Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Because some people wanted to configure this: https://github.com/huggingface/pytorch-pretrained-BERT/pull/585", "So I have to add `config.layer_norm_eps = 1e-12` if I am taking the config from the link above ?", "You don't need to, it's the default value when instantiating a `BertConfig` class.", "When I printed `dir(bert_config)` I am not able to see layer_norm_eps. ", "I have figured out the problem. I have to initialize like this `BertConfig.from_json_file(config_file_path)` but rather I have done like this `BertConfig(config_file_path)`. My bad :disappointed: \r\n\r\nThanks for the reply. Closing the issue." ]

[]

[]

[ "Hi Shubham,\r\nThis is a legacy from the original Tensorflow code (https://github.com/openai/finetune-transformer-lm/blob/master/train.py#L64-L69).", "Thanks for the link to the original reference. " ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/707?src=pr&el=h1) Report\n> Merging [#707](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/707?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/c304593d8fa93f25febe1458c63497a846749c89?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/707?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #707 +/- ##\n=======================================\n Coverage 62.27% 62.27% \n=======================================\n Files 18 18 \n Lines 3979 3979 \n=======================================\n Hits 2478 2478 \n Misses 1501 1501\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/707?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/707?src=pr&el=footer). Last update [c304593...620f2c1](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/707?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Yes, that's for clarity" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=h1) Report\n> Merging [#706](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/c304593d8fa93f25febe1458c63497a846749c89?src=pr&el=desc) will **decrease** coverage by `0.05%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #706 +/- ##\n==========================================\n- Coverage 62.27% 62.22% -0.06% \n==========================================\n Files 18 18 \n Lines 3979 3979 \n==========================================\n- Hits 2478 2476 -2 \n- Misses 1501 1503 +2\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `82.44% <0%> (-1.07%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=footer). Last update [c304593...8910034](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=h1) Report\n> Merging [#706](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/c304593d8fa93f25febe1458c63497a846749c89?src=pr&el=desc) will **decrease** coverage by `0.05%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #706 +/- ##\n==========================================\n- Coverage 62.27% 62.22% -0.06% \n==========================================\n Files 18 18 \n Lines 3979 3979 \n==========================================\n- Hits 2478 2476 -2 \n- Misses 1501 1503 +2\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `82.44% <0%> (-1.07%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=footer). Last update [c304593...8910034](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/706?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "FYI @roholazandie we are currently working on XLNet with pytorch over here\r\nhttps://github.com/pingpong-ai/XLNet-pytorch/tree/dev/poc", "I'll add it here also. I was working on a coming release this week anyway.\r\nIt's a mix of BERT/Transformer-XL and something I was also playing with (Two-Stream Self-Attention) so hopefully, it won't delay too much the release. ", "Awesome, downstream libraries like [flair](https://github.com/zalandoresearch/flair) are really looking forward to use XLNet 🤗 (So expect results on CoNLL and PoS tagging whenever XLNet is implemented here)", "and [jiant](https://jiant.info)!", "I had finished Simple XLNet implementation with Pytorch Wrapper here : https://github.com/graykode/xlnet-Pytorch", "How is this project going?\r\n\r\nI'm debating whether to try to build XLNet into jiant directly, but I'm not eager to replicate your hard work. Tx!", "Yes we're on track to finish the release this week I think (or next Monday in the worse case).\r\n\r\nWe reproduced the results of XLNet on STS-B (Pearson R > 0.918), the GLUE task showcased on the TF repo, with the same hyper-parameters (didn't try the others tasks but the model is the same for all).\r\n\r\nIt's taking a bit more time than planned because we took the opportunity to refactor the library's back-end. I'm really excited about the new release.\r\n\r\nWe will now have 6 different architectures (BERT, GPT, GPT-2, Transformer-XL, XLNet and XLM) and over 25 associated pretrained weights, all with the same API (for instance the GLUE training script is now the same for all the models!) and direct access to all the models' internals.\r\n\r\nAnd a lot of other things (a lot of tests to avoid future regressions, compatibility with TorchScript, easy serialization and loading of fine-tuned models...)", "> I'll add it here also. I was working on a coming release this week anyway.\r\n> It's a mix of BERT/Transformer-XL and something I was also playing with (Two-Stream Self-Attention) so hopefully, it won't delay too much the release.\r\n\r\n@thomwolf , very much looking forward to your implementation of BERT/Transformer-XL. Wondering if you were planning to release that too, and if so where you are planning to make it available. Thanks so much ", "Is there any direct wrapper for QA task using XLNet?", "I appriciate your hard work. I just saw it today, and you almost implemented.\r\n\r\nIs it really better than Bert, which i find to be amazing so far?", "Under the conditions in their paper, yes. Question is whether it always is: in what tasks, using the same data set; with same compute time (or energy consumption) for training (+on what hardware as same speed on hardware X doesn't imply same speed on hardware Y); with same number of parameters (model size); with architecture size for same prediction speed; consistently / on average (+confidence interval) for repeated runs with random initialisation..." ]

[]

[]

[]

[]

[]

[ "Yeah I've removed the memory-heavy tests", "Thanks!" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Thanks I don't think we'll add this since converted models are already provided." ]

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

[]

[ "@thomwolf \r\nHi, it seems there is something wrong with the training code in this repo.\r\nI used Google's official BERT training code and I could get decent results even with a small batch size: \r\n`{'EM': 73.80717341230668, 'F1': 77.11048422305339, 'AvNA': 80.78315235274762}`\r\n\r\nHere's the settings I used for Google's and this repo's:\r\n```\r\n--vocab_file=uncased_L-12_H-768_A-12/vocab.txt\r\n--bert_config_file=uncased_L-12_H-768_A-12/bert_config.json\r\n--init_checkpoint=uncased_L-12_H-768_A-12/bert_model.ckpt\r\n--do_train=True\r\n--train_file=../squad/data/train-v2.0.json\r\n--do_predict=True\r\n--predict_file=../squad/data/dev-v2.0.json\r\n--train_batch_size=5\r\n--learning_rate=3e-5\r\n--num_train_epochs=2.0\r\n--max_seq_length=384\r\n--doc_stride=128\r\n--output_dir=gugel_bert\r\n--version_2_with_negative=True\r\n--do_lower_case=True\r\n```\r\n```\r\n--bert_model=bert-base-uncased\r\n--output_dir=try_bert_2\r\n--train_file=data/train-v2.0.json\r\n--predict_file=data/dev-v2.0.json\r\n--do_train\r\n--do_predict\r\n--do_lower_case\r\n--train_batch_size=5\r\n--predict_batch_size=5\r\n--num_train_epochs=2.0\r\n--learning_rate=3e-5\r\n--version_2_with_negative\r\n```\r\nEDIT:\r\nForgot to say that I also can get the expected results using SQuAD v1.1 with this repo's code.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

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

[]

[ "Seems like this is now possible with last week's [merged PR](https://github.com/huggingface/pytorch-pretrained-BERT/pull/597), but I'm curious to see what the core devs say about this as well (btw, keep up the great work!)", "I have the same question :)\r\nI have tried the codes for BERT finetuning which is in lm-finetuning folder but looking for the same script for gpt-2.\r\n\r\nThanks", "Yes fine-tuning GPT-2 is fixed with #597 indeed.\r\nI'll see if I can add an example but basically changing `gpt` to `gpt-2` in the gpt example should be pretty much fine.", "@thomwolf Thanks for the great work.\r\njust wondering in order to do unsupervised LM fine-tuning (not classification) on a new dataset, should we just modify run_openai_gpt.py or is there an existing script for that?", "No existing script for that but you can start from run_openai indeed and use just the `OpenAIGPTLMHeadModel`.\r\nIf you want to supply another example, happy to welcome a PR", "I am still confused as to how to use the run_openai_gpt.py to finetune gpt2 model. A short example would be helpful", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Hey, I tried doing the same and was successful(in running the script, at least).\r\nWhat do you specify as your --gpt2_checkpoint_path as? (hope you have copied your finetuned checkpoints to preloaded 117M)?\r\n\r\nUpdate: Currently the output just stores a `config.json` and `pytorch_model.bin`. Suprisingly I don't see the vocab.txt and the merges.txt that they specify in the README. My tensorflow model checkpoint has a vocab.bpe but not merges.txt. So I'm stuck on how to proceed with the following error when running `run_gpt2.py`:\r\n` We assumed '/home/code-base/pytorch-pretrained-BERT/pytorch_pretrained_bert/temp-model/' was a path or url but couldn't find files /home/code-base/pytorch-pretrained-BERT/pytorch_pretrained_bert/temp-model/vocab.json and /home/code-base/pytorch-pretrained-BERT/pytorch_pretrained_bert/temp-model/merges.txt at this path or url.`\r\nSince I'm unable to generate samples using these weights now, any ideas would be great!", "I was able to fix my issue. I realized that the GPTConfig constructor used in convert_gpt2_checkpoint_to_pytorch.py is only for the 117M model, while I was trying to convert a 345M model. I ended up just making a new json file for the larger model. I never ran into your issues, however, as I didn't use run_gpt2.py.", "> I was able to fix my issue. I realized that the GPTConfig constructor used in convert_gpt2_checkpoint_to_pytorch.py is only for the 117M model, while I was trying to convert a 345M model. I ended up just making a new json file for the larger model. I never ran into your issues, however, as I didn't use run_gpt2.py.\r\n\r\nSo the new checkpoint files for you contain only a `config.json` and `pytorch_model.bin`?", "> > I was able to fix my issue. I realized that the GPTConfig constructor used in convert_gpt2_checkpoint_to_pytorch.py is only for the 117M model, while I was trying to convert a 345M model. I ended up just making a new json file for the larger model. I never ran into your issues, however, as I didn't use run_gpt2.py.\r\n> \r\n> So the new checkpoint files for you contain only a `config.json` and `pytorch_model.bin`?\r\n\r\nCorrect", "How did you end up fixing this problem? What json file did you have to create? I'd also like to convert a 345M model and am running into the problem described in this issue.", "@dsonbill Sorry, I honestly don't remember. My work was on the laptop I used at my old job, so I don't have access to the files anymore. You could check my really old fork though and maybe find something useful there.", "Hacky solution by modifying this script: https://raw.githubusercontent.com/huggingface/transformers/master/src/transformers/convert_gpt2_original_tf_checkpoint_to_pytorch.py\r\n\r\nBefore:\r\n```py\r\n # Construct model\r\n if gpt2_config_file == \"\":\r\n config = GPT2Config()\r\n else:\r\n config = GPT2Config.from_json_file(gpt2_config_file)\r\n model = GPT2Model(config)\r\n```\r\n\r\nAfter:\r\n\r\n```py\r\n # Construct model\r\n config = GPT2Config.from_pretrained('gpt2-medium') # Replace 'gpt2-medium' with whichever model spec you're converting\r\n model = GPT2Model(config)\r\n```", "> Hacky solution by modifying this script: https://raw.githubusercontent.com/huggingface/transformers/master/src/transformers/convert_gpt2_original_tf_checkpoint_to_pytorch.py\r\n> \r\n> Before:\r\n> \r\n> ```python\r\n> # Construct model\r\n> if gpt2_config_file == \"\":\r\n> config = GPT2Config()\r\n> else:\r\n> config = GPT2Config.from_json_file(gpt2_config_file)\r\n> model = GPT2Model(config)\r\n> ```\r\n> \r\n> After:\r\n> \r\n> ```python\r\n> # Construct model\r\n> config = GPT2Config.from_pretrained('gpt2-medium') # Replace 'gpt2-medium' with whichever model spec you're converting\r\n> model = GPT2Model(config)\r\n> ```\r\n\r\nHi, this doesn't work on me. I'm facing the same problem. Does anyone have any ideas?", "> > Hacky solution by modifying this script: https://raw.githubusercontent.com/huggingface/transformers/master/src/transformers/convert_gpt2_original_tf_checkpoint_to_pytorch.py\r\n> > Before:\r\n> > ```python\r\n> > # Construct model\r\n> > if gpt2_config_file == \"\":\r\n> > config = GPT2Config()\r\n> > else:\r\n> > config = GPT2Config.from_json_file(gpt2_config_file)\r\n> > model = GPT2Model(config)\r\n> > ```\r\n> > \r\n> > \r\n> > After:\r\n> > ```python\r\n> > # Construct model\r\n> > config = GPT2Config.from_pretrained('gpt2-medium') # Replace 'gpt2-medium' with whichever model spec you're converting\r\n> > model = GPT2Model(config)\r\n> > ```\r\n> \r\n> Hi, this doesn't work on me. I'm facing the same problem. Does anyone have any ideas?\r\n\r\nI could get past the initial error by setting the config to be hparams.json\r\n\r\n```export OPENAI_GPT2_CHECKPOINT_PATH=gpt2/355M\r\n\r\ntransformers-cli convert --model_type gpt2 \\\r\n --tf_checkpoint $OPENAI_GPT2_CHECKPOINT_PATH/model.ckpt \\\r\n --pytorch_dump_output $OPENAI_GPT2_CHECKPOINT_PATH \\\r\n --config $OPENAI_GPT2_CHECKPOINT_PATH/hparams.json\r\n```\r\n\r\nwhich seems to create the model checkpoint. I haven't check to see if the converted model works yet though. " ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=h1) Report\n> Merging [#697](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/3763f8944dc3fef8afb0c525a2ced8a04889c14f?src=pr&el=desc) will **decrease** coverage by `6%`.\n> The diff coverage is `62.5%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #697 +/- ##\n==========================================\n- Coverage 68.23% 62.22% -6.01% \n==========================================\n Files 18 18 \n Lines 3976 3979 +3 \n==========================================\n- Hits 2713 2476 -237 \n- Misses 1263 1503 +240\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/tokenization.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uLnB5) | `89.04% <ø> (-3.66%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfb3BlbmFpLnB5) | `68.11% <50%> (-12.11%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmcucHk=) | `79.49% <66.66%> (-9.06%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `69.79% <66.66%> (-12.05%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfdHJhbnNmb194bC5weQ==) | `52.93% <0%> (-6.29%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX29wZW5haS5weQ==) | `75.64% <0%> (-5.7%)` | :arrow_down: |\n| ... and [2 more](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=footer). Last update [3763f89...411981a](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/697?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=h1) Report\n> Merging [#696](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/a6f2511811f08c24184f8162f226f252cb6ceaa4?src=pr&el=desc) will **decrease** coverage by `0.13%`.\n> The diff coverage is `68.18%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #696 +/- ##\n==========================================\n- Coverage 68.37% 68.23% -0.14% \n==========================================\n Files 18 18 \n Lines 3990 3976 -14 \n==========================================\n- Hits 2728 2713 -15 \n- Misses 1262 1263 +1\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/modeling\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfb3BlbmFpLnB5) | `80.21% <100%> (-0.09%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `81.83% <100%> (-0.08%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfdHJhbnNmb194bC5weQ==) | `59.21% <100%> (-0.12%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/modeling.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmcucHk=) | `88.55% <56.25%> (-0.61%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/tokenization.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uLnB5) | `92.69% <0%> (+0.91%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=footer). Last update [a6f2511...f964753](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/696?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

[]

[ "As with all the other issues about Bert being not deterministic (#403, #679, #432, #475, #265, #278), it's likely because you didn't set the model in eval mode to desactivate the DropOut modules: `model.eval()`.\r\n\r\nI will try to emphasize this more in the examples of the readme because this issue keeps being raised.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> Epoch 1/6\r\n> loss: 2.0674 - bert_loss: 1.0283 - bert_1_loss: 1.0390 - bert_accuracy: 0.6604 - bert_1_accuracy: 0.6650\r\n> \r\n> Epoch 2/6\r\n> loss: 1.7190 - bert_loss: 0.8604 - bert_1_loss: 0.8586 - bert_accuracy: 0.7000 - bert_1_accuracy: 0.7081\r\n> \r\n> Epoch 3/6\r\n> loss: 1.5244 - bert_loss: 0.7715 - bert_1_loss: 0.7528 - bert_accuracy: 0.7250 - bert_1_accuracy: 0.7424\r\n> \r\n> Epoch 4/6\r\n> loss: 1.3203 - bert_loss: 0.6765 - bert_1_loss: 0.6438 - bert_accuracy: 0.7585 - bert_1_accuracy: 0.7741\r\n> \r\n> Epoch 5/6\r\n> loss: 1.1102 - bert_loss: 0.5698 - bert_1_loss: 0.5404 - bert_accuracy: 0.7936 - bert_1_accuracy: \r\n> 0.8082 - val_loss: 0.7052 - val_bert_loss: 0.3709 - val_bert_1_loss: 0.3343 - val_bert_accuracy: 0.8687 - val_bert_1_accuracy: 0.8803\r\n> Epoch 6/6\r\n> ETA: 0s - loss: 0.9269 - bert_loss: 0.4823 - bert_1_loss: 0.4446 - bert_accuracy: 0.8287 - bert_1_accuracy: 0.8452\r\n> bert_loss: 0.4823 - bert_1_loss: 0.4446 - bert_accuracy: 0.8287 - bert_1_accuracy: 0.8452`\r\n\r\nI have the same problem in tensorflow and I configured the model in order to consider Dropout only during the training phase (training=True). But I still have random outputs after each prediction.\r\n\r\nAs you can see during the training phase performance gets better so I guess that the problem is on the prediction" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=h1) Report\n> Merging [#694](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/80684f6f86c13a89fc1e4feac248ef96b013765c?src=pr&el=desc) will **increase** coverage by `1.17%`.\n> The diff coverage is `95.76%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #694 +/- ##\n==========================================\n+ Coverage 67.19% 68.37% +1.17% \n==========================================\n Files 18 18 \n Lines 3847 3990 +143 \n==========================================\n+ Hits 2585 2728 +143 \n Misses 1262 1262\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfdHJhbnNmb194bC5weQ==) | `59.33% <ø> (ø)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX29wZW5haS5weQ==) | `81.34% <ø> (ø)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX3RyYW5zZm9feGwucHk=) | `32.59% <ø> (ø)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/tokenization.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uLnB5) | `91.78% <ø> (-0.92%)` | :arrow_down: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `83.51% <ø> (+1.06%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfb3BlbmFpLnB5) | `80.3% <93.65%> (+2%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `81.91% <94.93%> (+2.52%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/modeling.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmcucHk=) | `89.16% <97.87%> (+0.59%)` | :arrow_up: |\n| ... and [2 more](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=footer). Last update [80684f6...4447f27](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/694?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "[You can train a tensorflow model using google colab for free](https://github.com/google-research/bert#using-bert-in-colab). After training it, you can [convert your tf model to pytorch](https://github.com/huggingface/pytorch-pretrained-BERT#command-line-interface). ", "Or use 300 usd credit for google cloud, that you get when you signup i believe.", "Thank you @oliverguhr and @Oxi84 for suggestions. \n\nI have tried both methods, using google colabs and GPU as runtime processor, it took about 240hours for every epoch (maybe if I use apex, it will be faster but I think still hundreds of hours), i think it's impossible to run google colabs dozens of days.\n\nI got free trial for GCP, unfortunately Google not provide GPU for free trial version. I try training use GCP with 2CPU and 13GB RAM, it take 200thousands of hours training, is sooo long time.\n\nMaybe I should reduce corpus size?\n\nThanks", "Or smaller vocabulary. I am pretty sure you can even use TPU on Google cloud, let someone else confirm that.", "@Oxi84 \r\nFor my classification task, I noticed that training the model with just 40 mb of data will give me already pretty good results. Training with the full 1,5 GB of my dataset improves the results by just 2-3% accuracy. \r\nSo you might start with a (random) subset of your data and improve the size step by step and see if your scores get better. ", "Oh nice insight @oliverguhr , thank you. I will try to reduce training data and train.\r\n" ]

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

[]

[ "Ah, thank you very much for this! I'll read over the paper and include it as a reference soon.", "Finally read it once I had some free time at the weekend and added PR #715. Thank you!", "Thank you @Rocketknight1 !", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=h1) Report\n> Merging [#691](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/commit/b3f9e9451b3f999118f2299229bb13f2f691c48f?src=pr&el=desc) will **increase** coverage by `0.16%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #691 +/- ##\n==========================================\n+ Coverage 67.08% 67.24% +0.16% \n==========================================\n Files 18 18 \n Lines 3846 3847 +1 \n==========================================\n+ Hits 2580 2587 +7 \n+ Misses 1266 1260 -6\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvX19pbml0X18ucHk=) | `100% <ø> (ø)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX3RyYW5zZm9feGwucHk=) | `32.59% <0%> (+0.18%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `83.51% <0%> (+0.53%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/optimization.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvb3B0aW1pemF0aW9uLnB5) | `74.26% <0%> (+0.73%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/tokenization.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uLnB5) | `92.69% <0%> (+0.91%)` | :arrow_up: |\n| [pytorch\\_pretrained\\_bert/file\\_utils.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvZmlsZV91dGlscy5weQ==) | `67.78% <0%> (+1.34%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=footer). Last update [b3f9e94...8289646](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/691?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks @vanche!" ]

[]

[]

[ "Perfect, thanks @shashwath94!" ]

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

[]

[ "Your server probably can't reach AWS to download the models.\r\nI need to make these error messages more clear, they currently gather several failure cases.\r\nWill do that in the coming release of next week.", "What should I do if I have downloaded the package manually?", "If you download them manually, you will have to figure out what is their file name in the cache. My problem was that on the cluster I'm using, worker nodes don't have access to the internet but the interactive nodes do. It's when I was tracing the code on an interactive node that I downloaded the models. Then, to run on a worker node, I define `export PYTORCH_PRETRAINED_BERT_CACHE=$BERT_MODEL_HOME/pytorch_pretrained_bert` which obviously is pointing to my cache.\r\n\r\nBack to downloading manually, you will have to \"properly\" name your model. Looking at my cache, I see\r\n`\r\na803ce83ca27fecf74c355673c434e51c265fb8a3e0e57ac62a80e38ba98d384.681017f415dfb33ec8d0e04fe51a619f3f01532ecea04edbfd48c5d160550d9c\r\n` which is actually `bert-base-cased.tar.gz`. The name of the model's file is base on some web id (sorry not familiar with this).", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "This looks great @Timoeller – do you have an estimate for the compute power you used to train your model?\r\n\r\nUPDATE. Ok the answer is in the blogpost: https://deepset.ai/german-bert\r\n> We trained using Google's Tensorflow code on a single cloud TPU v2 with standard settings. \r\n> We trained 840k steps with a batch size of 1024 for sequence length 128. Training took about 9 days. \r\n", "Sorry, I just realized we made a wrong oversimplification. We of course trained in the end for 30k steps on a longer batch size.\r\nI updated the article accordingly: We trained 810k steps with a batch size of 1024 for sequence length 128 and 30k steps with sequence length 512. Training took about 9 days. ", "Looks great, thanks a lot @Timoeller!", "@Timoeller (and @tholor also I guess): in the coming release 0.6.3, I'm switching to a split file format for Bert (like already done in GPT/GPT-2/Transformer-XL) in which we separately store config and weights files on the S3 to avoid having to untar an archive at each instantiation of the model.\r\n\r\nIn the short term I'll be storing your model's files on our s3 but you can also split the archive yourself and I can switch back to your s3 if you would like to.", "Correctly guessed. @tholor and me are working together. I also created another PR for the updated file locations.", "Hello, could you please share how did you generate the vocab list? ", "This is the code how we did it. There was a special \",\" symbol at index 0, which got used unintentionally as padding token. So we swapped the first two strings in the vocab.txt and created a \"[unused3001]\". See also the discussion in: https://github.com/huggingface/pytorch-transformers/issues/778\r\n\r\nHope that helps.\r\n\r\n```\r\n\r\nspm.SentencePieceTrainer.Train(\r\n f'--input={INPUT_FILE} --model_prefix={TEMP_FILE} --vocab_size={VOCAB_SIZE} --character_coverage=1.0 --model_type=bpe')\r\n\r\ndf = pd.read_csv(TEMP_FILE + \".vocab\", sep=\"\\t\", # use a char for separation that cannot be inside vocab\r\n header=None, names=[\"vocab\", \"unk\"],\r\n encoding='utf-8', dtype=str, quotechar=\"\\r\", # use a char for quoting that cannot be inside vocab\r\n engine='python')\r\nvocab = df.vocab.values\r\nprint(vocab.shape)\r\n\r\nprint(len(vocab))\r\nfor i, current in enumerate(vocab):\r\n current = str(current)\r\n if current.startswith(\"▁\"):\r\n vocab[i] = current[1:]\r\n else:\r\n vocab[i] = \"##\" + current\r\n\r\nunused = []\r\n\r\nfor i in range(1, UNUSED_TOKENS + 1):\r\n unused.append(\"[unused%i]\" % i)\r\ntoadd = np.array([\"[PAD]\", \"[UNK]\", \"[CLS]\", \"[SEP]\", \"[MASK]\"])\r\nvocab = np.concatenate((toadd, vocab[3:], unused), axis=0)\r\n```", "@Timoeller thanks for the open sourcing the model :+1: and the great [FARM](https://github.com/deepset-ai/FARM) library.\r\n\r\nI trained a German BERT model from scratch a while ago (16GB of text, incl. some WMT monolingual data for German) and here are some preliminary results:\r\n\r\n| Task | Result\r\n| --------- | -------\r\n| CoNLL-2003 | 85.49\r\n| GermEval | 84.38\r\n| GermEval18Coarse | 74.60 (reproduced result for German BERT was 74.06)\r\n\r\nSo on average the model is +0.48% better. My question to @Timoeller and @thomwolf does it make sense to include another cased German BERT model 🤔\r\n", "Hey @stefan-it \r\nThanks for linking our library and also sharing your results. We have made the experience that multiple downstream runs vary in performance by some degree. We are also in contact with people applying German Bert for germeval19, using an ensemble of multiple downstream runs with quite good results (the official ranking isnt out yet though). \r\n\r\nConcerning the performance of your Bert model: Your NER results seem to be consistently better than with our German Bert. Maybe it lacks behind in the other tasks? \r\n\r\nHow about we both have a call this week to see where the differences of our Berts come from and if it makes sense to include your model as well? Just pm me. ", "Awesome, I just send you an email :)" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=h1) Report\n> :exclamation: No coverage uploaded for pull request base (`master@cad88e1`). [Click here to learn what that means](https://docs.codecov.io/docs/error-reference#section-missing-base-commit).\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #687 +/- ##\n=========================================\n Coverage ? 67.14% \n=========================================\n Files ? 18 \n Lines ? 3847 \n Branches ? 0 \n=========================================\n Hits ? 2583 \n Misses ? 1264 \n Partials ? 0\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `79.39% <100%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=footer). Last update [cad88e1...44e9ddd](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=h1) Report\n> :exclamation: No coverage uploaded for pull request base (`master@cad88e1`). [Click here to learn what that means](https://docs.codecov.io/docs/error-reference#section-missing-base-commit).\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #687 +/- ##\n=========================================\n Coverage ? 67.14% \n=========================================\n Files ? 18 \n Lines ? 3847 \n Branches ? 0 \n=========================================\n Hits ? 2583 \n Misses ? 1264 \n Partials ? 0\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `79.39% <100%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=footer). Last update [cad88e1...44e9ddd](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/687?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

[]

[ "You would need an insertion-based transformer model like Google's recent KERMIT (http://arxiv.org/abs/1906.01604). But unfortunately, we currently don't have this model in the library.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Hi, I'm not convinced we need this additional option, see my [comment](https://github.com/huggingface/pytorch-pretrained-BERT/issues/676#issuecomment-502252962) in the associated issue thread.", "As mentioned in, https://github.com/huggingface/pytorch-pretrained-BERT/issues/676#issuecomment-506134506, I recognise that directly loading TF checkpoints is a rather niche use case and will close this PR.", "I don't think it's a niche but the `from_tf` option was there for importing from tf files so I would rather fix it to work in all cases rather than have several ways to import from a tf checkpoint.", "I'll have a look." ]

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

[]

[ "It should be fine. A bit of randomness in the pre-processing of the inputs is never bad when training a deep learning model.", "> It should be fine. A bit of randomness in the pre-processing of the inputs is never bad when training a deep learning model.\r\n\r\nI found the same problem that the implementation is different from tensorflow. But the key point is not the fixed 15% prob of all the token. If we use the implementation of pytorch will produce two extreme case especially for short sentences like article title,usually 10-20 characters.\r\ncase 1. sentence with too much '[MASK]'\r\ncase 2. sentence with none '[MASK]'\r\nboth case1 and case2 would cause the drop of performance. case 1 make the model difficult to predict and case2 would not produce the loss.\r\nGiven a corpus with an average sentence length of 10. The implementation of tensorflow would generate 1 '[MASK]' for the sentences, but the implementation of pytorch would have :\r\n0.85^10 = 0.19 to generate 0 '[MASK]'\r\n0.15 * 0.85^9 * 10 =0.34 to generate 1 '[MASK]'\r\n0.15^2 * 0.85^8 * 45 =0.27 to generate 2 '[MASK]'\r\n0.15^2 * 0.85^7 * 120 =0.12 to generate 3 '[MAKS]'\r\n... \r\n\r\nIf we roughly consider the sentence with 15% '[MASK]' is appropriate, only 1/2 '[MASK]' is useful for training models. So only 0.34 + 0.27 = 0.61 training case is useful.\r\n\r\nAnd we found it is this is a very serious problem for short text.\r\n\r\n\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[ "I got the solution.\r\n" ]

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

[]

[ "What do yuu get when you multiply the probabilities for words in these 2 places. Probability for b times probability for \"##oa\".\r\n['[CLS]', 'This', 'is', 'a', 'picture', 'of', 'a', '[MASK]', '[MASK]', '.']\r\n['[CLS]', 'This', 'is', 'a', 'picture', 'of', 'a', '[MASK]', '##oa', '.']\r\n\r\nBut there is also whole word masking model is realized by Google team, I hope it will be added here soon. It mask all the parts of these words at the same time, so it would give a better accuracy (1% for some tasks).\r\nhttps://github.com/google-research/bert\r\n\r\n", "Nothing interesting.\r\n\r\nMoreover, I would not like to mask words. I just want to get predictions with the word in clear.", "Well, another solution is to recreate bert with a large vocabulary. This would take days on TPU.\r\n", "Okay thanks for the answer", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Yes, 512 tokens for Bert.", "Thank you :) ", "Is there a way to bypass this limit? To increase the number of words?" ]

[]

[]

[ "They won't be able to help you if you don't provide a code for reproducing your issue, as this is not an expected behaviour.", "Thanks a lot for reminding. The issue is renewed with the code.", "That's true! I can reproduce it also on my computer... Really weird!", "You should use `model.eval()` to desactivate dropout like in the usage examples of the readme.", "It solves the problem, thanks!\r\n\r\n" ]

[]

[]

[ "Yes! We don't see that when we use the pre-trained model because the number of clusters is greater than zero anyway. Will fix.", "Thank you. I created a PR since it was a small bug. #690 " ]

[]

[]

[ "Is this what you want?\r\n\r\n```python\r\nPRETRAINED_MODEL_ARCHIVE_MAP = {\r\n 'bert-base-uncased': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz\",\r\n 'bert-large-uncased': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased.tar.gz\",\r\n 'bert-base-cased': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased.tar.gz\",\r\n 'bert-large-cased': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased.tar.gz\",\r\n 'bert-base-multilingual-uncased': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased.tar.gz\",\r\n 'bert-base-multilingual-cased': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased.tar.gz\",\r\n 'bert-base-chinese': \"https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese.tar.gz\",\r\n}\r\n```\r\n\r\nfrom https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling.py\r\n\r\nI feel like I'm not exactly understanding your question. They have hosted the pytorch dumps of the pretrained BERT models that were released by Google and hosted them on AWS. After downloading the model, what do you wish to do with it?", "Thank you for the answer.\r\n\r\nSo with your code, I now have the url. But, where should I put the files in my filesystem?\r\n\r\nSpacy provide a command to download the weights, put them at the good location in the filesystem, and use them. Is there an equivalent in your repository?\r\n\r\nFor example, [DeepPavlov](https://github.com/deepmipt/DeepPavlov) provides this command:\r\n\r\n```\r\npython -m deeppavlov install -d squad_bert\r\n```\r\n\r\nto install and download a model.", "Does my question make sense?", "Not really. What is the reason you are trying to do that?\r\nThis library will automatically download pre-trained weights, you don't need to do that yourself (even though you can).", "> Not really.\r\n\r\nDo you understand what Spacy & Deeppavlov enable to do? If yes, I am asking if there is something similar here. But, if you don't understand, it surely means it is not possible.\r\n\r\n> What is the reason you are trying to do that?\r\n\r\nBecause, when I put my code in production, I don't want to make the first query very long because it has to download the model.", "I happen to know quite well SpaCy (if you look at the Huggingface github, you will see we have developed a coreference resolution extension for SpaCy, [NeuralCoref](https://github.com/huggingface/neuralcoref), which interfaces directly with the cython internals of SpaCy) so I know the download process they use which is there mainly because they need the models to install as python packages (which we don't need to do here).\r\n\r\nYou actually shouldn't have to do anything special to avoid a long first query (and we don't do anything special at HuggingFace with the model in production) for the following reason:\r\nThe model weights are downloaded and cached when you instantiated the model for the first time and this should be done before the first query is even received. If you create and load the model at each query, you will experience a very heavy overhead, you should avoid that.\r\n\r\nIf you want to download the weights yourself (you can also do that) you will need to download the weights, configuration and vocabulary files manually from the url that @chrisgzf has pointed and put these in a folder. You can then load the model and tokenizer from that folder as indicated in the readme.", "Okay, thank you for your detailed and interesting answer. So there is not the feature I was asking for.\r\n\r\nAnd I miswrote: I didn't want to say \"I don't want to make the first query very long\", but rather \"I don't want to make the first server start very long\" for some reasons." ]

[]

[]

[ "Hello everyone,\r\n\r\nI have temporarily come up with a workaround for this. Not sure if it's the best solution but it works. What I did was I essentially merged what `load_tf_weights_in_bert()` and what part of `BertPreTrainedModel.from_pretrained()` was doing. `BertPreTrainedModel` is the parent class of `BertForTokenClassification`, so if you are trying to do something similar for `BertFor{TaskName}`, it should work too.\r\n\r\n```python\r\ndef load_BFTC_from_TF_ckpt(bert_config, ckpt_path, num_labels):\r\n config = BertConfig.from_json_file(bert_config)\r\n model = BertForPreTraining(config)\r\n load_tf_weights_in_bert(model, ckpt_path)\r\n state_dict=model.state_dict()\r\n model = BertForTokenClassification(config, num_labels=num_labels)\r\n\r\n # Load from a PyTorch state_dict\r\n old_keys = []\r\n new_keys = []\r\n for key in state_dict.keys():\r\n new_key = None\r\n if 'gamma' in key:\r\n new_key = key.replace('gamma', 'weight')\r\n if 'beta' in key:\r\n new_key = key.replace('beta', 'bias')\r\n if new_key:\r\n old_keys.append(key)\r\n new_keys.append(new_key)\r\n for old_key, new_key in zip(old_keys, new_keys):\r\n state_dict[new_key] = state_dict.pop(old_key)\r\n\r\n missing_keys = []\r\n unexpected_keys = []\r\n error_msgs = []\r\n # copy state_dict so _load_from_state_dict can modify it\r\n metadata = getattr(state_dict, '_metadata', None)\r\n state_dict = state_dict.copy()\r\n if metadata is not None:\r\n state_dict._metadata = metadata\r\n\r\n def load(module, prefix=''):\r\n local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})\r\n module._load_from_state_dict(\r\n state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)\r\n for name, child in module._modules.items():\r\n if child is not None:\r\n load(child, prefix + name + '.')\r\n start_prefix = ''\r\n if not hasattr(model, 'bert') and any(s.startswith('bert.') for s in state_dict.keys()):\r\n start_prefix = 'bert.'\r\n load(model, prefix=start_prefix)\r\n if len(missing_keys) > 0:\r\n logger.info(\"Weights of {} not initialized from pretrained model: {}\".format(\r\n model.__class__.__name__, missing_keys))\r\n if len(unexpected_keys) > 0:\r\n logger.info(\"Weights from pretrained model not used in {}: {}\".format(\r\n model.__class__.__name__, unexpected_keys))\r\n if len(error_msgs) > 0:\r\n raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n model.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n return model\r\n\r\nmodel = load_BFTC_from_TF_ckpt(CONFIG_FILE, \"/home/bert/pt_baseuncased/model.ckpt-98000\", num_labels)\r\n```", "Hi chrisgzf, thank you for the quick fix. I saw it's been integrated in the latest version.\r\n\r\nI am trying to do quite the same using **BertForSequenceClassification** but the \"**AttributeError**: 'BertForTokenClassification' object has no attribute 'bias'\" still shows up.\r\n\r\nAny ideas how I could use your fix for BertForSequenceClassification too ?\r\n\r\nThank you", "Hi @stormskidd, \r\n\r\nin my code snippet here (https://github.com/huggingface/pytorch-pretrained-BERT/issues/676#issuecomment-501526327), just change\r\n\r\n`model = BertForTokenClassification(config, num_labels=num_labels)`\r\nto\r\n`model = BertForSequenceClassification(config, num_labels=num_labels)`\r\n\r\nand it _should_ work.", "Thank you for the quick response.\r\nSo I tried this as you said :\r\n\r\n```\r\ndef load_BFTC_from_TF_ckpt(bert_config, ckpt_path, num_labels):\r\n config = BertConfig.from_json_file(bert_config)\r\n model = BertForSequenceClassification(config, num_labels=num_labels)\r\n load_tf_weights_in_bert(model, ckpt_path)\r\n state_dict=model.state_dict()\r\n model = **BertForSequenceClassification**(config, num_labels=num_labels)\r\n\r\n # Load from a PyTorch state_dict\r\n old_keys = []\r\n new_keys = []\r\n for key in state_dict.keys():\r\n new_key = None\r\n if 'gamma' in key:\r\n new_key = key.replace('gamma', 'weight')\r\n if 'beta' in key:\r\n new_key = key.replace('beta', 'bias')\r\n if new_key:\r\n old_keys.append(key)\r\n new_keys.append(new_key)\r\n for old_key, new_key in zip(old_keys, new_keys):\r\n state_dict[new_key] = state_dict.pop(old_key)\r\n\r\n missing_keys = []\r\n unexpected_keys = []\r\n error_msgs = []\r\n # copy state_dict so _load_from_state_dict can modify it\r\n metadata = getattr(state_dict, '_metadata', None)\r\n state_dict = state_dict.copy()\r\n if metadata is not None:\r\n state_dict._metadata = metadata\r\n\r\n def load(module, prefix=''):\r\n local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})\r\n module._load_from_state_dict(\r\n state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)\r\n for name, child in module._modules.items():\r\n if child is not None:\r\n load(child, prefix + name + '.')\r\n start_prefix = ''\r\n if not hasattr(model, 'bert') and any(s.startswith('bert.') for s in state_dict.keys()):\r\n start_prefix = 'bert.'\r\n load(model, prefix=start_prefix)\r\n if len(missing_keys) > 0:\r\n logger.info(\"Weights of {} not initialized from pretrained model: {}\".format(\r\n model.__class__.__name__, missing_keys))\r\n if len(unexpected_keys) > 0:\r\n logger.info(\"Weights from pretrained model not used in {}: {}\".format(\r\n model.__class__.__name__, unexpected_keys))\r\n if len(error_msgs) > 0:\r\n raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n model.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n return model\r\n```\r\n\r\nThen:\r\n\r\n```\r\nCONFIG_FILE = \"Bert/multi_cased_L-12_H-768_A-12/bert_config.json\"\r\nmodel = load_BFTC_from_TF_ckpt(CONFIG_FILE, \"model.ckpt-6032\", num_labels = 2)\r\n```\r\n\r\n\r\nBut still got the error:\r\n`\r\n<ipython-input-2-8b2ad52bd838> in load_BFTC_from_TF_ckpt(bert_config, ckpt_path, num_labels)\r\n 2 config = BertConfig.from_json_file(bert_config)\r\n 3 model = BertForSequenceClassification(config, num_labels=num_labels)\r\n----> 4 load_tf_weights_in_bert(model, ckpt_path)\r\n 5 state_dict=model.state_dict()\r\n 6 model = BertForSequenceClassification(config, num_labels=num_labels)\r\n\r\nC:\\ProgramData\\Anaconda3\\Lib\\site-packages\\pytorch_pretrained_bert\\modeling.py in load_tf_weights_in_bert(model, tf_checkpoint_path)\r\n 89 \r\n 90 elif l[0] == 'output_bias' or l[0] == 'beta':\r\n---> 91 pointer = getattr(pointer, 'bias')\r\n 92 \r\n 93 elif l[0] == 'output_weights':\r\n\r\nC:\\ProgramData\\Anaconda3\\Lib\\site-packages\\torch\\nn\\modules\\module.py in __getattr__(self, name)\r\n 537 return modules[name]\r\n 538 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n--> 539 type(self).__name__, name))\r\n 540 \r\n 541 def __setattr__(self, name, value):\r\n\r\nAttributeError: 'BertForSequenceClassification' object has no attribute 'bias''`\r\n\r\n\r\n\r\n\r\nAlso, I wandered in the source code where the error occurs :\r\n```\r\n\r\n 90 elif l[0] == 'output_bias' or l[0] == 'beta':\r\n---> 91 pointer = getattr(pointer, 'bias')\r\n```\r\n\r\n\r\nI tried to change getattr(pointer, 'bias') to getattr(pointer, 'beta') but then I got a slightly different error:\r\n\r\n```\r\nC:\\ProgramData\\Anaconda3\\Lib\\site-packages\\pytorch_pretrained_bert\\modeling.py in load_tf_weights_in_bert(model, tf_checkpoint_path)\r\n 89 \r\n 90 elif l[0] == 'output_bias' or l[0] == 'beta':\r\n---> 91 pointer = getattr(pointer, 'beta')\r\n 92 \r\n 93 elif l[0] == 'output_weights':\r\n\r\nC:\\ProgramData\\Anaconda3\\Lib\\site-packages\\torch\\nn\\modules\\module.py in __getattr__(self, name)\r\n 537 return modules[name]\r\n 538 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n--> 539 type(self).__name__, name))\r\n 540 \r\n 541 def __setattr__(self, name, value):\r\n\r\nAttributeError: 'BertLayerNorm' object has no attribute 'beta'\r\n```\r\n\r\nHope it helps. Please let me know you think of any workaround for this !\r\nBy the way, I'm on windows using Anaconda with Python 3.7.1\r\n\r\nGreetings,\r\nMaxime", "Hi Maxime (@stormskidd),\r\n\r\nDo read my comment carefully. Change\r\n\r\n`model = BertForTokenClassification(config, num_labels=num_labels)`\r\nto\r\n`model = BertForSequenceClassification(config, num_labels=num_labels)`\r\n\r\nPlease leave `model = BertForPreTraining(config)` (line 3) as is, and do not change it.\r\n\r\nEdit: you might want to check out #685 as well. I submitted a PR to make it easier to do something we are trying to do. Maybe the code examples will make it clearer to you.\r\n\r\nIn line 3 of the function body, the model has to be an instance of `BertForPreTraining` because that's what `load_tf_weights_in_bert()` is designed to work with. I'm not sure if what I'm doing is the proper way or just a jank way, but I'm just trying to copy out the state_dict after the weights are imported as a `BertForPreTraining` instances, then creating a brand new `BertFor[TaskName]` instance, and loading the state_dict into it.\r\n\r\nHope this clears things up.\r\n\r\nCheers.", "Oh, you're right I posted the wrong version of the many things I tried :(\r\n\r\nAnyway, I'm afraid the error still shows up with the following:\r\n\r\n```\r\ndef load_BFTC_from_TF_ckpt(bert_config, ckpt_path, num_labels):\r\n config = BertConfig.from_json_file(bert_config)\r\n model = BertForPreTraining(config)\r\n load_tf_weights_in_bert(model, ckpt_path)\r\n state_dict=model.state_dict()\r\n model = BertForSequenceClassification(config, num_labels=num_labels)\r\n\r\n # Load from a PyTorch state_dict\r\n old_keys = []\r\n new_keys = []\r\n for key in state_dict.keys():\r\n new_key = None\r\n if 'gamma' in key:\r\n new_key = key.replace('gamma', 'weight')\r\n if 'beta' in key:\r\n new_key = key.replace('beta', 'bias')\r\n if new_key:\r\n old_keys.append(key)\r\n new_keys.append(new_key)\r\n for old_key, new_key in zip(old_keys, new_keys):\r\n state_dict[new_key] = state_dict.pop(old_key)\r\n\r\n missing_keys = []\r\n unexpected_keys = []\r\n error_msgs = []\r\n # copy state_dict so _load_from_state_dict can modify it\r\n metadata = getattr(state_dict, '_metadata', None)\r\n state_dict = state_dict.copy()\r\n if metadata is not None:\r\n state_dict._metadata = metadata\r\n\r\n def load(module, prefix=''):\r\n local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})\r\n module._load_from_state_dict(\r\n state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)\r\n for name, child in module._modules.items():\r\n if child is not None:\r\n load(child, prefix + name + '.')\r\n start_prefix = ''\r\n if not hasattr(model, 'bert') and any(s.startswith('bert.') for s in state_dict.keys()):\r\n start_prefix = 'bert.'\r\n load(model, prefix=start_prefix)\r\n if len(missing_keys) > 0:\r\n logger.info(\"Weights of {} not initialized from pretrained model: {}\".format(\r\n model.__class__.__name__, missing_keys))\r\n if len(unexpected_keys) > 0:\r\n logger.info(\"Weights from pretrained model not used in {}: {}\".format(\r\n model.__class__.__name__, unexpected_keys))\r\n if len(error_msgs) > 0:\r\n raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n model.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n return model\r\n```\r\n\r\nThen:\r\n\r\n\r\n```\r\nCONFIG_FILE = \"/Bert/multi_cased_L-12_H-768_A-12/bert_config.json\"\r\nmodel = load_BFTC_from_TF_ckpt(CONFIG_FILE, \"model.ckpt-6032\", num_labels = 2)\r\n```\r\n\r\n\r\nAnd :\r\n```\r\nC:\\ProgramData\\Anaconda3\\Lib\\site-packages\\pytorch_pretrained_bert\\modeling.py in load_tf_weights_in_bert(model, tf_checkpoint_path)\r\n 89 \r\n 90 elif l[0] == 'output_bias' or l[0] == 'beta':\r\n---> 91 pointer = getattr(pointer, 'bias')\r\n 92 \r\n 93 elif l[0] == 'output_weights':\r\n\r\nC:\\ProgramData\\Anaconda3\\Lib\\site-packages\\torch\\nn\\modules\\module.py in __getattr__(self, name)\r\n 537 return modules[name]\r\n 538 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n--> 539 type(self).__name__, name))\r\n 540 \r\n 541 def __setattr__(self, name, value):\r\n\r\nAttributeError: 'BertForPreTraining' object has no attribute 'bias'\r\n```\r\n\r\nGreetings,\r\nMax", "@stormskidd,\r\n\r\nthis is odd.... I actually specifically tested my code above on `BertForSequenceClassification` as well and I was able to successfully import TF weights. The code snippet looks like it would work...\r\n\r\nJust want to check:\r\n- are you running the latest pytorch-pretrained-BERT from upstream master?\r\n- did you make any other changes to the source?\r\n- have you tried https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py and successfully converted your TF ckpt to a pytorch dump? if my code snippet above doesn't work for you, this should at least work for you.\r\n- not sure if this matters, but are you using CUDA? and do you have apex installed?\r\n\r\nDo let me know if you face any errors. I'm curious about this issue. Btw, check out my edit in the previous comment.\r\n\r\nChris", "Hi, not sure I fully understand the issue here. What is the kind of tensorflow checkpoint you are trying to convert? Is it a pretrained model like the original Bert checkpoints or is it a fine-tuned model with additional elements (like a classification layer on top)?", "Hi @thomwolf,\r\n\r\nI am trying to convert a pretrained model like the original Bert checkpoints, except that I did additional pretraining on top of the released models with `run_pretraining.py` from the BERT repo. I then wish to fine-tune these pretrained models in pytorch, which is why I had to do this conversion. I am not importing any fine-tuned TF models.", "Ok so you can just convert your Tensorflow model using the command line script (see [here in the readme](https://github.com/huggingface/pytorch-pretrained-BERT#Command-line-interface)) store the converted pytorch model in a folder with the configuration file and then load it in a `BertForTokenClassification` model as follow:\r\n`BertForTokenClassification.from_pretrained('PATH_TO_YOUR_CONVERTED_MODEL_FOLDER')`\r\n\r\nFor the tokenizer, you can use the one associated to the original TensorFlow model from which you did the fine-tuning since you probably didn't change the vocabulary itself.", "Hi guys,\r\nI saw your discussion and it gave me the idea to try the following:\r\nIt seems like I can use both the function 'load_BFTC_from_TF_ckpt' and the script 'convert_tf_checkpoint_to_pytorch.py' to load a pretrained Google model : multi_cased_L-12_H-768_A-12\\bert_model.ckpt.\r\n\r\nHowever the error occurs when I tried to do the same with a fined-tuned model (from Bert script run_classifier.py ran on GluON/MRPC -like data).\r\nIs there any ways I can load a TF checkpoint fine-tuned model directly in Pytorch ? Or do I have to re-finetune it with the pytorch_pretrained_bert library ?\r\n\r\nThank you !\r\nMax", "> Oh, you're right I posted the wrong version of the many things I tried :(\r\n> \r\n> Anyway, I'm afraid the error still shows up with the following:\r\n> \r\n> ```\r\n> def load_BFTC_from_TF_ckpt(bert_config, ckpt_path, num_labels):\r\n> config = BertConfig.from_json_file(bert_config)\r\n> model = BertForPreTraining(config)\r\n> load_tf_weights_in_bert(model, ckpt_path)\r\n> state_dict=model.state_dict()\r\n> model = BertForSequenceClassification(config, num_labels=num_labels)\r\n> \r\n> # Load from a PyTorch state_dict\r\n> old_keys = []\r\n> new_keys = []\r\n> for key in state_dict.keys():\r\n> new_key = None\r\n> if 'gamma' in key:\r\n> new_key = key.replace('gamma', 'weight')\r\n> if 'beta' in key:\r\n> new_key = key.replace('beta', 'bias')\r\n> if new_key:\r\n> old_keys.append(key)\r\n> new_keys.append(new_key)\r\n> for old_key, new_key in zip(old_keys, new_keys):\r\n> state_dict[new_key] = state_dict.pop(old_key)\r\n> \r\n> missing_keys = []\r\n> unexpected_keys = []\r\n> error_msgs = []\r\n> # copy state_dict so _load_from_state_dict can modify it\r\n> metadata = getattr(state_dict, '_metadata', None)\r\n> state_dict = state_dict.copy()\r\n> if metadata is not None:\r\n> state_dict._metadata = metadata\r\n> \r\n> def load(module, prefix=''):\r\n> local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})\r\n> module._load_from_state_dict(\r\n> state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs)\r\n> for name, child in module._modules.items():\r\n> if child is not None:\r\n> load(child, prefix + name + '.')\r\n> start_prefix = ''\r\n> if not hasattr(model, 'bert') and any(s.startswith('bert.') for s in state_dict.keys()):\r\n> start_prefix = 'bert.'\r\n> load(model, prefix=start_prefix)\r\n> if len(missing_keys) > 0:\r\n> logger.info(\"Weights of {} not initialized from pretrained model: {}\".format(\r\n> model.__class__.__name__, missing_keys))\r\n> if len(unexpected_keys) > 0:\r\n> logger.info(\"Weights from pretrained model not used in {}: {}\".format(\r\n> model.__class__.__name__, unexpected_keys))\r\n> if len(error_msgs) > 0:\r\n> raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n> model.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n> return model\r\n> ```\r\n> \r\n> Then:\r\n> \r\n> ```\r\n> CONFIG_FILE = \"/Bert/multi_cased_L-12_H-768_A-12/bert_config.json\"\r\n> model = load_BFTC_from_TF_ckpt(CONFIG_FILE, \"model.ckpt-6032\", num_labels = 2)\r\n> ```\r\n> \r\n> And :\r\n> \r\n> ```\r\n> C:\\ProgramData\\Anaconda3\\Lib\\site-packages\\pytorch_pretrained_bert\\modeling.py in load_tf_weights_in_bert(model, tf_checkpoint_path)\r\n> 89 \r\n> 90 elif l[0] == 'output_bias' or l[0] == 'beta':\r\n> ---> 91 pointer = getattr(pointer, 'bias')\r\n> 92 \r\n> 93 elif l[0] == 'output_weights':\r\n> \r\n> C:\\ProgramData\\Anaconda3\\Lib\\site-packages\\torch\\nn\\modules\\module.py in __getattr__(self, name)\r\n> 537 return modules[name]\r\n> 538 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n> --> 539 type(self).__name__, name))\r\n> 540 \r\n> 541 def __setattr__(self, name, value):\r\n> \r\n> AttributeError: 'BertForPreTraining' object has no attribute 'bias'\r\n> ```\r\n> \r\n> Greetings,\r\n> Max\r\n\r\nyou can try:\r\nmodel = BertForPreTraining.from_pretrained(BERT_DIR, from_tf=True)", "Hello guys,\r\nhere is something I found that seem to do the trick for now:\r\n\r\nin `modeling.py:`\r\n\r\nIn class `class BertForSequenceClassification(BertPreTrainedModel):`\r\n\r\nAdd :\r\n```\r\n self.weight = Variable(torch.ones(2, config.hidden_size), requires_grad=True) \r\n self.bias = Variable(torch.ones(2), requires_grad=True)\r\n```\r\nto the attributes.\r\n\r\nObviously juste change the right class regarding your needs. I'll let you know if it provokes another error, but for now I can load in memory the trained model I couldn't load before.\r\n\r\nGreetings,\r\nMax\r\n", "if use BertFor* class, will not initiate classifier layer.\r\nwhen i load tf model for predict/evaluate modify **load_tf_weights_in_bert**\r\n\r\n\r\n```\r\n if re.fullmatch(r'[A-Za-z]+_\\d+', m_name):\r\n l = re.split(r'_(\\d+)', m_name)\r\n else:\r\n l = [m_name]\r\n if l[0] == 'kernel' or l[0] == 'gamma':\r\n pointer = getattr(pointer, 'weight')\r\n elif l[0] == 'output_bias' or l[0] == 'beta':\r\n if pointer == model:\r\n pointer = getattr(pointer, 'classifier')\r\n pointer = getattr(pointer, 'bias')\r\n elif l[0] == 'output_weights':\r\n if pointer == model:\r\n pointer = getattr(pointer, 'classifier')\r\n pointer = getattr(pointer, 'weight')\r\n elif l[0] == 'squad':\r\n pointer = getattr(pointer, 'classifier')\r\n```\r\n", "> \r\n> \r\n> Ok so you can just convert your Tensorflow model using the command line script (see [here in the readme](https://github.com/huggingface/pytorch-pretrained-BERT#Command-line-interface)) store the converted pytorch model in a folder with the configuration file and then load it in a `BertForTokenClassification` model as follow:\r\n> `BertForTokenClassification.from_pretrained('PATH_TO_YOUR_CONVERTED_MODEL_FOLDER')`\r\n> \r\n> For the tokenizer, you can use the one associated to the original TensorFlow model from which you did the fine-tuning since you probably didn't change the vocabulary itself.\r\n\r\nHello @thomwolf,\r\n\r\nYes, I am aware of the conversion script and `from_pretrained()` being able to load full models (PyTorch dumps) from converted TF checkpoints. However, in my use case, I did pre-training with BERT using the script `run_pretraining.py` from the BERT repo, and I wanted to do fine-tuning on the many checkpoint steps that I have saved, so it would make more sense for me to load the checkpoints directly.\r\n\r\nHowever, I am aware that my use case is a very niche one, and the others here are talking about a different use case (loading TF finetuned models). Since my issue has been resolved, I will close this issue." ]

[]

[]

[ "Thanks @meetshah1995 " ]

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

[]

[ "I've tried a bit to play with these training schemes on a deep transformer for our [tutorial on Transfer Learning in Natural Language Processing](https://naacl2019.org/program/tutorials/#t4-transfer-learning-in-natural-language-processing) held at NAACL last week but I couldn't get gradual unfreezing and discriminative fine-tuning to out-perform a standard fine-tuning procedure (multi-tasking did help, however).\r\n\r\nYou can have a look at the results by reading the \"Hands-on\" parts of the tutorial here: https://tinyurl.com/NAACLTransfer.\r\n\r\nYou can give it a try your-self with the associated Colab notebook which is here: https://tinyurl.com/NAACLTransferColab (and a full stand-alone codebase is here: https://tinyurl.com/NAACLTransferCode).\r\n\r\nIt's possible that I just didn't spend enough time scanning the hyper-parameters or maybe these two training variants are better suited to LSTM than Transformers.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> I've tried a bit to play with these training schemes on a deep transformer for our [tutorial on Transfer Learning in Natural Language Processing](https://naacl2019.org/program/tutorials/#t4-transfer-learning-in-natural-language-processing) held at NAACL last week but I couldn't get gradual unfreezing and discriminative fine-tuning to out-perform a standard fine-tuning procedure (multi-tasking did help, however).\r\n> \r\n> You can have a look at the results by reading the \"Hands-on\" parts of the tutorial here: https://tinyurl.com/NAACLTransfer.\r\n> \r\n> You can give it a try your-self with the associated Colab notebook which is here: https://tinyurl.com/NAACLTransferColab (and a full stand-alone codebase is here: https://tinyurl.com/NAACLTransferCode).\r\n> \r\n> It's possible that I just didn't spend enough time scanning the hyper-parameters or maybe these two training variants are better suited to LSTM than Transformers.\r\n\r\nI find the standard fine-tuning procedure having a unstable issue, that different shuffle order affects a lot. I wander if unfreezing help the bert finetune get a relative stable result?\r\nThis is also mentioned in bert paper, they just use different random seeds to get a best result." ]

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

[]

[ "To your first question, the inputs will be almost identical, but the token_type_ids argument will be unused, as this is the vector that indicates the split between the two 'sentences' for the NextSentence objective. I'm not familiar with that part of the code - you might be able to just pass `None` for that argument, or maybe you'll need to pass a vector of zeros that has the right shape (i.e. the same length as the sequence). Also, I believe pre-training with just the masked LM objective works okay - it's by far the most important of the two objectives. The BERT paper didn't show any ablation studies, so I don't know what the effect of removing it will be, but I wouldn't worry too much for a fine-tuning task.\r\n\r\nSecondly, I wouldn't expect a large benefit from just fine-tuning the LM on a small labelled training corpus. The main benefit arises when you have a large corpus to fine-tune the LM on, but only a small fraction of it is labelled. \r\n\r\nAlso, I would guess that fine-tuning the LM on the dev/test examples will probably result in test performance that is optimistic compared to the performance on truly unseen data, but I'm not aware of any research in that area. However, this suggests a potential research direction - if you try including the dev/test data in the LM pre-training task and find that it significantly improves dev/test accuracy, compared to text that was not included in the LM pre-training task, that would be an interesting approach to improving the performance of language models!\r\n\r\nIt would be tricky to take advantage of this effect in practice, but you can imagine ways it might be done - for example, at inference time it might be possible to add new inputs to the LM fine-tuning corpus, and then fine-tune your language model on them followed by retraining the classifier (with your pre-existing labelled data) and only then labelling the new inputs! This would probably be too computationally expensive to be used in many production systems, especially when low latency is required, but for some purposes the improvement in data efficiency and accuracy might be worthwhile.", "Great, thank you!\r\n\r\nSo it's understood that language modeling (or masked LM) is an effective pre-training objective for learning a general representation of the language. But do you think it makes sense to use a MLM head during the fine-tuning phase? For example, if your main task of interest is sentence classification, perhaps you could fine-tune the pre-trained model on a sentence classification head as well as a LM head in a multi-task learning setting. Intuitively it would have a regularizing effect and potentially lead to better generalization. I'd love to hear your thoughts on that idea!", "Take a look at issue #692 , there's a recent Chinese paper where they tried fine-tuning with domain text as well as doing something very similar to what you're proposing (multi-task fine-tuning) and reported good results.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> \r\n> \r\n> Hello,\r\n> \r\n> I'm thinking about fine-tuning a BERT model using only the Masked LM pre-training objective, and I'd appreciate a bit of guidance. The most straightforward way is probably to modify the simple_lm_finetuning.py script to only do LM fine-tuning. Besides importing BertForMaskedLM instead of BertForPretraining (which has both objectives), what changes should I make, and what potential problems should I consider?\r\n> \r\n> Also, would it make sense to do MLM fine-tuning on a relatively small domain-specific corpus consisting of just the training examples from the datasets? In other words, does it make sense to do LM pretraining on a corpus that includes training examples taken from the datasets for which you want to make downstream predictions? I'm assuming including the dev/test examples in the corpus is out of the question since it'll likely overfit, but what about just the training examples? I'd like hear to people's thoughts about it.\r\n> \r\n> Lastly, was the run_lm_finetuning.py script replaced by the contents in the examples/lm_finetuning directory? If so, there are still 2 references to that script in the README that should be edited.\r\n> \r\n> @Rocketknight1\r\n\r\nHello @dchang56\r\n\r\nDid you try fine-tuning a BERT model using only the Masked LM pre-training objective ?\r\n\r\nWas that relatively as good as doing that along with next sentence prediction ?\r\n\r\nHow much was your block size (maximum sequence length) for that ?", "> \r\n> \r\n> To your first question, the inputs will be almost identical, but the token_type_ids argument will be unused, as this is the vector that indicates the split between the two 'sentences' for the NextSentence objective. I'm not familiar with that part of the code - you might be able to just pass `None` for that argument, or maybe you'll need to pass a vector of zeros that has the right shape (i.e. the same length as the sequence). Also, I believe pre-training with just the masked LM objective works okay - it's by far the most important of the two objectives. The BERT paper didn't show any ablation studies, so I don't know what the effect of removing it will be, but I wouldn't worry too much for a fine-tuning task.\r\n> \r\n> Secondly, I wouldn't expect a large benefit from just fine-tuning the LM on a small labelled training corpus. The main benefit arises when you have a large corpus to fine-tune the LM on, but only a small fraction of it is labelled.\r\n> \r\n> Also, I would guess that fine-tuning the LM on the dev/test examples will probably result in test performance that is optimistic compared to the performance on truly unseen data, but I'm not aware of any research in that area. However, this suggests a potential research direction - if you try including the dev/test data in the LM pre-training task and find that it significantly improves dev/test accuracy, compared to text that was not included in the LM pre-training task, that would be an interesting approach to improving the performance of language models!\r\n> \r\n> It would be tricky to take advantage of this effect in practice, but you can imagine ways it might be done - for example, at inference time it might be possible to add new inputs to the LM fine-tuning corpus, and then fine-tune your language model on them followed by retraining the classifier (with your pre-existing labelled data) and only then labelling the new inputs! This would probably be too computationally expensive to be used in many production systems, especially when low latency is required, but for some purposes the improvement in data efficiency and accuracy might be worthwhile.\r\n\r\nHello @Rocketknight1\r\n\r\nwhy do you think pre-training with just the masked LM objective works okay ?\r\n\r\nis there any article or study about that , have you got any link of it ?\r\n\r\nOr, have you down any training (BERT) without NSP (next sentence prediction) , yourself ?", "> LM fine-tuning\r\n\r\nCould be useful to you: https://colab.research.google.com/github/huggingface/notebooks/blob/master/examples/language_modeling.ipynb" ]

[]

[]

[ "Nice indeed, thanks @oliverguhr!" ]

[]

[]

[]

[]

[]

[ "Because we don't use BertAdam in fp16 mode but the optimizer of NVIDIA's apex library.", "OK thank you!" ]

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

[]

[ "Perhaps it's the problem of tokenizer...After stripping the lengths of two sequences changed, so `orig_text` is returned...", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Nice, thanks @jeonsworld " ]

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

[]

[ "Yes, probably overflow. Try a smaller batch size?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Have you tried the provided GPT-2 generation example? It's here: https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/examples/run_gpt2.py", "> Have you tried the provided GPT-2 generation example? It's here: https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/examples/run_gpt2.py\r\n\r\nHey, I encountered the same issue. I tried using the example you provided but it tends to produce repetitive text much more often than earlier versions of the library as well (from around 1-2 months back). Thank you very much for all the work! ", "> Have you tried the provided GPT-2 generation example? It's here: https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/examples/run_gpt2.py\r\n\r\nI tried the example. It is working properly. I think with `torch.argmax` there is chance of repetitive text generation. If we sample using `torch.multinomial`, there is always some variation.\r\n\r\n\r\n", "`torch.argmax` is basically top_k with 1 which is very bad for creating \"human like\" sentences. A better way to sample is using Nucleus Sampling [https://arxiv.org/abs/1904.09751](url).\r\nNot sure whether this is implemented in PyTorch yet.\r\nEDIT: I have found this code from @thomwolf [https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317](url) that implements Nucleus sampling.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", " generating repetitive text when using GPU, but this not happens when using CPU, anyone knows how to solve this weird issue?", "> when using GPU, but this not happens when using CPU, anyone knows how to solve this weird issue?\r\n\r\nI am facing the same issue. Can someone lead in this problem? " ]

[]

[]

[ "Take a look at the `attention` branch @g-karthik:\r\n\r\nhttps://github.com/huggingface/pytorch-pretrained-BERT/blob/attention/pytorch_pretrained_bert/modeling_gpt2.py#L42-L45", "Thanks @julien-c, I had not looked at the file in the `attention` branch!", "What is the recommended hardware setup for fine-tuning GPT2 medium?" ]

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

[]

[ "No padding implemented in GPT-2, you have to add implement your-self if you want e.g. by adding a special token but note that:\r\n- GPT-2 doesn't like left side padding (doesn't mix well with a causal transformer having absolute positions)\r\n- right-side padding is often not necessary (the causal mask means that right context is ignored anyway).", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> * GPT-2 doesn't like left side padding (doesn't mix well with a causal transformer having absolute positions)\r\n\r\nTo those stumbling on this issue, this doesn't seem to be a problem anymore. [#3021](https://github.com/huggingface/transformers/issues/3021#issuecomment-1232149031)\r\n\r\n" ]

[]

[]

[]

[]

[]

[ "Update: specify device works for at least 1 GPU\r\n\r\nexport CUDA_VISIBLE_DEVICES=0\r\npython run_classifier.py \\\r\n\r\n\r\nmore than 1 GPU still not working:\r\n\r\nexport CUDA_VISIBLE_DEVICES=0,1\r\npython run_classifier.py \\\r\n", "@AndreasFdev Your distributed training setting is False.", "Problem:\r\nP2P GPU traffic fails with enabled IOMMU, unless the cards are behind PLX switch.\r\n\r\nSolution:\r\nTo disable IOMMU edit /etc/default/grub \r\n#GRUB_CMDLINE_LINUX=\"\" <----- Original commented\r\nGRUB_CMDLINE_LINUX=\"iommu=soft\" <------ Change\r\n\r\nSource:\r\nhttps://github.com/pytorch/pytorch/issues/1637\r\n\r\nThanks to all who had a look" ]

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

[]

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

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "@okgrammer Larger batch size often means lower accuracy but faster epochs. You can try it by doing several runs of varying batch size while keeping other params constant. \r\n\r\nSee, especially, https://arxiv.org/pdf/1804.07612.pdf", "> In the original paper, BERT model is fine-tuned on downstream NLP tasks, where the number of instances for each task is in the order of thousands to hundreds of thousands. In my case, I have about 5 million samples. I'm curious whether there are recommended batch size and epochs for such training size? I'm fine-tuning bert-base-multilingual on 4 GPUs and there is a lot of unused GPU memory with the default batch size of 32. Even after increasing it to 128 there is still free available memory.\r\n\r\nI have exactly the same issue. Can anyone help? \r\nThe pretraining is really slow with more than 90% GPU memory available. No matter how I increase the batch size, the GPU memory usage is minimal." ]

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

[]

[ "It's not yet but thanks for the pointer, we can probably add it fairly easily. I'll have a look.", "+10000 This would be very helpful!", "Hi, \r\n\r\nI converted the cased and uncased whole-word-masking models using the command line tool. If you're interested in adding these to the repository, I've uploaded them to [this](https://www.kaggle.com/bkkaggle/bert-large-whole-word-masking) kaggle dataset. ", "Is this resolved? These seem to be available at head, and I don't see anything immediately wrong when I try them...", "Yes they are working fine, I've added them to master last week.\r\nThey will be advertised in the next release.\r\nWhen fine-tuned with run_squad they give pretty nice results: `exact_match: 86.91, f1: 93.15`.\r\nI've included a version fine-tuned on SQuAD as well.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Does Whole Word Masking support bert base as well?" ]

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

[]

[ "I took example code from https://github.com/huggingface/pytorch-pretrained-BERT#fine-tuning-bert-large-on-gpus (which has additional option `--loss_scale 128`). Still getting very low test scores:\r\n```\r\n$ python evaluate-v1.1.py dev-v1.1.json ../output/debug_squad_fp16/predictions.json \r\n{\"exact_match\": 0.5771050141911069, \"f1\": 8.853750220358535}\r\n```\r\n\r\nIs there any known bug in current (on `v0.6.2`) PyTorch BERT implementation or is this my setup?\r\n\r\n---\r\n\r\n**UPDATE**\r\nProbably there is a bug in current stable `v0.6.2` version. When running same command on latest `master` branch I'm getting good results:\r\n```\r\n$ python evaluate-v1.1.py dev-v1.1.json ../output/debug_squad_fp16/predictions.json\r\n{\"exact_match\": 81.45695364238411, \"f1\": 88.71433452234619}\r\n```", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Did anyone conduct different learning rate in different layers when fine-tuning BERT?\r\n\r\nThanks. ", "[This paper](https://arxiv.org/abs/1905.05583) suggests a \"layer-wise decreasing learning rate\" improves BERT for text classification tasks, but I can't see any option for it in the [AdamW](https://huggingface.co/transformers/main_classes/optimizer_schedules.html#adamw-pytorch) optimizer. If it is helpful for text classification, it would be great to see it implemented.", "Did anyone conduct different learning rate in different layers when fine-tuning BERT?\r\n\r\nThanks", "You can do this in PyTorch (not sure about tf) and my understanding from this thread is that `huggingface`'s `AdamW` is now equivalent to PT's `AdamW` (see https://github.com/huggingface/transformers/issues/3407) so it should be equivalent - it would be great to get confirmation of this from someone more familiar with the huggingface codebase. \r\n\r\nSee here for PT multiple rates: https://discuss.pytorch.org/t/how-to-set-a-different-learning-rate-for-a-single-layer-in-a-network/48552/4\r\n\r\n", "As an update to the above - it actually _is_ possible to use the `huggingface` `AdamW` directly with different learning rates. \r\n\r\nSay you wanted to train your new parameters at x10 the learning rate of the pre-trained bert-variant parameters (in this case held as `model.bert`) you would do:\r\n```python\r\nfrom transformers import AdamW\r\n# define model etc.\r\n...\r\n\r\npretrained = model.bert.parameters()\r\n# Get names of pretrained parameters (including `bert.` prefix)\r\npretrained_names = [f'bert.{k}' for (k, v) in model.bert.named_parameters()]\r\n\r\nnew_params= [v for k, v in model.named_parameters() if k not in pretrained_names]\r\n\r\noptimizer = AdamW(\r\n [{'params': pretrained}, {'params': new_params, 'lr': learning_rate * 10}],\r\n lr=learning_rate,\r\n)\r\n```" ]

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

[]

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

[]

[]

[]

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

[]

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

[]

[]

[ "Hi, no we didn't change the weights. Can you share a sample on which the results are different?", "Hi @thomwolf , thanks for your quick reply. I found even version 0.4.0 is different to version 0.2.0 and 0.3.0. I trained the model on v0.4.0, and then I tried to load the model using v0.2.0, here is the mismatch of keys:\r\n\r\n```\r\nRuntimeError: Error(s) in loading state_dict for BertForSequenceClassification:\r\n\r\nMissing key(s) in state_dict: \"bert.embeddings.LayerNorm.gamma\", \"bert.embeddings.LayerNorm.beta\",...\r\n\r\nUnexpected key(s) in state_dict: \"bert.embeddings.LayerNorm.weight\", \"bert.embeddings.LayerNorm.bias\",...\r\n```\r\nThis is also appears when I trained the model on v0.5.0+ and tried to load the model using v0.2.0 and v0.3.0.\r\n\r\nHowever, the first step loss are the same to v0.2.0, v0.3.0 and v0.4.0, loss=0.7228, but the first step loss are different to v0.5.0+ whose loss is 0.7091. And the final converge results are different too. I have set a seed to reproduce the results.", "Oh, sorry, I found the mismatch problem is from my loading scripts (the 2nd method in [Serialization best-practices](https://github.com/huggingface/pytorch-pretrained-BERT#serialization-best-practices)), because I didn't use the mapping from old_keys to new_keys as you did in 'from_pretrained()' function.", "Hi @thomwolf , I have found where the differences are. It is the different 'init_bert_weights' that makes the results different.\r\n\r\nIn version 0.2.0, 0.3.0 and 0.4.0, you use 'normal_' to initialize the 'BertLayerNorm':\r\n\r\n```\r\ndef init_bert_weights(self, module):\r\n \"\"\" Initialize the weights.\r\n \"\"\"\r\n if isinstance(module, (nn.Linear, nn.Embedding)):\r\n # Slightly different from the TF version which uses truncated_normal for initialization\r\n # cf https://github.com/pytorch/pytorch/pull/5617\r\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\r\n elif isinstance(module, BertLayerNorm):\r\n module.beta.data.normal_(mean=0.0, std=self.config.initializer_range)\r\n module.gamma.data.normal_(mean=0.0, std=self.config.initializer_range)\r\n if isinstance(module, nn.Linear) and module.bias is not None:\r\n module.bias.data.zero_()\r\n```\r\n\r\nbut in version 0.5.0+, you use 'zeros_' and 'ones_' to initialize 'BertLayerNorm':\r\n\r\n```\r\ndef init_bert_weights(self, module):\r\n \"\"\" Initialize the weights.\r\n \"\"\"\r\n if isinstance(module, (nn.Linear, nn.Embedding)):\r\n # Slightly different from the TF version which uses truncated_normal for initialization\r\n # cf https://github.com/pytorch/pytorch/pull/5617\r\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\r\n elif isinstance(module, BertLayerNorm):\r\n module.bias.data.zero_()\r\n module.weight.data.fill_(1.0)\r\n if isinstance(module, nn.Linear) and module.bias is not None:\r\n module.bias.data.zero_()\r\n```\r\n\r\nBy the way, after correct mapping from old_keys to new_keys, the old version pertained model could be loaded by new version with no different results! Thank you for sharing such great work with us!" ]

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

[]

[ "Rerun with environmental variable `CUDA_LAUNCH_BLOCKING=1` and see what line it crashed on.\r\n\r\nThis is almost always an out-of-bounds error on some embeddings lookup. Usually positional embeddings, but it could be word embeddings or segment embeddings.", "HI @stephenroller , I do set environmental variable `CUDA_LAUNCH_BLOCKING=1` and get the previous log. I will check my word embeddings or segment embeddings.", "Then it’s definitely that you’ve got a bad index into the positional embeddings.", "But when I removed the positional embeddings, it still posts the error.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "> But when I removed the positional embeddings, it still posts the error.\r\n\r\nI met the same problem. Did you find how to solve it?", "I met the same problem...", "> I met the same problem...\r\n\r\nI solved it myself. I met the problem when I distill the big model，but the vocab size of teacher model and student model is different. I modify the vocab size and it works.", "> > I met the same problem...\r\n> \r\n> I solved it myself. I met the problem when I distill the big model，but the vocab size of teacher model and student model is different. I modify the vocab size and it works.\r\n\r\nI'm experiencing the same problem. Can you please elaborate on what to do ?", "> But when I removed the positional embeddings, it still posts the error.\r\n\r\nHi! Could you please elaborate on how did you solve this error? I removed the positional embeddings and then this error is showing. Sorry about dragging a problem from years ago!", "Hi! Could you please elaborate on how did you solve this error? Thanks" ]

[]

[]

[ "Amazing, thanks a lot @VictorSanh!" ]

[]

[]

[]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "same not working for me. NSP loss is not converging even though MLM loss is converging." ]

[]

[]

[ "@thomwolf Thanks!", "line 604", "> line 604\r\n\r\nThanks so much. my mistake.\r\n\r\nDo you know why there is no dropout for the dev and eval?\r\n", "first of all, no one uses dropout at evaluation stage as it's a regularizer. The difference of implementation is due to the fact that a dropout layer in pytorch behaves differently once you turn on a evaluation model (i.e., model.eval()). So even though it has a dropout layer it doesn't take any effect (p=0) at evaluation.", "> first of all, no one uses dropout at evaluation stage as it's a regularizer. The difference of implementation is due to the fact that a dropout layer in pytorch behaves differently once you turn on a evaluation model (i.e., model.eval()). So even though it has a dropout layer it doesn't take any effect (p=0) at evaluation.\r\n\r\nThanks for help me clarify this problem." ]

[]

[]

[ "It's because `nn.CrossEntropyLoss` already has a Softmax integrated in the module:\r\nhttps://pytorch.org/docs/stable/nn.html?highlight=crossentropy#torch.nn.CrossEntropyLoss", "I see it now. Thanks!" ]

[]

[]

[ "Thanks!" ]

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

[]

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

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

[]

[ "Python version is 3.6, the cuda version is 10.1.105, cudnn version is 7.51, and pytorch version is 1.0.1, platform is ubuntu 14.04. And GPU is Titan GTX 1080Ti with 11g memory.\r\nThanks all of you!", "I am running into almost exactly the same issue. python3.6 cuda 10.0 ubuntu 18.04 and a 1080ti as well. Not sure what it is coming from. I see lots of google results from 2080ti's but those looked like architecture issues. ", "I tried running with CUDA_LAUNCH_BLOCKING=1 because others have stated this will get us a more accurate error. This led to RuntimeError: Creating MTGP constants failed. at /pytorch/aten/src/THC/THCTensorRandom.cu:33\r\n. Doing some searching as to why this might be. Some people said indexing error somewhere. seems to be happening in the dropout area though so I dont see how that is possible. ", "https://github.com/pytorch/pytorch/issues/20489\r\nhttps://github.com/pytorch/pytorch/pull/20886\r\nThese look related", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Python version is 3.6, the cuda version is 10.0 , cudnn version is 7.5, and pytorch version is 1.1.0, platform is ubuntu 16.04. And GPU is Titan GTX 2080Ti.\r\nSolve this problem." ]