Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4720?src=pr&el=h1) Report\n> Merging [#4720](https://codecov.io/gh/huggingface/transformers/pull/4720?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/47a551d17b6ed2eaf03301f049006d559fca5cf3&el=desc) will **decrease** coverage by `0.56%`.\n> The diff coverage is `96.92%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4720?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4720 +/- ##\n==========================================\n- Coverage 77.14% 76.57% -0.57% \n==========================================\n Files 128 128 \n Lines 21073 21089 +16 \n==========================================\n- Hits 16256 16149 -107 \n- Misses 4817 4940 +123 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4720?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4720/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `89.94% <ø> (ø)` | |\n| [src/transformers/modeling\\_reformer.py](https://codecov.io/gh/huggingface/transformers/pull/4720/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19yZWZvcm1lci5weQ==) | `88.21% <96.92%> (+0.26%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4720/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `31.73% <0.00%> (-40.39%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4720/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.50% <0.00%> (-0.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4720/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `81.78% <0.00%> (+1.37%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4720?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/4720?src=pr&el=footer). Last update [47a551d...6fe9553](https://codecov.io/gh/huggingface/transformers/pull/4720?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 }, { "id": 1834067346, "node_id": "MDU6TGFiZWwxODM0MDY3MzQ2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Documentation", "name": "Documentation", "color": "77cc3b", "default": false, "description": "" } ]

[]

[ "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", "there is a `BasicTokenizer` in previous `pytorch_transformers` package:\r\n\r\nThis must be what \"basic tokenize\" does" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4718?src=pr&el=h1) Report\n> Merging [#4718](https://codecov.io/gh/huggingface/transformers/pull/4718?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/e80d6c689bd62f805a5c8d77ec0cc3b09f240d14&el=desc) will **decrease** coverage by `0.64%`.\n> The diff coverage is `100.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4718?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4718 +/- ##\n==========================================\n- Coverage 77.10% 76.45% -0.65% \n==========================================\n Files 128 128 \n Lines 21723 21725 +2 \n==========================================\n- Hits 16749 16610 -139 \n- Misses 4974 5115 +141 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4718?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/data/data\\_collator.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL2RhdGFfY29sbGF0b3IucHk=) | `89.55% <100.00%> (+0.32%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_pytorch\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9weXRvcmNoX3V0aWxzLnB5) | `8.72% <0.00%> (-81.21%)` | :arrow_down: |\n| [src/transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19jdHJsLnB5) | `96.56% <0.00%> (-2.58%)` | :arrow_down: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `79.51% <0.00%> (-1.39%)` | :arrow_down: |\n| [src/transformers/trainer.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90cmFpbmVyLnB5) | `38.08% <0.00%> (-1.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `86.49% <0.00%> (-0.78%)` | :arrow_down: |\n| [src/transformers/benchmark/benchmark\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9iZW5jaG1hcmsvYmVuY2htYXJrX3V0aWxzLnB5) | `72.80% <0.00%> (-0.30%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.79% <0.00%> (ø)` | |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4718/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.61% <0.00%> (+0.11%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4718?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/4718?src=pr&el=footer). Last update [e80d6c6...6cceabd](https://codecov.io/gh/huggingface/transformers/pull/4718?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Closed the PR by mistake... Re-opening it.", "Just a quick question for @mfuntowicz, is `copy.deepcopy` a performant way to clone a tensor? (given that this is called at each training step)", "I saw `deepcopy` being used elsewhere in the code, so added it as is. Just looking at the new tensor [documentation](https://pytorch.org/docs/stable/tensors.html#torch.Tensor.new_tensor), and they recommend using `.clone().detach()`. Happy to change it to that.", "Bumping this since I haven't seen any activity in a few days.", "Yes `.clone().detach()` sounds good.", "LGTM but let's let @LysandreJik have a last check and merge this", "Thanks @julien-c!", "Hey @LysandreJik, can you please review when you have a chance? Thanks!", "Thanks @setu4993!" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4717?src=pr&el=h1) Report\n> Merging [#4717](https://codecov.io/gh/huggingface/transformers/pull/4717?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/76779363160a598f130433209a77f8a747351b61&el=desc) will **decrease** coverage by `1.23%`.\n> The diff coverage is `50.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4717?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4717 +/- ##\n==========================================\n- Coverage 77.38% 76.14% -1.24% \n==========================================\n Files 128 128 \n Lines 21071 21073 +2 \n==========================================\n- Hits 16305 16046 -259 \n- Misses 4766 5027 +261 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4717?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `89.59% <50.00%> (+0.02%)` | :arrow_up: |\n| [src/transformers/modeling\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19sb25nZm9ybWVyLnB5) | `18.70% <0.00%> (-74.83%)` | :arrow_down: |\n| [src/transformers/configuration\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2xvbmdmb3JtZXIucHk=) | `76.92% <0.00%> (-23.08%)` | :arrow_down: |\n| [src/transformers/modeling\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190NS5weQ==) | `77.18% <0.00%> (-6.35%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.44% <0.00%> (-0.42%)` | :arrow_down: |\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `86.77% <0.00%> (-0.19%)` | :arrow_down: |\n| [src/transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/4717/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3F1YWQucHk=) | `56.68% <0.00%> (+28.02%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4717?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/4717?src=pr&el=footer). Last update [7677936...8217ee0](https://codecov.io/gh/huggingface/transformers/pull/4717?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

[]

[ "As per my understanding, you can! \r\n\r\nIf you read the BERT paper by Devlin et.al., you can see that two suggested ways to extract said word embeddings would be to concatenate the last four hidden layers (9 to 12), generating a 4*768=3072 sized embedding for each token. Alternatively you can also sum or average out the last 4 layers to generate vectors of size 768.\r\n\r\nYou can also instead prefer storing sentence embeddings, then the CLS token serves as averaged representation of the sentence embedding (not a very good representation though, unless the model was fine-tuned well on the data or the LM was pretrained on similar data, and even then the sentence embeddings through that CLS token could be substandard). Although, averaging out the second to last hidden layer (i.e. averaging out token embeddings for the sentence at the 11th hidden layer) seems to generate decent sentence embeddings (of size 768, the same as one token, since its averaged and not concatenated). \r\n\r\nHope this helps!", "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" ]

[]

[]

[ "This might help: #3444", "`MODEL_WITH_LM_HEAD_MAPPING` is a mapping containing all the *Pytorch* models that have an LM head. Since you don't have PyTorch installed, you can't import them.\r\n\r\nWith TensorFlow you're probably looking for `TF_MODEL_WITH_LM_HEAD_MAPPING`.\r\n\r\nPlease note that the `run_language_modeling.py` script is currently PyTorch only. A TensorFlow version will be available in the future." ]

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

[]

[ "\"Finetuning\" a GPT-2 model is generally used for generation, to make it output text more similar to a given dataset.\r\n\r\nIf you are using it for other tasks (e.g. binary predictions), you may not have to finetune it.", "Hello,\r\n\r\nThank you for your reply.\r\nSo say if I am trying to solve multiple choice questions with GPT2DoubleHeadsModel, should I just use the pre-trained model without fine-tuning?\r\n\r\nThanks, ", "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" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4711?src=pr&el=h1) Report\n> Merging [#4711](https://codecov.io/gh/huggingface/transformers/pull/4711?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/6449c494d0f40f0b70442f3da9e61f042ff807a8&el=desc) will **decrease** coverage by `0.18%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4711?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4711 +/- ##\n==========================================\n- Coverage 77.32% 77.14% -0.19% \n==========================================\n Files 128 128 \n Lines 21071 21071 \n==========================================\n- Hits 16294 16256 -38 \n- Misses 4777 4815 +38 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4711?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_bart.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iYXJ0LnB5) | `96.79% <ø> (ø)` | |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `72.11% <ø> (-14.11%)` | :arrow_down: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `80.41% <ø> (-1.38%)` | :arrow_down: |\n| [src/transformers/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190cmFuc2ZvX3hsLnB5) | `77.00% <ø> (ø)` | |\n| [src/transformers/modeling\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG0ucHk=) | `89.46% <ø> (ø)` | |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `89.94% <0.00%> (-0.24%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.85% <0.00%> (+0.41%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4711/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.66% <0.00%> (+1.80%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4711?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/4711?src=pr&el=footer). Last update [6449c49...e7d6cb9](https://codecov.io/gh/huggingface/transformers/pull/4711?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "I can work on this if there is no one on it. Quick question though: what about the models that have both `lm_labels` *and* `masked_lm_labels`? encode_decoder is one of them for instance, don't know if there are more.", "Yes, that's the case for [`BertForMaskedLM` for example](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_bert.py#L850). I don't really know the best way to handle this.\r\n\r\nAs with this update we're trying to have the *exact* same API for all models so that the training/inference code is model agnostic, I'd say that we should look for the most natural on a case-by-case basis.\r\n\r\nFor example with the `BertForMaskedLM` example, I believe the `labels` should be the `masked_lm_labels`, as BERT should be used for MLM rather than CLM. ", "> the\r\n\r\nAs far as I know, `BertForMaskedLM` does not really use `lm_labels` at the moment. I think it was added to support a causal `Bert` in an encoder-decoder setting so that the decoder can be trained with a causal mask with the language model objective. Since the encoder-decoder framework is not really released yet, I think we can also add a new `BertWithLMHead` class so that each class only has one `labels` argument. It would be a breaking change in terms of the class name for people that already implemented Bert2Bert models, but I think it's worth it for consistency. What do you think? \r\n\r\n@sgugger - In the encoder-decoder model I added both `lm_labels` and `masked_lm_labels` because `Bert` has both `lm_labels` and `masked_lm_labels`. Normally, encoder-decoder models are trained with a CLM objective so not sure if we even need `masked_lm_lables` for the encoder-decoder model wrapper. ", "@patrickvonplaten good for me" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4710?src=pr&el=h1) Report\n> Merging [#4710](https://codecov.io/gh/huggingface/transformers/pull/4710?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/6449c494d0f40f0b70442f3da9e61f042ff807a8&el=desc) will **increase** coverage by `0.05%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4710?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4710 +/- ##\n==========================================\n+ Coverage 77.32% 77.38% +0.05% \n==========================================\n Files 128 128 \n Lines 21071 21071 \n==========================================\n+ Hits 16294 16305 +11 \n+ Misses 4777 4766 -11 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4710?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4710/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.85% <0.00%> (+0.41%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4710/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.50% <0.00%> (+1.64%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4710?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/4710?src=pr&el=footer). Last update [6449c49...6081929](https://codecov.io/gh/huggingface/transformers/pull/4710?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[ { "id": 1834054694, "node_id": "MDU6TGFiZWwxODM0MDU0Njk0", "url": "https://api.github.com/repos/huggingface/transformers/labels/TensorFlow", "name": "TensorFlow", "color": "FF6F00", "default": false, "description": "Anything TensorFlow" }, { "id": 1834056761, "node_id": "MDU6TGFiZWwxODM0MDU2NzYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Core:%20Modeling", "name": "Core: Modeling", "color": "FF8446", "default": false, "description": "Internals of the library; Models." }, { "id": 1862634478, "node_id": "MDU6TGFiZWwxODYyNjM0NDc4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Should%20Fix", "name": "Should Fix", "color": "FF0000", "default": false, "description": "This has been identified as a bug and should be fixed." } ]

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

[ "Hello,\r\n\r\nThe way the TF models behave has been recently updated. Can you please retry with the `master` version?", "Hi Julien,\r\n\r\nThanks. It's the same issue in the `master` currently:\r\n\r\nhttps://github.com/huggingface/transformers/blob/9f5d5a531d769d07403f59661884e254f8420afe/src/transformers/modeling_tf_roberta.py#L320-L324\r\n\r\n`config` is still passed as the first parameter to the `__init__` of parent class `tf.keras.layers.Layer` while the latter expects `trainable` as the first parameter. I fixed it on my side by simply removing that `config` parameter from the `super().__init__(` call. But I wasn't sure if this affects other parts of the repo. Otherwise, I would have submitted a PR.\r\n", "Ok, thanks for the feedback. Indeed, the `config` parameter is important. I will take some time to review this. Sorry for the inconvenience.", "You are totally right, `config` here is useless and the same appears in other models . Do you mind to do a PR? And I will help you to fix all this :)", "Thanks for making the checks. I submitted a PR: https://github.com/huggingface/transformers/pull/4884" ]

[ { "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" ]

[]

[]

[ "pinging @n1t0, lead on [huggingface/tokenizers](https://github.com/huggingface/tokenizers) :)", "Thank you! Closing it here, moved link: [https://github.com/huggingface/tokenizers/issues/290](url)" ]

[]

[]

[ "Hi @h56cho, if you add new tokens then yes, you'll need to train the model. If you just want `eos` and `pad` token then its a good idea to use the model defaults. `eos` and `pad` tokens are already available in GPT2 model ", "Hi,\r\n\r\nThank you for your reply.\r\nSo what I am getting is that if I add any new \"special token\" onto the existing pre-trained tokenizer, I will need to re-train the pre-trained Transformer to make it learn that new special token.\r\n\r\n But what if I just add extra non-special tokens? for example, a word \"paradox\" is not included in the existing GPT-2 tokenizer, so say I add the word \"paradox\" to the existing set of GPT-2 vocabulary. If I didn't make any changes to the special tokens in the GPT-2 tokenizer, do I still need to train the pre-trained GPT-2 because I added a new word to a set of vocabulary?\r\n\r\nThanks, " ]

[]

[]

[]

[]

[]

[ "Hello! If you're using 8 labels, you'll need to tell the model that it needs to do 8-way classification. You can do so by specifying the `num_labels` argument during instantiation:\r\n\r\n```py\r\nmodel = TFXLMRobertaForSequenceClassification.from_pretrained(\r\n \"jplu/tf-xlm-roberta-base\", \r\n num_labels=8\r\n)\r\n```\r\n\r\nLet me know if this fixes your issue.", "Yes! That fixed the issue. I apologize for my oversight of that argument. Thank you very much for your time.", "Sure, my pleasure!" ]

[]

[]

[ "Hi! What's your model? Do you have the configuration file?\r\n\r\nWhat is the `run_pretraining.py`? This doesn't seem to be one of our scripts.", "Hi, extremely sorry for not providing adequate information.\r\n\r\nI used google-research/bert for creating the model. The bert repo contains the run_pretraining.py, before that I created the vocab.txt file from my own data and after that I used create_pretraining_data.py and yes I do have the config.json file as well.\r\n\r\nI am trying to convert tf_checkpoint file to pytorch_model.bin where I am encountering the issue of memory error on Kaggle and Colab and fused_norm_layer_cuda error on linux server without GPU. Just want to know if it is possible to convert from tf_checkpoint to pytorch without GPU or any way I can reduce the model size to load weights on kaggle or colab without running out of memory", "Do you mind sharing the configuration file?\r\n\r\nSo you pre-trained a model using google-research/bert, and now you're trying to convert it to one of our models, is that correct?", "{\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 384,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"max_position_embeddings\": 256,\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 6,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 2800000\r\n}\r\nThis is the config.json file.\r\n\r\nNot exactly, I have used bert repo to create a model which is in tensorflow and have files model.ckpt-100000.data-00000-of-00001 which seems to be in tensorflow but I need model_pytorch.bin so converting from tf_checkpoint file to pytorch using this amazing library. Trying to use the code written in convert_bert_original_tf_checkpoint_to_pytorch.py", "Right, you're using the correct script!\r\n\r\nPlease note, that having a 2_800_000 vocabulary size is absolutely huge!\r\n\r\nUnfortunately it's hard to get around memory errors during conversion. The script should work whether you have a GPU available or not. Can you show me the command you used to convert the model on your linux server?", "I have little data and many unique terms which are written in English but are addresses from all over India due to which I have to keep the vocabulary size this high and due to which I compromised with the dimensions of the model and even the embedding size as well.\r\n\r\nHere is the code snippet:- \r\n\r\nimport torch\r\n\r\nfrom pytorch_transformers.modeling_bert import BertConfig, BertForPreTraining, load_tf_weights_in_bert\r\n\r\n\r\ntf_checkpoint_path=\"./distlang/\"\r\nbert_config_file = \"./config.json\"\r\npytorch_dump_path=\"./distlangpytorch\"\r\n\r\nconfig = BertConfig.from_json_file(bert_config_file)\r\nprint(\"Building PyTorch model from configuration: {}\".format(str(config)))\r\nmodel = BertForPreTraining(config)\r\n\r\n# Load weights from tf checkpoint\r\nload_tf_weights_in_bert(model, config, tf_checkpoint_path)\r\n\r\n# Save pytorch-model\r\nprint(\"Save PyTorch model to {}\".format(pytorch_dump_path))\r\ntorch.save(model.state_dict(), pytorch_dump_path)\r\n\r\nIts, throwing error while instantiation of the model model = BertForPreTraining(config) \r\n\r\nI have apex installed but without cuda_ext as my server is not having GPU, so it won't install 😅", "Hmm I can't reproduce on my setup. Two things:\r\n\r\n- Can you update your `transformers` library? It seems that you're using `pytorch-transformers` which is starting to be quite old now. We've patched quite a few bugs since then! > `pip install -U transformers` and update your imports to `from transformers import`\r\n- Do you mind pasting the stack trace?", "If I use thrasformers library will it work fine?\r\nAnd yes here the stack trace:-\r\nBuilding PyTorch 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\": 384,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 256,\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 6,\r\n \"num_labels\": 2,\r\n \"output_attentions\": false,\r\n \"output_hidden_states\": false,\r\n \"pruned_heads\": {},\r\n \"torchscript\": false,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 2800000\r\n}\r\n\r\nTraceback (most recent call last):\r\n File \"tftopytorch.py\", line 18, in <module>\r\n model = BertForPreTraining(config)\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/pytorch_transformers/modeling_bert.py\", line 761, in __init__\r\n self.bert = BertModel(config)\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/pytorch_transformers/modeling_bert.py\", line 651, in __init__\r\n self.embeddings = BertEmbeddings(config)\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/pytorch_transformers/modeling_bert.py\", line 240, in __init__\r\n self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/apex/normalization/fused_layer_norm.py\", line 133, in __init__\r\n fused_layer_norm_cuda = importlib.import_module(\"fused_layer_norm_cuda\")\r\n File \"/usr/lib/python3.6/importlib/__init__.py\", line 126, in import_module\r\n return _bootstrap._gcd_import(name[level:], package, level)\r\n File \"<frozen importlib._bootstrap>\", line 994, in _gcd_import\r\n File \"<frozen importlib._bootstrap>\", line 971, in _find_and_load\r\n File \"<frozen importlib._bootstrap>\", line 953, in _find_and_load_unlocked\r\nModuleNotFoundError: No module named 'fused_layer_norm_cuda'", "Yes, upgrading the library should solve your problems. Since [this commit](https://github.com/huggingface/transformers/commit/98dd19b96b351f481e1268ab6c7b035bb21d106e) we're not using apex for the LayerNorm anymore, so having PyTorch installed on a recent version alongside `transformers` on a recent version should solve your issue!", "Turns out the problem is solved. Thanks a lot man for saving the day. Amazing library as well 🤗 really appreciate your quick response and understanding my problem and solving it.", "My pleasure :hugs: ", "Hey the above problem is solved but I am running into this problem while writing the file to the directory, to my understanding I have to provide the path where I have to save the pyorch model and I am providing a path. Am I missing out something here? \r\n\r\nHere's the snippet of the error the code is same as mentioned above just used transformers instead of pytorch_transformer\r\n\r\nSave PyTorch model to ./distlangpytorch/\r\nTraceback (most recent call last):\r\n File \"tftopytorch.py\", line 25, in <module>\r\n torch.save(model.state_dict(), pytorch_dump_path)\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/torch/serialization.py\", line 369, in save\r\n with _open_file_like(f, 'wb') as opened_file:\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/torch/serialization.py\", line 234, in _open_file_like\r\n return _open_file(name_or_buffer, mode)\r\n File \"/home/rohit.anand/envs/data_augment/lib/python3.6/site-packages/torch/serialization.py\", line 215, in __init__\r\n super(_open_file, self).__init__(open(name, mode))\r\nIsADirectoryError: [Errno 21] Is a directory: './distlangpytorch/'", "We usually recommend saving using our method `from_pretrained`, as it saves the configuration as well as the model state dict.\r\n\r\nCan you try using the following?\r\n\r\n```py\r\nmodel = load_tf_weights_in_bert(model, config, tf_checkpoint_path)\r\nmodel.save_pretrained(pytorch_dump_path)\r\n```", "It worked fine, the model is saved now and the size of the model is 1/3rd the original size which was of the tf_checkpoint. Is the model quantized as well, is it something implemented in the library? Also thanks a ton have been struggling since evening and once posted the issue here got the resolution immediately. 🤗", "That's great! The model is not quantized, having a 2/3 reduction is surprising. It's possible that your checkpoint originally had optimizer states which do take quite a big amount of memory. We don't save that in our checkpoints.\r\n\r\nCool, glad I could help!", "Hey, its not related to the issue can you please help me with post training quantization. I only am able to see tutorials of dynamic quantization which indeed increases the size of my model. I am not able to find any decent tutorials on post training(static) quantization. Tried torch.quantization.quantize but the method asks for two positional arguments fn_run and fn_args and I am not sure how to define them or create a function to pass to these arguments", "So, it turns out I might have figured it out. The model size is 4.3 GB my vocab size is 28_000_00 and hidden layer size is 384. So the vocab part of model turn out to be 4X2800000X384 = 4.3GB. What I don't understand is why the model size only have vocabulary part and no part of weights strange. And is there any free resource where I can train this model for classification you might be aware of 🙈, On colab and kaggle the kernel restarts because they only have 12,16 GB RAM, the model needs a bit more 🙈" ]

[ { "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", "@makailove123 Have you figured out why?" ]

[ { "id": 1834081910, "node_id": "MDU6TGFiZWwxODM0MDgxOTEw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Usage", "name": "Usage", "color": "e28436", "default": false, "description": "General questions about the library" } ]

[]

[ "Have a look at the code for [`.from_pretrained()`](https://github.com/huggingface/transformers/blob/a9aa7456ac824c9027385b149f405e4f5649273f/src/transformers/modeling_utils.py#L490). What actually happens is something like this:\r\n\r\n- find the correct base model class to initialise\r\n- initialise that class with pseudo-random initialisation (by using the `_init_weights` function that you mention)\r\n- find the file with the pretrained weights\r\n- overwrite the weights of the model that we just created with the pretrained weightswhere applicable\r\n\r\nThis ensure that layers were not pretrained (e.g. in some cases the final classification layer) _do_ get initialised in `_init_weights` but don't get overridden.", "Great. Thanks. I also read through the code and that really clears my confusion. ", "Good. If the answer was sufficient on Stack Overflow as well, please close that too. ", "\r\n\r\n\r\n\r\n> Have a look at the code for [`.from_pretrained()`](https://github.com/huggingface/transformers/blob/a9aa7456ac824c9027385b149f405e4f5649273f/src/transformers/modeling_utils.py#L490). What actually happens is something like this:\r\n> \r\n> * find the correct base model class to initialise\r\n> * initialise that class with pseudo-random initialisation (by using the `_init_weights` function that you mention)\r\n> * find the file with the pretrained weights\r\n> * overwrite the weights of the model that we just created with the pretrained weightswhere applicable\r\n> \r\n> This ensure that layers were not pretrained (e.g. in some cases the final classification layer) _do_ get initialised in `_init_weights` but don't get overridden.\r\n\r\nwhen we construct BertForSequenceClassification from pre-trained model, didn't we overwrite the loaded weights with random initialisation?", "@sunersheng No, the random initialization happens [first](https://github.com/huggingface/transformers/blob/a9aa7456ac824c9027385b149f405e4f5649273f/src/transformers/modeling_utils.py#L659) and then the existing weights are loaded [into it](https://github.com/huggingface/transformers/blob/a9aa7456ac824c9027385b149f405e4f5649273f/src/transformers/modeling_utils.py#L732)." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4700?src=pr&el=h1) Report\n> Merging [#4700](https://codecov.io/gh/huggingface/transformers/pull/4700?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0866669e751bef636fa693b704a28c1fea9a17f3&el=desc) will **decrease** coverage by `1.41%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4700?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4700 +/- ##\n==========================================\n- Coverage 77.14% 75.72% -1.42% \n==========================================\n Files 128 128 \n Lines 21070 21070 \n==========================================\n- Hits 16255 15956 -299 \n- Misses 4815 5114 +299 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4700?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19sb25nZm9ybWVyLnB5) | `18.70% <0.00%> (-74.83%)` | :arrow_down: |\n| [src/transformers/configuration\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2xvbmdmb3JtZXIucHk=) | `76.92% <0.00%> (-23.08%)` | :arrow_down: |\n| [src/transformers/modeling\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190NS5weQ==) | `77.18% <0.00%> (-6.35%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.44% <0.00%> (-0.42%)` | :arrow_down: |\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `86.77% <0.00%> (-0.19%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.17% <0.00%> (+0.23%)` | :arrow_up: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `81.78% <0.00%> (+1.37%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4700/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `86.21% <0.00%> (+14.10%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4700?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/4700?src=pr&el=footer). Last update [0866669...34c9a46](https://codecov.io/gh/huggingface/transformers/pull/4700?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Nice webinar, and cool notebook :)\r\n\r\n@patrickvonplaten, do you want to take a look?", "Thanks @LysandreJik ! :smile:", "Awesome notebook @enzoampil! \r\n\r\nLGTM for merge!\r\n\r\nWhich dataset do you use exactly to fine-tune T5 here? ", "Thanks @patrickvonplaten ! 😄 \r\n\r\nFor the dataset, I got it from an ongoing Kaggle competition called [Tweet Sentiment Extraction](https://www.kaggle.com/c/tweet-sentiment-extraction/data).\r\n\r\n**The objective is to extract the span from a tweet that indicates its sentiment**\r\n\r\nExample input:\r\n```\r\nsentiment: negative\r\ntweet: How did we just get paid and still be broke as hell?! No shopping spree for me today\r\n```\r\n\r\nExample output:\r\n```\r\nbroke as hell?!\r\n```\r\n", "I was thinking about contributing this to the `nlp` library, but I'm not sure if Kaggle has policies regarding uploading their datasets to other public sources ...", "I see! Yeah no worries - I don't think we currently handle dataset processing from on-going kaggle competition links. " ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1834054694, "node_id": "MDU6TGFiZWwxODM0MDU0Njk0", "url": "https://api.github.com/repos/huggingface/transformers/labels/TensorFlow", "name": "TensorFlow", "color": "FF6F00", "default": false, "description": "Anything TensorFlow" }, { "id": 1834060867, "node_id": "MDU6TGFiZWwxODM0MDYwODY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Named%20Entity%20Recognition", "name": "Ex: Named Entity Recognition", "color": "06FFD8", "default": false, "description": "" } ]

[]

[ "Also, the example was missing this parameter in the python3 run_tf_ner.py command to work. Would be good to update the doc:\r\n\r\n`--logging_dir ./my-model/`", "@chuckabees I got the same issue, did you solve this problem?", "@YuqiShen Unfortunately no. I even wrote to the code's author but no response :(", "I got the same problem. This issue helped me https://github.com/huggingface/transformers/issues/4631#issuecomment-636063607\r\n\r\nI added `--mode token-classification` param in the shell command and now it works fine :)", "@donuzium thanks but when I try, I get this error now:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"run_tf_ner.py\", line 281, in <module>\r\n main()\r\n File \"run_tf_ner.py\", line 135, in main\r\n cache_dir=model_args.cache_dir,\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/configuration_auto.py\", line 203, in from_pretrained\r\n config_dict, _ = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/configuration_utils.py\", line 252, in get_config_dict\r\n raise EnvironmentError(msg)\r\nOSError: Can't load config for 'token-classification'. Make sure that:\r\n\r\n- 'token-classification' is a correct model identifier listed on 'https://huggingface.co/models'\r\n\r\n- or 'token-classification' is the correct path to a directory containing a config.json file\r\n```\r\n\r\nGoing to https://huggingface.co/models, I don't see 'token-classification' there.", "@chuckabees it's under `tags`", "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", "Closing as solved." ]

[]

[]

[ "Hi there!\r\n\r\n> I currently pass this data into the model like this:\r\n> \r\n> ```\r\n> output = model(input_ids=data, labels=target, mems=mems)\r\n> ```\r\n> \r\n> Is this correct?\r\n\r\nNo, this is not correct, because the labels are shifted inside the model (as the documentation suggests). This happens [here](https://github.com/huggingface/transformers/blob/ec8717d5d8f6edc2c595ff6954ffaa2078dcc97d/src/transformers/modeling_transfo_xl_utilities.py#L104) so in your example, the target vector will become\r\n```\r\ntensor([[3,4,5,6,7,8,9]])\r\n```\r\nto be matched with the predictions corresponding to\r\n```\r\ntensor([[1,2,3,4,5,6,7]])\r\n```\r\nso you'll try to predict the token that is two steps ahead of the current one.\r\n\r\nI am guessing that `lm_labels` is a typo for `labels`, and that you should either:\r\n- pass `labels = input_ids` as suggested by the doc string (in this case you will not compute any loss for the last prediction, but that's probably okay)\r\n- add something at the beginning of your target tensor (anything can work since it will be removed by the shift) : `target = tensor([[42,2,3,4,5,6,7,8,9]])`\r\n\r\nI'm still learning the library, so tagging @TevenLeScao (since he worked on the issue/PR you mentioned) to make sure I'm not saying something wrong (also, do we want to update `LMOrderedIterator` from tokenization_transfo_xl.py to return target tensors that can be used as labels?)", "Ah yes that does sound like a typo from another model's convention! You do have to pass `data` twice, once to `input_ids` and once to `labels` (in your case, `[1, ... , 8]` for both). The model will then attempt to predict `[2, ... , 8]` from `[1, ... , 7]`). I am not sure adding something at the beginning of the target tensor would work as that would probably cause size mismatches later down the line.\r\n\r\nPassing twice is the default way to do this in `transformers`; before the aforementioned PR, `TransfoXL` did not shift labels internally and you had to shift the labels yourself. The PR changed it to be consistent with the library and the documentation, where you have to pass the same data twice. I believe #4711 fixed the typo, you should be all set ! I'll also answer on StackOverflow in case someone finds that question there.", "Thanks @sgugger and @TevenLeScao for your help!\r\n\r\n@TevenLeScao \r\n> before the aforementioned PR, TransfoXL did not shift labels internally and you had to shift the labels yourself\r\n\r\nSo this means that in the versions before the fix my method with shifting the labels beforehand was correct? Because I'm currently using `transformers 2.6`.", "Yes, it was changed in 2.9.0. You should probably consider updating ;)", "> The model will then attempt to predict `[2, ... , 8]` from `[1, ... , 7]`).\r\n\r\nNote that if you are using the state, the memory returned is computed on the whole `[1, ... , 8]`, so you should use `[9,10,... , 16]` as your next batch.", "Thanks guys!\r\n\r\nSorry for asking this here, but maybe one of you can help me with my workaround in issue #3554 ? That would help me a lot!", "Hello again, sorry for bothering again, but I have update my code from version 2.6 to 2.11 as @TevenLeScao has suggested. Now I experience a drop in my model's performance, but I don't know why. I use the same code as before except passing `data` in twice as suggested.\r\n\r\nI know that this can have several other reasons but I just want to know if there where other breaking changes to `TransformerXLLMHeadModel` or to the generation process?\r\n\r\nI skipped through the changelog in the releases but could not find anything.\r\n\r\nThanks in advance!", "Sorry, by a drop in model performance you mean the loss is worse right? I've noticed discrepancies between CMU code performance (better) and ours in the past, so maybe a bug was introduced between 2.6 and 2.11 (never used 2.6 myself). I'm comparing the two.", "Well mainly I saw differences during text generation with `model.generate()` The sequences tend to be shorter and end more often with an <eos> in 2.11, where before in 2.6 they were just cut of at some point.\r\n\r\nBut I can't guarantee that there are no mistakes from my side.", "Could it be that this is also related to #4826 ?", "FYI: The issue regarding worse model performance on the newer version of `transformers` is solved. There were some errors on my side.\r\n\r\nNevertheless, I hope that the fix in the PR linked above will improve the generated texts, since I also experience low quality output despite proper finetuning." ]

[ { "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" ]

[]

[]

[ "Hello! The `d_head` is actually computed in the configuration: `self.d_head = d_model // n_head`. \r\n\r\nIt would probably be better to handle `d_head` directly, but currently the `d_model` and `d_head` are linked to each other. \r\n\r\nThis would crash in your case as `d_model // n_head != d_head`", "@LysandreJik Thanks a lot. Yes, I am setting d_head directly before loading my model. But it would be nice to see the model load its configuration from the given config file. Just my opinion though :)", "You're right! Raising an error if the `d_head` is wrong in #4747 " ]

[]

[]

[ "This would be good indeed (cc @julien-c). In the meantime I think you can specify \r\n```\r\n--tokenizer_name=$TOKENIZER_NAME_OR_PATH\r\n```\r\n so that it always loads the initial tokenizer (which does not change during training).", "> This would be good indeed (cc @julien-c). In the meantime I think you can specify\r\n> \r\n> ```\r\n> --tokenizer_name=$TOKENIZER_NAME_OR_PATH\r\n> ```\r\n> \r\n> so that it always loads the initial tokenizer (which does not change during training).\r\n\r\nThank you!", "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'm trying to figure out how to use a custom made or create a tokenizer using this script and having significant difficulty. There does not seem to be any documentation on how to do this. I attempted to follow [this ](https://huggingface.co/blog/how-to-train) example but it makes no mention of how the tokenizer gets loaded / used. I get errors like:\r\n\r\n```\r\nOSError: Can't load config for 'EsperBERTo-small'. Make sure that:\r\n\r\n- 'EsperBERTo-small' is a correct model identifier listed on 'https://huggingface.co/models'\r\n\r\n- or 'EsperBERTo-small' is the correct path to a directory containing a config.json file\r\n```", "Hi @EdwardRaff ,\r\nI'm facing the same issue while trying to train BERT model from scratch on my own dateset. Did you figure out how to solve it?\r\n```\r\n\r\nOSError: Can't load config for './models/BuckBERTer-small/'. Make sure that:\r\n\r\n- './models/BuckBERTer-small/' is a correct model identifier listed on 'https://huggingface.co/models'\r\n\r\n- or './models/BuckBERTer-small/' is the correct path to a directory containing a config.json file\r\n```", "@EdwardRaff @FerchichiNourchene `run_language_modeling.py` doesn't work if a tokenizer is specified, but it does not contain the model configuration files. [This](https://stackoverflow.com/a/64795300/3950710) workaround worked for me.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@vincentwen1995, did you manage to get the tokenizers somehow? One year later and it seems like the `Trainer` is still not saving `tokenizer_config.json` in the checkpoint folders. \r\nWhere is it even saved?" ]

[ { "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", "Hey all, any chance anyone else is working around this? I think a neutral label or a standard sentiment score would be great for such an extensive model. Neutral statements are not caught with this adjustment:\r\n\r\nclassifier('I do not know the answer.')\r\nOut[16]: [{'label': 'NEGATIVE', 'score': 0.9995205402374268}]\r\n\r\nclassifier('This is meant to be a very neutral statement.')\r\nOut[17]: [{'label': 'NEGATIVE', 'score': 0.987031102180481}]\r\n\r\nclassifier('The last president of US is Donald Trump.')\r\nOut[18]: [{'label': 'POSITIVE', 'score': 0.9963828325271606}]\r\n\r\nclassifier('There is going to be an election in two months.')\r\nOut[19]: [{'label': 'NEGATIVE', 'score': 0.9604763984680176}]\r\n\r\nJust raising this thread again to see if there is a common interest...\r\nCheers!" ]

[ { "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 } ]

[]

[ "\r\nI get this message every 2000 steps. \r\n\"Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 262144.0\"\r\n\r\nThe train run on rtx 2070 and seems to not have other problem but that message keeps going on every 2000 steps.\r\n\r\npytorch 1.5.0\r\npython 3.7\r\ncuda 10.1\r\n", "Do you have the same issue with opt level O1? Using O2 is discouraged. The issue may also be related to PyTorch 1.5, so if switching to O1 does not help, try a previous PyTorch version. Note that it is highly likely that this is an AMP problem, not a transformers issue. Have a look here https://github.com/NVIDIA/apex/issues/318", "@BramVanroy Yes i guess you 're right. It seems problem of the AMP. I asked and in APEX the same question and their answer was : \r\n\r\n\"The loss scaler tries to increase the loss scaling factor after a threshold of successful steps was reached. In your case it seems that the scaling factor is being downgraded to the same value, so it should be fine.\"\r\n\r\nSo, according to that answer, is not a problem. But still i am not sure. \r\n\r\n", "I guess that skipping one step every 2000 steps is not a problem. You can monitor the loss, and as long as it seems to decrease normally, then you should be fine. ", "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 } ]

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

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4691?src=pr&el=h1) Report\n> Merging [#4691](https://codecov.io/gh/huggingface/transformers/pull/4691?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0866669e751bef636fa693b704a28c1fea9a17f3&el=desc) will **increase** coverage by `0.19%`.\n> The diff coverage is `16.66%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4691?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4691 +/- ##\n==========================================\n+ Coverage 77.14% 77.34% +0.19% \n==========================================\n Files 128 128 \n Lines 21070 21087 +17 \n==========================================\n+ Hits 16255 16309 +54 \n+ Misses 4815 4778 -37 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4691?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19yb2JlcnRhLnB5) | `89.42% <16.66%> (-6.29%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.44% <0.00%> (-0.42%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.83% <0.00%> (+0.16%)` | :arrow_up: |\n| [src/transformers/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190cmFuc2ZvX3hsLnB5) | `77.22% <0.00%> (+0.21%)` | :arrow_up: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.29% <0.00%> (+0.35%)` | :arrow_up: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `81.78% <0.00%> (+1.37%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4691/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `86.21% <0.00%> (+14.10%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4691?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/4691?src=pr&el=footer). Last update [0866669...b0bbd24](https://codecov.io/gh/huggingface/transformers/pull/4691?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" ]

[]

[]

[ "@tchainzzz Can you tell how you resolved the particular issue?\r\n" ]

[ { "id": 1108649070, "node_id": "MDU6TGFiZWwxMTA4NjQ5MDcw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Need%20more%20information", "name": "Need more information", "color": "d876e3", "default": false, "description": "Further information is requested" }, { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1834052574, "node_id": "MDU6TGFiZWwxODM0MDUyNTc0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Sequence%20Classification", "name": "Ex: Sequence Classification", "color": "46FFCF", "default": false, "description": "" } ]

[]

[ "Can you post a minimal dataset that we can test this with? So for instance two sentences that give the same result for you.", "```\r\n[\r\n [\"Nah I don't think he goes to usf, he lives around here though\", 2],\r\n [\"URGENT! You have won a 1 week FREE membership in our $100,000 Prize Jackpot!\", 1]\r\n]\r\n```\r\n\r\n**The model outputs same results only after fine-tuning phase.**\r\n\r\nThis is my Dataset subclass:\r\n\r\n```\r\nclass SequenceClassificationDataset(Dataset):\r\n def __init__(self, df, tokenizer, max_length):\r\n encodings = tokenizer.batch_encode_plus(df.values[:, 0].tolist(), return_tensors=\"pt\", max_length=max_length, pad_to_max_length=True)\r\n self.input_ids = encodings.input_ids\r\n self.attention_masks = encodings.attention_mask\r\n self.y = torch.LongTensor(df.values[:,1].tolist())\r\n\r\n def __getitem__(self, index):\r\n return self.input_ids[index], self.attention_masks[index], self.y[index]\r\n \r\n def __len__(self):\r\n return self.input_ids.shape[0]\r\n```\r\n\r\nAm I passing the \"y\" value incorrectly? It's not one-hot encoded matrix, but it's vector with size (batch_size, ) where each element represents the category for that text.", "Yes, your labels are correct.\r\n\r\nI can't reproduce your problem, though. This seems to work correctly.\r\n\r\n\r\n```python\r\nimport torch\r\nfrom torch.utils.data import Dataset, DataLoader\r\nfrom transformers import AutoTokenizer, BartForSequenceClassification\r\n\r\nimport pandas as pd\r\n\r\n\r\nclass SequenceClassificationDataset(Dataset):\r\n def __init__(self, df, tokenizer, max_length):\r\n encodings = tokenizer.batch_encode_plus(df.values[:, 0].tolist(), return_tensors=\"pt\",\r\n max_length=max_length, pad_to_max_length=True)\r\n self.input_ids = encodings.input_ids\r\n self.attention_masks = encodings.attention_mask\r\n self.y = torch.LongTensor(df.values[:, 1].tolist())\r\n\r\n def __getitem__(self, index):\r\n return self.input_ids[index], self.attention_masks[index], self.y[index]\r\n\r\n def __len__(self):\r\n return self.input_ids.shape[0]\r\n\r\n\r\ndef main():\r\n df = pd.DataFrame([\r\n [\"Nah I don't think he goes to usf, he lives around here though\", 2],\r\n [\"URGENT! You have won a 1 week FREE membership in our $100,000 Prize Jackpot!\", 1]\r\n ])\r\n\r\n tokenizer = AutoTokenizer.from_pretrained(\"bart-large\")\r\n model = BartForSequenceClassification.from_pretrained(\"bart-large\")\r\n ds = SequenceClassificationDataset(df, tokenizer, 32)\r\n dl = DataLoader(ds)\r\n\r\n with torch.no_grad():\r\n for batch_index, (batch_input_ids, batch_attention_masks, batch_y) in enumerate(dl, start=1):\r\n batch_input_ids_cuda = batch_input_ids\r\n batch_attention_masks_cuda = batch_attention_masks\r\n batch_y_cuda = batch_y\r\n loss, logits, _ = model(batch_input_ids_cuda, attention_mask=batch_attention_masks_cuda,\r\n labels=batch_y_cuda)\r\n print(\"Input ids:\", batch_input_ids_cuda)\r\n print(\"Attention masks:\", batch_attention_masks_cuda)\r\n print(\"Loss:\", loss)\r\n print(\"Logits:\", logits)\r\n print(\"y:\", batch_y)\r\n\r\nif __name__ == '__main__':\r\n main()\r\n```\r\n\r\nAre you running this in a notebook? If so, try restarting the notebook. Having unexpected results like this can be a sign of cached cells.", "Everything works fine if I don't fine-tune the model, the problem occurs after fine-tuning. Can you please check my training subroutine?", "The training code seems okay. Even if there was a bug in the loop, you would not expect that any input gives the same output. I can't debug this unfortunately since I can't reproduce your issue. Can you share the dataset that you use for fine-tuning?", "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", "> # ❓ Questions & Help\r\n> ## Details\r\n> I use BartForSequenceClassification pre-trained model from the HuggingFace Transformers library. During training phase logits from classification head gets different values, but during validation phase all logits values are equal even for different input texts.\r\n> \r\n> I use BartTokenizer.batch_encode_plus to encode the text before feeding into the model.\r\n> \r\n> I fine-tuned the model for 1 epoch using the following code:\r\n> \r\n> ```\r\n> config = BartConfig.from_pretrained(model_name)\r\n> config.num_labels = 3\r\n> config.output_hidden_states = False\r\n> config.output_attentions = False\r\n> \r\n> transformer_model = BartForSequenceClassification.from_pretrained(model_name, config=config)\r\n> transformer_model.cuda();\r\n> optimizer = AdamW(transformer_model.parameters())\r\n> \r\n> NUM_TRAIN_EPOCHS = 1\r\n> print(\"Training:\")\r\n> for i in range(1, 1+NUM_TRAIN_EPOCHS):\r\n> transformer_model.train()\r\n> for batch_index, (batch_input_ids, batch_attention_masks, batch_y) in enumerate(train_dataloader, start=1):\r\n> batch_input_ids_cuda = batch_input_ids.to(device)\r\n> batch_attention_masks_cuda = batch_attention_masks.to(device)\r\n> batch_y_cuda = batch_y.to(device)\r\n> loss, logits, _ = transformer_model(batch_input_ids_cuda, attention_mask=batch_attention_masks_cuda, labels=batch_y_cuda)\r\n> transformer_model.zero_grad()\r\n> loss.backward()\r\n> optimizer.step()\r\n> ```\r\n> \r\n> And for validation I use the following code:\r\n> \r\n> ```\r\n> transformer_model.eval()\r\n> with torch.no_grad():\r\n> for batch_index, (batch_input_ids, batch_attention_masks, batch_y) in enumerate(test_dataloader, start=1):\r\n> batch_input_ids_cuda = batch_input_ids.to(device)\r\n> batch_attention_masks_cuda = batch_attention_masks.to(device)\r\n> batch_y_cuda = batch_y.to(device)\r\n> loss, logits, _ = transformer_model(batch_input_ids_cuda, attention_mask=batch_attention_masks_cuda, labels=batch_y_cuda)\r\n> print(\"Input ids:\", batch_input_ids_cuda)\r\n> print(\"Attention masks:\", batch_attention_masks_cuda)\r\n> print(\"Logits:\", logits)\r\n> ```\r\n> \r\n> Output of validation phase is:\r\n> \r\n> ```\r\n> Input ids: tensor([[ 0, 3655, 9, ..., 1, 1, 1],\r\n> [ 0, 31524, 347, ..., 1, 1, 1],\r\n> [ 0, 12806, 24220, ..., 1, 1, 1],\r\n> ...,\r\n> [ 0, 8518, 7432, ..., 1, 1, 1],\r\n> [ 0, 15006, 23613, ..., 1, 1, 1],\r\n> [ 0, 14729, 13178, ..., 1, 1, 1]], device='cuda:0')\r\n> Attention masks: tensor([[1, 1, 1, ..., 0, 0, 0],\r\n> [1, 1, 1, ..., 0, 0, 0],\r\n> [1, 1, 1, ..., 0, 0, 0],\r\n> ...,\r\n> [1, 1, 1, ..., 0, 0, 0],\r\n> [1, 1, 1, ..., 0, 0, 0],\r\n> [1, 1, 1, ..., 0, 0, 0]], device='cuda:0')\r\n> Logits: tensor([[-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862],\r\n> [-1.3014, -0.7394, 0.7862]], device='cuda:0')\r\n> ```\r\n> \r\n> Why all the logits have exact same value for every input?\r\n> \r\n> **A link to original question on Stack Overflow**: https://stackoverflow.com/questions/62097267/same-logits-value-for-different-input-in-huggingfaces-transformer\r\n\r\nsame problem occurs to me", "> Yes, your labels are correct.\r\n> \r\n> I can't reproduce your problem, though. This seems to work correctly.\r\n> \r\n> ```python\r\n> import torch\r\n> from torch.utils.data import Dataset, DataLoader\r\n> from transformers import AutoTokenizer, BartForSequenceClassification\r\n> \r\n> import pandas as pd\r\n> \r\n> \r\n> class SequenceClassificationDataset(Dataset):\r\n> def __init__(self, df, tokenizer, max_length):\r\n> encodings = tokenizer.batch_encode_plus(df.values[:, 0].tolist(), return_tensors=\"pt\",\r\n> max_length=max_length, pad_to_max_length=True)\r\n> self.input_ids = encodings.input_ids\r\n> self.attention_masks = encodings.attention_mask\r\n> self.y = torch.LongTensor(df.values[:, 1].tolist())\r\n> \r\n> def __getitem__(self, index):\r\n> return self.input_ids[index], self.attention_masks[index], self.y[index]\r\n> \r\n> def __len__(self):\r\n> return self.input_ids.shape[0]\r\n> \r\n> \r\n> def main():\r\n> df = pd.DataFrame([\r\n> [\"Nah I don't think he goes to usf, he lives around here though\", 2],\r\n> [\"URGENT! You have won a 1 week FREE membership in our $100,000 Prize Jackpot!\", 1]\r\n> ])\r\n> \r\n> tokenizer = AutoTokenizer.from_pretrained(\"bart-large\")\r\n> model = BartForSequenceClassification.from_pretrained(\"bart-large\")\r\n> ds = SequenceClassificationDataset(df, tokenizer, 32)\r\n> dl = DataLoader(ds)\r\n> \r\n> with torch.no_grad():\r\n> for batch_index, (batch_input_ids, batch_attention_masks, batch_y) in enumerate(dl, start=1):\r\n> batch_input_ids_cuda = batch_input_ids\r\n> batch_attention_masks_cuda = batch_attention_masks\r\n> batch_y_cuda = batch_y\r\n> loss, logits, _ = model(batch_input_ids_cuda, attention_mask=batch_attention_masks_cuda,\r\n> labels=batch_y_cuda)\r\n> print(\"Input ids:\", batch_input_ids_cuda)\r\n> print(\"Attention masks:\", batch_attention_masks_cuda)\r\n> print(\"Loss:\", loss)\r\n> print(\"Logits:\", logits)\r\n> print(\"y:\", batch_y)\r\n> \r\n> if __name__ == '__main__':\r\n> main()\r\n> ```\r\n> \r\n> Are you running this in a notebook? If so, try restarting the notebook. Having unexpected results like this can be a sign of cached cells.\r\n\r\nyou should set the mode of the model to 'eval'.\r\n\r\nhttps://huggingface.co/transformers/main_classes/model.html#transformers.PreTrainedModel.from_pretrained\r\n\r\n`The model is set in evaluation mode by default using model.eval() (Dropout modules are deactivated). To train the model, you should first set it back in training mode with model.train().`", "@xylcbd No. The _default_ is already eval, as the documentation writes. So you do not have to explicitly set it to eval() again. But if you want to train, you need to set it to train()." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "Interested in model weights too but currently not available. Author does mention releasing tf code here:\r\n\r\nhttps://news.ycombinator.com/item?id=22290227\r\n\r\nRequires tf 1.15+ and deepmind/sonnet ver 1.36. Link to python script here:\r\n\r\nhttps://github.com/deepmind/sonnet/blob/cd5b5fa48e15e4d020f744968f5209949ebe750f/sonnet/python/modules/nets/transformer.py#L915\r\n\r\nHave tried running as-is but doesn't appear to have options for training on custom data as per the paper and available data sets.", "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": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4687?src=pr&el=h1) Report\n> Merging [#4687](https://codecov.io/gh/huggingface/transformers/pull/4687?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0866669e751bef636fa693b704a28c1fea9a17f3&el=desc) will **decrease** coverage by `1.44%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4687?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4687 +/- ##\n==========================================\n- Coverage 77.14% 75.70% -1.45% \n==========================================\n Files 128 128 \n Lines 21070 21070 \n==========================================\n- Hits 16255 15950 -305 \n- Misses 4815 5120 +305 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4687?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19sb25nZm9ybWVyLnB5) | `18.70% <0.00%> (-74.83%)` | :arrow_down: |\n| [src/transformers/configuration\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2xvbmdmb3JtZXIucHk=) | `76.92% <0.00%> (-23.08%)` | :arrow_down: |\n| [src/transformers/modeling\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190NS5weQ==) | `77.18% <0.00%> (-6.35%)` | :arrow_down: |\n| [src/transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19jdHJsLnB5) | `95.94% <0.00%> (-2.71%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.44% <0.00%> (-0.42%)` | :arrow_down: |\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `86.77% <0.00%> (-0.19%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.50% <0.00%> (-0.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.29% <0.00%> (+0.35%)` | :arrow_up: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `81.78% <0.00%> (+1.37%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4687/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `86.21% <0.00%> (+14.10%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4687?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/4687?src=pr&el=footer). Last update [0866669...d473be4](https://codecov.io/gh/huggingface/transformers/pull/4687?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4686?src=pr&el=h1) Report\n> Merging [#4686](https://codecov.io/gh/huggingface/transformers/pull/4686?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/9c17256447b91cf8483c856cb15e95ed30ace538&el=desc) will **increase** coverage by `0.24%`.\n> The diff coverage is `75.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4686?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4686 +/- ##\n==========================================\n+ Coverage 77.23% 77.47% +0.24% \n==========================================\n Files 128 128 \n Lines 21050 21051 +1 \n==========================================\n+ Hits 16257 16309 +52 \n+ Misses 4793 4742 -51 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4686?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/pipelines.py](https://codecov.io/gh/huggingface/transformers/pull/4686/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9waXBlbGluZXMucHk=) | `76.15% <75.00%> (+0.04%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4686/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.66% <0.00%> (ø)` | |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4686/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.17% <0.00%> (+0.23%)` | :arrow_up: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4686/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.85% <0.00%> (+0.41%)` | :arrow_up: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4686/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `81.78% <0.00%> (+1.37%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4686/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `86.21% <0.00%> (+14.10%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4686?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/4686?src=pr&el=footer). Last update [9c17256...95cf209](https://codecov.io/gh/huggingface/transformers/pull/4686?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

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

[ "This is expected behaviour, but I understand your confusion.\r\n\r\n```python\r\nmodel_from_pretrained = AutoModel.from_pretrained(pretrained)\r\n```\r\n\r\nThis actually loads the pretrained weights. It looks up the mapping and locations of the config file and the weights, and loads both.\r\n\r\n```python\r\nmodel_from_config = AutoModel.from_config(AutoConfig.from_pretrained(pretrained))\r\n```\r\n\r\nHere, the pretrained weights are never requested. You request the pretrained _config_ (basically the pretraining settings for the architecture), and (randomly) initialise an AutoModel given that config - but the weights are never requested and, thus, never loaded. \r\n\r\nThis means that both initialised models will have the same architecture, the same config, but different weights. The former has pretrained weights, the latter is randomly initialised.\r\n\r\nI think that what you expected or wanted is actually this, which will load pretrained weights and taking into account a pretrained config (however, this is practically the same as the first option):\r\n\r\n```python\r\nmodel_from_config = AutoModel.from_pretrained(pretrained, config=AutoConfig.from_pretrained(pretrained))\r\n```\r\n\r\nHope that helps.", "Thank you very much for the fast response. \r\nI think that the documentation is not clear enough about this difference, especially when there are pre-trained models such as `bert-base-uncased` and `bert-base-cased`, and there is the `AutoModelForPreTraining` class (that now I'm not sure what is for).\r\n\r\n\r\n", "If I understand correctly, your confusion lies in \"well I called `.from_pretrained` so I would expect the model to have pretrained weights\". However, the distinction is that if you run .from_pretrained on Auto**Config** you are not loading weights but you are loading a pre-existing config file. Loading pre-existing weights can only be done in a **Model** by using its `from_pretrained` method. But I agree that this could be improved in the documentation. I'll reopen this, try to improve the documentation, and close the issue when it's done.\r\n\r\nThanks for bringing this to the attention!", "Hi, we tried to make it clear in the documentation by specifying [it in the `PretrainedConfig` class](https://huggingface.co/transformers/main_classes/configuration.html#pretrainedconfig). \r\n\r\nI think we could add this note to `AutoConfig` as well, as I doubt users using `AutoConfig` read the documentation of `PretrainedConfig` as well.", "Oh, sorry @BramVanroy I didn't see you assigned it to yourself. Do you want to add the documentation note? Maybe you have additional ideas of where it should be added?", "I think that another place to mention this note is under the [from-config](https://huggingface.co/transformers/model_doc/auto.html#transformers.AutoModel.from_config) method.", "You're right, it would be nice to specify it there as well!", "> Oh, sorry @BramVanroy I didn't see you assigned it to yourself. Do you want to add the documentation note? Maybe you have additional ideas of where it should be added?\r\n\r\nOh, go ahead! You know the library better than I do so your judgement of where to add a note is better." ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4684?src=pr&el=h1) Report\n> Merging [#4684](https://codecov.io/gh/huggingface/transformers/pull/4684?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0866669e751bef636fa693b704a28c1fea9a17f3&el=desc) will **increase** coverage by `0.00%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4684?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4684 +/- ##\n=======================================\n Coverage 77.14% 77.15% \n=======================================\n Files 128 128 \n Lines 21070 21070 \n=======================================\n+ Hits 16255 16256 +1 \n+ Misses 4815 4814 -1 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4684?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4684/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.83% <0.00%> (+0.16%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4684?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/4684?src=pr&el=footer). Last update [0866669...e12a141](https://codecov.io/gh/huggingface/transformers/pull/4684?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

[]

[ "I can reproduce this. cc @n1t0 @mfuntowicz Special \"unused\" tokens are not tokenised correctly. This happens in the fast tokenizers as well as the slow ones. See test case below.\r\n\r\n```python\r\nfrom transformers import BertTokenizer\r\nUSE_FAST = True\r\n\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-cased\", use_fast=USE_FAST)\r\nprint('\"[unused1]\" in vocab?', \"[unused1]\" in tokenizer.vocab)\r\nprint('\"[unused1]\" index in vocab', tokenizer.vocab[\"[unused1]\"] if \"[unused1]\" in tokenizer.vocab else \"NA\")\r\n\r\nidxs = tokenizer.encode(\"[unused1]\", add_special_tokens=False)\r\nprint(\"indices\", idxs)\r\nrecoded = tokenizer.decode(idxs)\r\nprint(\"recoded\", recoded)\r\n```", "Hi @jxyxiangyu, thanks for reporting this, thanks @BramVanroy to making a code to reproduce.\r\n\r\nSo far, the behavior you want to achieve needs to be done by deactivating the `do_basic_tokenize` feature on `BertTokenizer`, otherwise the input will be splitted on ponctuation chars before actually going through the wordpiece tokenizer.\r\n\r\n_I don't think we have an equivalent on the Rust implementation of Bert, let me check internally and get back to you on this point._\r\n\r\nHere a snippet of code which should achieve the desired behavior:\r\n\r\n```python\r\nfrom transformers import BertTokenizer\r\n\r\ntokenizer = BertTokenizer.from_pretrained(\"bert-base-cased\", do_basic_tokenize=False)\r\ntokenizer.tokenize(\"[unused1]\")\r\n\r\n>>> ['[unused1]']\r\n\r\ntokenizer.encode(\"[unused1]\", add_special_tokens=False)\r\n>>> [1]\r\n\r\ntokenizer.decode([1])\r\n>>> '[unused1]'\r\n```", "Thanks for responding to my query. After I tried the method you gave, ‘[unused1]’ could indeed be tokenized correctly, but I want to use '[unused1]' to concatenate two words with little relation. In my opinion, may I set other words do_basic_tokenize as True, and '[unused1]' as False?", "Hi @jxyxiangyu! Thank you @BramVanroy & @mfuntowicz for the help on this!\r\n\r\nI think in this case the easiest way to handle this, is by adding the tokens you plan to use as special tokens. After all, that's what they are. They are not added by default since only a handful of them are actually used so you need to do it manually with\r\n```python\r\ntokenizer.add_special_tokens({ \"additional_special_tokens\": [ \"[unused1]\" ] })\r\n```\r\n\r\nThen, it should work for both fast and slow tokenizers:\r\n```python\r\n>>> from transformers import AutoTokenizer\r\n\r\n>>> slow = AutoTokenizer.from_pretrained(\"bert-base-cased\", use_fast=False)\r\n>>> fast = AutoTokenizer.from_pretrained(\"bert-base-cased\", use_fast=True)\r\n\r\n>>> slow.add_special_tokens({ \"additional_special_tokens\": [ \"[unused1]\" ] })\r\n>>> fast.add_special_tokens({ \"additional_special_tokens\": [ \"[unused1]\" ] })\r\n\r\n>>> slow.encode(\"[unused1]\", add_special_tokens=False)\r\n[1]\r\n>>> fast.encode(\"[unused1]\", add_special_tokens=False)\r\n[1]\r\n```", "Thank you very much for your response, which solved my confusion", "Awesome! Closing the issue, do not hesitate to reopen if needed!", "Hi @n1t0 , I had a related question. I want to re-register a token's id in the vocab so that I don't need to add a new token and expand the size of the vocab. Specifically, for example I want to register \"qazwsx\" to id 1, which means I want to replace \"[unused1]\" : 1 by \"qazwsx\" : 1. Do you know how to achieve this?\r\nAnother question is how to synchronize two tokens with the same id, to avoid expanding the size of vocab. For example, instead of replacing [unused1]\" : 1 by \"qazwsx\" : 1, I want to keep both in the new vocab.\r\n\r\nThank you so much for the help!" ]

[ { "id": 1834059054, "node_id": "MDU6TGFiZWwxODM0MDU5MDU0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Generation", "name": "Ex: Generation", "color": "06EFF8", "default": false, "description": "Natural Language Generation" } ]

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

[ "Hi @thesamuel, \r\n\r\nIdeally, the padding text should not influence the outcome, but this is more a hack to make XLNet work with short prompts, than actual science.\r\n\r\nAlso note that it is recommended now to use the TextGeneration Pipeline instead of the `run_generation` script: \r\n\r\n```\r\nfrom transformers import pipeline\r\ngenerator = pipeline(\"text-generation\", model=\"xlnet-base-cased\")\r\nprint(generator(\"We propose a \"))\r\n```\r\n**Note**: This works well for XLNet only after merging PR: #4686. So for pipelines to work, you either have to wait a bit or work on the branch of the PR: #4686. \r\n\r\nAs a default the pipeline employs sampling instead of greedy search. You might also want to play around with the generation hyperparameters here a bit for better results. To learn more about how to effectively use the many parameters for text generation, you might want to take a look at: https://huggingface.co/blog/how-to-generate\r\n" ]

[ { "id": 1834060867, "node_id": "MDU6TGFiZWwxODM0MDYwODY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Named%20Entity%20Recognition", "name": "Ex: Named Entity Recognition", "color": "06FFD8", "default": false, "description": "" }, { "id": 1834067346, "node_id": "MDU6TGFiZWwxODM0MDY3MzQ2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Documentation", "name": "Documentation", "color": "77cc3b", "default": false, "description": "" } ]

[]

[ "Looks great!", "(I've relaunched the failing CI test that's unrelated)", "@julien-c do you think rebasing onto latest master would fix that problem?" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4680?src=pr&el=h1) Report\n> Merging [#4680](https://codecov.io/gh/huggingface/transformers/pull/4680?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/a801c7fd74f56a651ba43bfc93eba93c63e84766&el=desc) will **decrease** coverage by `0.01%`.\n> The diff coverage is `75.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4680?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4680 +/- ##\n==========================================\n- Coverage 78.02% 78.01% -0.02% \n==========================================\n Files 124 124 \n Lines 20626 20634 +8 \n==========================================\n+ Hits 16094 16098 +4 \n- Misses 4532 4536 +4 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4680?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_encoder\\_decoder.py](https://codecov.io/gh/huggingface/transformers/pull/4680/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19lbmNvZGVyX2RlY29kZXIucHk=) | `92.20% <75.00%> (-3.45%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4680/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.41% <0.00%> (-0.12%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4680?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/4680?src=pr&el=footer). Last update [a801c7f...a61742a](https://codecov.io/gh/huggingface/transformers/pull/4680?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "@thomwolf @LysandreJik @sshleifer - merging for now to solve a bunch of issues. On a side-note, we have not really released the `EncoderDecoderModel` feature of `transformers` yet, or? Are we planning on doing something semi-official for this?", ">I will add an encoder decoder notebook in the next 1-2 weeks showing in-detail how Bert2Bert can be used with EncoderDecoder\r\n\r\n@patrickvonplaten \r\n\r\nI'm already working on a EncoderDecoder notebook for summarization task using kaggle news summery dataset. Hope to finish it in a week :)", "> > I will add an encoder decoder notebook in the next 1-2 weeks showing in-detail how Bert2Bert can be used with EncoderDecoder\r\n> \r\n> @patrickvonplaten\r\n> \r\n> I'm already working on a EncoderDecoder notebook for summarization task using kaggle news summery dataset. Hope to finish it in a week :)\r\n\r\nThat's great news! Do you use a Bert2Bert implementation? ", "Yes, I am using Bert2Bert", "@patrickvonplaten I noticed that since this change was merged this issue happens #5826, is it possible I'm using the config api incorrectly or may be a real issue? \r\n\r\nSorry for tagging you, thanks in advance!", "Hey @afcruzs - will answer on the issue :-) ", "Hey, @patrickvonplaten , I wonder if there is any way to get the cross attention weights in the decoder from `EncoderDecoderModel` model. I looked at the document, it only will return the self-attention weights (`decoder_attentions `) in the decoder from `EncoderDecoderModel`? Thanks very much, it bothers me for a while.", "Hey @kimmo1019 - could you please open a new issue about this? :-) " ]

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

[]

[ "Thanks, looks like a duplicate of #4658 ", "Ups, haven't seen this on mobile.\r\nMy apologies" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1834067346, "node_id": "MDU6TGFiZWwxODM0MDY3MzQ2", "url": "https://api.github.com/repos/huggingface/transformers/labels/Documentation", "name": "Documentation", "color": "77cc3b", "default": false, "description": "" } ]

[ { "login": "srush", "id": 35882, "node_id": "MDQ6VXNlcjM1ODgy", "avatar_url": "https://avatars.githubusercontent.com/u/35882?v=4", "gravatar_id": "", "url": "https://api.github.com/users/srush", "html_url": "https://github.com/srush", "followers_url": "https://api.github.com/users/srush/followers", "following_url": "https://api.github.com/users/srush/following{/other_user}", "gists_url": "https://api.github.com/users/srush/gists{/gist_id}", "starred_url": "https://api.github.com/users/srush/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/srush/subscriptions", "organizations_url": "https://api.github.com/users/srush/orgs", "repos_url": "https://api.github.com/users/srush/repos", "events_url": "https://api.github.com/users/srush/events{/privacy}", "received_events_url": "https://api.github.com/users/srush/received_events", "type": "User", "site_admin": false } ]

[ "Since no one else answers this:\r\nCheck out the simpletransformers library, it's and wrapper of this one and there are some blog posts linked to examples.\r\nhttps://github.com/ThilinaRajapakse/simpletransformers\r\nThe author is very open to every idea and the work with him is pretty good.\r\nMaybe it will help you.\r\nFor most common tasks the distilroberta model gives good results and doesn't need as much computing power as longformer or t5 needs", "Thanks for the link, very helpful. I’m also finding that if I know what I need to do and which model to use, the API docs do have well coded examples.", "If you don't know which model to use you can even just checkout the model hub.\r\nFor example for question answering tasks just search for \"squad\" models, then you will find a lot of pretrained models for this task", "> Since no one else answers this:\r\n\r\nI marked this question as a \"Good First Issue\". The idea is that we encourage people who are new to contributing to add their input. Particularly, you can add new examples that are very entry-level to explain the basic principles of the library.\r\n\r\n", "@falconair Thanks for raising this issue. We think we can make things better here.\r\n\r\nIf any one is interested in helping out, let me know. In particular it would be helpful to have examples of (non-nlp) projects that do this well. We could also use some beta testers. \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", "> @falconair Thanks for raising this issue. We think we can make things better here.\r\n> \r\n> If any one is interested in helping out, let me know. In particular it would be helpful to have examples of (non-nlp) projects that do this well. We could also use some beta testers.\r\n\r\nPing me if you need beta testers. I can test for Windows, Linux, Linux DDP. ", "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 }, { "id": 1834059054, "node_id": "MDU6TGFiZWwxODM0MDU5MDU0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Generation", "name": "Ex: Generation", "color": "06EFF8", "default": false, "description": "Natural Language Generation" } ]

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

[ "After looking over `modeling_utils.generate` (the function used for generation by `run_generation.py`), I believe that the sliding window approach is not yet implemented.\r\n\r\nThis method was implemented for CTRL generation in their repo, so you may be able to adapt some of their code for your use case: https://github.com/salesforce/ctrl/blob/master/generation.py#L186-L189", "The iterative approach to generation (using `past`) may work better because you can control the sliding window manually.", "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", "@patrickvonplaten @rautnikita77 @minimaxir Has anyone attempted to implement this (using cached keys and values)?" ]

[ { "id": 1834052333, "node_id": "MDU6TGFiZWwxODM0MDUyMzMz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Ex:%20Question%20Answering", "name": "Ex: Question Answering", "color": "86FFCF", "default": false, "description": "" } ]

[]

[ "I can't reproduce on master. Do you mind specifying the context/question?", "Because you couldn't reproduce, I tried with the latest version (from git : 2.11.0) and the problem seems gone (tested with random articles from wikipedia). \r\n\r\nMaybe my first text had bad format or non utf-8 characters (but I remember to have tested with many differents inputs before openning an issue) or it was due to a bug fixed in 2.11\r\n\r\nSorry to have bothered you, thanks for the support ! \r\n", "Alright, no worries! Let me know if you have an issue down the road.", "@LysandreJik \r\nI got the exact same error using Camembert (\"illuin/camembert-large-fquad\") and the question answering pipeline.\r\n\r\nOpening a new issue.\r\n\r\nQuestions :\r\nLe loyer est-il révisé annuellement ou triennalemment ?\r\nQuel est la nature de l’indice de base ?\r\nLe bail est-il soumis à TVA ou non soumis à TVA ?\r\n\r\nContext : \r\n[context_mono.txt](https://github.com/huggingface/transformers/files/4752269/context_mono.txt)\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"qa.py\", line 164, in <module>\r\n main_file()\r\n File \"qa.py\", line 161, in main_file\r\n analayse(mode)\r\n File \"qa.py\", line 85, in analayse\r\n answer_C = nlp_camembert_gpu_f(question=question_text, context=context)\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/transformers/pipelines.py\", line 1229, in __call__\r\n for s, e, score in zip(starts, ends, scores)\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/transformers/pipelines.py\", line 1229, in <listcomp>\r\n for s, e, score in zip(starts, ends, scores)\r\nKeyError: 377\r\n```" ]

[]

[]

[ "Hi! If you're new to the library, I heavily recommend taking a look at the [glossary (position IDs in this case)](https://huggingface.co/transformers/glossary.html#position-ids), which explains how to use such inputs.\r\n\r\nIf you ignore the `position_ids`, then they're always automatically generated to be the same as the model's pre-training scheme. If you're fine-tuning and wish to keep the same position embeddings, then you don't need to pass them to the model.", "THanks @LysandreJik , your second statement pretty much answered it completely! Thanks", "@LysandreJik Hi, does that mean the position embeddings won't get updated if `position_ids` are not passed to the model (or with the default value of `None`)? Would you point me to the related lines in the code that implements this logic? Thanks!", "This means that the position IDs will be generated on the fly, and the position embeddings will be exactly the same than during the pre-training. You can check the code [here](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_bert.py#L192-L214)." ]

[]

[]

[ "Regarding the Multiple Choice `global_attention_mask`. Discussion taken from PR: https://github.com/huggingface/transformers/pull/4645#issuecomment-635429380\r\n\r\n> @patil-suraj, we can leave it to the user, or we can just do as you suggested earlier, put global attention on the question and all choices, which should work.\r\n> \r\n> @patrickvonplaten, what do you think?\r\n\r\nRegarding the multiple choice, I think we usually have the following tensor:\r\n```\r\n[\r\n[ context, choice_a],\r\n[ context, choice_b],\r\n[ context, choice_c], \r\n...\r\n]\r\n```\r\n\r\nsee here: https://github.com/huggingface/transformers/blob/9c17256447b91cf8483c856cb15e95ed30ace538/examples/multiple-choice/utils_multiple_choice.py#L529\r\n\r\nSo I'd suggest if no `global_attention_mask` is provided by the user, we initialize the `global_attention_mask` so that all choice contexts do global attention. If the user wants a different global attention he has now the possibility to define it himself.\r\n\r\n@ibeltagy ", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4672?src=pr&el=h1) Report\n> Merging [#4672](https://codecov.io/gh/huggingface/transformers/pull/4672?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/9c17256447b91cf8483c856cb15e95ed30ace538&el=desc) will **decrease** coverage by `0.05%`.\n> The diff coverage is `41.37%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4672?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4672 +/- ##\n==========================================\n- Coverage 77.23% 77.17% -0.06% \n==========================================\n Files 128 128 \n Lines 21050 21060 +10 \n==========================================\n- Hits 16257 16253 -4 \n- Misses 4793 4807 +14 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4672?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_longformer.py](https://codecov.io/gh/huggingface/transformers/pull/4672/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19sb25nZm9ybWVyLnB5) | `93.52% <41.37%> (-3.53%)` | :arrow_down: |\n| [src/transformers/file\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4672/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9maWxlX3V0aWxzLnB5) | `73.85% <0.00%> (+0.41%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4672?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/4672?src=pr&el=footer). Last update [9c17256...9ad5319](https://codecov.io/gh/huggingface/transformers/pull/4672?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Ok, checked the notebook: https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing and results are the same as before so no breaking changes. \r\n\r\nGood to merge for me!", "> Ok, checked the notebook: https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing and results are the same as before so no breaking changes.\r\n> \r\n> Good to merge for me!\r\n\r\nThis is great !\r\n\r\nJust noticed one typo in the first line\r\n> This notebook shows how `nlp` can be leveraged `nlp` to evaluate Longformer on TriviaQA " ]

[]

[]

[ "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", "Ran into this today, this is a major issue behind corporate proxy/CA self signed certs. @SinghB maybe you can reopen this, I'm not sure who to `@mention`?\r\n\r\n```python\r\nget_from_cache(\r\n # ...\r\n etag = None\r\n if not local_files_only:\r\n try:\r\n response = requests.head(url, allow_redirects=True, proxies=proxies, timeout=etag_timeout)\r\n if response.status_code == 200:\r\n etag = response.headers.get(\"ETag\")\r\n except (EnvironmentError, requests.exceptions.Timeout):\r\n # etag is already None\r\n # THIS ALSO SWALLOWS ALL OTHER NON \"404\" ERRORS (e.g. SSL, Proxy, etc.)\r\n pass\r\n # ....\r\n```", "Yes I want to track and solve that issue in the next couple of weeks.", "+1 from me.\r\n\r\nCurrently I get this error message:\r\n\r\n```\r\nOSError: Can't load weights for 'bert-base-uncased'. Make sure that:\r\n\r\n- 'bert-base-uncased' is a correct model identifier listed on 'https://huggingface.co/models'\r\n\r\n- or 'bert-base-uncased' is the correct path to a directory containing a file named one of pytorch_model.bin, tf_model.h5, model.ckpt.\r\n```\r\n\r\nBut if I add a `raise` after the `# etag is already None` comment in stadelmanma's snippet, I see:\r\n\r\n```\r\nrequests.exceptions.SSLError: HTTPSConnectionPool(host='cdn.huggingface.co', port=443): Max retries exceeded with url: /bert-base-uncased-pytorch_model.bin (Caused by SSLError(SSLCertVerificationError(1, '[SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: unable to get local issuer certificate (_ssl.c:1056)')))\r\n```\r\n\r\nThis leads to confusing debugging... the behavior of this error shouldn't be to indicate to the user that models like `bert-based-uncased` don't exist.", "Hi @julien-c , I'm having the same issue as @stadelmanma (I'm behind a coporate proxy as well)\r\n\r\n``` File \"/home/USER/anaconda3/envs/codebert/lib/python3.7/site-packages/transformers/configuration_utils.py\", line 376, in from_pretrained\r\n config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)\r\n File \"/home/USER/anaconda3/envs/codebert/lib/python3.7/site-packages/transformers/configuration_utils.py\", line 436, in get_config_dict\r\n raise EnvironmentError(msg)\r\nOSError: Can't load config for 'microsoft/deberta-base'. Make sure that:\r\n\r\n- 'microsoft/deberta-base' is a correct model identifier listed on 'https://huggingface.co/models'\r\n\r\n- or 'microsoft/deberta-base' is the correct path to a directory containing a config.json file\r\n```\r\nI even tried setting the proxy in `proxies`: \r\n```config = config_class.from_pretrained(args.config_name, proxies={'http://': '<HOST>:<PORT>'})```\r\n\r\nBut same thing happens. \r\nMaybe there's a workaround for this?\r\n\r\nMany thanks!" ]

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

[]

[ "GPT-2 and BERT have very different tokenization mechanism. What do you mean by \"convert between the tokenizers\"? What do you want to do?", "I want to use Bert as encoder and GPT2 as decoder. Then I want to evaluate the generated text with another Bert as discriminator (like the technique [here](https://arxiv.org/pdf/1703.00955.pdf)). I want the decoder to generate text just based on the context vector (refer to the above link). So I don't think I can use EncoderDecoderModel (am I right?)" ]

[ { "id": 1834056761, "node_id": "MDU6TGFiZWwxODM0MDU2NzYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Core:%20Modeling", "name": "Core: Modeling", "color": "FF8446", "default": false, "description": "Internals of the library; Models." } ]

[]

[ "I think I see where that could be a problem, indeed. Do you mind sharing your model configuration file so that I may take a closer look?", "Here it is:\r\n\r\n```\r\n{\r\n \"architectures\": [\r\n \"BertForTokenClassification\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.3,\r\n \"bos_token_id\": 0,\r\n \"do_sample\": false,\r\n \"eos_token_ids\": 0,\r\n \"finetuning_task\": null,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.3,\r\n \"hidden_size\": 768,\r\n \"id2label\": {\r\n \"0\": \"I-MISC\",\r\n \"1\": \"B-MISC\",\r\n \"2\": \"O\",\r\n \"3\": \"I-LOC\",\r\n \"4\": \"I-ORG\",\r\n \"5\": \"B-LOC\",\r\n \"6\": \"B-ORG\",\r\n \"7\": \"I-PER\",\r\n \"8\": \"B-PER\"\r\n },\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"is_decoder\": false,\r\n \"label2id\": {\r\n \"I-MISC\": 0,\r\n \"B-MISC\": 1,\r\n \"O\": 2,\r\n \"I-LOC\": 3,\r\n \"I-ORG\": 4,\r\n \"B-LOC\": 5,\r\n \"B-ORG\": 6,\r\n \"I-PER\": 7,\r\n \"B-PER\": 8\r\n },\r\n \"layer_norm_eps\": 1e-12,\r\n \"length_penalty\": 1.0,\r\n \"max_length\": 20,\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"bert\",\r\n \"num_attention_heads\": 12,\r\n \"num_beams\": 1,\r\n \"num_hidden_layers\": 12,\r\n \"num_labels\": 9,\r\n \"num_return_sequences\": 1,\r\n \"output_attentions\": false,\r\n \"output_hidden_states\": false,\r\n \"output_past\": true,\r\n \"pad_token_id\": 0,\r\n \"pruned_heads\": {},\r\n \"repetition_penalty\": 1.0,\r\n \"temperature\": 1.0,\r\n \"top_k\": 50,\r\n \"top_p\": 1.0,\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\n", "@GuillemGSubies Thanks for reporting, I can reproduce this issue.\r\n\r\nTo fix it on your side while we push a fix, you can just remove the `num_labels` attribute from your config.json, it's not needed anymore. Let me know if this solves your issue.", "Thanks you very much. I will do that :heart: ", "Should also be fixed on master by 751a1e08904fda197366e4b0033bdfb8b10d256c" ]

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

[]

[ "Hi! This is indeed a memory error. At which point does it crash?", "A similar problem like this has been reported when using the Lazy version of LinebyLineTextDataset. Colab deals badly with situation where you are using 90+% of memory - it'll kick you out or throw OOM errors - which you would not get on local machines. This is unfortunate and hard to get around.\r\n\r\nIn this case, I think you are simply running out of memory. The newsroom dataset is huge (1M+ news _articles_). So that is likely the issue.", "@LysandreJik It crashes when suddenly the ram usage increases to around 7-8 GB and the increase is also very sudden. Its like it stays at 2-3 GB usage for a minute or so and then suddenly it shoots to 8GB and crashes.\r\n@BramVanroy I tried reducing the dataset by half and running it but I am still getting the same error. So would you suggest running it on local machine?I will have to run this part on local as on my local I have a little better ram (16GB ) but then I will have to train in colab only as I dont have a GPU on my local laptop. Is there a better workaround\r\nAlso thanks guys for giving such a quick answer.", "The sudden increase in RAM may be due to a sudden very large sentence/text which results in the whole batch having to be very large, exponentially increasing the memory usage. \r\n\r\nHow large is the dataset in terms of GB/number of lines?\r\n\r\nUnfortunately sometimes you cannot do what you would want due to practical restrictions (mostly money). If you want to train or finetune a model with a huge dataset, it is likely that you need more hardware than is available in free plans. \r\n\r\nPerhaps you can try out https://github.com/huggingface/nlp and find out if it has the dataset that you need. If not you can open an issue there and ask whether the dataset can be included. That should solve some issues since it takes into account RAM issues.", "@BramVanroy I have tried with 26GB ram, but it still crashes, is there any minimum requirement of hardware mentioned?", "No. I fear that this might simply not work on Colab. Line cache loads as much of the file as it can in memory and goes from there but Colab is being annoying and locks you out because it thinks you are going to throw an OOM error (but you won't on a regular system). ", "@BramVanroy , I am using \"notebook\" platform from google AI platforms with 26GB ram, (without GPU) but after running 2% for the very 1st epoch, it says : \r\n\r\n**can't allocate memory: you tried to allocate 268435456 bytes. Error code 12 (Cannot allocate memory)** . \r\n\r\nAm I doing something wrong?", "Just read my previous post. This is a problem about how Google deals with increasing memory usage and thinks an OOM will occur even though it won't. The problem is not with the implementation. It seems that you cannot. Use this functionality in these kinds of VMs. ", "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", "Could anybody find a solution for the issue?" ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4666?src=pr&el=h1) Report\n> Merging [#4666](https://codecov.io/gh/huggingface/transformers/pull/4666?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/b5015a2a0f4ea63035a877f5626cb0c3ce97e25d&el=desc) will **increase** coverage by `0.63%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4666?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4666 +/- ##\n==========================================\n+ Coverage 77.19% 77.83% +0.63% \n==========================================\n Files 128 128 \n Lines 21021 21021 \n==========================================\n+ Hits 16228 16362 +134 \n+ Misses 4793 4659 -134 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4666?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4666/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.17% <0.00%> (+0.23%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4666/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `86.21% <0.00%> (+14.10%)` | :arrow_up: |\n| [src/transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/4666/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3F1YWQucHk=) | `56.68% <0.00%> (+28.02%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4666?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/4666?src=pr&el=footer). Last update [b5015a2...49068a5](https://codecov.io/gh/huggingface/transformers/pull/4666?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[ { "id": 1838412367, "node_id": "MDU6TGFiZWwxODM4NDEyMzY3", "url": "https://api.github.com/repos/huggingface/transformers/labels/model%20card", "name": "model card", "color": "92d5f4", "default": false, "description": "Related to pretrained model cards" } ]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4665?src=pr&el=h1) Report\n> Merging [#4665](https://codecov.io/gh/huggingface/transformers/pull/4665?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/b5015a2a0f4ea63035a877f5626cb0c3ce97e25d&el=desc) will **increase** coverage by `0.21%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4665?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4665 +/- ##\n==========================================\n+ Coverage 77.19% 77.41% +0.21% \n==========================================\n Files 128 128 \n Lines 21021 21021 \n==========================================\n+ Hits 16228 16274 +46 \n+ Misses 4793 4747 -46 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4665?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4665/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.66% <0.00%> (ø)` | |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4665/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `90.17% <0.00%> (+0.23%)` | :arrow_up: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/4665/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `86.21% <0.00%> (+14.10%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4665?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/4665?src=pr&el=footer). Last update [b5015a2...e727442](https://codecov.io/gh/huggingface/transformers/pull/4665?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "Hello,\r\n\r\nThis is because you have to specify `--logging_dir /path/to/logs` as parameter. There will be a default location for the next release of the TF Trainer.", "Right. I later passed a parameter in run_tf_ner.py and it worked. I felt somewhere in the Trainer or run_tf_ner.py, it needs to be indicated. \r\n\r\nThanks for your prompt response!" ]

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

[]

[ "I might be wrong, but I think the focus of this library is on NLP which at most is multimodal. Also including object detection transformers may fall out of the scope of this project." ]

[]

[]

[ "Delete. It may be a hardware issue. When I changed it to Colab, this problem disappeared" ]

[ { "id": 1108649070, "node_id": "MDU6TGFiZWwxMTA4NjQ5MDcw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Need%20more%20information", "name": "Need more information", "color": "d876e3", "default": false, "description": "Further information is requested" } ]

[]

[ "I cannot reproduce your problem. Can you try again?", "I tried again. \r\n\r\nConsole on dev tools shows:\r\n\r\n```\r\nDevTools failed to load SourceMap: Could not load content for chrome-extension://ibnejdfjmmkpcnlpebklmnkoeoihofec/dist/contentScript.js.map: HTTP error: status code 404, net::ERR_UNKNOWN_URL_SCHEME\r\nDevTools failed to load SourceMap: Could not load content for chrome-extension://nkbihfbeogaeaoehlefnkodbefgpgknn/sourcemaps/contentscript.js.map: HTTP error: status code 404, net::ERR_UNKNOWN_URL_SCHEME\r\nDevTools failed to load SourceMap: Could not load content for chrome-extension://nkbihfbeogaeaoehlefnkodbefgpgknn/sourcemaps/inpage.js.map: HTTP error: status code 404, net::ERR_UNKNOWN_URL_SCHEME\r\nDevTools failed to load SourceMap: Could not load content for chrome-extension://ibnejdfjmmkpcnlpebklmnkoeoihofec/dist/pageHook.js.map: HTTP error: status code 404, net::ERR_UNKNOWN_URL_SCHEME\r\n```", "Those are unrelated warnings related to source maps. I can;t reproduce your problem and I tried on different browsers. Can you clear the cache/try in private mode/incognito?", "<img width=\"2543\" alt=\"Screen Shot 2020-06-01 at 1 50 50 am\" src=\"https://user-images.githubusercontent.com/597346/83359074-6acfea00-a3aa-11ea-97e9-d9fa0ed872ea.png\">\r\nStill no luck", "D'oh, I didn't check the PPLM page. The other versions of Write With Transformer seem to work, but you are right that it doesn't seem to work for [PPLM](https://transformer.huggingface.co/doc/pplm). When you trigger \"autocomplete\", the web page seems to hang.\r\n\r\ncc @julien-c ", "Yes, we turned off the PPLM machine as it was costly to host. We need to add a notice to try it locally instead, and/or re-spawn a cheaper machine. Both are on our todo-list.", "Added a notice there: https://transformer.huggingface.co/doc/pplm" ]

[ { "id": 2052904485, "node_id": "MDU6TGFiZWwyMDUyOTA0NDg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/reformer", "name": "reformer", "color": "5319e7", "default": false, "description": "Everything related to the reformer model" } ]

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

[]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4659?src=pr&el=h1) Report\n> Merging [#4659](https://codecov.io/gh/huggingface/transformers/pull/4659?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/f4e1f022100834bd00d4f877a883b5946c4cac37&el=desc) will **decrease** coverage by `0.34%`.\n> The diff coverage is `50.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4659?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4659 +/- ##\n==========================================\n- Coverage 78.40% 78.06% -0.35% \n==========================================\n Files 138 138 \n Lines 23757 23766 +9 \n==========================================\n- Hits 18627 18552 -75 \n- Misses 5130 5214 +84 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4659?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `87.50% <44.44%> (-0.72%)` | :arrow_down: |\n| [src/transformers/configuration\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2JlcnQucHk=) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3F1YWQucHk=) | `34.07% <0.00%> (-22.62%)` | :arrow_down: |\n| [src/transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19vcGVuYWkucHk=) | `79.51% <0.00%> (-1.39%)` | :arrow_down: |\n| [src/transformers/tokenization\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYmVydC5weQ==) | `90.45% <0.00%> (-0.83%)` | :arrow_down: |\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `94.81% <0.00%> (-0.38%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `85.81% <0.00%> (-0.30%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `91.28% <0.00%> (+0.12%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4659?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/4659?src=pr&el=footer). Last update [f4e1f02...400070b](https://codecov.io/gh/huggingface/transformers/pull/4659?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "> we'll look at upstreaming it in the `transformers.PretrainedModel` if everyone's on board.\r\n\r\nThanks, @LysandreJik. It would be great to make `gradient_checkpointing` available to more models. While the configuration can be upstreamed in `transformers.PretrainedConfig`, the implementation is model specific, where you need to call `torch.utils.checkpoint.checkpoint` inside the layers loop as in [here](https://github.com/huggingface/transformers/blob/bf4342743ad2f5a5e1090818ecb72f2ebc6e4f73/src/transformers/modeling_bert.py#L404).", "I was thinking of having the implementation be model agnostic as well. I haven't really thought out the best way, but a possible way to achieve it would be with a decorator; for example, in `PretrainedModel` we could have something like:\r\n\r\n```py\r\n @staticmethod\r\n def gradient_checkpointing(layer):\r\n @functools.wraps(layer)\r\n def wrapper(*args):\r\n layer_instance = args[0]\r\n # Remove the wrapper to prevent infinite recursion on the wrapper\r\n layer_instance.forward = functools.partial(layer_instance.forward.__wrapped__, layer_instance)\r\n \r\n if args[0].config.gradient_checkpointing:\r\n return torch.utils.checkpoint.checkpoint(layer_instance, *args[1:])\r\n else:\r\n return layer(*args)\r\n return wrapper\r\n\r\n```\r\n\r\nThen we can very simply add that decorator on the layers where we want the checkpoint:\r\n\r\n```py\r\nclass BertLayer(nn.Module):\r\n\r\n ...\r\n\r\n @PreTrainedModel.gradient_checkpointing\r\n def forward(\r\n self,\r\n hidden_states,\r\n attention_mask=None,\r\n head_mask=None,\r\n encoder_hidden_states=None,\r\n encoder_attention_mask=None,\r\n ):\r\n\r\n ...\r\n```\r\n\r\nThis would require that these layers have access to the configuration so that they're aware of gradient check-pointing or not.\r\n\r\nPretty convenient, but pretty different from our coding style as well cc @thomwolf ", "neat ", "A model agnostic approach might be best. In my research for isolating https://github.com/minimaxir/aitextgen/issues/6 for finetuning larger GPT-2 models, it appeared that checkpointing would have to be implemented at the model level, as this PR does for BERT.", "torch.utils.checkpoint.checkpoint works well in single GPU. But it causes OOM in multi-gpu with torch.nn.DataParallel.", "> I was thinking of having the implementation be model agnostic as well. I haven't really thought out the best way, but a possible way to achieve it would be with a decorator; for example, in `PretrainedModel` we could have something like:\r\n> \r\n> ```python\r\n> @staticmethod\r\n> def gradient_checkpointing(layer):\r\n> @functools.wraps(layer)\r\n> def wrapper(*args):\r\n> layer_instance = args[0]\r\n> # Remove the wrapper to prevent infinite recursion on the wrapper\r\n> layer_instance.forward = functools.partial(layer_instance.forward.__wrapped__, layer_instance)\r\n> \r\n> if args[0].config.gradient_checkpointing:\r\n> return torch.utils.checkpoint.checkpoint(layer_instance, *args[1:])\r\n> else:\r\n> return layer(*args)\r\n> return wrapper\r\n> ```\r\n> \r\n> Then we can very simply add that decorator on the layers where we want the checkpoint:\r\n> \r\n> ```python\r\n> class BertLayer(nn.Module):\r\n> \r\n> ...\r\n> \r\n> @PreTrainedModel.gradient_checkpointing\r\n> def forward(\r\n> self,\r\n> hidden_states,\r\n> attention_mask=None,\r\n> head_mask=None,\r\n> encoder_hidden_states=None,\r\n> encoder_attention_mask=None,\r\n> ):\r\n> \r\n> ...\r\n> ```\r\n> \r\n> This would require that these layers have access to the configuration so that they're aware of gradient check-pointing or not.\r\n> \r\n> Pretty convenient, but pretty different from our coding style as well cc @thomwolf\r\n\r\nI like idea of having a decorator function! Would it be enough to have this wrapper only at all \"Model\" forward functions, like `BertModel.forward()`? ", "> torch.utils.checkpoint.checkpoint works well in single GPU. But it causes OOM in multi-gpu with torch.nn.DataParallel.\r\n\r\nI haven't tried `torch.nn.DataParallel` but it works well with `torch.nn.DistributedDataParallel` on a single or multiple machines. ", "> I like idea of having a decorator function! Would it be enough to have this wrapper only at all \"Model\" forward functions, like `BertModel.forward()`?\r\n\r\nI don't think so. Even with the decorator, it is still model-specific; the decorator just makes the syntax easier. You still need to decide where to call it because too few calls (e.g. only on `BertModel.forward`), and you won't save enough memory, too many calls (e.g. on every `.forward` function) and the backward pass will be very slow.", "Pinging @julien-c so he can take a look.", "> > torch.utils.checkpoint.checkpoint works well in single GPU. But it causes OOM in multi-gpu with torch.nn.DataParallel.\r\n> \r\n> I haven't tried `torch.nn.DataParallel` but it works well with `torch.nn.DistributedDataParallel` on a single or multiple machines.\r\n\r\nThanks for the advice. But I try `torch.nn.DistributedDataParallel` and meet the same problem in https://github.com/pytorch/pytorch/issues/24005. The pytorch version is 1.2.0.\r\n\r\nThe code is:\r\n```\r\nif n_gpu > 1:\r\n # model = torch.nn.DataParallel(model)\r\n torch.distributed.init_process_group(backend=\"nccl\")\r\n model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)\r\n```\r\nBoth `find_unused_parameters=True` and `find_unused_parameters=False` get errors.\r\n\r\n\r\n\r\n", "@ibeltagy, after some back and forth offline with @julien-c and @thomwolf, the way you implemented it is preferred as it's simpler to understand and adheres better to the library's philosophy.\r\n\r\nI think we can merge this and then in a following PR add it to all the other models. Would you like to take care of that? No worries if not, I can definitely take care of it.", "@LysandreJik, glad this will be merged. \r\n\r\n> Would you like to take care of that? No worries if not, I can definitely take care of it.\r\n\r\nI will pass :D \r\n\r\n", "> > > torch.utils.checkpoint.checkpoint works well in single GPU. But it causes OOM in multi-gpu with torch.nn.DataParallel.\r\n> > \r\n> > \r\n> > I haven't tried `torch.nn.DataParallel` but it works well with `torch.nn.DistributedDataParallel` on a single or multiple machines.\r\n> \r\n> Thanks for the advice. But I try `torch.nn.DistributedDataParallel` and meet the same problem in [pytorch/pytorch#24005](https://github.com/pytorch/pytorch/issues/24005). The pytorch version is 1.2.0.\r\n> \r\n> The code is:\r\n> \r\n> ```\r\n> if n_gpu > 1:\r\n> # model = torch.nn.DataParallel(model)\r\n> torch.distributed.init_process_group(backend=\"nccl\")\r\n> model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)\r\n> ```\r\n> \r\n> Both `find_unused_parameters=True` and `find_unused_parameters=False` get errors.\r\n> \r\n> \r\n\r\nI encounter the same issue with torch 1.5.0 and latest transformers", "@ewrfcas, @schinger, do you have a small example that reproduces the error?\r\n\r\nI don't think we can fix this issue (needs a PyTorch fix https://github.com/pytorch/pytorch/issues/24005), but I think we can work around it by removing the unused parameters mentioned in the error message. ", "> @ewrfcas, @schinger, do you have a small example that reproduces the error?\r\n> \r\n> I don't think we can fix this issue (needs a PyTorch fix [pytorch/pytorch#24005](https://github.com/pytorch/pytorch/issues/24005)), but I think we can work around it by removing the unused parameters mentioned in the error message.\r\n\r\nsquad example training can reproduce this error: https://github.com/huggingface/transformers/tree/master/examples/question-answering\r\n\r\npython -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \\\r\n --model_type bert \\\r\n --model_name_or_path bert-large-uncased-whole-word-masking \\\r\n --do_train \\\r\n --do_eval \\\r\n --do_lower_case \\\r\n --train_file SQUAD_DIR/dev-v1.1.json \\\r\n --predict_file SQUAD_DIR/dev-v1.1.json \\\r\n --learning_rate 3e-5 \\\r\n --num_train_epochs 2 \\\r\n --max_seq_length 384 \\\r\n --doc_stride 128 \\\r\n --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \\\r\n --per_gpu_eval_batch_size=1 \\\r\n --per_gpu_train_batch_size=1 \\\r\n\r\nno matter find_unused_parameters is ture or false", "Thanks. It would be more helpful if you provide a simpler and smaller example that I can easily run.", "> Thanks. It would be more helpful if you provide a simpler and smaller example that I can easily run.\r\n\r\nyou can change --train_file to SQUAD_DIR/dev-v1.1.json, dev set is small for quickly run", "> > torch.utils.checkpoint.checkpoint works well in single GPU. But it causes OOM in multi-gpu with torch.nn.DataParallel.\r\n> \r\n> I haven't tried `torch.nn.DataParallel` but it works well with `torch.nn.DistributedDataParallel` on a single or multiple machines.\r\n\r\ncould you show me a example about gradient checkpoint works successfully with `torch.nn.DistributedDataParallel` on multi-gpu?", "> @ewrfcas, @schinger, do you have a small example that reproduces the error?\r\n> \r\n> I don't think we can fix this issue (needs a PyTorch fix [pytorch/pytorch#24005](https://github.com/pytorch/pytorch/issues/24005)), but I think we can work around it by removing the unused parameters mentioned in the error message.\r\n\r\nI have trained a base model instead of large to delay this problem.\r\nThe only differences in the code are \r\n```\r\nclass BertEncoder(nn.Module):\r\n def forward(...):\r\n ...\r\n for i, layer_module in enumerate(self.layer):\r\n ...\r\n if self.use_grad_ckpt:\r\n layer_outputs = torch.utils.checkpoint.checkpoint(layer_module, hidden_states, attention_mask, head_mask[i],\r\n encoder_hidden_states, encoder_attention_mask)\r\n else:\r\n layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i],\r\n encoder_hidden_states, encoder_attention_mask)\r\n ...\r\n ...\r\n```\r\nand \r\n```\r\ntorch.distributed.init_process_group(backend=\"nccl\")\r\nmodel = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)\r\n```\r\nOther codes are the same as normal finetuning codes.", "Here's a small example to replicate the error\r\n```\r\nimport os\r\nimport torch\r\nfrom transformers import BertForPreTraining\r\nos.environ['MASTER_ADDR'] = 'localhost'\r\nos.environ['MASTER_PORT'] = '12355'\r\ntorch.distributed.init_process_group(backend=\"nccl\", rank=0, world_size=1)\r\n\r\nmodel = BertForPreTraining.from_pretrained('bert-base-uncased', gradient_checkpointing=True).cuda()\r\nmodel = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)\r\noutputs = model(torch.tensor([[1, 2, 3]]).cuda())\r\noutputs[0].sum().backward()\r\n```\r\n\r\nUse `find_unused_parameters=True` and you will get \r\n```\r\nRuntimeError: Expected to mark a variable ready only once. This error is caused by use of a module parameter outside the `forward` function.\r\n```\r\n\r\nUse `find_unused_parameters=False`, and things will work just fine. \r\n\r\nI couldn't replicate the other error, \r\n```\r\nRuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. \r\n```\r\n@ewrfcas, do you know how to modify the example above to reproduce it?\r\n\r\n@schinger, can you try `find_unused_parameters=False` see if it fixes your problem.", "> Here's a small example to replicate the error\r\n> \r\n> ```\r\n> import os\r\n> import torch\r\n> from transformers import BertForPreTraining\r\n> os.environ['MASTER_ADDR'] = 'localhost'\r\n> os.environ['MASTER_PORT'] = '12355'\r\n> torch.distributed.init_process_group(backend=\"nccl\", rank=0, world_size=1)\r\n> \r\n> model = BertForPreTraining.from_pretrained('bert-base-uncased').cuda()\r\n> model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)\r\n> outputs = model(torch.tensor([[1, 2, 3]]).cuda())\r\n> outputs[0].sum().backward()\r\n> ```\r\n> \r\n> Use `find_unused_parameters=True` and you will get\r\n> \r\n> ```\r\n> RuntimeError: Expected to mark a variable ready only once. This error is caused by use of a module parameter outside the `forward` function.\r\n> ```\r\n> \r\n> Use `find_unused_parameters=False`, and things will work just fine.\r\n> \r\n> I couldn't replicate the other error,\r\n> \r\n> ```\r\n> RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. \r\n> ```\r\n> \r\n> @ewrfcas, do you know how to modify the example above to reproduce it?\r\n> \r\n> @schinger, can you try `find_unused_parameters=False` see if it fixes your problem.\r\n\r\nI have tried this code. Although it works in the first, the second forword will be failed. You can try to repeat the loss.backward for few times.", "@ewrfcas, I get this error with `gradient_checkpointing=True` and `gradient_checkpointing=False` (btw, `gradient_checkpointing` was set to `False` in the example above and I just updated it), so it is a problem with the example, not gradient checkpointing. Can you try to fix the example? or can you try it in a training loop that uses DDP correctly, either with pytorch-lightning or hugginface trainer?", "> @ewrfcas, I get this error with `gradient_checkpointing=True` and `gradient_checkpointing=False` (btw, `gradient_checkpointing` was set to `False` in the example above and I just updated it), so it is a problem with the example, not gradient checkpointing. Can you try to fix the example? or can you try it in a training loop that uses DDP correctly, either with pytorch-lightning or hugginface trainer?\r\n\r\nI have solved this problem by removing the self.pooler layer in BertModel because it did not forward any thing during the training. As the error saied, all parameters should be used in loss for DistributedDataParallel with find_unused_parameters=False, and find_unused_parameters=True is incompatible with gradient_checkpointing.\r\n\r\nMaybe we need this code after the first backward\r\n```\r\n# check parameters with no grad\r\nfor n, p in model.named_parameters():\r\n if p.grad is None and p.requires_grad is True:\r\n print('No forward parameters:', n, p.shape)\r\n```", "Nice finding, @ewrfcas. \r\n\r\n@LysandreJik, what is the best way to address this problem? do we need to fix it or can we leave it to the user to make sure all the model params are used? maybe in a separate PR we can find a way to automatically remove unused model params?\r\n\r\nAlso, aside from this issue, anything else we need to merge the PR? ", "Right, I think this should be looked at in a separate PR. Will take a final look and merge this PR tomorrow, and then look at implementing gradient checkpointing to the rest of the models." ]

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

[]

[ "My god, the paper hasn't even been up for a day...\r\n\r\nSaid being, +1", "So who can run 175B parameters and what do I have to do for a favor?", "The full model will be at least 350 GB (16-bit parameters). You'd need to partition it across more than (350 GB) / (16 GB) ~ **22 GPUs** just to run it! Not to mention the egress costs of making a model that size available.\r\n\r\nOf course, the paper shows **8** different-sized models, **4 of which are smaller than GPT-2**, so some of those could be practical. 🙂 \r\n\r\n", "Is there any Colab to test at least GPT-3 XL ?", "> Is there any Colab to test at least GPT-3 XL ?\r\n\r\nThey haven't released any code or pretrained models yet. See the issue on the official repo: https://github.com/openai/gpt-3/issues/1", "Note that the released models may be FP16, which may require forcing FP16 for use/finetuning (and therefore hardware-limited), or casting up to FP32.", "> Of course, the paper shows **8** different-sized models, **4 of which are smaller than GPT-2**, so some of those could be practical. slightly_smiling_face\r\n\r\nOne of the main benefits of the smaller gpt-3 models compared to their gpt-2 counterparts could be the increased context length of 2048 tokens.", "Yeah, personally, I wouldn't be able to use the models that won't fit in a Tesla P100", "The [GPT-3 repo](https://github.com/openai/gpt-3) is now archived (read-only) so perhaps OpenAI isn't planning on releasing anything this time around.", "> The [GPT-3 repo](https://github.com/openai/gpt-3) is now archived (read-only) so perhaps OpenAI isn't planning on releasing anything this time around.\r\n\r\nThat is a crying shame, because my system could do-er... :(", "Hopefully they have a better excuse than last time.", "> Hopefully they have a better excuse than last time.\r\n\r\n@flarn2006 You mean the....ooohhhh we created something scary and have soggy diapers excuse with GPT-3?", "@flarn2006 If they don't make excuses or drag their feet, and I finish my system build in a relatively congruent time frame, hopefully I can help...", "A little update: OpenAI's now running their own API with GPT-3 on it. https://beta.openai.com\r\nYou can apply for access, but seems like they're aiming mostly for big companies, not researchers. Sad, way too sad.", "But who put the \"Open\" in OpenAI then 🤔", "I guess we will need to \"fundraise\" enough GPU-compute to run the GPT3 model. :smile: ", "It should be possible to run lower-models on regular GPUs, like 1b model. But we don't have the model itself, and seems that OpenAI is against releasing it and would rather commercialize it :(", "I wonder if you could hardcode the 175B model into an electronic chip(like an ASIC but more specific)", "> I wonder if you could hardcode the 175B model into an electronic chip(like an ASIC but more specific)\r\n\r\nVery interesting as an idea. @StealthySemicolon do you have reference to other similar work done in the past?", "> > I wonder if you could hardcode the 175B model into an electronic chip(like an ASIC but more specific)\r\n> \r\n> Very interesting as an idea. @StealthySemicolon do you have reference to other similar work done in the past?\r\n\r\nNo, just a hunch. Even if I did know how to do this, it's not like OpenAI would publicly release the model weights...", "Guys when is this gonna be integrated!?", "When OpenAI decides to release GPT-3 open-sourcely, but this won't happen it seems, they just want to sell access to big corporations.", "https://bdtechtalks.com/2020/08/17/openai-gpt-3-commercial-ai/amp/\n\nHere it goes...", "https://arxiv.org/abs/2009.07118\r\nhttps://github.com/timoschick/pet", "> Hopefully they have a better excuse than last time.\r\n\r\nBecause Microsoft [gave us money.](https://openai.com/blog/openai-licenses-gpt-3-technology-to-microsoft/) ", "GPT-3 is not coming out anytime soon :(", "this thread signifies capitalism's pros and cons at the same time...😅", "> The full model will be at least 350 GB (16-bit parameters). You'd need to partition it across more than (350 GB) / (16 GB) ~ **22 GPUs** just to run it! Not to mention the egress costs of making a model that size available.\r\n> \r\n> Of course, the paper shows **8** different-sized models, **4 of which are smaller than GPT-2**, so some of those could be practical. 🙂\r\n> \r\n> \r\n\r\n@AdamDanielKing is there a way to estimate the size of the GPT-3 XL 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", "we're still waiting.. :(" ]

[ { "id": 1834053813, "node_id": "MDU6TGFiZWwxODM0MDUzODEz", "url": "https://api.github.com/repos/huggingface/transformers/labels/PyTorch", "name": "PyTorch", "color": "a12bef", "default": false, "description": "Anything PyTorch" } ]

[]

[ "I've tried `transformers 2.10.0` under `CUDA 10.2` with `PyTorch 1.5.0` and apex compiled for that environment, as well as under `CUDA 10.1` with both PyTorch 1.5.0 and 1.4.1, as well as apex compiled for both of those. However, I get pretty much the same issue. Should I down convert to a different version of transformers?\r\n```\r\nEpoch: 0%| | 0/2 [00:00<?, ?it/s] Traceback (most recent call last):\r\nptcc_1 | File \"/ptcc/run_language_modeling.py\", line 281, in <module>\r\nptcc_1 | main()\r\nptcc_1 | File \"/ptcc/run_language_modeling.py\", line 245, in main\r\nptcc_1 | trainer.train(model_path=model_path)\r\nptcc_1 | File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 470, in train\r\nptcc_1 | tr_loss += self._training_step(model, inputs, optimizer)\r\nptcc_1 | File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 577, in _training_step\r\nptcc_1 | scaled_loss.backward()\r\nptcc_1 | File \"/usr/lib/python3.6/contextlib.py\", line 88, in __exit__\r\nptcc_1 | next(self.gen)\r\nptcc_1 | File \"/usr/local/lib/python3.6/dist-packages/apex-0.1-py3.6-linux-x86_64.egg/apex/amp/handle.py\", line 127, in scale_loss\r\nptcc_1 | should_skip = False if delay_overflow_check else loss_scaler.update_scale()\r\nptcc_1 | File \"/usr/local/lib/python3.6/dist-packages/apex-0.1-py3.6-linux-x86_64.egg/apex/amp/scaler.py\", line 200, in update_scale\r\nptcc_1 | self._has_overflow = self._overflow_buf.item()\r\nptcc_1 | RuntimeError: CUDA error: an illegal memory access was encountered\r\nptcc_1 | Gradient overflow. Skipping step, loss scaler 0 reducing loss scale to 32768.0\r\nptcc_1 | /usr/local/lib/python3.6/dist-packages/torch/optim/lr_scheduler.py:114: UserWarning: Seems like `optimizer.step()` has been overridden after learning rate scheduler initialization. Please, make sure to call `optimizer.step()` before `lr_scheduler.step()`. See more details at https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate\r\nptcc_1 | \"https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate\", UserWarning)\r\nptcc_1 | terminate called after throwing an instance of 'c10::Error'\r\nptcc_1 | what(): CUDA error: an illegal memory access was encountered (insert_events at /pytorch/c10/cuda/CUDACachingAllocator.cpp:771)\r\nptcc_1 | frame #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x46 (0x7f2ededfd536 in /usr/local/lib/python3.6/dist-packages/torch/lib/libc10.so)\r\nptcc_1 | frame #1: c10::cuda::CUDACachingAllocator::raw_delete(void*) + 0x7ae (0x7f2edf040fbe in /usr/local/lib/python3.6/dist-packages/torch/lib/libc10_cuda.so)\r\nptcc_1 | frame #2: c10::TensorImpl::release_resources() + 0x4d (0x7f2edededabd in /usr/local/lib/python3.6/dist-packages/torch/lib/libc10.so)\r\nptcc_1 | frame #3: std::vector<c10d::Reducer::Bucket, std::allocator<c10d::Reducer::Bucket> >::~vector() + 0x1d9 (0x7f2f26356d99 in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #4: c10d::Reducer::~Reducer() + 0x23a (0x7f2f2634c6ea in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #5: std::_Sp_counted_ptr<c10d::Reducer*, (__gnu_cxx::_Lock_policy)2>::_M_dispose() + 0x12 (0x7f2f2632b662 in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #6: std::_Sp_counted_base<(__gnu_cxx::_Lock_policy)2>::_M_release() + 0x46 (0x7f2f25cee306 in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #7: <unknown function> + 0x87130b (0x7f2f2632c30b in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #8: <unknown function> + 0x2403a0 (0x7f2f25cfb3a0 in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #9: <unknown function> + 0x2415ee (0x7f2f25cfc5ee in /usr/local/lib/python3.6/dist-packages/torch/lib/libtorch_python.so)\r\nptcc_1 | frame #10: /usr/bin/python3() [0x572a27]\r\nptcc_1 | frame #11: /usr/bin/python3() [0x54eef2]\r\nptcc_1 | frame #12: /usr/bin/python3() [0x588948]\r\nptcc_1 | frame #13: /usr/bin/python3() [0x5ad438]\r\nptcc_1 | frame #14: /usr/bin/python3() [0x5ad44e]\r\nptcc_1 | frame #15: /usr/bin/python3() [0x5ad44e]\r\nptcc_1 | frame #16: /usr/bin/python3() [0x56b276]\r\nptcc_1 | frame #17: PyDict_SetItemString + 0x153 (0x5709f3 in /usr/bin/python3)\r\nptcc_1 | frame #18: PyImport_Cleanup + 0x76 (0x4f2fc6 in /usr/bin/python3)\r\nptcc_1 | frame #19: Py_FinalizeEx + 0x5e (0x637e2e in /usr/bin/python3)\r\nptcc_1 | frame #20: Py_Main + 0x395 (0x638e95 in /usr/bin/python3)\r\nptcc_1 | frame #21: main + 0xe0 (0x4b0d00 in /usr/bin/python3)\r\nptcc_1 | frame #22: __libc_start_main + 0xe7 (0x7f2f2b53cb97 in /lib/x86_64-linux-gnu/libc.so.6)\r\nptcc_1 | frame #23: _start + 0x2a (0x5b250a in /usr/bin/python3)``` ", "I've also tried 3 different machines. All ubuntu 18.04, but with different GPUs sets. 2 Tesla V100-SXM2, 2 P100-SXM2, and 2 Tesla M40, but still get the same error.", "Can you install the repo from source and try again? There have been some issues with PyTorch upstream that Julien addressed here: https://github.com/huggingface/transformers/pull/4300. So you can try with the latest master branch.", "@BramVanroy, that merge request appears to have been merged prior to v2.10.0 release. I've installed both `v2.10.0` and `master` from source and unfortunately get the same error above when I tried to train a model distributed using mixed precision. ", "The one thing I can think of that you can try is specifically setting the current device for each process.\r\n\r\nCan you try cloning the library and installing in dev mode, and adding a line here:\r\n\r\nhttps://github.com/huggingface/transformers/blob/0866669e751bef636fa693b704a28c1fea9a17f3/examples/language-modeling/run_language_modeling.py#L134-L136\r\n\r\nSo that it looks like this:\r\n\r\n```python\r\n model_args, data_args, training_args = parser.parse_args_into_dataclasses()\r\n torch.cuda.set_device(training_args.device)\r\n if data_args.eval_data_file is None and training_args.do_eval:\r\n```\r\n\r\n", "Thanks @BramVanroy , you suggestion worked. I really appreciate it.", "Re-opening so that we can close this in a PR.", "@BramVanroy, while your suggestion works for multiple GPUs. I get the following error when trying to use a single GPU.\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/ptcc/run_language_modeling.py\", line 283, in <module>\r\n main()\r\n File \"/ptcc/run_language_modeling.py\", line 136, in main\r\n torch.cuda.set_device(training_args.device)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/cuda/__init__.py\", line 243, in set_device\r\n device = _get_device_index(device)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/cuda/_utils.py\", line 34, in _get_device_index\r\n 'or an integer, but got: '.format(device))\r\nValueError: Expected a cuda device with a specified index or an integer, but got:\r\n```\r\nand \r\n```\r\nTraceback (most recent call last):\r\n File \"/ptcc/run_glue.py\", line 230, in <module>\r\n main()\r\n File \"/ptcc/run_glue.py\", line 78, in main\r\n torch.cuda.set_device(training_args.device)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/cuda/__init__.py\", line 243, in set_device\r\n device = _get_device_index(device)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/cuda/_utils.py\", line 34, in _get_device_index\r\n 'or an integer, but got: '.format(device))\r\nValueError: Expected a cuda device with a specified index or an integer, but got:\r\n```", "@CMobley7 Thanks for the update! I pushed another update to my PR, can you try that one out? When we are not using DDP (and local_rank is -1), we do not specify the GPU id to use. It's best to strictly select that main device, so now we select it by using index 0. (This will still work if you set different devices with CUDA_VISIBLE_DEVICES, it'll just select the first device available _in that environment_).", "@BramVanroy , I can confirm that the changes made in https://github.com/huggingface/transformers/pull/4728 successfully fix the apex issues with both a single and multiple GPUs. I've tested on 3 different machines. All ubuntu 18.04, but with different GPUs sets. 2 Tesla V100-SXM2, 2 P100-SXM2, and 2 Tesla M40. Thanks for your help.", "Thank you @CMobley7 for the extensive testing, this is very valuable. \r\n\r\nAnd thanks @BramVanroy for fixing! " ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/4656?src=pr&el=h1) Report\n> Merging [#4656](https://codecov.io/gh/huggingface/transformers/pull/4656?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/76779363160a598f130433209a77f8a747351b61&el=desc) will **decrease** coverage by `0.03%`.\n> The diff coverage is `52.38%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/4656?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #4656 +/- ##\n==========================================\n- Coverage 77.38% 77.34% -0.04% \n==========================================\n Files 128 128 \n Lines 21071 21096 +25 \n==========================================\n+ Hits 16305 16316 +11 \n- Misses 4766 4780 +14 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/4656?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/trainer.py](https://codecov.io/gh/huggingface/transformers/pull/4656/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90cmFpbmVyLnB5) | `39.37% <48.27%> (-0.11%)` | :arrow_down: |\n| [src/transformers/modeling\\_electra.py](https://codecov.io/gh/huggingface/transformers/pull/4656/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19lbGVjdHJhLnB5) | `76.42% <100.00%> (ø)` | |\n| [src/transformers/trainer\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4656/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90cmFpbmVyX3V0aWxzLnB5) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/training\\_args.py](https://codecov.io/gh/huggingface/transformers/pull/4656/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90cmFpbmluZ19hcmdzLnB5) | `76.66% <100.00%> (+0.26%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/4656/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `88.66% <0.00%> (+0.16%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/4656?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/4656?src=pr&el=footer). Last update [7677936...e50654c](https://codecov.io/gh/huggingface/transformers/pull/4656?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hi @Frozenwords, thanks for your comments! These were introduced in the latest commit, reverting part of that commit now.", "It runs correctly now. If you run it, please let me know of the results!", "Hey @LysandreJik !\r\n\r\nI'm in the process of training an electra-small model and I'm having some issues. \r\n\r\nI'm running the `run_electra_pretraining` script with the following command:\r\n<img width=\"694\" alt=\"Screenshot 2020-06-08 at 19 11 24\" src=\"https://user-images.githubusercontent.com/32191917/84064906-bf372300-a9c3-11ea-9d61-68d1c9644e2a.png\">\r\n\r\nHere are my generator and discriminator configurations:\r\n<img width=\"291\" alt=\"Screenshot 2020-06-08 at 18 35 30\" src=\"https://user-images.githubusercontent.com/32191917/84064927-c78f5e00-a9c3-11ea-9feb-fe3ead8f5afc.png\">\r\n<img width=\"290\" alt=\"Screenshot 2020-06-08 at 18 35 07\" src=\"https://user-images.githubusercontent.com/32191917/84064946-ce1dd580-a9c3-11ea-9707-8ce65712291d.png\">\r\n\r\nAfter making the fixes I suggested the script is running just fine but I suspect either an error with my configurations/tokenizer setup or a case of silent failure. Indeed, the training loss quickly goes down but then slightly increases and plateaus after only 2000 training steps (which is quite different from the training curves shown in the original electra repository).\r\n\r\n<img width=\"992\" alt=\"Screenshot 2020-06-04 at 09 58 47\" src=\"https://user-images.githubusercontent.com/32191917/84065306-79c72580-a9c4-11ea-9e65-cf6c3a5ec470.png\">\r\n\r\nThe only changes I've made to the script are the following:\r\n- Replace the `OpenWebTextDataset` by a `LineByLineTextDataset` \r\n- Set the CLS token ID to 5 in line 424 since I'm using Camembert's tokenizer\r\n\r\nI've been through the script a few times and I can't seem to find the potential issue. Maybe you've observed that behaviour in your end-to-end tests ?\r\n\r\n\r\n", "Do you have any graph of your generator/discriminator accuracies?", "I don't have any validation metrics at the moment since I'm using a `LineByLineTextDataset`. Indeed, `_prediction_loop` in the `Trainer` concatenates predictions per batches of sequences (line 804). However in my case sequences have variable lengths (dynamic padding is done at the batch level by the `DataCollatorForLanguageModeling`) thus a variable number of tokens are being masked resulting in a mismatch of shapes. I believe a fix would be to flatten predictions before concatenation.", "Hey @LysandreJik, any news on the testing process ? If the model performs as expected on your end then there must be something wrong with my setup 🤔 ", "I can't reproduce yet either unfortunately, still trying to find out what's wrong.", "I am not sure if this is the right place to ask this question, so apologies in advance.\r\n\r\nwhy are position_ids fed twice here? \r\n1) [generator](https://github.com/huggingface/transformers/blob/e50654c03cd28e79bded1276774abe7572793a2c/examples/language-modeling/run_electra_pretraining.py#L434)\r\n2) [discriminator](https://github.com/huggingface/transformers/blob/e50654c03cd28e79bded1276774abe7572793a2c/examples/language-modeling/run_electra_pretraining.py#L457) \r\n\r\n```python\r\ngenerator_loss, generator_output = self.generator(\r\n masked_lm_inputs,\r\n attention_mask,\r\n token_type_ids,\r\n position_ids, # <--\r\n head_mask,\r\n position_ids, # <--\r\n masked_lm_labels=masked_lm_labels,\r\n )[:2]\r\n```\r\n", "Is this planned for the next release of transformers lib?", "@LysandreJik \r\nHello, I have a little confusion. In ELECTRA paper, word embedding, type embedding, position embedding are all shared. However in this pretraining code, it seems only word embedding shared. I'm not very sure, so is it a correct way to set the embedding?\r\nThank you.", "> It runs correctly now. If you run it, please let me know of the results!\r\n\r\nHi. I’m not sure if we should use the call masked_lm_inputs.scatter_(-1, replace_with_mask_positions, masked_tokens) or use the mask_fill method in line 425. It seems that the current version is using the scatter call, is it ok or we should switch to the mask_fill call as suggested by @Frozenwords ? Thanks!", "> Hey @LysandreJik !\r\n> \r\n> I'm in the process of training an electra-small model and I'm having some issues.\r\n\r\n@Frozenwords - any chance you could share the exact script/text you're running? I'd be happy to test this on my own dataset (sentences without spaces, requiring tokenization from scratch, followed by finetuning), but i'd like to make sure i'm using the right \"script\". (I haven't used HF much before, i'm a keras person :))", "# Training a small Electra on custom data\r\n\r\nI have changed my data to have a fixed length (128), and now everything seems to work. However, I do not have anything to compare with. I am training on a small ARM device (Xavier AGX), and it will take a few days before training is done and I can benchmark the model :)\r\n\r\n## Traning with variable length tensors\r\n\r\nI think that the current version can not support dynamically padded/variable length tensors. I suspect this is what was giving me issues with earlier runs.\r\n\r\nAt least the ```mask_inputs``` function would probably have to change a bit. As I mentioned above, the fake tokens will get scattered to the padding tokens. Moreover, when choosing how many tokens will be masked in ```mask_inputs```, a single number, ```number_of_tokens_to_be_masked``` is calculated, based on the longest tensor. I think this would need to vary along the batch dimension, and the sample probabilities would also need to vary.\r\n\r\n## Data\r\n\r\nI use my own data. It is around 18.5gb raw and 34gb in precomputed tensors. I precompute and save tensors using the ```tokenizers``` library, much like the OpenWebText data loader. The data has been transliterated to ASCII characters, with the exception of special danish letters. The tensors all have length 128. The tokenizer is the word piece tokenizer. Evaluation is run on the same 1024 tensors each time.\r\n\r\n## Changes to the code\r\n\r\nAside from the custom data loader, I have only changed line 424, to use the token_id for my [CLS] token, instead of 101.\r\n\r\n\r\n## Script parameters and model configs\r\n\r\n```\r\npython3 ~/nvme/elektra/transformers/examples/language-modeling/run_electra_pretraining.py \\\r\n --output_dir ./model_electra/models_dense_128 \\\r\n --logging_dir ./model_electra/logging \\\r\n --generator_config_name generator_config.json \\\r\n --discriminator_config_name discriminator_config.json \\\r\n --tokenizer_name ./tokenizer/ \\\r\n --do_train \\\r\n --do_eval \\\r\n --evaluate_during_training \\\r\n --max_eval_steps 10 \\\r\n --danish_corpus_directory ./features_dense_128 \\\r\n --overwrite_output_dir \\\r\n --block_size 128 \\\r\n --num_tensors_per_file 65536 \\\r\n --fp16 \\\r\n --seed 31 \\\r\n --max_steps -1 \\\r\n --logging_steps 100 \\\r\n --save_steps 32768 \\\r\n --save_total_limit 20 \\\r\n --learning_rate 5e-4 \\\r\n --adam_epsilon 1e-6 \\\r\n --per_device_train_batch_size=64 \\\r\n --per_device_eval_batch_size=64 \\\r\n --num_train_epochs=1 \\\r\n --warmup_steps 10000\r\n```\r\n\r\n## Discriminator config\r\n```\r\n{\r\n \"architectures\": [\r\n \"ElectraForPreTraining\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"embedding_size\": 128,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 256,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 1024,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"electra\",\r\n \"num_attention_heads\": 4,\r\n \"num_hidden_layers\": 12,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 46997\r\n}\r\n```\r\n\r\n## Generator config \r\n\r\n```\r\n{\r\n \"architectures\": [\r\n \"ElectraForMaskedLM\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"embedding_size\": 128,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 64,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 256,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"electra\",\r\n \"num_attention_heads\": 1,\r\n \"num_hidden_layers\": 12,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 46997\r\n}\r\n```\r\n\r\n## Graphs\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "@EmilLaursen I had same requirement, i.e. to have different number of masked tokens per sample in batch, because of huge variation in my (non-padded) sentence lengths. This is how I modified the electra pretraining model:\r\n(I am training on a non-NLP application, so I won't share any metrics or losses as they won't mean much)\r\n\r\n```python\r\nclass CombinedModel(nn.Module):\r\n\r\n ...\r\n \r\n def mask_inputs_by_row(\r\n self, input_ids: torch.Tensor, tokens_to_ignore, proposal_distribution=1.0,\r\n ):\r\n input_ids = input_ids.clone()\r\n inputs_which_can_be_masked = torch.ones_like(input_ids)\r\n for token in tokens_to_ignore:\r\n inputs_which_can_be_masked -= torch.eq(input_ids, token).long()\r\n\r\n total_number_of_tokens = input_ids.shape[-1]\r\n\r\n # Identify the number of tokens to be masked, which should be: 1 < num < max_predictions per seq.\r\n # It is set to be: n_tokens * mask_probability, but is truncated if it goes beyond bounds.\r\n num_mask_per_row = (inputs_which_can_be_masked.sum(dim=1) *\r\n self.mask_probability).type(torch.long)\r\n\r\n device = inputs_which_can_be_masked.device\r\n\r\n number_of_tokens_to_be_masked = torch.max(\r\n torch.tensor(1).to(device),\r\n torch.min(\r\n torch.min(\r\n torch.tensor(self.max_predictions_per_sequence,\r\n dtype=torch.long),\r\n torch.tensor(int(total_number_of_tokens *\r\n self.mask_probability), dtype=torch.long),\r\n ).to(device),\r\n num_mask_per_row\r\n )\r\n )\r\n\r\n # The probability of each token being masked\r\n sample_prob = proposal_distribution * inputs_which_can_be_masked\r\n sample_prob /= torch.sum(sample_prob, dim=1).view(-1, 1)\r\n # At this point each row should sum to 1. \r\n # i.e. all maskable tokens treated equally (equal opportunity)\r\n \r\n masked_lm_positions = torch.full_like(sample_prob, False).type(torch.bool)\r\n # Not sure if there is a way around using a for loop\r\n for i in range(sample_prob.size(0)):\r\n masked = sample_prob[i].multinomial(\r\n number_of_tokens_to_be_masked[i])\r\n masked_lm_positions[i, masked] = True\r\n \r\n return masked_lm_positions\r\n\r\n def forward(\r\n self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, labels=None,\r\n ):\r\n # get the masked positions as well as their original values\r\n masked_lm_positions = self.mask_inputs_by_row(\r\n input_ids, [self.tokenizer.cls_token_id,\r\n self.tokenizer.sep_token_id,\r\n self.tokenizer.mask_token_id,\r\n self.tokenizer.pad_token_id],\r\n )\r\n\r\n # masked_lm_ids = masked_lm_positions * input_ids\r\n\r\n # # Of the evaluated tokens, 15% of those will keep their original tokens\r\n # replace_with_mask_positions = masked_lm_positions * (\r\n # torch.rand(masked_lm_positions.shape, device=masked_lm_positions.device) < (\r\n # 1 - self.mask_probability)\r\n # )\r\n\r\n masked_lm_inputs = input_ids.clone()\r\n # use a bool mask of positions we want to MASKed to insert MASK token ids\r\n # bool_masked_positions = (masked_lm_positions > 0)\r\n masked_lm_inputs[masked_lm_positions] = self.tokenizer.mask_token_id\r\n\r\n # create MASKED labels with real token ids in positions where the MASKed tokens\r\n # were inserted, -100 otherwise \r\n masked_lm_labels = torch.full_like(input_ids, -100)\r\n masked_lm_labels[masked_lm_positions] = input_ids[masked_lm_positions]\r\n\r\n generator_loss, generator_output = self.generator(\r\n masked_lm_inputs,\r\n attention_mask,\r\n token_type_ids,\r\n position_ids,\r\n head_mask,\r\n None, # position_ids,\r\n masked_lm_labels=masked_lm_labels,\r\n )[:2]\r\n\r\n # softmax the predictions\r\n fake_softmaxed = torch.softmax(generator_output, dim=-1)\r\n # At this point if we sum 3rd dim, we should get a tensor of ones, the same size as input\r\n # i.e. fake_softmaxed.sum(dim=2) == torch.ones_like(input_ids)\r\n\r\n # for each position in sentence, sample ONE token from the generator probability distribution\r\n # this is why the 3rd dim is ONE.\r\n fake_sampled = torch.zeros_like(input_ids).view(\r\n input_ids.shape[0], input_ids.shape[1], 1)\r\n\r\n # multinomial cannot be applid to 3d array. so loop over examples\r\n # what we are doing here is for each position in a sentence,\r\n # we will sample a token using Generator's learned probability \r\n # distribution.\r\n for i in range(fake_softmaxed.shape[0]):\r\n fake_sampled[i] = fake_softmaxed[i,:,:].multinomial(1)\r\n\r\n # At this point we have generator samples for ALL the positions in the sentence.\r\n # But we only need the predictions for the positions corresponding to MASKED tokens\r\n\r\n # First, align shape with the input. Get rid of 3rd dim which was created to make\r\n # sampling easier\r\n fake_sampled = fake_sampled.view(input_ids.shape[0], input_ids.shape[1])\r\n\r\n # Discriminator input is same as generator, except instead of masked tokens\r\n # we insert tokens sampled from the generator distribution.\r\n fake_tokens = input_ids.clone()\r\n fake_tokens[masked_lm_positions] = fake_sampled[masked_lm_positions]\r\n\r\n # D labels are binary labels indicating whether the \r\n discriminator_labels = (labels != fake_tokens).int()\r\n\r\n discriminator_loss, discriminator_output = self.discriminator(\r\n fake_tokens,\r\n attention_mask,\r\n token_type_ids,\r\n position_ids,\r\n head_mask,\r\n None, # position_ids,\r\n labels=discriminator_labels,\r\n )[:2]\r\n\r\n discriminator_predictions = torch.round(\r\n (torch.sign(discriminator_output) + 1.0) * 0.5)\r\n\r\n total_loss = (self.discriminator_weight * discriminator_loss) + \\\r\n (self.generator_weight * generator_loss)\r\n\r\n\r\n # For evaluation, pass tensors of masked tokens and sampled tokens\r\n masked_input_ids = input_ids[masked_lm_positions]\r\n fake_sampled_ids = fake_sampled[masked_lm_positions]\r\n\r\n return (\r\n total_loss,\r\n (generator_output, discriminator_output),\r\n (masked_input_ids, fake_sampled_ids),\r\n (discriminator_labels, discriminator_predictions),\r\n )\r\n\r\n```", "@EmilLaursen \r\nCan you train your model on glue to see the dev accuracy ? (especially matthew correlation for CoLA task)\r\nI used another training script and found even train loss 1x may still result in very bad glue accuracy, so I am wondering if your model score good acc on GLUE. ", "@richarddwang \r\nMy model is trained on danish text, so I suspect the glue score would be terrible. I have assembled my own danish benchmark suite, with 3 objectives: pos-tag, NER, and text classification.\r\n\r\nFor what it is worth, my model scores about 4-5% F1 lower than BERT multilingual and the Danish BERT model (uncased and same size as bert-base) on the NER task (0.845 vs 0.795). On the pos tagging task, it is about 0.5 F1 lower, (0.977 vs 0.972). I have not tried the text classification task yet.\r\n\r\nTo me, this seems comparable with what is stated in the Electra paper, i.e. glue score about 4-5 points lower on the small model compared to the base-sized models.", "Thanks! It is very kind of you to share such detailed results.", "@LysandreJik This would be very helpful! Is there any plan to get this merged soon?", "@LysandreJik I and Phil have implemented and verified Electra on the Glue downstream task.\r\n\r\nhttps://github.com/lucidrains/electra-pytorch\r\n\r\nThe forward pass is based on your replica of the TF forward pass (HF 3.0.2).\r\n\r\nThe remaining code is written to closely replicate the TF reference code.\r\n\r\nThat is, data-preprocessing (including ExampleBuilder), learning rate schedule, gradient clipping, etc. might differ.\r\n\r\nI believe the two main difference between this PR and our code might be:\r\n\r\n(1) For the \"small\" setting, for the generator, there is a discrepancy of the configuration reported in the paper and used in the TF reference code.\r\n(2) Data preprocessing with example builder.\r\n\r\nFor the TF reference, after 1M updates, the Glue MRPC accuracy is ~87%. Note, there is high variance in these measurements.\r\n\r\nFor our code, after 200k updates, the Glue MRPC accuracy is ~82%, which might approach ~87% accuracy after 1M updates.", "Hi @enijkamp , nick work !\r\nCould you share both dev and test score for every task in GLUE ?\r\nIt will help a lot, thanks !", "Hi, can I use `electra-trainer` branch to run pre-trianing, then save the model `checkpoint`.\r\n\r\nIs there problem to use the saved `checkpoint ` in the latest master branch?\r\n\r\nThanks~", "Hi @LysandreJik @enijkamp and all !\r\n\r\nAfter develop and debug for a long time. My implementaion of ELECTRA training and finetuning finally successfully pretrains a model from scratch and replicates the results in the paper.\r\n\r\n|Model|CoLA|SST|MRPC|STS|QQP|MNLI|QNLI|RTE|Avg.|\r\n|---|---|---|---|---|---|---|---|---|---|\r\n|ELECTRA-Small|54.6|89.1|83.7|80.3|88.0|79.7|87.7|60.8|78.0|\r\n|ELECTRA-Small (electra_pytorch)|57.2|87.1|82.1|80.4|88|78.9|87.9|63.1|78.08\r\n\r\n💻Code: https://github.com/richarddwang/electra_pytorch\r\n📰Post: https://discuss.huggingface.co/t/electra-training-reimplementation-and-discussion/1004\r\n🐦Tweet: https://twitter.com/_RichardWang_/status/1302545459374772224?s=20 \r\n\r\nI've listed details that are easy to be overlooked when reimplementing ELECTRA, includes a bug in the original official implementation. I would be glad if this helps 😊", "Hey @LysandreJik and all\r\nI was wondering if you have tried to run this with pytorch 1.6 ? I am currently on a device, where i'd have to reinstall from scratch to downgrade to a lower version of pytorch. I am getting some strange result, and I am considering if it is a compatibility issue, after trying with different learning rates, tokenizers and batch sizes. It seems that my generator does not learn anything, and does not converge. I used fixed-length tensors as @EmilLaursen of size 256. My discriminator does seem to converge, however performs poorly on downstream tasks (Danish NER and POSTAG tasks), which is suspect is caused by the generator not converging.\r\n### Update (18 October)\r\nI can confirm that this behavior is **not** observed when using PyTorch 1.5.0. Perhaps it has something to do with PyTorch \"Automatic Mixed Precision\" feature released in version 1.6. If anyone else is experiencing the same issue, then i recommend you use PyTorch < 1.6.\r\n### Changes to code\r\nCustom data loader (danish corpus), and changed the [CLS] token in line 426 to the one of my vocab.\r\n### Configs\r\n#### Discriminator\r\n```json \r\n{\r\n \"architectures\": [\r\n \"ElectraForPreTraining\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"embedding_size\": 128,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 256,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 1024,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"electra\",\r\n \"num_attention_heads\": 4,\r\n \"num_hidden_layers\": 12,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 45000\r\n}\r\n```\r\n#### Generator\r\n```json\r\n{\r\n \"architectures\": [\r\n \"ElectraForMaskedLM\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"embedding_size\": 128,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 64,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 256,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"electra\",\r\n \"num_attention_heads\": 1,\r\n \"num_hidden_layers\": 12,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 45000\r\n}\r\n```\r\n#### Run config\r\n```shell\r\npython3 pretraining/run_electra_pretraining.py \\\r\n\t--output_dir ./models/electra/small_256 \\\r\n\t--discriminator_config_name ./pretraining/config/discriminator_config.json \\\r\n\t--generator_config_name ./pretraining/config/generator_config.json \\\r\n\t--tokenizer_name ./models/tokenizers/ \\\r\n\t--do_train \\\r\n\t--do_eval \\\r\n\t--evaluate_during_training \\\r\n\t--max_eval_steps 16 \\\r\n\t--danish_feature_directory ./data/features_dense_256 \\\r\n\t--overwrite_output_dir \\\r\n\t--block_size 256 \\\r\n\t--num_tensors_per_file 65536 \\\r\n\t--fp16 \\\r\n\t--seed 1337 \\\r\n\t--max_steps -1 \\\r\n\t--logging_steps 200 \\\r\n\t--save_steps 20000 \\\r\n\t--save_total_limit 20 \\\r\n\t--learning_rate 2.5e-4 \\\r\n\t--adam_epsilon 1e-6 \\\r\n\t--per_device_train_batch_size=64 \\\r\n\t--per_device_eval_batch_size=64 \\\r\n\t--num_train_epochs=1 \\\r\n\t--warmup_steps 10000\r\n```\r\n### Logged metrics\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n", "I am wondering when will \"electra training from scratch\" feature be released? ", "This PR is unfortunately in a stale state, with no projects to work on it further in the near future. You can take a look at this discussion: https://discuss.huggingface.co/t/electra-training-reimplementation-and-discussion/1004 or at the above comment by @richarddwang for a PyTorch implementation of the ELECTRA training from scratch.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "If you are using `pytorch-lightning` then you won't need to transfer data on GPU or TPU manually. `lightning` takes care of that for you", "I have copied the method, which transfer data from pytorch-lightning, so you can reproduce the error.", "Okay, sorry I misunderstood the question.", "It is my bad English, I guess. Do you need more explanation, or the question is clear now?", "The reason is in latest version `batch_encode_plus` returns an instance of `BatchEncoding` and in 2.8.0 it return a `dict`. So you can just do it like this\r\nin your collate function\r\n```\r\nreturn dict(tokens), torch.tensor(labels, dtype=torch.long)\r\n```\r\n\r\nI think tokenizer should handle it itself, so tagging @mfuntowicz ", "BatchEncoding is indeed a UserDict, if you want to access the actual dict, you can use the data attribute:\n\n```python\nbe = tokenizer.batch_encode_plus(...)\nbe.data\n```", "Thank you! So it's not a bug, but an expected behaviour" ]

[]

[]

[ "@LysandreJik this should work (assuming tf >=2.2.0) let me know if I can do anything to help.", "Hello !\r\n\r\nThanks a lot for this PR!! Can you rebase on master plz? It will be easier to review :)", "@jplu (I think/hope) I just did it. Let me know if everything is ok. Great work by the way, congrats to all of Huggingface team!", "Well no, you did a merge not a rebase :smile: can you revert your merge with the following command line:\r\n\r\n```\r\ngit reset --hard e4741ef\r\ngit fetch upstream\r\ngit pull --rebase upstream master\r\ngit push --force\r\n```\r\n\r\nAlso be careful because the tests are broken.\r\n", ":blush: yes sorry, I had done all the above but with a commit before the final push. Now it should (hopefully) be ok. I saw the tests are failing, but I was able to successfully use electra for some text classification tasks of interest.", "Awesome!! Thanks for having re-pushed ^^\r\n\r\nNow there are several things to change, the output of the `call` method should be like the PyTorch one, it means `(loss), logits, (hidden_states), (attentions)`. Can you add the following updates:\r\n\r\n1) just before the return add:\r\n```python\r\nif labels is not None:\r\n loss = self.compute_loss(labels, logits)\r\n outputs = (loss,) + outputs\r\n```\r\n\r\n2) add the `labels` parameter to the method.\r\n\r\n3) The `TFElectraForSequenceClassification` should inherit from the `TFSequenceClassificationLoss` class.", "Sorry my bad, the signature of the `call` method should look like this:\r\n```python\r\ndef call(\r\n self,\r\n input_ids=None,\r\n attention_mask=None,\r\n token_type_ids=None,\r\n position_ids=None,\r\n head_mask=None,\r\n inputs_embeds=None,\r\n labels=None,\r\n training=False,\r\n ):\r\n```", "I also changed the call to self.electra", "Here are the errors raised by the tests:\r\n\r\n```\r\n=================================== FAILURES ===================================\r\n__________________ TFElectraModelTest.test_attention_outputs ___________________\r\n[gw0] linux -- Python 3.7.7 /usr/local/bin/python\r\n\r\nself = <tests.test_modeling_tf_electra.TFElectraModelTest testMethod=test_attention_outputs>\r\n\r\n def test_attention_outputs(self):\r\n config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()\r\n \r\n decoder_seq_length = (\r\n self.model_tester.decoder_seq_length\r\n if hasattr(self.model_tester, \"decoder_seq_length\")\r\n else self.model_tester.seq_length\r\n )\r\n encoder_seq_length = (\r\n self.model_tester.encoder_seq_length\r\n if hasattr(self.model_tester, \"encoder_seq_length\")\r\n else self.model_tester.seq_length\r\n )\r\n decoder_key_length = (\r\n self.model_tester.key_length if hasattr(self.model_tester, \"key_length\") else decoder_seq_length\r\n )\r\n encoder_key_length = (\r\n self.model_tester.key_length if hasattr(self.model_tester, \"key_length\") else encoder_seq_length\r\n )\r\n \r\n for model_class in self.all_model_classes:\r\n config.output_attentions = True\r\n config.output_hidden_states = False\r\n model = model_class(config)\r\n outputs = model(inputs_dict)\r\n attentions = [t.numpy() for t in outputs[-1]]\r\n self.assertEqual(model.config.output_attentions, True)\r\n self.assertEqual(model.config.output_hidden_states, False)\r\n> self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)\r\nE AssertionError: 13 != 5\r\n\r\ntests/test_modeling_tf_common.py:324: AssertionError\r\n_________________ TFElectraModelTest.test_hidden_states_output _________________\r\n[gw0] linux -- Python 3.7.7 /usr/local/bin/python\r\n\r\nself = <tests.test_modeling_tf_electra.TFElectraModelTest testMethod=test_hidden_states_output>\r\n\r\n def test_hidden_states_output(self):\r\n config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()\r\n \r\n for model_class in self.all_model_classes:\r\n config.output_hidden_states = True\r\n config.output_attentions = False\r\n model = model_class(config)\r\n outputs = model(inputs_dict)\r\n hidden_states = [t.numpy() for t in outputs[-1]]\r\n self.assertEqual(model.config.output_attentions, False)\r\n self.assertEqual(model.config.output_hidden_states, True)\r\n> self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)\r\nE AssertionError: 13 != 6\r\n\r\ntests/test_modeling_tf_common.py:369: AssertionError\r\n______________________ TFElectraModelTest.test_save_load _______________________\r\n[gw0] linux -- Python 3.7.7 /usr/local/bin/python\r\n\r\nself = <tests.test_modeling_tf_electra.TFElectraModelTest testMethod=test_save_load>\r\n\r\n def test_save_load(self):\r\n config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()\r\n \r\n for model_class in self.all_model_classes:\r\n model = model_class(config)\r\n outputs = model(inputs_dict)\r\n \r\n with tempfile.TemporaryDirectory() as tmpdirname:\r\n model.save_pretrained(tmpdirname)\r\n model = model_class.from_pretrained(tmpdirname)\r\n after_outputs = model(inputs_dict)\r\n \r\n> self.assert_outputs_same(after_outputs, outputs)\r\n\r\ntests/test_modeling_tf_common.py:93: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\ntests/test_modeling_tf_common.py:154: in assert_outputs_same\r\n self.assertLessEqual(max_diff, 1e-5)\r\nE AssertionError: 0.24899402 not less than or equal to 1e-05\r\n```", "Hello! Any news on this PR? :)", "> Hello! Any news on this PR? :)\r\n\r\nHi, sorry I am a bit pressed at the moment, I'd be glad if you would want to take over, otherwise it might take some time for me to re-grab this.", "Ok, no problem I will try to retake what you have done. Thanks a lot for the update.", "Any news on this PR @ypapanik @jplu ? Can I help in some way ?", "Sorry, no time on my side to work on this for now.", "@maxibor The code should be working I have used it successfully to get a better acc than BERT, but some tests had failed. Meanwhile the main codebase has evolved and there should be some conflicts, easy to resolve probably. \r\nI don't have time to complete those two things now (tests failing, resolve conflicts), perhaps someone from HF should spare a few minutes? @LysandreJik ", "Updated version right here https://github.com/huggingface/transformers/pull/6227", "Thanks @jplu for opening the new PR, closing this." ]