> attribute, so that a user can use to swap in a new function at will, rather\n> than adding new arguments?\n>\n> 1. You're saying that users may need to pass more args to\n> compute_metrics, you can do that via closure, e.g. how it's done here:\n>\n> https://github.com/huggingface/transformers/blob/cbe63949d76efd153a1f389f38fe9ce1287e06b0/examples/seq2seq/utils.py#L80\n> so you build your compute_metrics on the fly, getting whatever data\n> you need into the closure function and then it'll have access to whatever\n> other data you may need at run time.\n>\n> Here is a silly example:\n>\n> def make_compute_metrics():\n> extra_input = 1\n> def compute_metrics(pred):\n> print(f\"Look ma, I can pass my own args: {extra_input}\")\n> return compute_metrics\n>\n> compute_metrics_fn = make_compute_metrics()\n> compute_metrics_fn()\n>\n> So your custom compute_metrics_fn now can access whatever other data you\n> want besides the EvalPrediction object.\n>\n> As I have shown in (1) you can now assign this to trainer.compute_metrics.\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/9264#issuecomment-749827774>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ARPXHH2C7MLFOY6V3MWST5LSWERUDANCNFSM4VGBUKSQ>\n> .\n>\n", "Yes, `partial` would do the trick.\r\n\r\nI've just shared my take on it. and that there might be a need for a public API to override `trainer.compute_metrics` post-`__init__`, \r\n\r\nIn my limited experience `partial` or a manual closure is how some projects implement such functions.\r\n\r\nI will let others comment though on this feature request.", "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." ]