Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "TF or pytorch?\r\n\r\nIf what you want is TF, you can check #1936." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2011?src=pr&el=h1) Report\n> Merging [#2011](https://codecov.io/gh/huggingface/transformers/pull/2011?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/b0ee7c7df3d49a819c4d6cef977214bd91f5c075?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2011?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2011 +/- ##\n=======================================\n Coverage 84.05% 84.05% \n=======================================\n Files 105 105 \n Lines 15555 15555 \n=======================================\n Hits 13075 13075 \n Misses 2480 2480\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2011?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/transformers/pull/2011?src=pr&el=footer). Last update [b0ee7c7...c356290](https://codecov.io/gh/huggingface/transformers/pull/2011?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks!", "Thanks a lot to the creators (and contributors) of this amazing lib for making our lives easier!" ]

[ { "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" ]

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

[]

[ "Indeed, here are my 2 cents on that:\r\n- ctrl: easy to add (should work out of the box)\r\n- xlm: should also work out of the box (but need to check if the model is an mlm or a clm model to finetune)\r\n- albert: should work out of the box\r\n- transfo-xl: need to take care of history => a little more work\r\n- xlnet: need to take care of history + permutations => quite more work.\r\n\r\nDo you want to give it a try? We don't have that in our short term roadmap until the end of the year. ", "Okay, I'm gonna try to add `ctrl`, `xlm` and `albert`. Then I'll make pull request in order to discuss on it. \r\n\r\nIsn't there any example of how to train `transfo-xl` and `xlnet`?", "You have to look at both original repos", "Out of curiosity, has any progress been made on a pull request for this?", "+1 for this request, especially `transfo-xl` :)", "Is this issue addressed with https://github.com/huggingface/transformers/commit/a8e3336a850e856188350a93e67d77c07c85b8af?", "a8e3336a850e856188350a93e67d77c07c85b8af makes all those models accessible from `run_language_modeling.py`, but does not do anything special for models whose training has peculiarities, like `transfo-xl` or `xlnet`. I'm not familiar with those two so maybe someone else (@patrickvonplaten?) can chime in.", "As far as I know: \r\n\r\nCurrently the `lun_language_modeling.py` script is not really made to train `transfo-xl` or `xlnet`\r\n\r\nFirst as @thomwolf already said, the `mems` parameter (the \"history\") of the models is not taken care of during training. During training the model \"caches\" past sequences to effectively reuse them afterwards. It's described quite well in Figure 2 in the [Transfo-XL paper](https://arxiv.org/pdf/1901.02860.pdf). This should be rather easy to add though. \r\n\r\nSecond, `XLNet` samples from a permutation mask during training, which is one of the core ideas of the paper, see https://github.com/huggingface/transformers/issues/2822 or Equation 5 in the official [paper](https://arxiv.org/pdf/1906.08237.pdf) This is a very special case for `XLNet` and is not yet implemented in `run_language_modeling.py` (shouldn't be too hard though to implement it since there is only one additional sum per training sample). \r\n\r\nThird, `Transfo-XL` uses adaptive word embeddings and adaptive softmax which also leads to some specialties when training. See also this issue #3310. This should be implemented in the model class itself though. ", "I'm assuming that `Albert` is fine out of the box. What about `T5`?", "Is anybody still working on this currently?", "We are currently working on it. Might still take ~2 weeks.", "Any update?", "I'd like to try this (#4739). I'd like to start with XLNet since that's relevant to my work right now.", "I think you would just need to add a XLNet data collator to this file so that the trainer can be used with XLNet :-) So I would add a new XLNetLanguageModelingCollator here: https://github.com/huggingface/transformers/blob/1b5820a56540a2096daeb43a0cd8247c8c94a719/src/transformers/data/data_collator.py#L76", "Thanks so much! I'll look into it :)", "Any progress on XLNet? @shngt", "Any updates regarding XLNet ?", "@patrickvonplaten I added the data collator as you suggested - please review :) You also mentioned earlier \"the `mems` parameter (the \"history\") of the models is not taken care of during training\" - has that been taken care of, or does the logic need to be implemented separately?", "I was looking into the other models requested:\r\n\r\n- CTRL -> CLM, works out of the box, already added comments\r\n- XLM -> can be trained with three different objectives - CLM, MLM and Translation LM, which is a supervised multilingual extension of MLM. The example script does not seem to require any changes (except for maybe a warning somewhere to use the right flag with the right checkpoint?). TLM does require a lot of data-specific preprocessing, but it seems relevant only in light of the multilingual setting. I feel it would be better to incorporate those in a separate `mulitlingual_language_modeling` example script if others would like an end-to-end example of how this would be properly done.\r\n- Albert -> Instead of the random masking in BERT, the authors use a span-based masking system first seen in SpanBERT (section 3.1 of https://arxiv.org/pdf/1907.10529.pdf). It seems to be a mix of what I implemented in XLNet and the masking procedure in BERT, so should be kept in another function in the main `DataCollatorForLanguageModeling` class in my opinion\r\n- TransformerXL -> seems to be CLM with reuse of previous states. I think this functionality has been added, so no additional work should be needed\r\n\r\nIn summary, I think all that needs to be done right now for XLM and TransformerXL is to add a line or two in the starting docstring mentioning which type of LM to use. For Albert, I think we need to incorporate the masking scheme as a separate procedure in `DataCollatorForLanguageModeling`, but am not sure if this is the cleanest way to do it. Let me know what you would like.\r\n\r\n@patrickvonplaten ", "I agree very much with what you say. For `XLM` and `TransformerXL` the script should work pretty much out of the box, so we would just have to adapt some comments in `examples/language-modeling/run_language_modeling.py`.\r\n\r\nFor Albert, it would be nice to create a new `SpanMaskLanguageModeling` Data collator.", "Great, I'll get started then. I'll try to finish it over the weekend :)", "Awesome, no rush though ;-)", "Maybe a stupid question, but where should I find `run_lm_finetuning.py`? [The docs](https://huggingface.co/transformers/v2.0.0/examples.html) point to a dead link, as the file doesn't exist in the master branch. ", "it's renamed and moved [there](https://github.com/huggingface/transformers/tree/master/examples/language-modeling).", "Thanks for the notice @KristenMoore - The documentation was quite old. The new documentation should have fixed it :-) ", "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/transformers/pull/2007?src=pr&el=h1) Report\n> Merging [#2007](https://codecov.io/gh/huggingface/transformers/pull/2007?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/b0ee7c7df3d49a819c4d6cef977214bd91f5c075?src=pr&el=desc) will **increase** coverage by `0.03%`.\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2007?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2007 +/- ##\n==========================================\n+ Coverage 84.05% 84.09% +0.03% \n==========================================\n Files 105 105 \n Lines 15555 15570 +15 \n==========================================\n+ Hits 13075 13093 +18 \n+ Misses 2480 2477 -3\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2007?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/tests/modeling\\_xlnet\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/2007/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3hsbmV0X3Rlc3QucHk=) | `96.42% <100%> (+0.29%)` | :arrow_up: |\n| [transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2007/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3hsbmV0LnB5) | `74.22% <100%> (+0.61%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2007?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/transformers/pull/2007?src=pr&el=footer). Last update [b0ee7c7...76c0bc0](https://codecov.io/gh/huggingface/transformers/pull/2007?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "That's great, thanks for fixing the issue! Looks good to me.", "Yes, this is great, thanks a lot @roskoN!" ]

[]

[]

[ "Same issue here. I did slightly different steps, but same result.\r\n\r\n```\r\nmodel = AlbertModel(config=config)\r\nmodel = load_tf_weights_in_albert(model,config,'sample_tf_checkpoint/model.ckpt-100000')\r\n```\r\nThen I get,\r\n\r\n```\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n<ipython-input-5-a47f5e7bff26> in <module>\r\n----> 1 model = load_tf_weights_in_albert(model,config,'sample_tf_checkpoint/model.ckpt-100000')\r\n\r\n~/anaconda3/envs/pytorch_py37/lib/python3.7/site-packages/transformers/modeling_albert.py in load_tf_weights_in_albert(model, config, tf_checkpoint_path)\r\n 90 pointer = getattr(pointer, 'weight')\r\n 91 elif l[0] == 'output_bias' or l[0] == 'beta':\r\n---> 92 pointer = getattr(pointer, 'bias')\r\n 93 elif l[0] == 'output_weights':\r\n 94 pointer = getattr(pointer, 'weight')\r\n\r\n~/anaconda3/envs/pytorch_py37/lib/python3.7/site-packages/torch/nn/modules/module.py in __getattr__(self, name)\r\n 589 return modules[name]\r\n 590 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n--> 591 type(self).__name__, name))\r\n 592 \r\n 593 def __setattr__(self, name, value):\r\n\r\nAttributeError: 'AlbertModel' object has no attribute 'bias'\r\n```\r\n\r\nMiserably waiting for the solution :( \r\nThe pretrained tensorflow checkpoints were generated using the codes in https://github.com/google-research/google-research/tree/master/albert\r\n\r\nIt seems the latest code update was 3 days ago (Nov. 27). My training was initiated after that.\r\n\r\nPlease help us.", "Same issue here. ", "You can Try my repo convert Albert tf to torch .py\n\nOn Mon, Dec 2, 2019 at 11:28 SunYan <[email protected]> wrote:\n\n> Same issue here.\n>\n> —\n> You are receiving this because you are subscribed to this thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/2006?email_source=notifications&email_token=AIEAE4BXOVAOQN7RGG35JHLQWR6GRA5CNFSM4JTGZEWKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFSCWFA#issuecomment-560212756>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AIEAE4DCFSKTP2GFFB3GYDTQWR6GRANCNFSM4JTGZEWA>\n> .\n>\n", "> You can Try my repo convert Albert tf to torch .py\r\n> […](#)\r\n> On Mon, Dec 2, 2019 at 11:28 SunYan ***@***.***> wrote: Same issue here. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2006?email_source=notifications&email_token=AIEAE4BXOVAOQN7RGG35JHLQWR6GRA5CNFSM4JTGZEWKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFSCWFA#issuecomment-560212756>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIEAE4DCFSKTP2GFFB3GYDTQWR6GRANCNFSM4JTGZEWA> .\r\n\r\n秀！", "Hi, this should have been fixed with b3d834a, you can load the changes by installing from source.\r\n\r\nLet me know if you still have an error.", "@LysandreJik Thank you for your help. I am getting a different error saying that object Embedding doesn't have 'shape'\r\n\r\nIt seems the module is expecting numpy array, while the checkpoint contains object called Embedding, thus has no attribute \"shape\"\r\n\r\nI am not sure how to correct it though.\r\n\r\nThank you again!\r\n\r\n```\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n<ipython-input-4-a47f5e7bff26> in <module>\r\n----> 1 model = load_tf_weights_in_albert(model,config,'sample_tf_checkpoint/model.ckpt-100000')\r\n\r\n~/anaconda3/envs/pytorch_py37/lib/python3.7/site-packages/transformers/modeling_albert.py in load_tf_weights_in_albert(model, config, tf_checkpoint_path)\r\n 130 array = np.transpose(array)\r\n 131 try:\r\n--> 132 assert pointer.shape == array.shape\r\n 133 except AssertionError as e:\r\n 134 e.args += (pointer.shape, array.shape)\r\n\r\n~/anaconda3/envs/pytorch_py37/lib/python3.7/site-packages/torch/nn/modules/module.py in __getattr__(self, name)\r\n 589 return modules[name]\r\n 590 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n--> 591 type(self).__name__, name))\r\n 592 \r\n 593 def __setattr__(self, name, value):\r\n\r\nAttributeError: 'Embedding' object has no attribute 'shape'\r\n```\r\n", "Hi @hansaimlim, what is the size of the model you are loading? Could you paste here the 5-10 lines output by the conversion before the error was raised? ", "I could also reproduce that error:\r\n\r\n```bash\r\nglobal_step\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'beta'] from bert/embeddings/LayerNorm/beta\r\nINFO:transformers.modeling_albert:Skipping albert/embeddings/LayerNorm/beta/adam_m\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'beta', 'adam_m'] from bert/embeddings/LayerNorm/beta/adam_m\r\nINFO:transformers.modeling_albert:Skipping albert/embeddings/LayerNorm/beta/adam_v\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'beta', 'adam_v'] from bert/embeddings/LayerNorm/beta/adam_v\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'gamma'] from bert/embeddings/LayerNorm/gamma\r\nINFO:transformers.modeling_albert:Skipping albert/embeddings/LayerNorm/gamma/adam_m\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'gamma', 'adam_m'] from bert/embeddings/LayerNorm/gamma/adam_m\r\nINFO:transformers.modeling_albert:Skipping albert/embeddings/LayerNorm/gamma/adam_v\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'gamma', 'adam_v'] from bert/embeddings/LayerNorm/gamma/adam_v\r\nInitialize PyTorch weight ['albert', 'embeddings', 'position_embeddings'] from bert/embeddings/position_embeddings\r\nINFO:transformers.modeling_albert:Skipping albert/embeddings/position_embeddings/adam_m\r\nTraceback (most recent call last):\r\n File \"convert_albert_original_tf_checkpoint_to_pytorch.py\", line 66, in <module>\r\n args.pytorch_dump_path)\r\n File \"convert_albert_original_tf_checkpoint_to_pytorch.py\", line 37, in convert_tf_checkpoint_to_pytorch\r\n load_tf_weights_in_albert(model, config, tf_checkpoint_path)\r\n File \"/mnt/transformers/transformers/modeling_albert.py\", line 134, in load_tf_weights_in_albert\r\n assert pointer.shape == array.shape\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py\", line 585, in __getattr__\r\n type(self).__name__, name))\r\nAttributeError: 'Embedding' object has no attribute 'shape'\r\n```", "@LysandreJik Sure. Thanks for prompt feedback!\r\n\r\nmy_albert_config.json\r\n```\r\nattention_probs_dropout_prob:0\r\nhidden_act:\"gelu\"\r\nhidden_dropout_prob:0\r\nembedding_size:128\r\nhidden_size:312\r\ninitializer_range:0.02\r\nintermediate_size:1248\r\nmax_position_embeddings:512\r\nnum_attention_heads:12\r\nnum_hidden_layers:4\r\nnum_hidden_groups:1\r\nnet_structure_type:0\r\ngap_size:0\r\nnum_memory_blocks:0\r\ninner_group_num:1\r\ndown_scale_factor:1\r\ntype_vocab_size:2\r\nln_type:\"postln\"\r\nvocab_size:19686\r\n```\r\n\r\n```\r\nbert/embeddings/LayerNorm/beta\r\nbert/embeddings/LayerNorm/beta/adam_m\r\nbert/embeddings/LayerNorm/beta/adam_v\r\nbert/embeddings/LayerNorm/gamma\r\nbert/embeddings/LayerNorm/gamma/adam_m\r\nbert/embeddings/LayerNorm/gamma/adam_v\r\nbert/embeddings/position_embeddings\r\nbert/embeddings/position_embeddings/adam_m\r\nbert/embeddings/position_embeddings/adam_v\r\nbert/embeddings/token_type_embeddings\r\nbert/embeddings/token_type_embeddings/adam_m\r\nbert/embeddings/token_type_embeddings/adam_v\r\nbert/embeddings/word_embeddings\r\nbert/embeddings/word_embeddings/adam_m\r\nbert/embeddings/word_embeddings/adam_v\r\nbert/encoder/embedding_hidden_mapping_in/bias\r\nbert/encoder/embedding_hidden_mapping_in/bias/adam_m\r\nbert/encoder/embedding_hidden_mapping_in/bias/adam_v\r\nbert/encoder/embedding_hidden_mapping_in/kernel\r\nbert/encoder/embedding_hidden_mapping_in/kernel/adam_m\r\nbert/encoder/embedding_hidden_mapping_in/kernel/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm/beta\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm/beta/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm/beta/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm/gamma\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm/gamma/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm/gamma/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm_1/beta\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm_1/beta/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm_1/beta/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm_1/gamma\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm_1/gamma/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/LayerNorm_1/gamma/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/output/dense/bias\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/output/dense/bias/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/output/dense/bias/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/output/dense/kernel\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/output/dense/kernel/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/output/dense/kernel/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/key/bias\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/key/bias/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/key/bias/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/key/kernel\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/key/kernel/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/key/kernel/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/query/bias\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/query/bias/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/query/bias/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/query/kernel\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/query/kernel/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/query/kernel/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/value/bias\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/value/bias/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/value/bias/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/value/kernel\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/value/kernel/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/attention_1/self/value/kernel/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/dense/bias\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/dense/bias/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/dense/bias/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/dense/kernel\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/dense/kernel/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/dense/kernel/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/output/dense/bias\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/output/dense/bias/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/output/dense/bias/adam_v\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/output/dense/kernel\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/output/dense/kernel/adam_m\r\nbert/encoder/transformer/group_0/inner_group_0/ffn_1/intermediate/output/dense/kernel/adam_v\r\nbert/pooler/dense/bias\r\nbert/pooler/dense/bias/adam_m\r\nbert/pooler/dense/bias/adam_v\r\nbert/pooler/dense/kernel\r\nbert/pooler/dense/kernel/adam_m\r\nbert/pooler/dense/kernel/adam_v\r\ncls/predictions/output_bias\r\ncls/predictions/output_bias/adam_m\r\ncls/predictions/output_bias/adam_v\r\ncls/predictions/transform/LayerNorm/beta\r\ncls/predictions/transform/LayerNorm/beta/adam_m\r\ncls/predictions/transform/LayerNorm/beta/adam_v\r\ncls/predictions/transform/LayerNorm/gamma\r\ncls/predictions/transform/LayerNorm/gamma/adam_m\r\ncls/predictions/transform/LayerNorm/gamma/adam_v\r\ncls/predictions/transform/dense/bias\r\ncls/predictions/transform/dense/bias/adam_m\r\ncls/predictions/transform/dense/bias/adam_v\r\ncls/predictions/transform/dense/kernel\r\ncls/predictions/transform/dense/kernel/adam_m\r\ncls/predictions/transform/dense/kernel/adam_v\r\ncls/seq_relationship/output_bias\r\ncls/seq_relationship/output_bias/adam_m\r\ncls/seq_relationship/output_bias/adam_v\r\ncls/seq_relationship/output_weights\r\ncls/seq_relationship/output_weights/adam_m\r\ncls/seq_relationship/output_weights/adam_v\r\nglobal_step\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'beta'] from bert/embeddings/LayerNorm/beta\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'beta', 'adam_m'] from bert/embeddings/LayerNorm/beta/adam_m\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'beta', 'adam_v'] from bert/embeddings/LayerNorm/beta/adam_v\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'gamma'] from bert/embeddings/LayerNorm/gamma\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'gamma', 'adam_m'] from bert/embeddings/LayerNorm/gamma/adam_m\r\nInitialize PyTorch weight ['albert', 'embeddings', 'LayerNorm', 'gamma', 'adam_v'] from bert/embeddings/LayerNorm/gamma/adam_v\r\nInitialize PyTorch weight ['albert', 'embeddings', 'position_embeddings'] from bert/embeddings/position_embeddings\r\n---------------------------------------------------------------------------\r\nAttributeError Traceback (most recent call last)\r\n<ipython-input-4-a47f5e7bff26> in <module>\r\n----> 1 model = load_tf_weights_in_albert(model,config,'sample_tf_checkpoint/model.ckpt-100000')\r\n\r\n~/anaconda3/envs/pytorch_py37/lib/python3.7/site-packages/transformers/modeling_albert.py in load_tf_weights_in_albert(model, config, tf_checkpoint_path)\r\n 130 array = np.transpose(array)\r\n 131 try:\r\n--> 132 assert pointer.shape == array.shape\r\n 133 except AssertionError as e:\r\n 134 e.args += (pointer.shape, array.shape)\r\n\r\n~/anaconda3/envs/pytorch_py37/lib/python3.7/site-packages/torch/nn/modules/module.py in __getattr__(self, name)\r\n 589 return modules[name]\r\n 590 raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n--> 591 type(self).__name__, name))\r\n 592 \r\n 593 def __setattr__(self, name, value):\r\n\r\nAttributeError: 'Embedding' object has no attribute 'shape'\r\n```", "Alright, I see where the issue stems from, I'm patching it and will get back to you soon.", "Alright, please let me know if e85855f fixed it. I tested it with models saved from `run_pretraning.py` (with `AlbertForMaskedLM` as the host model) and `run_classifier_sp.py` (with `AlbertForSequenceClassifiication`) and both seem to work fine now.\r\n\r\nPlease keep in mind that we have no albert model that can do next sentence prediction so the weights from `cls/seq_relationship` are dropped. ", "@LysandreJik \r\n\r\nWorks fine!! :)))) Thank you so much! 👍 ", "Glad I could help!", "Thanks @LysandreJik ! I can also confirm that the conversion script is working now :+1: ", "Short update: I used the converted ALBERT model to perform NER. F-score was ~0.1%. I've seen this strange behaviour for v2 ALBERT models but still have no solution for that.\r\n\r\n@hansaimlim have you done some evaluations with your trained model? Would be great to know if this problem also occurs for non-NER tasks!\r\n\r\n", "@stefan-it I'm working on drug activity prediction. In my case, I used v2 ALBERT as well, and its performance for masked LM was fine, and I haven't done downstream prediction tasks yet. Assuming you're working on human language, I believe our tasks are very different. How was it when you use BERT?", "I used my trained model for predicting a masked token, and the model always returns `<unk>` (which is not the case for the English v1 and v2 models), so I guess I did something wrong in the pre-training steps... ", "Dear All,\r\nI still ha ve an issue by converting an albert checkpoint to pytorch binary using this script. Here is the error:\r\n```Traceback (most recent call last):\r\n File \"$WORK/Tools/miniconda3/envs/py309/lib/python3.9/site-packages/transformers/models/albert/convert_albert_original_tf_checkpoint_to_pytorch.py\", line 63, in <module>\r\n convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.albert_config_file, args.pytorch_dump_path)\r\n File \"$WORK/Tools/miniconda3/envs/py309/lib/python3.9/site-packages/transformers/models/albert/convert_albert_original_tf_checkpoint_to_pytorch.py\", line 36, in convert_tf_checkpoint_to_pytorch\r\n load_tf_weights_in_albert(model, config, tf_checkpoint_path)\r\n File \"$WORK/Tools/miniconda3/envs/py309/lib/python3.9/site-packages/transformers/models/albert/modeling_albert.py\", line 163, in load_tf_weights_in_albert\r\n pointer = getattr(pointer, \"bias\")\r\n File \"$WORK/Tools/miniconda3/envs/py309/lib/python3.9/site-packages/torch/nn/modules/module.py\", line 1269, in __getattr__\r\n raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\nAttributeError: 'AlbertEmbeddings' object has no attribute 'bias'\r\n```\r\nAny idea? \r\nUsing python 3.9\r\ntransformers 4.26.1\r\nunder linux (ubuntu)" ]

[]

[]

[ "Sorry, I created this duplicated issue as the previous one. Please delete this one, 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 } ]

[]

[ "For now we need to use:\r\n\r\n```python\r\ntf.config.optimizer.set_experimental_options({\"auto_mixed_precision\": True})\r\n```\r\n\r\nPlease see [example here](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).", "Thanks. I tried it during waiting the answer, and it doesn't speed up the training. I probably can post my model later.", "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" ]

[]

[]

[ "If you want to have `xlnet_config.json`, which is a JSON file which specifies the hyper-parameters of the XLNet model, you can download the .zip file from [here](https://github.com/zihangdai/xlnet/blob/5cd50bc451436e188a8e7fea15358d5a8c916b72/README.md) which contains the pre-trained weights of XLNet model.\r\n\r\n> ## Questions & Help\r\n> I notice in tokenizer_xlnet.py there is not the vocab,json only spiece model. So I want to know where I could find the vocab.json? And what I should rename the file ?", "I downloaded the pre-trained model, the config file and the sentence piece model, but when I run the code I found the vocab_size = -1. Did I miss something?", "An example of the content of `xlnet_config.json` is the following:\r\n```\r\n{\r\n \"d_head\": 64, \r\n \"d_inner\": 3072, \r\n \"d_model\": 768, \r\n \"ff_activation\": \"gelu\", \r\n \"n_head\": 12, \r\n \"n_layer\": 12, \r\n \"n_token\": 32000, \r\n \"untie_r\": true\r\n}\r\n```\r\n\r\n> When i run the code I found the vocab_size=-1\r\n\r\nWhich code you are talking about?\r\n", "You can either download from the S3 repo or the script is supposed to automatically download the vocab file. Make sure you have a working internet connection." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2002?src=pr&el=h1) Report\n> Merging [#2002](https://codecov.io/gh/huggingface/transformers/pull/2002?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/b0ee7c7df3d49a819c4d6cef977214bd91f5c075?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2002?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2002 +/- ##\n=======================================\n Coverage 84.05% 84.05% \n=======================================\n Files 105 105 \n Lines 15555 15555 \n=======================================\n Hits 13075 13075 \n Misses 2480 2480\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2002?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/transformers/pull/2002?src=pr&el=footer). Last update [b0ee7c7...508f939](https://codecov.io/gh/huggingface/transformers/pull/2002?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Yes, this works! Thank you @ethanjperez." ]

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

[]

[ "> I am a little confused about how to prepare input bathces for GPT2LMHeadModel. I want to use GPT2 as an LM. For instance, I want to generate probability distributions over the vocabulary at each timestep, as well as computing the perplexities of sentences. It is important to note that I am working with sentences and not documents, so I will have to pad the inputs in the batch.\r\n> \r\n> ```python\r\n> from transformers import GPT2Tokenizer, GPT2LMHeadModel\r\n> \r\n> # Prepare model\r\n> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')\r\n> model = GPT2LMHeadModel.from_pretrained('gpt2')\r\n> model.eval()\r\n> model.to('cuda')\r\n> \r\n> # input sentences\r\n> batch = ['this is a sentence.',\r\n> 'this is another sentence.',\r\n> 'this is another even longer sentence.']\r\n> ```\r\n> \r\n> ## Question 1: Special tokens\r\n> a) Do I have to add a bos token id on my own or is it handled internally by GPT2Tokenizer? Same for the eos token.\r\n> \r\n> ```\r\n> # tokenize\r\n> tokens = [tokenizer.encode(x) for x in batch]\r\n> \r\n> # add BOS and EOS\r\n> tokens = [[tokenizer.bos_token_id] + x + [tokenizer.eos_token_id] for x in tokens]\r\n> ```\r\n> \r\n> ```\r\n> [[50256, 428, 318, 257, 6827, 13, 50256],\r\n> [50256, 428, 318, 1194, 6827, 13, 50256],\r\n> [50256, 428, 318, 1194, 772, 2392, 6827, 13, 50256]]\r\n> ```\r\n> \r\n> b) `tokenizer.encode(x)` gives me a warning \"This tokenizer does not make use of special tokens. Input is returned with no modification.\" I replaced it with `tokenizer.convert_tokens_to_ids(tokenizer.tokenize(x, add_prefix_space=True))` and the warning went away, but I am not sure what is the difference. Which tokenization should I use?\r\n> \r\n> c) By looking at the properties of an instance of GPT2Tokenizer, I see that `bos_token` and `eos_token` are the same. Is this correct?\r\n> \r\n> ## Question 2: Padding\r\n> I want to pad based on the longest sentence in the batch. This is how I usually do it.\r\n> \r\n> ```\r\n> batch = pad_sequence([torch.LongTensor(x) for x in tokens], batch_first=True).to('cuda')\r\n> ```\r\n> \r\n> ```\r\n> tensor([[50256, 428, 318, 257, 6827, 13, 50256, 0, 0],\r\n> [50256, 428, 318, 1194, 6827, 13, 50256, 0, 0],\r\n> [50256, 428, 318, 1194, 772, 2392, 6827, 13, 50256]])\r\n> ```\r\n> \r\n> a) What id does the model expect for the padded tokens? Do I have to pass the token id as an argument to the model or the tokenizer or you have a predifined one?\r\n> \r\n> ## Question 3: Model input\r\n> How is GPT2 made aware of the padded steps? For instance for an RNN I would do something like this:\r\n> \r\n> ```\r\n> lengths = (batch != 0).sum(-1) # tensor([7, 7, 9])\r\n> packed = pack_padded_sequence(x, lengths, batch_first=True)\r\n> out_packed, hn = rnn(packed)\r\n> ```\r\n> \r\n> but for GPT2 I havent found an example. The only ones i found are with batch size 1. So something like this wont work as expected:\r\n> \r\n> ```\r\n> outputs = model(x, labels=x) # labels are shifted inside the model, right?\r\n> loss, logits = outputs[:2]\r\n> ```\r\n> \r\n> # Update\r\n> So there are still things unclear, but from reading other issues this is my current understanding:\r\n> \r\n> * GPT2 has no padding token, as it was trained on documents and not sentences.\r\n> * In order to use GPT2 with variable length inputs, we can apply padding with an arbitrary token and ensure that those tokens are not used by the model with an `attention_mask`.\r\n> * As for the labels, we should replace **only** on the `labels` variable the padded token ids with `-1`.\r\n> So based on that, here is my current toy implementation:\r\n> \r\n> ```python\r\n> inputs = [\r\n> 'this is a sentence.',\r\n> 'this is another sentence.',\r\n> 'this is another even longer sentence.', ]\r\n> \r\n> # tokenize\r\n> # tokens = [tokenizer.encode(x) for x in batch]\r\n> tokens = [tokenizer.convert_tokens_to_ids(\r\n> tokenizer.tokenize(x, add_prefix_space=True))\r\n> for x in inputs]\r\n> \r\n> # add BOS and EOS\r\n> tokens = [[tokenizer.bos_token_id] + x + [tokenizer.eos_token_id]\r\n> for x in tokens]\r\n> \r\n> # padding_value can be whatever...\r\n> inputs = pad_sequence([torch.LongTensor(x) for x in tokens], batch_first=True, padding_value=0).to('cuda')\r\n> # 1 for real tokens and 0 for padded tokens\r\n> mask = (inputs != 0).float()\r\n> # replace the ids of the padded tokens (where token_id==padded_id) with `-1`\r\n> labels = inputs.masked_fill(inputs == 0, -1)\r\n> \r\n> outputs = model(inputs, attention_mask=mask, labels=labels)\r\n> loss, logits = outputs[:2]\r\n> ```\r\n> \r\n> Is this correct??\r\n> \r\n> ### Bug: Padded tokens are not excluded from the loss\r\n> However, I computed the loss on my own and found your implementation does not take into account the padded tokens when averaging, unless I am missing something\r\n> \r\n> ```python\r\n> _logits = logits.view(-1, logits.size(-1)) # flatten logits\r\n> _labels = torch.cat([inputs[:, 1:], inputs[:, :1] * 0], dim=1).view(-1) # shift inputs one position to the left and flatten\r\n> loss_real_avg = F.cross_entropy(_logits, _labels, ignore_index=0, reduction='sum') / mask.sum() # ignore padded timesteps\r\n> loss_naive_avg = F.cross_entropy(_logits, _labels, ignore_index=0, reduction='mean')\r\n> \r\n> print(\"GPT2 loss:\", loss.item())\r\n> print(\"loss_naive_avg:\", loss_naive_avg.item())\r\n> print(\"loss_real_avg:\", loss_real_avg.item())\r\n> ```\r\n> \r\n> ```\r\n> GPT2 loss: 4.664564609527588\r\n> loss_naive_avg: 4.664564609527588\r\n> loss_real_avg: 4.056143283843994\r\n> ```\r\n\r\nWhat you proposed seems a valid walk-around to me. Also, look at #1464, which talked about adding `pad_token` to `tokenizer` and `embedding`. Perhaps that will help 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", "Check out [#3311](https://github.com/huggingface/transformers/issues/3311#issuecomment-601264426). GPT2 doesn't add BOS or EOS token, you must do it manually or use a tokenizer to do so. ", "> Bug: Padded tokens are not excluded from the loss\r\n> \r\n> However, I computed the loss on my own and found your implementation does not take into account the padded tokens when averaging, unless I am missing something\r\n> ```py\r\n> _logits = logits.view(-1, logits.size(-1)) # flatten logits\r\n> _labels = torch.cat([inputs[:, 1:], inputs[:, :1] * 0], dim=1).view(-1) # shift inputs one position to the left and flatten\r\n> loss_real_avg = F.cross_entropy(_logits, _labels, ignore_index=0, reduction='sum') / mask.sum() # ignore padded timesteps\r\n> loss_naive_avg = F.cross_entropy(_logits, _labels, ignore_index=0, reduction='mean')\r\n> print(\"GPT2 loss:\", loss.item())\r\n> print(\"loss_naive_avg:\", loss_naive_avg.item())\r\n> print(\"loss_real_avg:\", loss_real_avg.item())\r\n> GPT2 loss: 4.664564609527588\r\n> loss_naive_avg: 4.664564609527588\r\n> loss_real_avg: 4.056143283843994\r\n\r\nFor this bug, you may need to set `ignore_index` to -1 instead of 0 in `F.cross_entropy` according to this line:\r\n> ```py\r\n> # replace the ids of the padded tokens (where token_id==padded_id) with `-1`\r\n> labels = inputs.masked_fill(inputs == 0, -1)\r\n> ```" ]

[]

[]

[ "Indeed, do you want to fix this in a PR?", "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 } ]

[]

[ "> I've noticed that in the run_lm_finetuning example the model has an additional argument masked_lm_labels\r\n\r\nYes, I have the same issue here. Did you manage to port the example code to TF?\r\n\r\nIn the torch models the argument is interpreted as follows:\r\n\r\n```\r\n if masked_lm_labels is not None:\r\n loss_fct = CrossEntropyLoss(ignore_index=-1) # -1 index = padding token\r\n masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))\r\n outputs = (masked_lm_loss,) + outputs\r\n```\r\n\r\nwhich means that one has to define a custom cross-entropy loss in Tensorflow.", "Unfortunately no, I had a look around in order to implement the custom cross-entropy we are talking about. I switched to Pytorch since it wasn't clear to me whether switching to the custom loss would solve all the problems I had.", "I see. I guess I will take the same road ;-) At least I can do the finetuning in torch and later convert the model to TF. Thanks for sharing the info!\r\n\r\nBTW I found the implementation of the custom loss that we are talking about in google repo:\r\n\r\n```python\r\n # The `positions` tensor might be zero-padded (if the sequence is too\r\n # short to have the maximum number of predictions). The `label_weights`\r\n # tensor has a value of 1.0 for every real prediction and 0.0 for the\r\n # padding predictions.\r\n per_example_loss = -tf.reduce_sum(log_probs * one_hot_labels, axis=[-1])\r\n numerator = tf.reduce_sum(label_weights * per_example_loss)\r\n denominator = tf.reduce_sum(label_weights) + 1e-5\r\n loss = numerator / denominator\r\n```\r\n\r\nHere is the link to the original code: \r\n\r\nhttps://github.com/google-research/bert/blob/cc7051dc592802f501e8a6f71f8fb3cf9de95dc9/run_pretraining.py#L273-L280\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", "While everyone correctly pointed out that you need a loss function which handles masks, the original error message posted here is actually unrelated to that.\r\n\r\n> ```python\r\n> model.compile(optimizer=tf.optimizers.Adam(lr=params['learning_rate']), loss='binary_crossentropy')\r\n> ```\r\n\r\nYour model is compiled with binary crossenropy, e.g. one hot encoded binary labels of shape (batch size x len x len(dict)), while you provide the labels as integers representing the token values (5673 etc) with shape (batchsize x len). This leads to a shape mismatch.\r\nThe error message comes from the comparison of the last values of the shapes len(dict) vs textlen.\r\n\r\n> tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimensions must be equal, but are 10 and 119547 for 'loss/output_1_loss/mul' (op: 'Mul') with input shapes: [?,10], [?,?,119547].\r\n\r\nUsing tf.keras.losses.SparseCategoricalCrossentropy solves the error message, but of course you will still need to implement a masked loss function to use it properly.\r\n", "Is there anyone who went on with tensorflow? I don't want to switch to pytorch. I will try to implement a masked loss function. If there is anyone already did this, I would be happy to know.", "I made an attempt on kaggle: https://www.kaggle.com/riblidezso/finetune-xlm-roberta-on-jigsaw-test-data-with-mlm", "> I made an attempt on kaggle: https://www.kaggle.com/riblidezso/finetune-xlm-roberta-on-jigsaw-test-data-with-mlm\r\n\r\nIt is a very interesting and usefull notebook. Thanks for sharing", "> I made an attempt on kaggle: https://www.kaggle.com/riblidezso/finetune-xlm-roberta-on-jigsaw-test-data-with-mlm\r\n\r\nSuper useful, thank you!" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/1998?src=pr&el=h1) Report\n> Merging [#1998](https://codecov.io/gh/huggingface/transformers/pull/1998?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/1ab8dc44b3d84ed1894f5b6a6fab58fb39298fc7?src=pr&el=desc) will **increase** coverage by `1.28%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1998?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1998 +/- ##\n==========================================\n+ Coverage 82.83% 84.11% +1.28% \n==========================================\n Files 105 105 \n Lines 15545 15545 \n==========================================\n+ Hits 12877 13076 +199 \n+ Misses 2668 2469 -199\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1998?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/1998/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX29wZW5haS5weQ==) | `82% <0%> (+1.33%)` | :arrow_up: |\n| [transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/1998/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2N0cmwucHk=) | `96.47% <0%> (+2.2%)` | :arrow_up: |\n| [transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/1998/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3hsbmV0LnB5) | `73.61% <0%> (+2.43%)` | :arrow_up: |\n| [transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/1998/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3JvYmVydGEucHk=) | `71.76% <0%> (+12.35%)` | :arrow_up: |\n| [transformers/tests/modeling\\_tf\\_common\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/1998/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3RmX2NvbW1vbl90ZXN0LnB5) | `97.08% <0%> (+15.53%)` | :arrow_up: |\n| [transformers/modeling\\_tf\\_pytorch\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1998/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3B5dG9yY2hfdXRpbHMucHk=) | `95.13% <0%> (+85.41%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1998?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/transformers/pull/1998?src=pr&el=footer). Last update [1ab8dc4...a80f3cd](https://codecov.io/gh/huggingface/transformers/pull/1998?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks!" ]

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

[]

[ "Have you taken a look at [sentencepiece](https://github.com/google/sentencepiece)?", "请问这个问题您解决了吗", "I have the same problem. Have you solved it?", "> Have you taken a look at [sentencepiece](https://github.com/google/sentencepiece)?\r\n\r\nI have taken a look at sentencepiece documents, but found nothing to build a spiece.model from customized chinese vocab.txt in ALBERT. Do you have any solution to solve this problem?", "I think that the chinese version of albert uses wordpiece model instead of sentencepiece model.\r\n[https://github.com/google-research/ALBERT/issues/58](url)\r\n\r\n> For Chinese models, we use word piece model provided by Jacob as sentence piece get worse performance on reading comprehension tasks for Chinese.\r\n\r\nhttps://github.com/google-research/ALBERT/blob/master/tokenization.py\r\n```python\r\nclass FullTokenizer(object):\r\n \"\"\"Runs end-to-end tokenziation.\"\"\"\r\n\r\n def __init__(self, vocab_file, do_lower_case=True, spm_model_file=None):\r\n self.vocab = None\r\n self.sp_model = None\r\n if spm_model_file:\r\n self.sp_model = spm.SentencePieceProcessor()\r\n tf.logging.info(\"loading sentence piece model\")\r\n self.sp_model.Load(spm_model_file)\r\n # Note(mingdachen): For the purpose of consisent API, we are\r\n # generating a vocabulary for the sentence piece tokenizer.\r\n self.vocab = {self.sp_model.IdToPiece(i): i for i\r\n in range(self.sp_model.GetPieceSize())}\r\n else:\r\n self.vocab = load_vocab(vocab_file)\r\n self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)\r\n self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)\r\n self.inv_vocab = {v: k for k, v in self.vocab.items()}\r\n```\r\nWhen the sentencepiece model is None, the full tokenizer is initialized with a basic tokenizer and a workpiece tokenizer.", "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 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/transformers/pull/1995?src=pr&el=h1) Report\n> Merging [#1995](https://codecov.io/gh/huggingface/transformers/pull/1995?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/1ab8dc44b3d84ed1894f5b6a6fab58fb39298fc7?src=pr&el=desc) will **increase** coverage by `1.22%`.\n> The diff coverage is `31.25%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1995?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1995 +/- ##\n==========================================\n+ Coverage 82.83% 84.06% +1.22% \n==========================================\n Files 105 105 \n Lines 15545 15561 +16 \n==========================================\n+ Hits 12877 13081 +204 \n+ Misses 2668 2480 -188\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1995?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2F1dG8ucHk=) | `30.61% <20%> (-1.21%)` | :arrow_down: |\n| [transformers/tokenization\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rva2VuaXphdGlvbl9hdXRvLnB5) | `45% <33.33%> (-0.95%)` | :arrow_down: |\n| [transformers/configuration\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fYXV0by5weQ==) | `60% <66.66%> (+0.54%)` | :arrow_up: |\n| [transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX29wZW5haS5weQ==) | `82% <0%> (+1.33%)` | :arrow_up: |\n| [transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2N0cmwucHk=) | `96.47% <0%> (+2.2%)` | :arrow_up: |\n| [transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3hsbmV0LnB5) | `73.61% <0%> (+2.43%)` | :arrow_up: |\n| [transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3JvYmVydGEucHk=) | `71.76% <0%> (+12.35%)` | :arrow_up: |\n| [transformers/tests/modeling\\_tf\\_common\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3RmX2NvbW1vbl90ZXN0LnB5) | `97.08% <0%> (+15.53%)` | :arrow_up: |\n| [transformers/modeling\\_tf\\_pytorch\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1995/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3B5dG9yY2hfdXRpbHMucHk=) | `95.13% <0%> (+85.41%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1995?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/transformers/pull/1995?src=pr&el=footer). Last update [1ab8dc4...a415156](https://codecov.io/gh/huggingface/transformers/pull/1995?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thank you, that's great!" ]

[]

[]

[ "The issue is fixed in #2007 ." ]

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

[]

[ "The token embedding matrix and the linear layer of the **language modeling head** are indeed tied. The embedding matrix is used to map the vocabulary to vectors of last dimension `hidden_size`. \r\n\r\nThe linear layer is used to do the exact same thing, just the other way around -> mapping the model output of last dimension `hidden_size` to the vocabulary, so that the output may be converted into vocabulary tokens.", "First of all, thanks for the replay!\r\n\r\nI know that the last linear layer is to mapping `hidden_state` with the size `hidden_size` to the size of the vocabulary, but the linear layer does not need to output concrete tokens, right? It just needs to output a group of probabilities (with the size of vocabulary) with softmax, and these probabilities seem to have nothing to do with the token embedding matrix?\r\n\r\nI have read some other papers, like the CBOW model in word2vec, which uses a linear layer with separate parameters before softmax to train the language model. As a result, the way that GPT does makes me feel confused.", "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 } ]

[]

[ "cc @VictorSanh ", "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 two GPUs install but I've not passed any argument to utilize both GPUs\r\n\r\n\r\n+-----------------------------------------------------------------------------+\r\n| NVIDIA-SMI 418.87.01 Driver Version: 418.87.01 CUDA Version: 10.1 |\r\n|-------------------------------+----------------------+----------------------+\r\n| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |\r\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\r\n|===============================+======================+======================|\r\n| 0 Tesla V100-PCIE... On | 00000000:00:05.0 Off | Off |\r\n| N/A 39C P0 43W / 250W | 4976MiB / 32480MiB | 0% Default |\r\n+-------------------------------+----------------------+----------------------+\r\n| 1 Tesla V100-PCIE... On | 00000000:00:06.0 Off | Off |\r\n| N/A 35C P0 26W / 250W | 11MiB / 32480MiB | 0% Default |\r\n+-------------------------------+----------------------+----------------------+\r\n\r\n+-----------------------------------------------------------------------------+\r\n| Processes: GPU Memory |\r\n| GPU PID Type Process name Usage |\r\n|=============================================================================|\r\n| 0 26820 C python 1653MiB |\r\n| 0 69641 C python 1657MiB |\r\n| 0 114902 C python 1655MiB |\r\n+-----------------------------------------------------------------------------+\r\n", "I ran into the same issue. It is not fixed in the newest released version, 2.2.1.", "Same issue when using multi-gpu. Single gpu case works.\r\n\r\ne.g. running the command as `CUDA_VISIBLE_DEVICES=1 python run_lm_finetuning.py ...`", "As mentioned by @kalpitdixit using Single GPU works fine but on multiple GPUs, problem persists. ", "Ok should be fixed on master, thanks!" ]

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

[]

[ "\r\n\r\n\r\n\r\nThe parameters of Albert XXLarge is much less than that of Bert large, because albert used shared parameters in all transformer layers. But it does not reduce computation，Albert xlarge is 1.5 times lower than bert large and Albert xxlarge is 3 times", "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", "ALBERT repeats the same parameters for each layer but increases each layer size, so even though it has fewer parameters than BERT, the memory needs are greater due to the much larger activations in each layer." ]

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

[]

[ "Not in the short term", "@thomwolf Thanks a lot", "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 } ]

[]

[ "The documentation seems correct to me. Have you tried out the example code that's been provided? If you run it and check the resulting `lm_prediction_scores`, you'll see its shape is `torch.Size([1, 2, 7, 50258])`, 2 being the length of `choices`.\r\n\r\nThis comment, and the linked blog post, explains the *DoubleHeadsModel pretty well:\r\nhttps://github.com/huggingface/transformers/issues/1794#issuecomment-552627190", "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/transformers/pull/1987?src=pr&el=h1) Report\n> Merging [#1987](https://codecov.io/gh/huggingface/transformers/pull/1987?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0cb163865a4c761c226b151283309eedb2b1ca4d?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1987?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1987 +/- ##\n=======================================\n Coverage 82.67% 82.67% \n=======================================\n Files 111 111 \n Lines 16162 16162 \n=======================================\n Hits 13362 13362 \n Misses 2800 2800\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1987?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/transformers/pull/1987?src=pr&el=footer). Last update [0cb1638...bea4947](https://codecov.io/gh/huggingface/transformers/pull/1987?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thank you for your contribution! I just had a quick look an will give it the time it deserves later. Were you aware of [Pytorch's guidelines](https://pytorch.org/tutorials/beginner/saving_loading_models.html#saving-loading-a-general-checkpoint-for-inference-and-or-resuming-training) when it comes to saving models? I reckon it might simplify your solution. ", "Did this pull request check the epochs AND steps?\r\nFor instance, my training process stopped in epoch 2 step 6423.\r\nIf I run again it will continue from epoch 2 step 6423?", "I am no expert, but looking at PyTorch schedulers' code they do keep the current global step in the state. The schedulers should thus continue from the last step. The thing I am wondering about is how to continue from the same data sample.", "Hi guys,\r\nJust wanted to ask this, Wouldn't too frequent caching to disks slow down the training overall?\r\nWe can have a flag added if the user wants to save every epoch, like ```file_name_{epoch}.pt```.\r\nPlus we can save optimizer etc on the same weights file as well. \r\nPlus allowing users to specify those file names as well should be considered.\r\nThanks.", "Hi,\r\n\r\n@rlouf, \r\n\r\nSaving the all the checkpoints (model, tokenizer, optimizer, and scheduler) in one file like the pytorch example does would break the `from_pretrained` method, but I could change it to save the optimizer and scheduler in one file instead of two. \r\n\r\nI could change these lines:\r\n\r\n```\r\n# Saving\r\ntorch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))\r\ntorch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))\r\n\r\n# Loading\r\noptimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))\r\nscheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))\r\n```\r\n\r\nto something like\r\n\r\n```\r\n# Saving\r\ntorch.save({\r\n 'optimizer_state_dict': optimizer.state_dict(),\r\n 'scheduler_state_dict': scheduler.state_dict()\r\n }, os.path.join(output_dir, 'training_state.pt'))\r\n\r\n# Loading\r\ncheckpoint = torch.load(os.path.join(output_dir, 'training_state.pt'))\r\noptimizer.load_state_dict(checkpoint['optimizer_state_dict'])\r\nscheduler.load_state_dict(checkpoint['scheduler_state_dict'])\r\n```\r\n\r\n@marcoaleixo Yes, the code will resume training from the last saved checkpoint. The code saves the model every `--save_steps` training steps and saves the checkpoint in the format `checkpoint-global_step`, so we know exactly which global step the last checkpoint was on. From the global step we can figure out how many epochs have been trained and how many batches should be skipped in the current epoch to continue training from where we left off. The code for this is in [these](https://github.com/bkkaggle/transformers/blob/saving-and-resuming/examples/run_lm_finetuning.py#L230) lines\r\n\r\n@rlouf To make sure you also continue from the last data sample, you would probably have to set a constant random seed in the dataloader and then `continue` through all the epochs and batches until you get to the saved checkpoint, which would take longer, especially if your dataset is very large.\r\n\r\n@AdityaSoni19031997 You can set how often checkpoints are saved using the `--save_steps` flag. Saving the all the checkpoints (model, tokenizer, optimizer, and scheduler) in one file like the pytorch example does would break the `from_pretrained` method. Letting the user choose the file names might make it harder to automatically find and load in the files, and would require new command line parameters.", "Tell me when this is not WIP anymore and ready to review, ok?", "Hi @thomwolf, yes this is ready to review", "@bkkaggle have you compared the training curves of a single run with the training curve of two half-runs with saving/reloading in the middle?", "Hi @thomwolf, I ran some quick experiments to compare the loss curves. \r\n\r\nThe loss curves are almost, but not exactly identical - probably because the `RandomSampler` doesn't accept a random seed.\r\n\r\nOne more thing: The mean loss at the end of a continued training run won't be the same as a training run completed in one go because the mean loss gets reset to 0 when continuing training. Do you want to find some way around this, maybe by also saving the running loss at each checkpoint? or would doing this add too much complexity for no benefit?\r\n\r\nwandb dashboard with the loss curves: https://app.wandb.ai/bkkaggle/saving-and-resuming?workspace=default\r\n\r\n- `vague_valley_29` is the original 1 epoch training run\r\n- `vital_bird_30` is the same training run, but cancelled after step 100\r\n- `fresh-feather-34` is the training run resumed from `vital_bird_30`'s step 100 checkpoint", "Ok, yes full determinism can be a rabbit-hole.\r\nI think it's good to merge for me.\r\nOk for you as well @LysandreJik?", "Yes, looks good to me!", "I just finished updating the other pytorch examples: `run_xnli.py`, `run_ner.py`, `run_squad.py`, and `run_glue`. I pushed them up to my [branch](https://github.com/bkkaggle/transformers/commits/saving-and-resuming) and If you want, I can open another pull request for them." ]

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

[]

[ "A quick workaround would be appending your data to the SQUAD training dataset and doing the fine-tuning as usual.\r\n", "Yes thats an approach. I've been reading a bit more about GPT and GPT-2 ... i'm wondering if I could use the generative approach with fine-tuning on a specific task that would help with SQUAD Q&A ability for my target domain? What are people's thoughts. Does the math work out?", "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 is more of a theoretical question. I would like to use a Bert Model trained on SQUAD 2.0 then train it further on my domain's Q&A dataset.\r\n> \r\n> How would I do that. I've read through the code. As I understand it, I see that the BertforQuestionAnswering would be what I would need to use loaded with a model that is fine tuned on Squad so the weights match the architecture.\r\n> \r\n> But now I want to further fine tune this model to include Q&A training data from my target domain. How would I do that?\r\n\r\nHi @priteshpatel15 I am interested in this problem. Have you found a more elegant solution than what Matthew suggested above?\r\n" ]

[]

[]

[ "Not yet, but it's on the roadmap." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/1984?src=pr&el=h1) Report\n> Merging [#1984](https://codecov.io/gh/huggingface/transformers/pull/1984?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0cb163865a4c761c226b151283309eedb2b1ca4d?src=pr&el=desc) will **decrease** coverage by `3.11%`.\n> The diff coverage is `17.42%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1984?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1984 +/- ##\n==========================================\n- Coverage 82.67% 79.56% -3.12% \n==========================================\n Files 111 113 +2 \n Lines 16162 16969 +807 \n==========================================\n+ Hits 13362 13501 +139 \n- Misses 2800 3468 +668\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1984?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/data/metrics/squad\\_metrics.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2RhdGEvbWV0cmljcy9zcXVhZF9tZXRyaWNzLnB5) | `0% <0%> (ø)` | |\n| [transformers/tests/tokenization\\_tests\\_commons.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL3Rva2VuaXphdGlvbl90ZXN0c19jb21tb25zLnB5) | `100% <100%> (ø)` | :arrow_up: |\n| [transformers/data/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2RhdGEvX19pbml0X18ucHk=) | `100% <100%> (ø)` | :arrow_up: |\n| [transformers/tokenization\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rva2VuaXphdGlvbl94bG5ldC5weQ==) | `90.4% <100%> (+0.15%)` | :arrow_up: |\n| [transformers/data/processors/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2RhdGEvcHJvY2Vzc29ycy9fX2luaXRfXy5weQ==) | `100% <100%> (ø)` | :arrow_up: |\n| [transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rva2VuaXphdGlvbl91dGlscy5weQ==) | `91.02% <100%> (+0.55%)` | :arrow_up: |\n| [transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/1984/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2RhdGEvcHJvY2Vzc29ycy9zcXVhZC5weQ==) | `14.7% <14.7%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1984?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/transformers/pull/1984?src=pr&el=footer). Last update [0cb1638...2a4ef09](https://codecov.io/gh/huggingface/transformers/pull/1984?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Should be about ready to merge now. ~I'm reproducing the results paper results on XLNet, BERT, XLM, RoBERTa, DistilBERT and ALBERT to make sure it works as expected, then I'll rebase and call for review.~\r\n\r\nI made sure that I obtain somewhat the same results with fine-tuning + evaluating models with the old and new scripts. I could confirm it is the case for DistilBERT, BERT and XLNet.", "Ok this is great, merging!\r\n\r\n@LysandreJik do you want to have a quick look at the two comments I wrote about doc/docstring and the incomplete warning?\r\n\r\nMerging now because these are small tweak and @mfuntowicz need this PR for his PR so I'll let you push a doc commit on master directly maybe." ]

[]

[]

[ "Do you load everything (model, data) on GPU?\r\n\r\n> Hello\r\n> I tried to add special tokens to bert tokenizer via add_special_tokens:\r\n> \r\n> ```\r\n> tokenizer.add_special_tokens({'additional_special_tokens':['SS']})\r\n> ```\r\n> \r\n> But I got CUDA error\r\n> \r\n> ```\r\n> CUDA error: device-side assert triggered\r\n> ```\r\n> \r\n> The code runs without adding additional_special_tokens!\r\n> Any idea?", "Yes, I did. The code runs without this line : \r\n```\r\ntokenizer.add_special_tokens({'additional_special_tokens':['SS']})\r\n```\r\nDo you think it is a resourcse issue? \r\nThanks ", "I got this lately: \r\n```\r\n/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py in linear(input, weight, bias)\r\n 1370 ret = torch.addmm(bias, input, weight.t())\r\n 1371 else:\r\n-> 1372 output = input.matmul(weight.t())\r\n 1373 if bias is not None:\r\n 1374 output += bias\r\n\r\nRuntimeError: cublas runtime error : resource allocation failed at /pytorch/aten/src/THC/THCGeneral.cpp:216\r\n```", "Are you changing the #vocab count in the config as well?\r\n", "No, I did not. ", "you can check them with``` tokenizer.convert_ids_to_tokens(id)``` (I don’t remember exactly, but I think, from 100 to 1000 are free, maybe, from 5 to 1000 even free, crosscheck please, also it depends on the \"case\" of the model)\r\nGenerally there's a pack in the beginning and then somewhere in the between we have these free unused tokens..!", "Try this,\r\n\r\n```\r\n### Let's load a model and tokenizer\r\nmodel = BertForSequenceClassification.from_pretrained('bert-base-uncased')\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n\r\n### Do some stuff to our model and tokenizer\r\n# Ex: add new tokens to the vocabulary and embeddings of our model\r\ntokenizer.add_tokens(['[SPECIAL_TOKEN_1]', '[SPECIAL_TOKEN_2]'])\r\nmodel.resize_token_embeddings(len(tokenizer))\r\n# Train our model\r\ntrain(model)\r\n\r\n### Now let's save our model and tokenizer to a directory\r\nmodel.save_pretrained('./my_saved_model_directory/')\r\ntokenizer.save_pretrained('./my_saved_model_directory/')\r\n\r\n### Reload the model and the tokenizer\r\nmodel = BertForSequenceClassification.from_pretrained('./my_saved_model_directory/')\r\ntokenizer = BertTokenizer.from_pretrained('./my_saved_model_directory/')\r\n```", "Thank you, \r\nI missed this line. Silly mistake \r\n```\r\nmodel.resize_token_embeddings(len(tokenizer))\r\n```\r\nit worked!", "You can close the issue :)" ]

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

[]

[ "Actually I'm working on this right now. Interested to know as well if anyone else has done it.\r\nMost probably this is possible.", "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/transformers/pull/1980?src=pr&el=h1) Report\n> :exclamation: No coverage uploaded for pull request base (`master@49a69d5`). [Click here to learn what that means](https://docs.codecov.io/docs/error-reference#section-missing-base-commit).\n> The diff coverage is `90.69%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1980?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1980 +/- ##\n=========================================\n Coverage ? 84.05% \n=========================================\n Files ? 105 \n Lines ? 15533 \n Branches ? 0 \n=========================================\n Hits ? 13056 \n Misses ? 2477 \n Partials ? 0\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1980?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_tf\\_transfo\\_xl\\_utilities.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3RyYW5zZm9feGxfdXRpbGl0aWVzLnB5) | `85.57% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX2N0cmwucHk=) | `97.86% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3RyYW5zZm9feGwucHk=) | `92.54% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX2Rpc3RpbGJlcnQucHk=) | `98.66% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3JvYmVydGEucHk=) | `90.43% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3hsbmV0LnB5) | `88.16% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX2dwdDIucHk=) | `94.75% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX29wZW5haS5weQ==) | `95.92% <100%> (ø)` | |\n| [transformers/modeling\\_tf\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX2FsYmVydC5weQ==) | `85.26% <81.81%> (ø)` | |\n| [transformers/modeling\\_tf\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/1980/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX2JlcnQucHk=) | `96.31% <84.61%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1980?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/transformers/pull/1980?src=pr&el=footer). Last update [49a69d5...255516a](https://codecov.io/gh/huggingface/transformers/pull/1980?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Very nice!" ]

[]

[]

[ "Hi! The `albert` checkpoints only include the base model (the transformer model), and not the separate heads for each task (classification/question answering/...).\r\n\r\nFor question answering, you would have to first fine-tune the model to this specific task, as the question answering head is initialized randomly. You can do so with the `run_squad.py` example.", "It should be explained in that example, thank you for raising this issue! I'll change that.", "Ok! Thanks a lot!", "It would be really nice is you can release pretrained checkpoints for the specific tasks... I know its a big ask but it would save so many watts of energy all over the world....", "The model need to finetune for downstream task is very general and\ntask-agnostic, if released the specific task , what extra thing u need to\ndo ? Also if released, it is not called “pretrained model” , all training\nprocess finished.....\n\nOn Sat, Nov 30, 2019 at 15:58 mosheliv <[email protected]> wrote:\n\n> It would be really nice is you can release pretrained checkpoints for the\n> specific tasks... I know its a big ask but it would save so many watts of\n> energy all over the world....\n>\n> —\n> You are receiving this because you are subscribed to this thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/1979?email_source=notifications&email_token=AIEAE4ESDYPMYNVNJWN266DQWIMLFA5CNFSM4JSVQEGKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFP377Y#issuecomment-559923199>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AIEAE4ASDKKVFD3ZVMA4WYDQWIMLFANCNFSM4JSVQEGA>\n> .\n>\n", "I mean fine tuned for squad 2, for example. I would like to play with its\ncapabilities but the fine tuning process is a tad daunting....\n\nOn Sat, Nov 30, 2019, 21:22 pohan <[email protected]> wrote:\n\n> The model need to finetune for downstream task is very general and\n> task-agnostic, if released the specific task , what extra thing u need to\n> do ? Also if released, it is not called “pretrained model” , all training\n> process finished.....\n>\n> On Sat, Nov 30, 2019 at 15:58 mosheliv <[email protected]> wrote:\n>\n> > It would be really nice is you can release pretrained checkpoints for the\n> > specific tasks... I know its a big ask but it would save so many watts of\n> > energy all over the world....\n> >\n> > —\n> > You are receiving this because you are subscribed to this thread.\n> > Reply to this email directly, view it on GitHub\n> > <\n> https://github.com/huggingface/transformers/issues/1979?email_source=notifications&email_token=AIEAE4ESDYPMYNVNJWN266DQWIMLFA5CNFSM4JSVQEGKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFP377Y#issuecomment-559923199\n> >,\n> > or unsubscribe\n> > <\n> https://github.com/notifications/unsubscribe-auth/AIEAE4ASDKKVFD3ZVMA4WYDQWIMLFANCNFSM4JSVQEGA\n> >\n> > .\n> >\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/1979?email_source=notifications&email_token=AC7IWC66W7KKINHNFXJETBLQWIPD5A5CNFSM4JSVQEGKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEFP4Q7A#issuecomment-559925372>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AC7IWC65E5R33BTHMHSYJWLQWIPD5ANCNFSM4JSVQEGA>\n> .\n>\n", "I totally agree that it would be nice to have the weights for Albert finetuned on Squad available. ", "I have found a facebook model pretrained (oh sorry, fine tuned :) on squad2.0 in https://github.com/facebookresearch/SpanBERT.\r\nit is compatible with the huggingface models, so you can get get it with:\r\n`wget http://dl.fbaipublicfiles.com/fairseq/models/spanbert_squad2.tar.gz`\r\nand extract it into say, directory spanbert\r\nI use it something like:\r\n```\r\nimport torch \r\nfrom transformers import BertTokenizer, BertForQuestionAnswering\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-cased')\r\nmodel = BertForQuestionAnswering.from_pretrained('./spanbert')\r\nq = \"who am i?\"\r\ndoc = \"my name is slim shady\"\r\ninput_text = \"[CLS] \" + q+ \" [SEP] \" + doc + \" [SEP]\"\r\ninput_ids = tokenizer.encode(input_text)\r\ntoken_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]\r\nstart_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))\r\nall_tokens = tokenizer.convert_ids_to_tokens(input_ids)\r\nres = all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]\r\nif not res or res[0] == \"[CLS]\":\r\n print(\"MISSING\")\r\nelse:\r\n prev_token = \"\"\r\n for i, t in enumerate(res):\r\n if t.startswith(\"##\"):\r\n res[i-1] += t[2:]\r\n res[i] = \"\"\r\n print(\" \".join([x for x in res if x != \"\"]))\r\n```\r\nI am including the snipped here as it is so hard to find minimal activations of bert on single entries, especially for Q&A\r\n", "Thanks a lot!", "@mosheliv - isn't that just for bert, not albert?\r\n", "Yes, it is, but it was the only squad2 pre-trained i could find.\n\nOn Thu, Dec 5, 2019, 07:40 Mark Feblowitz <[email protected]> wrote:\n\n> @mosheliv <https://github.com/mosheliv> - isn't that just for bert, not\n> albert?\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/1979?email_source=notifications&email_token=AC7IWC3RQUZWCNR34BFZVHTQW72QFA5CNFSM4JSVQEGKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEF6B5II#issuecomment-561782433>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AC7IWC3TJXCV3RN46RWFRX3QW72QFANCNFSM4JSVQEGA>\n> .\n>\n", "> I have found a facebook model pretrained (oh sorry, fine tuned :) on squad2.0 in https://github.com/facebookresearch/SpanBERT.\r\n> it is compatible with the huggingface models, so you can get get it with:\r\n> `wget http://dl.fbaipublicfiles.com/fairseq/models/spanbert_squad2.tar.gz`\r\n> and extract it into say, directory spanbert\r\n> I use it something like:\r\n> \r\n> ```\r\n> import torch \r\n> from transformers import BertTokenizer, BertForQuestionAnswering\r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')\r\n> model = BertForQuestionAnswering.from_pretrained('./spanbert')\r\n> q = \"who am i?\"\r\n> doc = \"my name is slim shady\"\r\n> input_text = \"[CLS] \" + q+ \" [SEP] \" + doc + \" [SEP]\"\r\n> input_ids = tokenizer.encode(input_text)\r\n> token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]\r\n> start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))\r\n> all_tokens = tokenizer.convert_ids_to_tokens(input_ids)\r\n> res = all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]\r\n> if not res or res[0] == \"[CLS]\":\r\n> print(\"MISSING\")\r\n> else:\r\n> prev_token = \"\"\r\n> for i, t in enumerate(res):\r\n> if t.startswith(\"##\"):\r\n> res[i-1] += t[2:]\r\n> res[i] = \"\"\r\n> print(\" \".join([x for x in res if x != \"\"]))\r\n> ```\r\n> \r\n> I am including the snipped here as it is so hard to find minimal activations of bert on single entries, especially for Q&A\r\n\r\nCan we assume that whenever there's a `[CLS]` in the answer, it basically means no answer? I'm asking since I know depending on how we treat such cases, it can affect the performance evaluation. Please take a look at my question asked [here on SO](https://stackoverflow.com/questions/60133236/what-does-berts-special-characters-appearance-in-squads-qa-answers-mean).\r\n\r\nAlso for folks who might be looking for a running example of fine-tuned ALBERT on SQuAD v2.0, you might find this helpful:\r\n\r\n```\r\nfrom transformers import AutoTokenizer, AutoModelForQuestionAnswering\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(\"ktrapeznikov/albert-xlarge-v2-squad-v2\")\r\nmodel = AutoModelForQuestionAnswering.from_pretrained(\"ktrapeznikov/albert-xlarge-v2-squad-v2\")\r\nquestion = \"Where is the capital of the USA?\"\r\ntext = \"Capital of the USA is the beautiful Washington D.C.\"\r\n\r\ninput_dict = tokenizer.encode_plus(question, text, return_tensors=\"pt\")\r\ninput_ids = input_dict[\"input_ids\"].tolist()\r\nstart_scores, end_scores = model(**input_dict)\r\n\r\nall_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])\r\nanswer = ''.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]).replace('▁', ' ').strip()\r\nprint(answer)\r\n```\r\n", "No expert on this model but yes, this is how I used it.\nThanks for the albert, will try it later on!\n\nOn Sun, Feb 9, 2020, 16:56 Pedram <[email protected]> wrote:\n\n> I have found a facebook model pretrained (oh sorry, fine tuned :) on\n> squad2.0 in https://github.com/facebookresearch/SpanBERT.\n> it is compatible with the huggingface models, so you can get get it with:\n> wget http://dl.fbaipublicfiles.com/fairseq/models/spanbert_squad2.tar.gz\n> and extract it into say, directory spanbert\n> I use it something like:\n>\n> import torch\n>\n> from transformers import BertTokenizer, BertForQuestionAnswering\n>\n> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')\n>\n> model = BertForQuestionAnswering.from_pretrained('./spanbert')\n>\n> q = \"who am i?\"\n>\n> doc = \"my name is slim shady\"\n>\n> input_text = \"[CLS] \" + q+ \" [SEP] \" + doc + \" [SEP]\"\n>\n> input_ids = tokenizer.encode(input_text)\n>\n> token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]\n>\n> start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))\n>\n> all_tokens = tokenizer.convert_ids_to_tokens(input_ids)\n>\n> res = all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]\n>\n> if not res or res[0] == \"[CLS]\":\n>\n> print(\"MISSING\")\n>\n> else:\n>\n> prev_token = \"\"\n>\n> for i, t in enumerate(res):\n>\n> if t.startswith(\"##\"):\n>\n> res[i-1] += t[2:]\n>\n> res[i] = \"\"\n>\n> print(\" \".join([x for x in res if x != \"\"]))\n>\n>\n> I am including the snipped here as it is so hard to find minimal\n> activations of bert on single entries, especially for Q&A\n>\n> Can we assume that whenever there's a [CLS] in the answer, it basically\n> means no answer? I'm asking since I know depending on how we treat such\n> cases, it can affect the performance evaluation. Please see take a look at\n> my question asked here on SO\n> <https://stackoverflow.com/questions/60133236/what-does-berts-special-characters-appearance-in-squads-qa-answers-mean>\n> .\n>\n> Also for folks who might be looking for a running example of fine-tuned\n> ALBERT on SQuAD v2.0, you might find this helpful:\n>\n> from transformers import AutoTokenizer, AutoModelForQuestionAnswering\n>\n>\n>\n> tokenizer = AutoTokenizer.from_pretrained(\"ktrapeznikov/albert-xlarge-v2-squad-v2\")\n>\n> model = AutoModelForQuestionAnswering.from_pretrained(\"ktrapeznikov/albert-xlarge-v2-squad-v2\")\n>\n> question = \"Where is the capital of the USA?\"\n>\n> text = \"The capital of the USA is beautiful Washington D.C.\"\n>\n>\n>\n> input_dict = tokenizer.encode_plus(question, text, return_tensors=\"pt\")\n>\n> input_ids = input_dict[\"input_ids\"].tolist()\n>\n> start_scores, end_scores = model(**input_dict)\n>\n>\n>\n> all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])\n>\n> answer = ' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]).replace('▁', '')\n>\n> print(answer)\n>\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/1979?email_source=notifications&email_token=AC7IWCYV3DSSF2HRGDQDFWLRB55FLA5CNFSM4JSVQEGKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOELGBZ2A#issuecomment-583802088>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AC7IWC6KKB4S4UBYSQGQ3M3RB55FLANCNFSM4JSVQEGA>\n> .\n>\n", "> > I have found a facebook model pretrained (oh sorry, fine tuned :) on squad2.0 in https://github.com/facebookresearch/SpanBERT.\r\n> > it is compatible with the huggingface models, so you can get get it with:\r\n> > `wget http://dl.fbaipublicfiles.com/fairseq/models/spanbert_squad2.tar.gz`\r\n> > and extract it into say, directory spanbert\r\n> > I use it something like:\r\n> > ```\r\n> > import torch \r\n> > from transformers import BertTokenizer, BertForQuestionAnswering\r\n> > tokenizer = BertTokenizer.from_pretrained('bert-base-cased')\r\n> > model = BertForQuestionAnswering.from_pretrained('./spanbert')\r\n> > q = \"who am i?\"\r\n> > doc = \"my name is slim shady\"\r\n> > input_text = \"[CLS] \" + q+ \" [SEP] \" + doc + \" [SEP]\"\r\n> > input_ids = tokenizer.encode(input_text)\r\n> > token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]\r\n> > start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))\r\n> > all_tokens = tokenizer.convert_ids_to_tokens(input_ids)\r\n> > res = all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]\r\n> > if not res or res[0] == \"[CLS]\":\r\n> > print(\"MISSING\")\r\n> > else:\r\n> > prev_token = \"\"\r\n> > for i, t in enumerate(res):\r\n> > if t.startswith(\"##\"):\r\n> > res[i-1] += t[2:]\r\n> > res[i] = \"\"\r\n> > print(\" \".join([x for x in res if x != \"\"]))\r\n> > ```\r\n> > \r\n> > \r\n> > I am including the snipped here as it is so hard to find minimal activations of bert on single entries, especially for Q&A\r\n> \r\n> Can we assume that whenever there's a `[CLS]` in the answer, it basically means no answer? I'm asking since I know depending on how we treat such cases, it can affect the performance evaluation. Please see take a look at my question asked [here on SO](https://stackoverflow.com/questions/60133236/what-does-berts-special-characters-appearance-in-squads-qa-answers-mean).\r\n> \r\n> Also for folks who might be looking for a running example of fine-tuned ALBERT on SQuAD v2.0, you might find this helpful:\r\n> \r\n> ```\r\n> from transformers import AutoTokenizer, AutoModelForQuestionAnswering\r\n> \r\n> tokenizer = AutoTokenizer.from_pretrained(\"ktrapeznikov/albert-xlarge-v2-squad-v2\")\r\n> model = AutoModelForQuestionAnswering.from_pretrained(\"ktrapeznikov/albert-xlarge-v2-squad-v2\")\r\n> question = \"Where is the capital of the USA?\"\r\n> text = \"Capital of the USA is the beautiful Washington D.C.\"\r\n> \r\n> input_dict = tokenizer.encode_plus(question, text, return_tensors=\"pt\")\r\n> input_ids = input_dict[\"input_ids\"].tolist()\r\n> start_scores, end_scores = model(**input_dict)\r\n> \r\n> all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])\r\n> answer = ''.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]).replace('▁', ' ').strip()\r\n> print(answer)\r\n> ```\r\n\r\nHi! Thanks for this, I'm only a beginner and this really saved me a lot of trouble! I had a small question however. Apparently the page for this model [https://huggingface.co/ktrapeznikov/albert-xlarge-v2-squad-v2] shows there is a way to get the 'scores' of the spans in addition to getting an answer but I couldn't get it to work myself. The code is supposed to be on the lines of: \r\n\r\n```\r\nstart_scores, end_scores = model(input_ids) \r\nspan_scores = start_scores.softmax(dim=1).log()[:,:,None] + end_scores.softmax(dim=1).log()[:,None,:]\r\nignore_score = span_scores[:,0,0] #no answer scores\r\n```\r\n\r\nBut this doesn't return a single score. What am I missing? ", "@desaibhargav probably a little late for this but you can get the answers scores like so:\r\n\r\n answer_start = torch.argmax(start_scores) # get the most likely beginning of answer with the argmax of the score\r\n answer_end = torch.argmax(end_scores) + 1\r\n answer_span = inputs[\"input_ids\"][0][answer_start:answer_end]\r\n answer_tokens = tokenizer.convert_ids_to_tokens(answer_span)\r\n tokenizer.convert_tokens_to_string(answer_tokens)\r\n\r\nThis converts the answer spans to the answer. However, I'm not sure how to ignore score my best guess is that its filtered based off of some threshold\r\n" ]

[]

[]

[ "I think you **cannot** change this parameter because doing so you're trying to load weights with (512, 768) shape into an architecture with (1024, 768), and it's not possible.\r\nIf my statement is true (maybe some authors of Transformers can confirm or deny my statement), maybe a way to avoid that end users like you try to change this parameter would be to make this variable private, such as `_max_position_embeddings`.\r\n\r\n> ## Questions & Help\r\n> When I load a model like below:\r\n> `model1 = BertForSequenceClassification.from_pretrained('bert-base-uncased')`\r\n> \r\n> ```\r\n> BertForSequenceClassification(\r\n> (bert): BertModel(\r\n> (embeddings): BertEmbeddings(\r\n> (word_embeddings): Embedding(30522, 768, padding_idx=0)\r\n> (position_embeddings): Embedding(512, 768)\r\n> (token_type_embeddings): Embedding(2, 768)\r\n> (LayerNorm): LayerNorm((768,), eps=1e-12, elementwise_affine=True)\r\n> (dropout): Dropout(p=0.1, inplace=False)\r\n> )\r\n> (encoder): BertEncoder(\r\n> (layer): ModuleList(...\r\n> ```\r\n> \r\n> I want to change Embedding size from 512 to 1024, but when I try to add like this and get an error:\r\n> `model = BertForSequenceClassification.from_pretrained('bert-base-uncased', max_position_embeddings=1024)`\r\n> \r\n> > RuntimeError: Error(s) in loading state_dict for BertForSequenceClassification:\r\n> > size mismatch for bert.embeddings.position_embeddings.weight: copying a param with shape torch.Size([512, 768]) from checkpoint, the shape in current model is torch.Size([1024, 768]).\r\n> \r\n> May I know how to change configs of pre-trained model layers?", "> I think you **cannot** change this parameter because doing so you're trying to load weights with (512, 768) shape into an architecture with (1024, 768), and it's not possible.\r\n> If my statement is true (maybe some authors of Transformers can confirm or deny my statement), maybe a way to avoid that end users like you try to change this parameter would be to make this variable private, such as `_max_position_embeddings`.\r\n> \r\nAs I check with `vars(BertForSequenceClassification.from_pretrained('bert-base-uncased'))`:\r\n```\r\n{'_backend': <torch.nn.backends.thnn.THNNFunctionBackend at 0x1e557269400>,\r\n '_parameters': OrderedDict(),\r\n '_buffers': OrderedDict(),\r\n '_backward_hooks': OrderedDict(),\r\n '_forward_hooks': OrderedDict(),\r\n '_forward_pre_hooks': OrderedDict(),\r\n '_state_dict_hooks': OrderedDict(),\r\n '_load_state_dict_pre_hooks': OrderedDict(),\r\n '_modules': ...\r\n 'training': False,\r\n 'config': {\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"finetuning_task\": null,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 768,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"is_decoder\": false,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512, <============= This is the one\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 12,\r\n \"num_labels\": 2,\r\n \"output_attentions\": false,\r\n \"output_hidden_states\": false,\r\n \"output_past\": true,\r\n \"pruned_heads\": {},\r\n \"torchscript\": false,\r\n \"type_vocab_size\": 2,\r\n \"use_bfloat16\": false,\r\n \"vocab_size\": 30522\r\n },\r\n 'num_labels': 2}\r\n```\r\nSo I decided to replace `config` with `config=BertConfig(max_position_embeddings=1024)`:\r\n```\r\n{\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"finetuning_task\": null,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 768,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"is_decoder\": false,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 1024, <============== It changed\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 12,\r\n \"num_labels\": 2,\r\n \"output_attentions\": false,\r\n \"output_hidden_states\": false,\r\n \"output_past\": true,\r\n \"pruned_heads\": {},\r\n \"torchscript\": false,\r\n \"type_vocab_size\": 2,\r\n \"use_bfloat16\": false,\r\n \"vocab_size\": 30522\r\n}\r\n```\r\nBut the same error is occurred when `BertForSequenceClassification.from_pretrained('bert-base-uncased', config=config)`:\r\n```\r\nRuntimeError Traceback (most recent call last)\r\n<ipython-input-9-cfc2c553c1d9> in <module>\r\n 1 config=BertConfig(max_position_embeddings=1024)\r\n----> 2 BertForSequenceClassification.from_pretrained('bert-base-uncased', config=config)\r\n\r\nC:\\ProgramData\\Anaconda3\\lib\\site-packages\\transformers\\modeling_utils.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)\r\n 457 if len(error_msgs) > 0:\r\n 458 raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n--> 459 model.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n 460 \r\n 461 if hasattr(model, 'tie_weights'):\r\n\r\nRuntimeError: Error(s) in loading state_dict for BertForSequenceClassification:\r\n\tsize mismatch for bert.embeddings.position_embeddings.weight: copying a param with shape torch.Size([512, 768]) from checkpoint, the shape in current model is torch.Size([1024, 768]).\r\n```\r\n... 🚶 \r\n", "Sorry, but it is obvious that it doesn't work. As a said before, BERT was trained with a **particular** architecture (i.e. with 512 as max positional embeddings), and it was saved with this shape. You cannot load weights that doesn't match your architecture!\r\n\r\n> > I think you **cannot** change this parameter because doing so you're trying to load weights with (512, 768) shape into an architecture with (1024, 768), and it's not possible.\r\n> > If my statement is true (maybe some authors of Transformers can confirm or deny my statement), maybe a way to avoid that end users like you try to change this parameter would be to make this variable private, such as `_max_position_embeddings`.\r\n> \r\n> As I check with `vars(BertForSequenceClassification.from_pretrained('bert-base-uncased'))`:\r\n> \r\n> ```\r\n> {'_backend': <torch.nn.backends.thnn.THNNFunctionBackend at 0x1e557269400>,\r\n> '_parameters': OrderedDict(),\r\n> '_buffers': OrderedDict(),\r\n> '_backward_hooks': OrderedDict(),\r\n> '_forward_hooks': OrderedDict(),\r\n> '_forward_pre_hooks': OrderedDict(),\r\n> '_state_dict_hooks': OrderedDict(),\r\n> '_load_state_dict_pre_hooks': OrderedDict(),\r\n> '_modules': ...\r\n> 'training': False,\r\n> 'config': {\r\n> \"attention_probs_dropout_prob\": 0.1,\r\n> \"finetuning_task\": null,\r\n> \"hidden_act\": \"gelu\",\r\n> \"hidden_dropout_prob\": 0.1,\r\n> \"hidden_size\": 768,\r\n> \"initializer_range\": 0.02,\r\n> \"intermediate_size\": 3072,\r\n> \"is_decoder\": false,\r\n> \"layer_norm_eps\": 1e-12,\r\n> \"max_position_embeddings\": 512, <============= This is the one\r\n> \"num_attention_heads\": 12,\r\n> \"num_hidden_layers\": 12,\r\n> \"num_labels\": 2,\r\n> \"output_attentions\": false,\r\n> \"output_hidden_states\": false,\r\n> \"output_past\": true,\r\n> \"pruned_heads\": {},\r\n> \"torchscript\": false,\r\n> \"type_vocab_size\": 2,\r\n> \"use_bfloat16\": false,\r\n> \"vocab_size\": 30522\r\n> },\r\n> 'num_labels': 2}\r\n> ```\r\n> \r\n> So I decided to replace `config` with `config=BertConfig(max_position_embeddings=1024)`:\r\n> \r\n> ```\r\n> {\r\n> \"attention_probs_dropout_prob\": 0.1,\r\n> \"finetuning_task\": null,\r\n> \"hidden_act\": \"gelu\",\r\n> \"hidden_dropout_prob\": 0.1,\r\n> \"hidden_size\": 768,\r\n> \"initializer_range\": 0.02,\r\n> \"intermediate_size\": 3072,\r\n> \"is_decoder\": false,\r\n> \"layer_norm_eps\": 1e-12,\r\n> \"max_position_embeddings\": 1024, <============== It changed\r\n> \"num_attention_heads\": 12,\r\n> \"num_hidden_layers\": 12,\r\n> \"num_labels\": 2,\r\n> \"output_attentions\": false,\r\n> \"output_hidden_states\": false,\r\n> \"output_past\": true,\r\n> \"pruned_heads\": {},\r\n> \"torchscript\": false,\r\n> \"type_vocab_size\": 2,\r\n> \"use_bfloat16\": false,\r\n> \"vocab_size\": 30522\r\n> }\r\n> ```\r\n> \r\n> But the same error is occurred when `BertForSequenceClassification.from_pretrained('bert-base-uncased', config=config)`:\r\n> \r\n> ```\r\n> RuntimeError Traceback (most recent call last)\r\n> <ipython-input-9-cfc2c553c1d9> in <module>\r\n> 1 config=BertConfig(max_position_embeddings=1024)\r\n> ----> 2 BertForSequenceClassification.from_pretrained('bert-base-uncased', config=config)\r\n> \r\n> C:\\ProgramData\\Anaconda3\\lib\\site-packages\\transformers\\modeling_utils.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)\r\n> 457 if len(error_msgs) > 0:\r\n> 458 raise RuntimeError('Error(s) in loading state_dict for {}:\\n\\t{}'.format(\r\n> --> 459 model.__class__.__name__, \"\\n\\t\".join(error_msgs)))\r\n> 460 \r\n> 461 if hasattr(model, 'tie_weights'):\r\n> \r\n> RuntimeError: Error(s) in loading state_dict for BertForSequenceClassification:\r\n> \tsize mismatch for bert.embeddings.position_embeddings.weight: copying a param with shape torch.Size([512, 768]) from checkpoint, the shape in current model is torch.Size([1024, 768]).\r\n> ```\r\n> \r\n> ..." ]

[]

[]

[ "_PyTorch-pretrained-BERT_ is a older name of this library; now its name is **Transformers**.\r\nYou can check the latest docs of the library and install it from PyPi with `pip install transformers` (you have to install manually TensorFlow 2.0 and PyTorch as well through `pip install tensorflow==2.0.0` and `pip install torch`).\r\n\r\nSaid this, you can read [this](https://github.com/huggingface/transformers/blob/17ea43cf985829634bd86b36b44e5410c6f83e36/transformers/convert_bert_original_tf_checkpoint_to_pytorch.py) Python script whose goal is to convert BERT original TensorFlow checkpoint to PyTorch. The input of this script are the following three:\r\n- the path to the TensorFlow checkpoint, through `--tf_checkpoint_path` parameter\r\n- a JSON file which specifies the model architecture through `--bert_config_file` parameter\r\n- the path to the output converted PyTorch model through `--pytorch_dump_path` parameter\r\n\r\n> Hi all,\r\n> I pre-trained BERT base model on my domain-specific corpus using `https://github.com/google-research/bert` `create_pretraining_data.py` and `run_pretraining.py`.\r\n> Now, I want to use it with this `pytorch-transformers`. I saw from this page https://devhub.io/repos/huggingface-pytorch-pretrained-BERT that there is conversion script from tf checkpoints to `pytorch_model.bin` called `convert_tf_checkpoint_to_pytorch.py` but the file no longer exists.\r\n> Does anyone have solution? Thanks!", "Thanks @TheEdoardo93!" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/1976?src=pr&el=h1) Report\n> Merging [#1976](https://codecov.io/gh/huggingface/transformers/pull/1976?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/96e7ee72380a135bfd07b8fdc2018bcbea65b086?src=pr&el=desc) will **increase** coverage by `0.2%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1976?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1976 +/- ##\n=========================================\n+ Coverage 84.06% 84.26% +0.2% \n=========================================\n Files 105 104 -1 \n Lines 15537 15431 -106 \n=========================================\n- Hits 13061 13003 -58 \n+ Misses 2476 2428 -48\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1976?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/configuration\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fdXRpbHMucHk=) | `92.1% <0%> (-0.11%)` | :arrow_down: |\n| [transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3V0aWxzLnB5) | `92.68% <0%> (-0.05%)` | :arrow_down: |\n| [transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3V0aWxzLnB5) | `89.43% <0%> (-0.03%)` | :arrow_down: |\n| [transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2N0cmwucHk=) | `96.46% <0%> (-0.02%)` | :arrow_down: |\n| [transformers/configuration\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fZGlzdGlsYmVydC5weQ==) | `89.74% <0%> (ø)` | :arrow_up: |\n| [transformers/configuration\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fYXV0by5weQ==) | `59.45% <0%> (ø)` | :arrow_up: |\n| [transformers/tests/tokenization\\_tests\\_commons.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL3Rva2VuaXphdGlvbl90ZXN0c19jb21tb25zLnB5) | `100% <0%> (ø)` | :arrow_up: |\n| [transformers/modeling\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2Rpc3RpbGJlcnQucHk=) | `95.87% <0%> (ø)` | :arrow_up: |\n| [transformers/modeling\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2F1dG8ucHk=) | `31.81% <0%> (ø)` | :arrow_up: |\n| [transformers/tokenization\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rva2VuaXphdGlvbl9hdXRvLnB5) | `45.94% <0%> (ø)` | :arrow_up: |\n| ... and [8 more](https://codecov.io/gh/huggingface/transformers/pull/1976/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1976?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/transformers/pull/1976?src=pr&el=footer). Last update [96e7ee7...5a3f240](https://codecov.io/gh/huggingface/transformers/pull/1976?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "We're in the process of building better versioned documentation with easier links to follow, but at the moment the different versions are accessible in the [README](https://github.com/huggingface/transformers#state-of-the-art-natural-language-processing-for-tensorflow-20-and-pytorch), right before the `installation` section.", "Great, 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 } ]

[]

[ "Wondering this as well but for GLUE tasks. There don't seem to be a good consensus on hyperparameters such as weight decay and such", "Results using hyperparameters from my first post above, varying only batch size:\r\n```\r\nalbert_xxlargev1_squad2_512_bs32:\r\n{\r\n \"exact\": 83.67725090541565,\r\n \"f1\": 87.51235434089064,\r\n \"total\": 11873,\r\n \"HasAns_exact\": 81.86572199730094,\r\n \"HasAns_f1\": 89.54692697189559,\r\n \"HasAns_total\": 5928,\r\n \"NoAns_exact\": 85.48359966358284,\r\n \"NoAns_f1\": 85.48359966358284,\r\n \"NoAns_total\": 5945\r\n}\r\n\r\nalbert_xxlargev1_squad2_512_bs48:\r\n{\r\n \"exact\": 83.65198349195654,\r\n \"f1\": 87.4736247587816,\r\n \"total\": 11873,\r\n \"HasAns_exact\": 81.73076923076923,\r\n \"HasAns_f1\": 89.38501126197984,\r\n \"HasAns_total\": 5928,\r\n \"NoAns_exact\": 85.5677039529016,\r\n \"NoAns_f1\": 85.5677039529016,\r\n \"NoAns_total\": 5945\r\n}\r\n```\r\n\r\n\r\n", "@ahotrod There is a table in the appendix section of the ALBERT paper, which shows hyperparameters for ALBERT in downstream tasks:\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" ]

[ { "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 } ]

[]

[ "Are you trying to serialize which type of model to .hdf5 file? A pre-trained model such as BertFor* or a custom model trained with Transformers? Are you using GCP or Azure or AWS?\r\n\r\n> ## Questions & Help\r\n> I'm using this repo in the cloud and attempting to persist the model fails as it seems HDF5 and thus h5py doesn't support that paradigm per [h5py/h5py#925](https://github.com/h5py/h5py/issues/925)\r\n> \r\n> What is the recommended method of saving the model in this scenario? Thanks!", "Hi @TheEdoardo93. I'm trying to serialize TFBertFor* that has been fine-tuned. I'm on Azure.", "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/transformers/pull/1971?src=pr&el=h1) Report\n> Merging [#1971](https://codecov.io/gh/huggingface/transformers/pull/1971?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/5afca00b4732f57329824e1538897e791e02e894?src=pr&el=desc) will **increase** coverage by `0.18%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1971?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1971 +/- ##\n==========================================\n+ Coverage 84.06% 84.24% +0.18% \n==========================================\n Files 105 104 -1 \n Lines 15536 15431 -105 \n==========================================\n- Hits 13060 13000 -60 \n+ Misses 2476 2431 -45\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1971?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2JlcnQucHk=) | `87.77% <ø> (ø)` | :arrow_up: |\n| [transformers/tests/modeling\\_tf\\_common\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3RmX2NvbW1vbl90ZXN0LnB5) | `95.63% <0%> (-1.46%)` | :arrow_down: |\n| [transformers/configuration\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fdXRpbHMucHk=) | `92.1% <0%> (-0.11%)` | :arrow_down: |\n| [transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3V0aWxzLnB5) | `92.68% <0%> (-0.05%)` | :arrow_down: |\n| [transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3V0aWxzLnB5) | `89.43% <0%> (-0.03%)` | :arrow_down: |\n| [transformers/configuration\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fZGlzdGlsYmVydC5weQ==) | `89.74% <0%> (ø)` | :arrow_up: |\n| [transformers/configuration\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fYXV0by5weQ==) | `59.45% <0%> (ø)` | :arrow_up: |\n| [transformers/tests/tokenization\\_tests\\_commons.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL3Rva2VuaXphdGlvbl90ZXN0c19jb21tb25zLnB5) | `100% <0%> (ø)` | :arrow_up: |\n| [transformers/modeling\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2Rpc3RpbGJlcnQucHk=) | `95.87% <0%> (ø)` | :arrow_up: |\n| [transformers/modeling\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2F1dG8ucHk=) | `31.81% <0%> (ø)` | :arrow_up: |\n| ... and [9 more](https://codecov.io/gh/huggingface/transformers/pull/1971/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1971?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/transformers/pull/1971?src=pr&el=footer). Last update [5afca00...d5dd44e](https://codecov.io/gh/huggingface/transformers/pull/1971?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "This is less ambiguous, indeed! Thank you for taking the time to open a PR." ]

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

[]

[ "Hi, thanks for letting us know. Is there any way you could provide a script that reproduces the error in a few lines so that we may see what is wrong on our end?", "model.predict function fails with output_hidden_states=True in constructor", "I'm failing to reproduce what you're mentioning with the following snippet:\r\n\r\n```py\r\nfrom transformers import TFBertModel, BertTokenizer, BertConfig\r\nimport tensorflow as tf\r\n\r\nconfig = BertConfig.from_pretrained(\"bert-base-cased\", output_hidden_states=True)\r\nmodel = TFBertModel.from_pretrained(\"bert-base-cased\", config=config)\r\n\r\ntok = BertTokenizer.from_pretrained(\"bert-base-cased\")\r\ntext = tok.encode(\"Ain't this [MASK] best thing you've ever seen?\")\r\n\r\ninputs = tf.constant(text)\r\noutputs = model.predict(inputs)\r\n\r\nprint(outputs)\r\n```\r\nIs there any way you could provide a script that reproduces the error in a few lines so that we may see what is wrong on our end?\r\n", "With this piece of code you've posted, I'm encountered the same problem highlighted by @halidziya .\r\n\r\n**ENVIRONMENT**:\r\n- Python 3.6.9\r\n- OS: Ubuntu 16.04 ('Linux-4.15.0-70-generic-x86_64-with-debian-buster-sid')\r\n- Transformers: 2.2.2 (installed with `pip install transformers` the day after the release)\r\n- PyTorch: 1.3.1\r\n- TensorFlow: 2.0.0\r\n\r\nThe stack trace is reported below:\r\n```\r\nPython 3.6.9 |Anaconda, Inc.| (default, Jul 30 2019, 19:07:31) \r\n[GCC 7.3.0] on linux\r\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\r\n>>> from transformers import TFBertModel, BertTokenizer, BertConfig\r\n/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:541: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\r\n _np_qint8 = np.dtype([(\"qint8\", np.int8, 1)])\r\n/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:542: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\r\n _np_quint8 = np.dtype([(\"quint8\", np.uint8, 1)])\r\n/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:543: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\r\n _np_qint16 = np.dtype([(\"qint16\", np.int16, 1)])\r\n/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:544: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\r\n _np_quint16 = np.dtype([(\"quint16\", np.uint16, 1)])\r\n/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:545: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\r\n _np_qint32 = np.dtype([(\"qint32\", np.int32, 1)])\r\n/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorboard/compat/tensorflow_stub/dtypes.py:550: FutureWarning: Passing (type, 1) or '1type' as a synonym of type is deprecated; in a future version of numpy, it will be understood as (type, (1,)) / '(1,)type'.\r\n np_resource = np.dtype([(\"resource\", np.ubyte, 1)])\r\n2019-12-03 09:46:35.606174: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\r\n2019-12-03 09:46:35.610775: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 3408000000 Hz\r\n2019-12-03 09:46:35.611320: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55fa0a097860 executing computations on platform Host. Devices:\r\n2019-12-03 09:46:35.611341: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): Host, Default Version\r\n>>> import tensorflow as tf\r\n>>> config = BertConfig.from_pretrained(\"bert-base-cased\", output_hidden_states=True)\r\n>>> model = TFBertModel.from_pretrained(\"bert-base-cased\", config=config)\r\n>>> tok = BertTokenizer.from_pretrained(\"bert-base-cased\")\r\n>>> text = tok.encode(\"Ain't this [MASK] best thing you've ever seen?\")\r\n>>> inputs = tf.constant(text)\r\n>>> outputs = model.predict(inputs)\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 909, in predict\r\n use_multiprocessing=use_multiprocessing)\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training_arrays.py\", line 715, in predict\r\n x, check_steps=True, steps_name='steps', steps=steps)\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 2419, in _standardize_user_data\r\n all_inputs, y_input, dict_inputs = self._build_model_with_inputs(x, y)\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 2622, in _build_model_with_inputs\r\n self._set_inputs(cast_inputs)\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/training.py\", line 2709, in _set_inputs\r\n outputs = self(inputs, **kwargs)\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/base_layer.py\", line 842, in __call__\r\n outputs = call_fn(cast_inputs, *args, **kwargs)\r\n File \"/home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages/tensorflow_core/python/autograph/impl/api.py\", line 237, in wrapper\r\n raise e.ag_error_metadata.to_exception(e)\r\nValueError: in converted code:\r\n relative to /home/vidiemme/anaconda3/envs/huggingface/lib/python3.6/site-packages:\r\n\r\n transformers/modeling_tf_bert.py:684 call *\r\n outputs = self.bert(inputs, **kwargs)\r\n tensorflow_core/python/keras/engine/base_layer.py:842 __call__\r\n outputs = call_fn(cast_inputs, *args, **kwargs)\r\n transformers/modeling_tf_bert.py:512 call *\r\n attention_mask = tf.fill(input_shape, 1)\r\n tensorflow_core/python/ops/array_ops.py:171 fill\r\n result = gen_array_ops.fill(dims, value, name=name)\r\n tensorflow_core/python/ops/gen_array_ops.py:3602 fill\r\n \"Fill\", dims=dims, value=value, name=name)\r\n tensorflow_core/python/framework/op_def_library.py:545 _apply_op_helper\r\n (input_name, err))\r\n\r\n ValueError: Tried to convert 'dims' to a tensor and failed. Error: Cannot convert a partially known TensorShape to a Tensor: (None, 1)\r\n```\r\n\r\n> I'm failing to reproduce what you're mentioning with the following snippet:\r\n> \r\n> ```python\r\n> from transformers import TFBertModel, BertTokenizer, BertConfig\r\n> import tensorflow as tf\r\n> \r\n> config = BertConfig.from_pretrained(\"bert-base-cased\", output_hidden_states=True)\r\n> model = TFBertModel.from_pretrained(\"bert-base-cased\", config=config)\r\n> \r\n> tok = BertTokenizer.from_pretrained(\"bert-base-cased\")\r\n> text = tok.encode(\"Ain't this [MASK] best thing you've ever seen?\")\r\n> \r\n> inputs = tf.constant(text)\r\n> outputs = model.predict(inputs)\r\n> \r\n> print(outputs)\r\n> ```\r\n> \r\n> Is there any way you could provide a script that reproduces the error in a few lines so that we may see what is wrong on our end?", "Cannot reproduce on release 2.2.1.\r\nCan you check with the latest release or master?", "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 } ]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/1969?src=pr&el=h1) Report\n> Merging [#1969](https://codecov.io/gh/huggingface/transformers/pull/1969?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/49108288ba6e6dcfe554d1af98699ae7a1e6f39c?src=pr&el=desc) will **increase** coverage by `0.29%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1969?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1969 +/- ##\n==========================================\n+ Coverage 83.97% 84.26% +0.29% \n==========================================\n Files 105 104 -1 \n Lines 15529 15431 -98 \n==========================================\n- Hits 13040 13003 -37 \n+ Misses 2489 2428 -61\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1969?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/configuration\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fdXRpbHMucHk=) | `92.1% <0%> (-0.11%)` | :arrow_down: |\n| [transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rva2VuaXphdGlvbl91dGlscy5weQ==) | `92.14% <0%> (-0.09%)` | :arrow_down: |\n| [transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3V0aWxzLnB5) | `92.68% <0%> (-0.05%)` | :arrow_down: |\n| [transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3V0aWxzLnB5) | `89.43% <0%> (-0.03%)` | :arrow_down: |\n| [transformers/configuration\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fZGlzdGlsYmVydC5weQ==) | `89.74% <0%> (ø)` | :arrow_up: |\n| [transformers/configuration\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fYXV0by5weQ==) | `59.45% <0%> (ø)` | :arrow_up: |\n| [transformers/tests/tokenization\\_tests\\_commons.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL3Rva2VuaXphdGlvbl90ZXN0c19jb21tb25zLnB5) | `100% <0%> (ø)` | :arrow_up: |\n| [transformers/modeling\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2Rpc3RpbGJlcnQucHk=) | `95.87% <0%> (ø)` | :arrow_up: |\n| [transformers/modeling\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2F1dG8ucHk=) | `31.81% <0%> (ø)` | :arrow_up: |\n| [transformers/tokenization\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rva2VuaXphdGlvbl9hdXRvLnB5) | `45.94% <0%> (ø)` | :arrow_up: |\n| ... and [20 more](https://codecov.io/gh/huggingface/transformers/pull/1969/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1969?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/transformers/pull/1969?src=pr&el=footer). Last update [4910828...265dbe8](https://codecov.io/gh/huggingface/transformers/pull/1969?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\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" ]

[]

[]

[ "Hi, indeed there was a mistake with `TFBertForPretraining` referenced under the ALBERT documentation. This was fixed with 3616209.\r\n\r\nThe `PreTrainedModel` class is not available for ALBERT, which is the same for CTRL, GPT, GPT-2, DistilBERT, CamemBERT, TransformerXL, XLM and XLNet. \r\n\r\nWhy is that class useful for your use-case, seeing as it's a simple wrapper over `PreTrainedModel` with a few overridden attributes?\r\n", "> Hi, indeed there was a mistake with `TFBertForPretraining` referenced under the ALBERT documentation. This was fixed with [3616209](https://github.com/huggingface/transformers/commit/361620954acf16b27727d763a591257b03f90b5d).\r\n> \r\n> The `PreTrainedModel` class is not available for ALBERT, which is the same for CTRL, GPT, GPT-2, DistilBERT, CamemBERT, TransformerXL, XLM and XLNet.\r\n> \r\n> Why is that class useful for your use-case, seeing as it's a simple wrapper over `PreTrainedModel` with a few overridden attributes?\r\n\r\nI completely agree we can use PreTrainedModel or BertPreTrainedModel instead.\r\n \r\nThe question is that I do see the implementation of the class AlbertPreTrainedModel(PreTrainedModel) in the source code (transformers/modeling_albert.py line 313) but I cannot import it. It seems that it is not included in the released version. I just feel it is weird.\r\n", "Yes, it is not importable as it is an internal used by different models, but I fail to see a use-case where it would be useful for the library users.\r\n\r\nWhy is that class useful for your use-case, seeing as it's a simple wrapper over `PreTrainedModel` with a few overridden attributes?", "I think we could add all the `XXXPretrainedModel` in `__init__` indeed. Would make it easier for people to build custom-made models that can load pretrained checkpoints as well.", "Fixed on master" ]

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

[]

[ "I think the error was because text is greater than 512 tokens.\r\nI got no error when text is smaller than 512 tokens.", "Hi! Indeed the models have a maximum input size, which is 512 for BERT. You should have received a warning when tokenizing your sequence, but unfortunately, there isn't much more we can do to clarify this error further.", "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/transformers/pull/1966?src=pr&el=h1) Report\n> Merging [#1966](https://codecov.io/gh/huggingface/transformers/pull/1966?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/cc7968227e08858df4a5c618c739e1a3ca050196?src=pr&el=desc) will **decrease** coverage by `0.01%`.\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1966?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1966 +/- ##\n==========================================\n- Coverage 84.26% 84.24% -0.02% \n==========================================\n Files 104 104 \n Lines 15431 15431 \n==========================================\n- Hits 13003 13000 -3 \n- Misses 2428 2431 +3\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1966?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_tf\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/1966/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX2FsYmVydC5weQ==) | `85.49% <100%> (ø)` | :arrow_up: |\n| [transformers/tests/modeling\\_tf\\_common\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/1966/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3RmX2NvbW1vbl90ZXN0LnB5) | `95.63% <0%> (-1.46%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1966?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/transformers/pull/1966?src=pr&el=footer). Last update [cc79682...a1aec9c](https://codecov.io/gh/huggingface/transformers/pull/1966?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Great, thank you for this @billpku !" ]

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

[]

[ "Hi @alexvaca0 ,\r\n\r\nTry to remove the `d_model` parameter in the constructor. Use `config.emb_dim` (`emb_dim` is specified in the [xlm config](https://s3.amazonaws.com/models.huggingface.co/bert/xlm-mlm-en-2048-config.json)), so this should work:\r\n\r\n```python\r\nself.classifier = nn.Linear(config.emb_dim, config.num_labels)\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" ]

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

[]

[ "**You haven't to modify the source code in this script**. When you call `run_squad.py` script, you have to pass the `--save_steps` parameter and set its value to 1000 (as you can see [here](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py#L429).) So, the entire command would be something like that: `python run_squad.py ... --save_steps 1000`\r\n\r\n> ## Questions & Help\r\n> When I run the script `run_squad.py` , it creates model check points every 50th iteration. How to increase model saving checkpoint from 50 to 1000. Where should I edit the 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", "Related to this question, how about making the default value (50) bigger (e.g., 1000) in scripts such as `run_squad.py` and `run_ner.py`?\r\nIf a `--save_steps` option is not specified, and the default value is used, many checkpoints are saved.", "@tomohideshibata you're right that 50 is too low, I bumped it to 500 in 335dd5e.", "Thanks!", "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" ]

[]

[]

[ "Total user-error here. Made a change to a function that writes the TF Records to storage (the name of one of the features) and didn't propagate that info to the function i wrote that reads the TF Records back in, so it wasn't loading my input masks because it was looking for a key that didn't exist." ]

[]

[]

[ "> The code below was fine when transformers 2.1.1\r\n> \r\n> but after I update to transformers 2.2.0\r\n> \r\n> ```\r\n> model = TFBertForSequenceClassification.from_pretrained('bert-base-chinese', num_labels=5)\r\n> model.summary()\r\n> \r\n> optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)\r\n> loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\r\n> metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')\r\n> model.compile(optimizer=optimizer, loss=loss, metrics=[metric])\r\n> \r\n> model_fit = model.fit(train_input_ids, train_label, \r\n> batch_size=8, epochs=1, \r\n> validation_data=(validation_input_ids, validation_label)\r\n> )\r\n> ```\r\n> \r\n> ```\r\n> ValueError: Tried to convert 'dims' to a tensor and failed. \r\n> Error: Cannot convert a partially known TensorShape to a Tensor: (None, 512)\r\n> ```\r\n\r\nThe problem seem to cause by input's shape.\r\n\r\nFor transformers 2.2.0\r\nI fix it myself by modifying file \"transformers/modeling_tf_albert.py\"\r\nOn line 648, I change:\r\ninput_shape = input_ids.shape\r\nInto:\r\ninput_shape = tf.shape(input_ids)\r\nThen the problem fixed.\r\n\r\nFeel free to leave a comment if it work for you.", "@billpku \r\nSorry to test it so lately.\r\n\r\nIt did fix my problem above.\r\nbut it didn't fix the code below which worked fine in 2.1.1 for custom the layer.\r\n\r\n```\r\ninput_layer = Input(shape = (512,), dtype='int64') \r\nbert = TFBertModel.from_pretrained('bert-base-chinese')(input_layer)\r\nbert = bert[0] \r\ndropout = Dropout(0.1)(bert)\r\nflat = Flatten()(dropout)\r\nclassifier = Dense(units=5)(flat) \r\nmodel = Model(inputs=input_layer, outputs=classifier)\r\nmodel.summary()\r\n```\r\n```\r\nValueError: Tried to convert 'dims' to a tensor and failed. \r\nError: Cannot convert a partially known TensorShape to a Tensor: (None, 512)\r\n```", "Facing the same issue here.\r\n\r\n\r\n", "@AmalVijayan \r\nI just checked my problem was fixed somehow.\r\n\r\nHow's yours now?\r\n\r\n\r\n> ```\r\n> input_layer = Input(shape = (512,), dtype='int64') \r\n> bert = TFBertModel.from_pretrained('bert-base-chinese')(input_layer)\r\n> bert = bert[0] \r\n> dropout = Dropout(0.1)(bert)\r\n> flat = Flatten()(dropout)\r\n> classifier = Dense(units=5)(flat) \r\n> model = Model(inputs=input_layer, outputs=classifier)\r\n> model.summary()\r\n> ```\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 } ]

[]

[ "@LysandreJik do we support this one?", "We do support BERT in `run_lm_finetuning`, however, we do not support loading BERT checkpoints from the original BERT implementation.\r\n\r\nIf you wish to load a checkpoint that was pre-trained/fine-tuned using the original implementation (which seems to be what you're doing here), you can first convert to our implementation using [convert_bert_original_tf_checkpoint_to_pytorch](https://github.com/huggingface/transformers/blob/master/transformers/convert_bert_original_tf_checkpoint_to_pytorch.py), it will then be usable by `run_lm_finetuning`.\r\n\r\nIf you wish to use TensorFlow with the outputted model, you can use the script [convert_pytorch_checkpoint_to_tf2](https://github.com/huggingface/transformers/blob/master/transformers/convert_pytorch_checkpoint_to_tf2.py) which will convert the pytorch model back to tensorflow 2.", "@LysandreJik - not related to fine-tuning but converting one of the fine-tuned (rum_lm_finetuning.py) model to tensorflow checkpoint. \r\n\r\nHere is the command I used:\r\nbin/python3.6 convert_pytorch_checkpoint_to_tf2.py --tf_dump_path=\"../tf_test/\" --model_type=\"bert\" --pytorch_checkpoint_path=\"../pytorch_model.bin\" --config_file='../config.json'\r\n\r\nHowever, it was throwing the below error (log and stack trace)\r\n\r\n Converting model type 1/1 bert\r\n Converting checkpoint 1/15: ../pytorch_model.bin - model_type bert\r\n\r\nBuilding TensorFlow model from configuration: {\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"finetuning_task\": null,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 768,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"is_decoder\": false,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 12,\r\n \"num_labels\": 2,\r\n \"output_attentions\": true,\r\n \"output_hidden_states\": true,\r\n \"output_past\": true,\r\n \"pruned_heads\": {},\r\n \"torchscript\": false,\r\n \"type_vocab_size\": 2,\r\n \"use_bfloat16\": false,\r\n \"vocab_size\": 28996\r\n}\r\n\r\nTraceback (most recent call last):\r\n File \"/home/imagen/skc/bert/transformers-2.2.1/transformers/convert_pytorch_checkpoint_to_tf2.py\", line 248, in <module>\r\n only_convert_finetuned_models=args.only_convert_finetuned_models)\r\n File \"/home/imagen/skc/bert/transformers-2.2.1/transformers/convert_pytorch_checkpoint_to_tf2.py\", line 194, in convert_all_pt_checkpoints_to_tf\r\n compare_with_pt_model=compare_with_pt_model)\r\n File \"/home/imagen/skc/bert/transformers-2.2.1/transformers/convert_pytorch_checkpoint_to_tf2.py\", line 115, in convert_pt_checkpoint_to_tf\r\n tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)\r\n File \"/home/imagen/skc/environments/.virtualenvs/lstm_dev_tf2x/lib/python3.6/site-packages/transformers/modeling_tf_pytorch_utils.py\", line 82, in load_pytorch_checkpoint_in_tf2_model\r\n return load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=tf_inputs, allow_missing_keys=allow_missing_keys)\r\n File \"/home/imagen/skc/environments/.virtualenvs/lstm_dev_tf2x/lib/python3.6/site-packages/transformers/modeling_tf_pytorch_utils.py\", line 145, in load_pytorch_weights_in_tf2_model\r\n assert name in pt_state_dict, \"{} not found in PyTorch model\".format(name)\r\n**AssertionError: cls.seq_relationship.weight not found in PyTorch model**\r\n\r\nCan you please explain what did go wrong with the conversion? This is one of the BERT-base fine-tuned model. ", "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 think it's very useful this feature because, as you highlight, VMs and other variables could stop the (long) training process.\r\nTechnically, how could you implement this feature? Surrounding all the training code in a `try/except` statement and when it occurs a particular Exception (which one?) you ask the user whether he/she wants to save till now and after that you saved a file?\r\n\r\nIt could be necessary to write a method for saving the optimizer state, the scheduler and the tokenizer in a **standardized** way. Reading #1925 and #839, I understand that @thomwolf suggests to use standard PyTorch method for saving and loading e.g. scheduler.\r\n\r\n> ## Feature\r\n> The official transformer examples should make it easier to continue training a model that suddenly stopped (e.g. vm gets preempted in the middle of a training run).\r\n> \r\n> To do this, the examples should be updated to save the `optimizer` and `scheduler` states to the `output_dir` as well as the current epoch to disk at the end of each epoch in a `training_state.pt` file. This way, the user could choose to continue training from a previous model checkpoint, but would continue training from the saved epoch and would use the saved tokenizer, optimizer, and scheduler.\r\n> \r\n> ## Motivation\r\n> If your VM gets preempted in the middle of a training run, you won't be able to properly continue training the model since the scheduler will be reset and the current learning rate will be lost.\r\n> \r\n> ## Additional context\r\n> If anyone is interested in this, I can implement the feature and start a pull request.\r\n> \r\n> Related issues:\r\n> \r\n> * #1925\r\n> * #839", "This would be very useful indeed. \r\n\r\nI would guess that when the VM gets preempted the process in which your program runs is sent a `SIGTERM` or `SIGKILL` signal from the OS. You would need to catch this signal and act accordingly. Look at the [signal module](https://docs.python.org/3/library/signal.html) in python's standard library.\r\n\r\nAn elegant and very general solution would be to define a context manager (`with ... do`) in which we would execute the training and that handles all the backup logic on `SIGTERM` or `SIGKILL`.\r\n\r\nDo you want to give it a shot and make a Pull Request? You can @ me when you have a first draft and I can have a look and give you feedback.", "@rlouf I was thinking to just save the optimizer and scheduler whenever the model is saved. As for resuming, you could just load in the optimizer state, scheduler state, and current epoch from the checkpoint file when passing in `--model_name_or_path`\r\n\r\nI've got a basic implementation of this on my [fork](https://github.com/bkkaggle/transformers/tree/saving-and-resuming)\r\n", "I've seen your changes in source code. I think this is the \"easiest\" way to handle this feature and I like it. Do you have tested your code with unit testing? I don't see any test suite. Only for being sure that it works as expected.\r\nN.B: I've left some comments under your changes in your repo, please read them.\r\n\r\n> @rlouf I was thinking to just save the optimizer and scheduler whenever the model is saved. As for resuming, you could just load in the optimizer state, scheduler state, and current epoch from the checkpoint file when passing in `--model_name_or_path`\r\n> \r\n> I've got a basic implementation of this on my [fork](https://github.com/bkkaggle/transformers/tree/saving-and-resuming)", "@bkkaggle Definitely the easiest solution if you don't mind resuming from the last checkpoint---mine was a bit heavy-duty :)\r\n\r\nI also agree with @TheEdoardo93; Can you rebase your branch on the repo's `master` and open a pull request (if you haven't done so already)? I'm happy to have a closer look.\r\n", "I've updated my branch and submitted a [WIP] [pull request](https://github.com/huggingface/transformers/pull/1987)" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/1959?src=pr&el=h1) Report\n> Merging [#1959](https://codecov.io/gh/huggingface/transformers/pull/1959?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/5e289f69bc564c94132f77c89a34e5f1dd69a592?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1959?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1959 +/- ##\n=======================================\n Coverage 81.47% 81.47% \n=======================================\n Files 122 122 \n Lines 18342 18342 \n=======================================\n Hits 14945 14945 \n Misses 3397 3397\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1959?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/transformers/pull/1959?src=pr&el=footer). Last update [5e289f6...5190320](https://codecov.io/gh/huggingface/transformers/pull/1959?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks a lot @armancohan \r\n\r\nI think we'll want to keep the possibility to import both types of models for backward compatibility. Can you add a switch based on an identification of the model type?", "@armancohan (rebased on top of master so I force-pushed to your fork)", "So, the fairseq weights themselves didn't change, it's the multiheadattention API that did, in fairseq `0.9.0`. So I'll just check that the fairseq version is >= 0.9 in the script.\r\n\r\nI've also updated the script to not hardcode the vocab length, which should make it compatible with other roberta-like fairseq models such as CamemBERT + XLM-R out of the box.\r\n\r\ncc @myleott @louismartin ", "thanks @julien-c " ]

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

[]

[ "Three questions:\r\n- the error occurs at line 522, so at line 507 you've saved the file called `test_results.txt`. Do you see the content of this file and whether is it correct?\r\n- the input file has been formatted as CoNLL-2003?\r\n- **N.B**: moreover, the code block from line 512 to line 525 save to .txt file the predictions obtained from the NER. But already at line 505 you have the `predictions` variable. Have you seen the content of this variable? Maybe, it is only a saving problem.\r\n\r\n> ## Questions & Help\r\n> Hello again,\r\n> \r\n> I'm here to bother you one more time.\r\n> \r\n> I fine-tuned preloaded BioBERT weights on a custom dataset to run biomedical NER.\r\n> \r\n> Now I want to use the model for inference mode on a 'raw' set of documents. I renamed this set 'test.txt' and formatted it the following way (documents are separated by '-DOCSTART- (num_doc)' lines):\r\n> \r\n> ```\r\n> to O\r\n> be O\r\n> referred O\r\n> to O\r\n> the O\r\n> location O\r\n> of O\r\n> the O\r\n> disease O\r\n> in O\r\n> the O\r\n> skeletal O\r\n> structures O\r\n> examined O\r\n> ; O\r\n> \r\n> unchanged O\r\n> the O\r\n> areas O\r\n> of O\r\n> bone O\r\n> rarefaction O\r\n> reported O\r\n> to O\r\n> the O\r\n> sternum O\r\n> as O\r\n> a O\r\n> result O\r\n> of O\r\n> median O\r\n> sternotomy O\r\n> . O\r\n> ```\r\n> \r\n> I had to add the 'fake' labels on the right and place a space \" \" between col1 and col2.\r\n> \r\n> The error I now get is:\r\n> \r\n> ```\r\n> Traceback (most recent call last):\r\n> File \"run_ner.py\", line 531, in <module>\r\n> main()\r\n> File \"run_ner.py\", line 522, in main\r\n> output_line = line.split()[0] + \" \" + predictions[example_id].pop(0) + \"\\n\"\r\n> IndexError: list index out of range\r\n> ```\r\n> \r\n> Many thanks again.", "Ciao @TheEdoardo93 ,\r\nThanks for your support!\r\n\r\n - I formatted the test set trying to follow the indications from the tutorial on the german-eval, with the first column being the token and the second being the B-I-O tags (in this set it's just a pile of Os to fill the column). They are space-separated.\r\n- `test_results.txt` is saved and shows precision, recall, f-1, and loss. All are terrible of course, as the test set was actually filled with the dummy BIO tags.\r\n- `test_predictions.txt` is truncated after about 50 lines of token+BIO prediction.\r\n\r\nI'm now trying to print the content of `predictions`, I'll let you know.", "I wait your `predictions` variable content :D\r\nWe can implement a saving method that works as we expect (and not using the code lines in the `run_ner.py` script and see what happens!)\r\n\r\n> Ciao @TheEdoardo93 ,\r\n> Thanks for your support!\r\n> \r\n> * I formatted the test set trying to follow the indications from the tutorial on the german-eval, with the first column being the token and the second being the B-I-O tags (in this set it's just a pile of Os to fill the column). They are space-separated.\r\n> * `test_results.txt` is saved and shows precision, recall, f-1, and loss. All are terrible of course, as the test set was actually filled with the dummy BIO tags.\r\n> * `test_predictions.txt` is truncated after about 50 lines of token+BIO prediction.\r\n> \r\n> I'm now trying to print the content of `predictions`, I'll let you know.", "I'm back.\r\n\r\nWhat I did was: changed columns' separation from tab to space (I was wrong in the previous comment, I thought I already changed it).\r\n\r\nNow the code runs properly and `test_predictions.txt` is complete.\r\nThis is a snapshot of `print(predictions)`:\r\n```\r\n[['O', 'O', 'B-Organism_subdivision', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Organ', 'O', 'B-Organ', 'O', 'B-Multi-tissue_structure', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Organism_subdivision', 'I-Organism_subdivision', 'O', 'O', 'O', 'O', 'O', 'B-Cancer', 'I-Cancer', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'B-Immaterial_anatomical_entity', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'B-Multi-tissue_structure', 'B-Multi-tissue_structure', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Immaterial_anatomical_entity', 'O', 'O'], ..., ['O', 'O', 'O', 'O', 'B-Organ', 'O', 'B-Organ', 'O', 'B-Organ', 'O', 'B-Organ', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Tissue', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Tissue', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Tissue', 'O', 'O', 'O', 'O', 'B-Multi-tissue_structure', 'O', 'O', 'O', 'O', 'O', 'O', 'O']]\r\n```\r\nThere is another minor issue I guess, i.e. a long series of warnings about no predictions because of exceeded maximum sequence length. The non-predicted tokens appear to be not really relevant for my final interest, but I'd like to have a complete output nonetheless.\r\nI will try to place a newline not only after usual end-of-sentence punctuation (.!?), but also after semi-colons and colons, in order to split each document in more pieces.\r\nIs it a strategy that makes sense or have I misinterpreted the meaning of the maximum sequence length?\r\n", "Do you have sentences with length greater than **512 tokens**? BioBERT allows to have sentences with 512 tokens length maximum, as stated in the [paper](https://watermark.silverchair.com/btz682.pdf?token=AQECAHi208BE49Ooan9kkhW_Ercy7Dm3ZL_9Cf3qfKAc485ysgAAApcwggKTBgkqhkiG9w0BBwagggKEMIICgAIBADCCAnkGCSqGSIb3DQEHATAeBglghkgBZQMEAS4wEQQMYCT-oIlIkXngiEprAgEQgIICSsx4GD3tbVdWpu1VPO7avpF_9v-YeT2MtTj6ysC7p_RRxtqG74n5C_tpst2LvM8UKCDfiuvU4bWi8PfJiJfKoQiWwR4AH4K0JC2m-Q8YC2V8_Jfuk-AL_CR8oWJc2U40FdB2fyV2tdoYxV0v1A35Qjg6PEujCkc3ztZqcGctW1_awddqskqkGF8Oz02kiwFQyHHlMPuRAewMopnxii_Pqo-nSNynCl03WCCGUKPbkC-vbPwIjo7vjz-opJQaNcTrNOLI8xPzQm3qT5_R85w3mm-CpHbo2rj4LW7YkJrswc7Z4KOlEfdq7AC5WkiIYhYqyauVLTDNzVYwSYJ_L6RsPeNlfxv3rm71J7fppWO_fu4Mbn8vnzmjKS0nqxdEbRcI4bGkpkjCvW-sVa3FIcRbNlOp_fH_PTeMf3VwIR_wGR0Nrw_80_BMzqy774SB1LitxarWsA7h3dU7Gp1f162TloTdqISAsTzfJJSTa4YVU2qHDp2iRzghvsBlXGhtuuiNkLQ_TblRFq3hdMpLtpHH5KlfahZ0tMvfBvbc_YGLi-9U5NmQbUnM0unhb73mQ5SneLAAD9JlLQv-4pXwYDIGi9ekn5G2RwueTOKSiKji8dm1rCtmUFXVL56WsPUdNkgJROoqGCC87_iVdV95TjpL7jVvNfOX8Bvh1eF_iCGyfrsKyK1aDpvY8B4vt3uUJowPlFjDo21AXOe53aAgnb9yay-t53WzmTNw-Q6lfZNiWsSQn9H1cUi7g8P5bRruZkmL8HaYlZje8TVNIn4).\r\n> The maximum sequence length was fixed to 512\r\n\r\n If you have sentences with more than 512 tokens, you have to apply different workaround, e.g. splitting a sentence length 1024 in two different sentences of 512 length and combine in some manner their output.\r\n\r\nHowever, the strategy you've proposed (e.g. split by comma, dot, semi-column, etc.) works! Try to follow this approach and share the results with us! I suggest you to do a visual evaluation/comparison between the current output and the output you'll obtain with the strategy highlight by you.\r\n\r\n> I'm back.\r\n> \r\n> What I did was: changed columns' separation from tab to space (I was wrong in the previous comment, I thought I already changed it).\r\n> \r\n> Now the code runs properly and `test_predictions.txt` is complete.\r\n> This is a snapshot of `print(predictions)`:\r\n> \r\n> ```\r\n> [['O', 'O', 'B-Organism_subdivision', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Organ', 'O', 'B-Organ', 'O', 'B-Multi-tissue_structure', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Organism_subdivision', 'I-Organism_subdivision', 'O', 'O', 'O', 'O', 'O', 'B-Cancer', 'I-Cancer', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'B-Immaterial_anatomical_entity', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'B-Multi-tissue_structure', 'B-Multi-tissue_structure', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Immaterial_anatomical_entity', 'O', 'O'], ..., ['O', 'O', 'O', 'O', 'B-Organ', 'O', 'B-Organ', 'O', 'B-Organ', 'O', 'B-Organ', 'O'], ['O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Tissue', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Tissue', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-Tissue', 'O', 'O', 'O', 'O', 'B-Multi-tissue_structure', 'O', 'O', 'O', 'O', 'O', 'O', 'O']]\r\n> ```\r\n> \r\n> There is another minor issue I guess, i.e. a long series of warnings about no predictions because of exceeded maximum sequence length. The non-predicted tokens appear to be not really relevant for my final interest, but I'd like to have a complete output nonetheless.\r\n> I will try to place a newline not only after usual end-of-sentence punctuation (.!?), but also after semi-colons and colons, in order to split each document in more pieces.\r\n> Is it a strategy that makes sense or have I misinterpreted the meaning of the maximum sequence length?", "Quite funnily, now a lot more tokens are without predictions.\r\nWhat I did was just adding a newline after each semicolon with `sed`.\r\n\r\nA question that I thought was easy to answer to: what constitutes a sequences in BERT relative to this task? Is it a sequence of tokens between empty lines? Or between defined punctuation?", "Taken from the official BERT [paper](https://arxiv.org/pdf/1810.04805.pdf):\r\n\r\n> Throughout this work, a “sentence” can be an arbitrary span of contiguous text, rather than an actual linguistic sentence. A “sequence” refers to the input token sequence to BERT, which may be a single sentence or two sentences packed together.", "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" ]

[]

[]

[ "like this ? \r\n```\r\ninput_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\", add_special_tokens=True)).unsqueeze(0) # Batch size 1 \r\n```", "The output **without** `add_special_tokens=True`:\r\n```\r\nimport torch\r\nfrom transformers import BertTokenizer, BertForMaskedLM\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\ntext = 'Hello, my dog is cute'\r\ninput_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)\r\n>>> tensor([[ 101, 7592, 1010, 2026, 3899, 2003, 10140, 102]])\r\n```\r\nIf you inspect the vocabulary built by BertTokenizer (accessible through `tokenizer.vocab`), you can see that the token [CLS] and [SEP] have ID 101 and 102, respectively. So, `tokenizer.encode` already add these two tokens at the start and at the end of each encoded sentence.\r\n\r\nThe output **with** `add_special_tokens=True`:\r\n```\r\nimport torch\r\nfrom transformers import BertTokenizer\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased', add_special_tokens=True)\r\ntext = 'Hello, my dog is cute'\r\ninput_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)\r\n>>> tensor([[ 101, 7592, 1010, 2026, 3899, 2003, 10140, 102]])\r\n```\r\n\r\nAs you can see, **the output obtained is the same**.\r\n\r\nMoreover, this night Transformers passed from 2.1.1 to 2.2.0 version, and reading [here](https://github.com/huggingface/transformers/releases) we can see the statement **Tokenizers now add special tokens by default.**.\r\n\r\n> ## Questions & Help\r\n> https://github.com/huggingface/transformers/blob/cc7968227e08858df4a5c618c739e1a3ca050196/transformers/modeling_bert.py#L837-L841\r\n> \r\n> Seems like the example is wrong?", "Yes, my version v2.1.1 will not set add_special_tokens to True by default. Thanks for your comment.\r\n\r\n> like this ?\r\n> \r\n> ```\r\n> input_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\", add_special_tokens=True)).unsqueeze(0) # Batch size 1 \r\n> ```", "> The output **without** `add_special_tokens=True`:\r\n> \r\n> ```\r\n> import torch\r\n> from transformers import BertTokenizer, BertForMaskedLM\r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> text = 'Hello, my dog is cute'\r\n> input_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)\r\n> >>> tensor([[ 101, 7592, 1010, 2026, 3899, 2003, 10140, 102]])\r\n> ```\r\n> \r\n> If you inspect the vocabulary built by BertTokenizer (accessible through `tokenizer.vocab`), you can see that the token [CLS] and [SEP] have ID 101 and 102, respectively. So, `tokenizer.encode` already add these two tokens at the start and at the end of each encoded sentence.\r\n> \r\n> The output **with** `add_special_tokens=True`:\r\n> \r\n> ```\r\n> import torch\r\n> from transformers import BertTokenizer\r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', add_special_tokens=True)\r\n> text = 'Hello, my dog is cute'\r\n> input_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)\r\n> >>> tensor([[ 101, 7592, 1010, 2026, 3899, 2003, 10140, 102]])\r\n> ```\r\n> \r\n> As you can see, **the output obtained is the same**.\r\n> \r\n> Moreover, this night Transformers passed from 2.1.1 to 2.2.0 version, and reading [here](https://github.com/huggingface/transformers/releases) we can see the statement **Tokenizers now add special tokens by default.**.\r\n> \r\n> > ## Questions & Help\r\n> > https://github.com/huggingface/transformers/blob/cc7968227e08858df4a5c618c739e1a3ca050196/transformers/modeling_bert.py#L837-L841\r\n> > \r\n> > Seems like the example is wrong?\r\n\r\ncreate pull request #1971 to make this less ambiguous across different versions. " ]

[]

[]

[ "You might be using an old version of the library, try updating it to v2.2.0", "You're trying to import a method only available in more recent Transformers versions from a (very old) Transformers version called _pytorch-transformers_.\r\n\r\nWith Transformers 2.1.1 (the second recent one) and the new version 2.2.0, you can import correctly the `get_linear_schedule_with_warmup`. In fact, Transformers modifies its source code for what concern the optimization process (e.g. learning rate). You can see the changes [here](https://github.com/huggingface/transformers/commit/022525b0031bcdbbb62d1223f75919983f2ac426).\r\n\r\n> Hello,\r\n> \r\n> When I try to execute the line of code below, Python gives me an import error:\r\n> \r\n> ```js\r\n> from pytorch_transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n> AdamW, get_linear_schedule_with_warmup)\r\n> \r\n> ImportError: cannot import name 'get_linear_schedule_with_warmup' from 'pytorch_transformers' (/Users/hyunjindominiquecho/opt/anaconda3/lib/python3.7/site-packages/pytorch_transformers/__init__.py)\r\n> ```\r\n> \r\n> What should I then import to use the linear scheduler with warm up?\r\n> \r\n> Thank you,", "You should use the `transformers` library instead of the `pytorch_transformers`. The `get_linear_schedule_with_warmup` is only defined in the former, in its latest version.", "Thank you all,", "Hello,\r\n\r\nSo I installed transformers 2.2.0,\r\n```\r\npip install transformers\r\n```\r\nand tried to import the same things:\r\n```js\r\nfrom transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n AdamW, get_linear_schedule_with_warmup) \r\n```\r\n\r\nand it's still giving me the same error:\r\n```\r\nfrom transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n AdamW, get_linear_schedule_with_warmup)\r\n2019-11-27 08:40:15.940560: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\r\n2019-11-27 08:40:15.954925: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7fe20d2f5a50 executing computations on platform Host. Devices:\r\n2019-11-27 08:40:15.954938: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): Host, Default Version\r\nTraceback (most recent call last):\r\n\r\n File \"<ipython-input-1-99af21631e15>\", line 1, in <module>\r\n from transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n\r\nImportError: cannot import name 'get_linear_schedule_with_warmup' from 'transformers' (/Users/hyunjindominiquecho/opt/anaconda3/lib/python3.7/site-packages/transformers/__init__.py)\r\n```\r\n\r\nWhat should I do to be able to use the linear warm up scheduler?\r\n\r\nThank you,", "It's very strange.. In my environment works as expected the import statement.\r\n```\r\n> import transformers\r\n> transformers.__version__\r\n>>> '2.2.0'\r\n> from transformers import get_linear_schedule_with_warmup\r\n> ...\r\n```\r\nPlease, share with us your **OS** and your **Python version**.\r\n\r\n> Hello,\r\n> \r\n> So I installed transformers 2.2.0,\r\n> \r\n> ```\r\n> pip install transformers\r\n> ```\r\n> \r\n> and tried to import the same things:\r\n> \r\n> ```js\r\n> from transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n> AdamW, get_linear_schedule_with_warmup) \r\n> ```\r\n> \r\n> and it's still giving me the same error:\r\n> \r\n> ```\r\n> from transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n> AdamW, get_linear_schedule_with_warmup)\r\n> 2019-11-27 08:40:15.940560: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA\r\n> 2019-11-27 08:40:15.954925: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7fe20d2f5a50 executing computations on platform Host. Devices:\r\n> 2019-11-27 08:40:15.954938: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): Host, Default Version\r\n> Traceback (most recent call last):\r\n> \r\n> File \"<ipython-input-1-99af21631e15>\", line 1, in <module>\r\n> from transformers import (GPT2Config, GPT2LMHeadModel, GPT2DoubleHeadsModel,\r\n> \r\n> ImportError: cannot import name 'get_linear_schedule_with_warmup' from 'transformers' (/Users/hyunjindominiquecho/opt/anaconda3/lib/python3.7/site-packages/transformers/__init__.py)\r\n> ```\r\n> \r\n> What should I do to be able to use the linear warm up scheduler?\r\n> \r\n> Thank you,", "Hello,\r\n\r\nI tried uninstalling transformers and install the module again and it works now.\r\n\r\nThank you for all your help," ]

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

[]

[ "same issue", "I tracked it down further this morning and found the problem, you cannot run do_eval in pytorch distributed mode, do_eval works completely fine when there is no pytorch distributed in the equation. This should probably result in a change to the README\r\n\r\n", "👍Just ran it, I confirm that do_eval runs well without distributed mode", "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 } ]

[]

[ "Hi, could you try and test this on a more recent version of the library and let me know if it works?", "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 is because in a bert pretraining progress, there are two tasks: masked token prediction and next sentence predition . The first needs hidden state of each tokens ( shape: [batch_size, sequence_length, hidden_size]) the second needs the embedding of the whole sequence (shape : [batch_size, hidden_size] ) . \r\n\r\nAnd there is also position left for some one who want to get all the hidden state from each level inside the model ( may represent different level of abstraction besides the last one ) or the attention matrix. ", "> This is because in a bert pretraining progress, there are two tasks: masked token prediction and next sentence predition . The first needs hidden state of each tokens ( shape: [batch_size, sequence_length, hidden_size]) the second needs the embedding of the whole sequence (shape : [batch_size, hidden_size] ) .\r\n\r\n\r\nBecause of this\r\nif I want use tf.keras to custom the layer below TFBertModel\r\nI have to add this particular line\r\nbert = bert[0]\r\n\r\n```\r\ninput_layer = Input(shape = (512,), dtype='int64') \r\nbert = TFBertModel.from_pretrained('bert-base-chinese')(input_layer)\r\n\r\nbert = bert[0] # I have to add this particular line\r\n\r\ndropout = Dropout(0.1)(bert)\r\nflat = Flatten()(dropout)\r\nclassifier = Dense(units=5)(flat) \r\nmodel = Model(inputs=input_layer, outputs=classifier)\r\nmodel.summary()\r\n```\r\n```\r\nModel: \"model\"\r\n_________________________________________________________________\r\nLayer (type) Output Shape Param # \r\n=================================================================\r\ninput_1 (InputLayer) [(None, 512)] 0 \r\n_________________________________________________________________\r\ntf_bert_model (TFBertModel) ((None, 512, 768), (None, 102267648 \r\n_________________________________________________________________\r\ndropout_37 (Dropout) (None, 512, 768) 0 \r\n_________________________________________________________________\r\nflatten (Flatten) (None, 393216) 0 \r\n_________________________________________________________________\r\ndense (Dense) (None, 5) 1966085 \r\n=================================================================\r\nTotal params: 104,233,733\r\nTrainable params: 104,233,733\r\nNon-trainable params: 0\r\n```", "> > This is because in a bert pretraining progress, there are two tasks: masked token prediction and next sentence predition . The first needs hidden state of each tokens ( shape: [batch_size, sequence_length, hidden_size]) the second needs the embedding of the whole sequence (shape : [batch_size, hidden_size] ) .\r\n> \r\n> Because of this\r\n> if I want use tf.keras to custom the layer below TFBertModel\r\n> I have to add this particular line\r\n> bert = bert[0]\r\n> \r\n> ```\r\n> input_layer = Input(shape = (512,), dtype='int64') \r\n> bert = TFBertModel.from_pretrained('bert-base-chinese')(input_layer)\r\n> \r\n> bert = bert[0] # I have to add this particular line\r\n> \r\n> dropout = Dropout(0.1)(bert)\r\n> flat = Flatten()(dropout)\r\n> classifier = Dense(units=5)(flat) \r\n> model = Model(inputs=input_layer, outputs=classifier)\r\n> model.summary()\r\n> ```\r\n> \r\n> ```\r\n> Model: \"model\"\r\n> _________________________________________________________________\r\n> Layer (type) Output Shape Param # \r\n> =================================================================\r\n> input_1 (InputLayer) [(None, 512)] 0 \r\n> _________________________________________________________________\r\n> tf_bert_model (TFBertModel) ((None, 512, 768), (None, 102267648 \r\n> _________________________________________________________________\r\n> dropout_37 (Dropout) (None, 512, 768) 0 \r\n> _________________________________________________________________\r\n> flatten (Flatten) (None, 393216) 0 \r\n> _________________________________________________________________\r\n> dense (Dense) (None, 5) 1966085 \r\n> =================================================================\r\n> Total params: 104,233,733\r\n> Trainable params: 104,233,733\r\n> Non-trainable params: 0\r\n> ```\r\n\r\nThat's right. But for sentence level classification , I recommend you to use the embedding of whole sequence .\r\n```\r\nbert = bert[1] # instead of bert = bert[0] \r\n```\r\n\r\nJust like what the official sequence classificiation does in **TFBertForSequenceClassification** class at\r\nhttps://github.com/huggingface/transformers/blob/master/transformers/modeling_tf_bert.py", "> ```\r\n> bert = bert[1] # instead of bert = bert[0] \r\n> ```\r\n\r\nMay I ask why?\r\nIt looks like it reduce the features of flatten layer.\r\nIt doesn't look like whole.\r\n```\r\ninput_layer = Input(shape = (512,), dtype='int64') \r\nbert = TFBertModel.from_pretrained('bert-base-chinese')(input_layer)\r\nbert = bert[1] \r\ndropout = Dropout(0.1)(bert)\r\nflat = Flatten()(dropout)\r\nclassifier = Dense(units=5)(flat) \r\nmodel = Model(inputs=input_layer, outputs=classifier)\r\nmodel.summary()\r\n```\r\n```\r\nModel: \"model_5\"\r\n_________________________________________________________________\r\nLayer (type) Output Shape Param # \r\n=================================================================\r\ninput_6 (InputLayer) [(None, 512)] 0 \r\n_________________________________________________________________\r\ntf_bert_model_5 (TFBertModel ((None, 512, 768), (None, 102267648 \r\n_________________________________________________________________\r\ndropout_225 (Dropout) (None, 768) 0 \r\n_________________________________________________________________\r\nflatten_3 (Flatten) (None, 768) 0 \r\n_________________________________________________________________\r\ndense_5 (Dense) (None, 5) 3845 \r\n=================================================================\r\nTotal params: 102,271,493\r\nTrainable params: 102,271,493\r\nNon-trainable params: 0\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", "It is still a problem", "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/transformers/pull/1952?src=pr&el=h1) Report\n> Merging [#1952](https://codecov.io/gh/huggingface/transformers/pull/1952?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/8e5d84fcc1a645d3c13b8a2f64fa995637440dad?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1952?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1952 +/- ##\n======================================\n Coverage 84% 84% \n======================================\n Files 97 97 \n Lines 14340 14340 \n======================================\n Hits 12047 12047 \n Misses 2293 2293\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1952?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/transformers/pull/1952?src=pr&el=footer). Last update [8e5d84f...c6edc47](https://codecov.io/gh/huggingface/transformers/pull/1952?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 } ]

[]

[ "Hello @Pointy-Hat \r\n\r\nPlease tell us more, as explained [here](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md).", "Technical data of the cluster used for computation:\r\nSystem: CentOS Linux release 7.5.1804 (Core)\r\nPython: 3.7.4\r\nPytorch: 1.3.1\r\n\r\nCode run:\r\n```\r\npython3 -m torch.distributed.launch --nproc_per_node 4 ./examples/run_glue.py \\\r\n --model_type bert \\\r\n --model_name_or_path bert-large-uncased-whole-word-masking \\\r\n --task_name MRPC \\\r\n --do_train \\\r\n --do_eval \\\r\n --do_lower_case \\\r\n --data_dir $GLUE_DIR/MRPC/ \\\r\n --max_seq_length 128 \\\r\n --per_gpu_eval_batch_size=8 \\\r\n --per_gpu_train_batch_size=8 \\\r\n --learning_rate 2e-5 \\\r\n --num_train_epochs 3.0 \\\r\n --output_dir /tmp/mrpc_single/ \\\r\n --overwrite_output_dir \\\r\n --overwrite_cache \r\n\r\n```\r\n\r\nExpected results (as per README.txt):\r\n\r\n```\r\nacc = 0.8823529411764706\r\nacc_and_f1 = 0.901702786377709\r\nf1 = 0.9210526315789473\r\n```\r\n\r\nObtained results:\r\n```\r\nacc = 0.8725490196078431\r\nacc_and_f1 = 0.888829254329469\r\nf1 = 0.9051094890510949\r\n```\r\n\r\nGLUE data obtained via their `download_glue_data.py` script, as recommended in README.", "did you try multiple random seeds?", "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 } ]

[]

[ "You can use [`BertModel`](https://huggingface.co/transformers/model_doc/bert.html#transformers.BertModel), it'll return the hidden states for the input sentence.", "Found it, thanks @bkkaggle . Just for others who are looking for the same information. \r\n\r\nUsing Pytorch:\r\n```\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\nmodel = BertModel.from_pretrained('bert-base-uncased')\r\ninput_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\")).unsqueeze(0) # Batch size 1\r\noutputs = model(input_ids)\r\nlast_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n```\r\n\r\n\r\nUsing Tensorflow:\r\n```\r\nimport tensorflow as tf\r\nfrom transformers import BertTokenizer, TFBertModel\r\n\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\nmodel = TFBertModel.from_pretrained('bert-base-uncased')\r\ninput_ids = tf.constant(tokenizer.encode(\"Hello, my dog is cute\"))[None, :] # Batch size 1\r\noutputs = model(input_ids)\r\nlast_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n```", "This is a bit different for `...ForSequenceClassification` models. I've found that the item at `outputs[0]` are the logits and the only way to get the `hidden_states` is to set `config.output_hidden_states=True` when initializing the model. Only then was I able to get the `hidden_states` which are located at `outputs[1]`. \r\n\r\nExample:\r\n\r\n```python3\r\ninputs = {\r\n \"input_ids\": batch[0],\r\n \"attention_mask\": batch[1]\r\n}\r\n\r\noutput = bertmodel(**inputs)\r\nlogits = output[0]\r\nhidden_states = output[1]\r\n```\r\n\r\n", "By using this code, you can obtain a PyTorch tensor of (1, N, 768) shape, where _N_ is the number of different tokens extracted from `BertTokenizer`. If you want to build the sentence vector by exploiting these N tensors, how do you do that? @engrsfi\r\n\r\n> Found it, thanks @bkkaggle . Just for others who are looking for the same information.\r\n> \r\n> Using Pytorch:\r\n> \r\n> ```\r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> model = BertModel.from_pretrained('bert-base-uncased')\r\n> input_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\")).unsqueeze(0) # Batch size 1\r\n> outputs = model(input_ids)\r\n> last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> ```\r\n> \r\n> Using Tensorflow:\r\n> \r\n> ```\r\n> import tensorflow as tf\r\n> from transformers import BertTokenizer, TFBertModel\r\n> \r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> model = TFBertModel.from_pretrained('bert-base-uncased')\r\n> input_ids = tf.constant(tokenizer.encode(\"Hello, my dog is cute\"))[None, :] # Batch size 1\r\n> outputs = model(input_ids)\r\n> last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> ```", "> This is a bit different for `...ForSequenceClassification` models. I've found that the item at `outputs[0]` are the logits and the only way to get the `hidden_states` is to set `config.output_hidden_states=True` when initializing the model. Only then was I able to get the `hidden_states` which are located at `outputs[1]`.\r\n> \r\n> Example:\r\n> \r\n> ```python\r\n> inputs = {\r\n> \"input_ids\": batch[0],\r\n> \"attention_mask\": batch[1]\r\n> }\r\n> \r\n> output = bertmodel(**inputs)\r\n> logits = output[0]\r\n> hidden_states = output[1]\r\n> ```\r\n\r\nI am interested in the last hidden states which are seen as kind of embeddings. I think you are referring to all hidden states including the output of the embedding layer.\r\n\r\n```\r\n\"**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)\r\n list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)\r\n of shape ``(batch_size, sequence_length, hidden_size)``:\r\n Hidden-states of the model at the output of each layer plus the initial embedding outputs\r\n```.", "> By using this code, you can obtain a PyTorch tensor of (1, N, 768) shape, where _N_ is the number of different tokens extracted from `BertTokenizer`. If you want to build the sentence vector by exploiting these N tensors, how do you do that? @engrsfi\r\n> \r\n> > Found it, thanks @bkkaggle . Just for others who are looking for the same information.\r\n> > Using Pytorch:\r\n> > ```\r\n> > tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> > model = BertModel.from_pretrained('bert-base-uncased')\r\n> > input_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\")).unsqueeze(0) # Batch size 1\r\n> > outputs = model(input_ids)\r\n> > last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> > ```\r\n> > \r\n> > \r\n> > Using Tensorflow:\r\n> > ```\r\n> > import tensorflow as tf\r\n> > from transformers import BertTokenizer, TFBertModel\r\n> > \r\n> > tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> > model = TFBertModel.from_pretrained('bert-base-uncased')\r\n> > input_ids = tf.constant(tokenizer.encode(\"Hello, my dog is cute\"))[None, :] # Batch size 1\r\n> > outputs = model(input_ids)\r\n> > last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> > ```\r\n\r\nYou can take an average of them. However, I think the embeddings at first position [CLS] are considered a kind of sentence vector because only those are fed to a further classifier if any for downstream tasks. Disclaimer: I am not sure about it.", "> > This is a bit different for `...ForSequenceClassification` models. I've found that the item at `outputs[0]` are the logits and the only way to get the `hidden_states` is to set `config.output_hidden_states=True` when initializing the model. Only then was I able to get the `hidden_states` which are located at `outputs[1]`.\r\n> > Example:\r\n> > ```python\r\n> > inputs = {\r\n> > \"input_ids\": batch[0],\r\n> > \"attention_mask\": batch[1]\r\n> > }\r\n> > \r\n> > output = bertmodel(**inputs)\r\n> > logits = output[0]\r\n> > hidden_states = output[1]\r\n> > ```\r\n> \r\n> I am interested in the last hidden states which are seen as kind of embeddings. I think you are referring to all hidden states including the output of the embedding layer.\r\n> \r\n> ```\r\n> \"**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)\r\n> list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)\r\n> of shape ``(batch_size, sequence_length, hidden_size)``:\r\n> Hidden-states of the model at the output of each layer plus the initial embedding outputs\r\n> ```.\r\n> ```\r\n\r\nShould be as simple as grabbing the last element in the list:\r\n\r\n```python3\r\nlast_layer = hidden_states[-1]\r\n```\r\n", "@maxzzze According to the documentation, one can get the last hidden states directly without setting this flag to True. See below.\r\nhttps://huggingface.co/transformers/_modules/transformers/modeling_bert.html#BertModel\r\n\r\n```\r\nOutputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:\r\n **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``\r\n Sequence of hidden-states at the output of the last layer of the model.\r\n **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``\r\n Last layer hidden-state of the first token of the sequence (classification token)\r\n further processed by a Linear layer and a Tanh activation function. The Linear\r\n layer weights are trained from the next sentence prediction (classification)\r\n objective during Bert pretraining. This output is usually *not* a good summary\r\n of the semantic content of the input, you're often better with averaging or pooling\r\n the sequence of hidden-states for the whole input sequence.\r\n **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)\r\n list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)\r\n of shape ``(batch_size, sequence_length, hidden_size)``:\r\n Hidden-states of the model at the output of each layer plus the initial embedding outputs.\r\n **attentions**: (`optional`, returned when ``config.output_attentions=True``)\r\n list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:\r\n Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.\r\n```\r\n\r\nBTW, for me, the shape of hidden_states in the below code is `(batch_size, 768)` when I set this Flag to True, not sure if I can extract last hidden states from that.\r\n\r\n```\r\noutput = bertmodel(**inputs)\r\nlogits = output[0]\r\nhidden_states = output[1]\r\n```", "> @maxzzze According to the documentation, one can get the last hidden states directly without setting this flag to True. See below.\r\n> \r\n> ```\r\n> Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:\r\n> **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``\r\n> Sequence of hidden-states at the output of the last layer of the model.\r\n> **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``\r\n> Last layer hidden-state of the first token of the sequence (classification token)\r\n> further processed by a Linear layer and a Tanh activation function. The Linear\r\n> layer weights are trained from the next sentence prediction (classification)\r\n> objective during Bert pretraining. This output is usually *not* a good summary\r\n> of the semantic content of the input, you're often better with averaging or pooling\r\n> the sequence of hidden-states for the whole input sequence.\r\n> **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)\r\n> list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)\r\n> of shape ``(batch_size, sequence_length, hidden_size)``:\r\n> Hidden-states of the model at the output of each layer plus the initial embedding outputs.\r\n> **attentions**: (`optional`, returned when ``config.output_attentions=True``)\r\n> list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:\r\n> Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.\r\n> ```\r\n> \r\n> BTW, for me, the shape of hidden_states in the below code is `(batch_size, 768)` whereas it should be `(batch_size, num_heads, sequence_length, sequence_length)`.\r\n> \r\n> ```\r\n> output = bertmodel(**inputs)\r\n> logits = output[0]\r\n> hidden_states = output[1]\r\n> ```\r\n\r\nI believe your comment is in reference to the standard models, but its hard to tell without a link. Can you link where to where in the documentation the pasted doc string is from? \r\n\r\nI dont know if you saw my original comment but I was providing an example for how to get `hidden_states` from the `..ForSequenceClassification` models, not the standard ones. The `..ForSequenceClassification` models do not output `hidden_states` by default: https://huggingface.co/transformers/model_doc/bert.html#bertforsequenceclassification", "Sorry, I missed that part :) I am referring to the standard BERTMODEL. Doc link:\r\nhttps://huggingface.co/transformers/model_doc/bert.html#bertmodel\r\n\r\n> > @maxzzze According to the documentation, one can get the last hidden states directly without setting this flag to True. See below.\r\n> > ```\r\n> > Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:\r\n> > **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``\r\n> > Sequence of hidden-states at the output of the last layer of the model.\r\n> > **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``\r\n> > Last layer hidden-state of the first token of the sequence (classification token)\r\n> > further processed by a Linear layer and a Tanh activation function. The Linear\r\n> > layer weights are trained from the next sentence prediction (classification)\r\n> > objective during Bert pretraining. This output is usually *not* a good summary\r\n> > of the semantic content of the input, you're often better with averaging or pooling\r\n> > the sequence of hidden-states for the whole input sequence.\r\n> > **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)\r\n> > list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)\r\n> > of shape ``(batch_size, sequence_length, hidden_size)``:\r\n> > Hidden-states of the model at the output of each layer plus the initial embedding outputs.\r\n> > **attentions**: (`optional`, returned when ``config.output_attentions=True``)\r\n> > list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:\r\n> > Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.\r\n> > ```\r\n> > \r\n> > \r\n> > BTW, for me, the shape of hidden_states in the below code is `(batch_size, 768)` whereas it should be `(batch_size, num_heads, sequence_length, sequence_length)`.\r\n> > ```\r\n> > output = bertmodel(**inputs)\r\n> > logits = output[0]\r\n> > hidden_states = output[1]\r\n> > ```\r\n> \r\n> I believe your comment is in reference to the standard models, but its hard to tell without a link. Can you link where to where in the documentation the pasted doc string is from?\r\n> \r\n> I dont know if you saw my original comment but I was providing an example for how to get `hidden_states` from the `..ForSequenceClassification` models, not the standard ones. The `..ForSequenceClassification` models do not output `hidden_states` by default: https://huggingface.co/transformers/model_doc/bert.html#bertforsequenceclassification\r\n\r\n", "@engrsfi @maxzzze @bkkaggle \r\nPlease, look [here](https://jalammar.github.io/a-visual-guide-to-using-bert-for-the-first-time/). I hope it can help :)", "@TheEdoardo93 is this example taking the first element in each of the `hidden_states`?", "@engrsfi You can process the hidden states of BERT (all layers or only the last layer) in whatever way you want.\r\n\r\nMost people usually only take the hidden states of the [CLS] token of the last layer - using the hidden states for all tokens or from multiple layers doesn't usually help you that much.\r\n\r\nIf you want to get the embeddings for classification, just do something like:\r\n\r\n```\r\ninput_sentence = torch.tensor(tokenizer.encode(\"[CLS] My sentence\")).unsqueeze(0)\r\nout = model(input_sentence)\r\nembeddings_of_last_layer = out[0]\r\ncls_embeddings = embeddings_of_last_layer[0]\r\n```", "> @engrsfi You can process the hidden states of BERT (all layers or only the last layer) in whatever way you want.\r\n> \r\n> Most people usually only take the hidden states of the [CLS] token of the last layer - using the hidden states for all tokens or from multiple layers doesn't usually help you that much.\r\n> \r\n> If you want to get the embeddings for classification, just do something like:\r\n> \r\n> ```\r\n> input_sentence = torch.tensor(tokenizer.encode(\"[CLS] My sentence\")).unsqueeze(0)\r\n> out = model(input_sentence)\r\n> embeddings_of_last_layer = out[0]\r\n> cls_embeddings = embeddings_of_last_layer[0]\r\n> ```\r\n\r\nDo you have any reference as to \"people usually only take the hidden states of the [CLS] token of the last layer\"?", "Here are a few related links: [1](https://github.com/google-research/bert/issues/196), [2](https://github.com/hanxiao/bert-as-service#q-what-are-the-available-pooling-strategies), [3](https://yashuseth.blog/2019/06/12/bert-explained-faqs-understand-bert-working/)\r\n\r\nThe [CLS] token isn't the only (or necessarily the best) way to finetune, but it is the easiest and is Bert's default", "There is some clarification about the use of the last hidden states in the BERT Paper.\r\n According to the paper, the last hidden state for [CLS] is mainly used for classification tasks and the last hidden states for all tokens are used for token level tasks such as sequence tagging or question answering. \r\n\r\nFrom the paper:\r\n\r\n> At the output, the token representations are fed into an output layer for token level tasks, such as sequence tagging or question answering, and the [CLS] representation is fed into an output layer for classification, such as entailment or sentiment analysis.\r\n\r\nReference: \r\nBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (https://arxiv.org/pdf/1810.04805.pdf)", "What about ALBERT? The output of the last hidden state isn't the same of the embedding because in the doc they say that the embedding have a size of 128 for every model (https://arxiv.org/pdf/1909.11942.pdf page 6).\r\nBut I'm not sure if the 128-embedding referenced in the table is something internally used to represent words or the final word embedding.", "> Found it, thanks @bkkaggle . Just for others who are looking for the same information.\r\n> \r\n> Using Pytorch:\r\n> \r\n> ```\r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> model = BertModel.from_pretrained('bert-base-uncased')\r\n> input_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\")).unsqueeze(0) # Batch size 1\r\n> outputs = model(input_ids)\r\n> last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> ```\r\n> \r\n> Using Tensorflow:\r\n> \r\n> ```\r\n> import tensorflow as tf\r\n> from transformers import BertTokenizer, TFBertModel\r\n> \r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> model = TFBertModel.from_pretrained('bert-base-uncased')\r\n> input_ids = tf.constant(tokenizer.encode(\"Hello, my dog is cute\"))[None, :] # Batch size 1\r\n> outputs = model(input_ids)\r\n> last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> ```\r\n\r\nif batch size is N, how to convert?", "> What about ALBERT? The output of the last hidden state isn't the same of the embedding because in the doc they say that the embedding have a size of 128 for every model (https://arxiv.org/pdf/1909.11942.pdf page 6).\r\n> But I'm not sure if the 128-embedding referenced in the table is something internally used to represent words or the final word embedding.\r\n\r\n128 is used internally by Albert. The output of the model (last hidden state) is your actual word embeddings. In order to understand this better, you should read the following blog from Google.\r\nhttps://ai.googleblog.com/2019/12/albert-lite-bert-for-self-supervised.html\r\n\r\nQuote: \r\n\"The key to optimizing performance, captured in the design of ALBERT, is to allocate the model’s capacity more efficiently. Input-level embeddings (words, sub-tokens, etc.) need to learn context-independent representations, a representation for the word “bank”, for example. In contrast, hidden-layer embeddings need to refine that into context-dependent representations, e.g., a representation for “bank” in the context of financial transactions, and a different representation for “bank” in the context of river-flow management.\r\n\r\n**This is achieved by factorization of the embedding parametrization — the embedding matrix is split between input-level embeddings with a relatively-low dimension (e.g., 128), while the hidden-layer embeddings use higher dimensionalities (768 as in the BERT case, or more).** With this step alone, ALBERT achieves an 80% reduction in the parameters of the projection block, at the expense of only a minor drop in performance — 80.3 SQuAD2.0 score, down from 80.4; or 67.9 on RACE, down from 68.2 — with all other conditions the same as for BERT.\"\r\n", "> > Found it, thanks @bkkaggle . Just for others who are looking for the same information.\r\n> > Using Pytorch:\r\n> > ```\r\n> > tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> > model = BertModel.from_pretrained('bert-base-uncased')\r\n> > input_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\")).unsqueeze(0) # Batch size 1\r\n> > outputs = model(input_ids)\r\n> > last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> > ```\r\n> > \r\n> > \r\n> > Using Tensorflow:\r\n> > ```\r\n> > import tensorflow as tf\r\n> > from transformers import BertTokenizer, TFBertModel\r\n> > \r\n> > tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> > model = TFBertModel.from_pretrained('bert-base-uncased')\r\n> > input_ids = tf.constant(tokenizer.encode(\"Hello, my dog is cute\"))[None, :] # Batch size 1\r\n> > outputs = model(input_ids)\r\n> > last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> > ```\r\n> \r\n> if batch size is N, how to convert?\r\n\r\nIf I understand you correctly, you are asking for how to get the last hidden states for all entries in a batch of size N. If that's the case, then here is the explanation. \r\n\r\nYour model expect input of the following shape:\r\n\r\n`(batch_size, sequence_length)`\r\n\r\nand returns last hidden states of the following shape:\r\n\r\n`(batch_size, sequence_length, hidden_size)`\r\n\r\nYou can just go through the last hidden states to get the individual last hidden state for each input in the batch size of N.\r\n\r\nReference:\r\nhttps://huggingface.co/transformers/model_doc/bert.html", "@engrsfi : What if I want to use bert embedding vector of each token as an input to an LSTM network? Can I get the embedding of each token of the sentence from the last hidden layer of the bert model? In this case I think I can't just use the embedding for [CLS] token as I need word embedding of each token?\r\n I used the code below to get bert's word embedding for all tokens of my sentences. I padded all my sentences to have maximum length of 80 and also used attention mask to ignore padded elements. in this case the shape of last_hidden_states element is of size (batch_size ,80 ,768). However, when I see my embeddings, I can see that embedding vectors for padded elements are not the same? like I have a vector of size 768 for each token of the sentence(most of them are padded tokens). but vectors for padded element are not equal. is it natural?\r\n\r\nimport tensorflow as tf\r\nimport numpy as np \r\nfrom transformers import BertTokenizer, TFBertModel\r\n\r\nbert_model = TFBertModel.from_pretrained(\"bert-base-uncased\")\r\nbert_tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\r\ntokenized = x_train['token'].apply((lambda x: bert_tokenizer.encode(x, add_special_tokens=True, max_length=80)))\r\npadded = np.array([i + [0]*(80-len(i)) for i in tokenized.values])\r\nattention_mask = np.where(padded != 0, 1, 0)\r\ninput_ids = tf.constant(padded) \r\nattention_mask = tf.constant(attention_mask)\r\noutput= bert_model(input_ids, attention_mask=attention_mask)\r\nlast_hidden_states=output[0]", "> How can I extract embeddings for a sentence or a set of words directly from pre-trained models (Standard BERT)? For example, I am using Spacy for this purpose at the moment where I can do it as follows:\r\n> \r\n> sentence vector:\r\n> `sentence_vector = bert_model(\"This is an apple\").vector`\r\n> \r\n> word_vectors:\r\n> \r\n> ```\r\n> words = bert_model(\"This is an apple\")\r\n> word_vectors = [w.vector for w in words]\r\n> ```\r\n> \r\n> I am wondering if this is possible directly with huggingface pre-trained models (especially BERT).\r\n\r\nHi, could I ask how you would use Spacy to do this? Is there a link? Thanks a lot. ", "> > How can I extract embeddings for a sentence or a set of words directly from pre-trained models (Standard BERT)? For example, I am using Spacy for this purpose at the moment where I can do it as follows:\r\n> > sentence vector:\r\n> > `sentence_vector = bert_model(\"This is an apple\").vector`\r\n> > word_vectors:\r\n> > ```\r\n> > words = bert_model(\"This is an apple\")\r\n> > word_vectors = [w.vector for w in words]\r\n> > ```\r\n> > \r\n> > \r\n> > I am wondering if this is possible directly with huggingface pre-trained models (especially BERT).\r\n> \r\n> Hi, could I ask how you would use Spacy to do this? Is there a link? Thanks a lot.\r\n\r\nHere is the link:\r\nhttps://spacy.io/usage/vectors-similarity", "> @engrsfi You can process the hidden states of BERT (all layers or only the last layer) in whatever way you want.\r\n> \r\n> Most people usually only take the hidden states of the [CLS] token of the last layer - using the hidden states for all tokens or from multiple layers doesn't usually help you that much.\r\n> \r\n> If you want to get the embeddings for classification, just do something like:\r\n> \r\n> ```\r\n> input_sentence = torch.tensor(tokenizer.encode(\"[CLS] My sentence\")).unsqueeze(0)\r\n> out = model(input_sentence)\r\n> embeddings_of_last_layer = out[0]\r\n> cls_embeddings = embeddings_of_last_layer[0]\r\n> ```\r\n\r\nThank you for sharing the code. It really helped in understanding tokenization in BERT. I ran this and had a minor problem. Shouldn't it be: \r\n\r\n```cls_embeddings = embeddings_of_last_layer[0][0]```? This is because embeddings_of_last_layer is of the dimension: 1*#tokens*#hidden-units. Then, since [CLS] is the first token (and usually have 101 as id), we want embedding corresponding to just [CLS]. ```embeddings_of_last_layer[0]``` is of shape #tokens*#hidden-units and contains embeddings of all the tokens.", "@sahand91 \r\npooled_output, sequence_output = bert_model(input_)\r\npooled_output.shape = (1, 768), one vector on 768 entries (represent the whole sentence)\r\nsequence_output.shape = (batch_size, max_len, dim), (1, 256, 768) bs = 1, n_tokens = 256\r\nsequence output gives the vector for each token of the sentence. \r\n\r\nI have used the sequence output for classification task like sentiment analysis. As the paper mentions that the pooled output is not a good representation of the whole sentence so we use the sequence output and feed it further in a CNN or LSTM. \r\n\r\nSo I don't see any problem in using the sequence output for classification tasks as we get to see the actual vector representation of the word say \"bank\" in both contexts \"commercial\" and \"location\" (bank of a river) ", "> > @engrsfi You can process the hidden states of BERT (all layers or only the last layer) in whatever way you want.\r\n> > Most people usually only take the hidden states of the [CLS] token of the last layer - using the hidden states for all tokens or from multiple layers doesn't usually help you that much.\r\n> > If you want to get the embeddings for classification, just do something like:\r\n> > ```\r\n> > input_sentence = torch.tensor(tokenizer.encode(\"[CLS] My sentence\")).unsqueeze(0)\r\n> > out = model(input_sentence)\r\n> > embeddings_of_last_layer = out[0]\r\n> > cls_embeddings = embeddings_of_last_layer[0]\r\n> > ```\r\n> \r\n> Thank you for sharing the code. It really helped in understanding tokenization in BERT. I ran this and had a minor problem. Shouldn't it be:\r\n> \r\n> `cls_embeddings = embeddings_of_last_layer[0][0]`? This is because embeddings_of_last_layer is of the dimension: 1*#tokens*#hidden-units. Then, since [CLS] is the first token (and usually have 101 as id), we want embedding corresponding to just [CLS]. `embeddings_of_last_layer[0]` is of shape #tokens*#hidden-units and contains embeddings of all the tokens.\r\n\r\nYes i think the same. @sumitsidana \r\nembeddings_of_last_layer[0][0].shape\r\nOut[179]: torch.Size([144]) # where 144 in my case is the hidden_size\r\n\r\nAnyone confirming that embeddings_of_last_layer[0][0] is the embedding related to CLS token for each sequence?", "> > > @engrsfi You can process the hidden states of BERT (all layers or only the last layer) in whatever way you want.\r\n> > > Most people usually only take the hidden states of the [CLS] token of the last layer - using the hidden states for all tokens or from multiple layers doesn't usually help you that much.\r\n> > > If you want to get the embeddings for classification, just do something like:\r\n> > > ```\r\n> > > input_sentence = torch.tensor(tokenizer.encode(\"[CLS] My sentence\")).unsqueeze(0)\r\n> > > out = model(input_sentence)\r\n> > > embeddings_of_last_layer = out[0]\r\n> > > cls_embeddings = embeddings_of_last_layer[0]\r\n> > > ```\r\n> > \r\n> > \r\n> > Thank you for sharing the code. It really helped in understanding tokenization in BERT. I ran this and had a minor problem. Shouldn't it be:\r\n> > `cls_embeddings = embeddings_of_last_layer[0][0]`? This is because embeddings_of_last_layer is of the dimension: 1*#tokens*#hidden-units. Then, since [CLS] is the first token (and usually have 101 as id), we want embedding corresponding to just [CLS]. `embeddings_of_last_layer[0]` is of shape #tokens*#hidden-units and contains embeddings of all the tokens.\r\n> \r\n> Yes i think the same. @sumitsidana\r\n> embeddings_of_last_layer[0][0].shape\r\n> Out[179]: torch.Size([144]) # where 144 in my case is the hidden_size\r\n> \r\n> Anyone confirming that embeddings_of_last_layer[0][0] is the embedding related to CLS token for each sequence?\r\n\r\nYes it is. but it is only for first batch. you will have to loop through all the batches and get the first element (CLS) for each sentence.", "Yes gotcha. Thanks", "> This is a bit different for `...ForSequenceClassification` models. I've found that the item at `outputs[0]` are the logits and the only way to get the `hidden_states` is to set `config.output_hidden_states=True` when initializing the model. Only then was I able to get the `hidden_states` which are located at `outputs[1]`.\r\n> \r\n> Example:\r\n> \r\n> ```python\r\n> inputs = {\r\n> \"input_ids\": batch[0],\r\n> \"attention_mask\": batch[1]\r\n> }\r\n> \r\n> output = bertmodel(**inputs)\r\n> logits = output[0]\r\n> hidden_states = output[1]\r\n> ```\r\nlogtis = output[0] means the word embedding. So, does hidden_states = output[1] means the sentence level embedding ?", "> Found it, thanks @bkkaggle . Just for others who are looking for the same information.\r\n> \r\n> Using Pytorch:\r\n> \r\n> ```\r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> model = BertModel.from_pretrained('bert-base-uncased')\r\n> input_ids = torch.tensor(tokenizer.encode(\"Hello, my dog is cute\")).unsqueeze(0) # Batch size 1\r\n> outputs = model(input_ids)\r\n> last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> ```\r\n> \r\n> Using Tensorflow:\r\n> \r\n> ```\r\n> import tensorflow as tf\r\n> from transformers import BertTokenizer, TFBertModel\r\n> \r\n> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')\r\n> model = TFBertModel.from_pretrained('bert-base-uncased')\r\n> input_ids = tf.constant(tokenizer.encode(\"Hello, my dog is cute\"))[None, :] # Batch size 1\r\n> outputs = model(input_ids)\r\n> last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple\r\n> ```\r\n\r\noutputs[0] is sentence embedding for \"Hello, my dog is cute\" right?\r\nthen what is outputs[1]?" ]

[]

[]

[ "You can initialize the weights of your model as the ones of e.g. BERT, and after that you can fine-tune your model with _your own data_ (**transfer learning**). Please see [here](https://github.com/huggingface/transformers/blob/master/examples/run_lm_finetuning.py) for fine-tuning for a particular task.\r\n\r\nHave i answered to you?\r\n\r\n> ## Questions & Help\r\n> I get the idea of using `from_pretrained` but Can i train my own text corpus and then get the weights? If So how?", "> You can initialize the weights of your model as the ones of e.g. BERT, and after that you can fine-tune your model with _your own data_ (**transfer learning**). Please see [here](https://github.com/huggingface/transformers/blob/master/examples/run_lm_finetuning.py) for fine-tuning for a particular task.\r\n> \r\n> Have i answered to you?\r\n> \r\n> > ## Questions & Help\r\n> > I get the idea of using `from_pretrained` but Can i train my own text corpus and then get the weights? If So how?\r\n\r\nHey Thanks! I'll go through the source code you referred, also just wanted to confirm the same goes with gpt-2 model right?", "Yeah, this particular script works with OpenAI GPT-2 too.\r\nIn general, the most part of the code is the same even changing the model chosen.\r\n\r\n> > You can initialize the weights of your model as the ones of e.g. BERT, and after that you can fine-tune your model with _your own data_ (**transfer learning**). Please see [here](https://github.com/huggingface/transformers/blob/master/examples/run_lm_finetuning.py) for fine-tuning for a particular task.\r\n> > Have i answered to you?\r\n> > > ## Questions & Help\r\n> > > I get the idea of using `from_pretrained` but Can i train my own text corpus and then get the weights? If So how?\r\n> \r\n> Hey Thanks! I'll go through the source code you referred, also just wanted to confirm the same goes with gpt-2 model right?", "Thanks for the quick response, i'll close this one then." ]

[]

[]

[ "Hi,\r\n\r\nThe attention mask is useful when there are padding indices on which you do not want to perform attention. You should use it when you're using padding, which should only happen when having a batch size superior to one with different sized sequences in those batches.", "> Hi,\r\n> \r\n> The attention mask is useful when there are padding indices on which you do not want to perform attention. You should use it when you're using padding, which should only happen when having a batch size superior to one with different sized sequences in those batches.\r\n\r\nThanks for your reply! @LysandreJik \r\nSo, you mean I should use it when I use batch data and there are different sized sequences in batch, otherwise I shouldn’t. \r\nBut why the official demo is ok? I guess it will implement PAD again no matter what I have already implemented PAD manually or not? Right?", "> So, you mean I should use it when I use batch data and there are different sized sequences in batch, otherwise I shouldn’t.\r\n\r\nExactly. You can use them, but you don't need to. You probably shouldn't because of the performance cost.\r\n\r\n> But why the official demo is ok? I guess it will implement PAD again no matter what I have already implemented PAD manually or not? Right?\r\n\r\nIt is ok because there is only one sentence.", "> > So, you mean I should use it when I use batch data and there are different sized sequences in batch, otherwise I shouldn’t.\r\n> \r\n> Exactly. You can use them, but you don't need to. You probably shouldn't because of the performance cost.\r\n> \r\n> > But why the official demo is ok? I guess it will implement PAD again no matter what I have already implemented PAD manually or not? Right?\r\n> \r\n> It is ok because there is only one sentence.\r\n\r\nThank you! @rlouf \r\n1. I think those tokens that have been padded should not be paid attention. So I don't know why I don't need to.\r\n2. The most models have a fixed length(max_length), the sentences should pad to it before feed. I mean why the official demo doesn't pad.", "> 1. I think those tokens that have been padded should not be paid attention. So I don't know why I don't need to.\r\n\r\nSorry, now I re-read my response I realize it was not very clear. I meant that if they are the same size you can still pad, but you'd hit the performance. If they are not, then you should absolutely pad and pass the appropriate attention mask.\r\n\r\n> 2. The most models have a fixed length(max_length), the sentences should pad to it before feed. I mean why the official demo doesn't pad.\r\n\r\nYes but sentences don't all have the same length:\r\n\r\n```\r\n// No need to pad this\r\n[[1, 2, 3],\r\n[5, 6, 7]]\r\n```\r\n\r\nBUT\r\n\r\n```\r\n// Here you should pad\r\n[[1, 2, 3, pad_token_id],\r\n[5, 6, 7, 8]]\r\n```", "Thank you very much! I got it. @rlouf " ]

[]

[]

[ "in probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)", "Please, describe your environment, post the source code for reproducibility and the error.\r\n\r\n> in probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)", "Please upgrade your Pytorch version to 1.2.0+.", "You're probably passing in a boolean tensor (true or false) instead of a byte tensor (0 or 1) for your attention mask.\r\n\r\nTry changing\r\n```\r\nprobability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)\r\n```\r\n\r\nto\r\n```\r\nprobability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.uint8), value=0.0)\r\n```" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/1946?src=pr&el=h1) Report\n> Merging [#1946](https://codecov.io/gh/huggingface/transformers/pull/1946?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/5d3b8daad2cc6287d30f03f8a96d0a1f7bc8d0dc?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/1946?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #1946 +/- ##\n======================================\n Coverage 84% 84% \n======================================\n Files 97 97 \n Lines 14340 14340 \n======================================\n Hits 12047 12047 \n Misses 2293 2293\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/1946?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/1946/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2JlcnQucHk=) | `87.77% <100%> (ø)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/1946?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/transformers/pull/1946?src=pr&el=footer). Last update [5d3b8da...e1d116d](https://codecov.io/gh/huggingface/transformers/pull/1946?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]