Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[ { "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 } ]

[ "Reopened the previous issue: https://github.com/huggingface/transformers/issues/789 and will take a look next week :-) ", "Closing this as well. Reasons are explained in #789." ]

[]

[ { "login": "TevenLeScao", "id": 26709476, "node_id": "MDQ6VXNlcjI2NzA5NDc2", "avatar_url": "https://avatars.githubusercontent.com/u/26709476?v=4", "gravatar_id": "", "url": "https://api.github.com/users/TevenLeScao", "html_url": "https://github.com/TevenLeScao", "followers_url": "https://api.github.com/users/TevenLeScao/followers", "following_url": "https://api.github.com/users/TevenLeScao/following{/other_user}", "gists_url": "https://api.github.com/users/TevenLeScao/gists{/gist_id}", "starred_url": "https://api.github.com/users/TevenLeScao/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/TevenLeScao/subscriptions", "organizations_url": "https://api.github.com/users/TevenLeScao/orgs", "repos_url": "https://api.github.com/users/TevenLeScao/repos", "events_url": "https://api.github.com/users/TevenLeScao/events{/privacy}", "received_events_url": "https://api.github.com/users/TevenLeScao/received_events", "type": "User", "site_admin": false } ]

[]

[ { "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": 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/3709?src=pr&el=h1) Report\n> Merging [#3709](https://codecov.io/gh/huggingface/transformers/pull/3709?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/6435b9f908e7361330db89e263a65b0a58060d11&el=desc) will **decrease** coverage by `0.00%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3709?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3709 +/- ##\n==========================================\n- Coverage 78.13% 78.12% -0.01% \n==========================================\n Files 104 104 \n Lines 17723 17723 \n==========================================\n- Hits 13847 13846 -1 \n- Misses 3876 3877 +1 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3709?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3709/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.79% <0.00%> (-0.17%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3709?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/3709?src=pr&el=footer). Last update [6435b9f...be477cf](https://codecov.io/gh/huggingface/transformers/pull/3709?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" } ]

[]

[]

[]

[ { "login": "mfuntowicz", "id": 2241520, "node_id": "MDQ6VXNlcjIyNDE1MjA=", "avatar_url": "https://avatars.githubusercontent.com/u/2241520?v=4", "gravatar_id": "", "url": "https://api.github.com/users/mfuntowicz", "html_url": "https://github.com/mfuntowicz", "followers_url": "https://api.github.com/users/mfuntowicz/followers", "following_url": "https://api.github.com/users/mfuntowicz/following{/other_user}", "gists_url": "https://api.github.com/users/mfuntowicz/gists{/gist_id}", "starred_url": "https://api.github.com/users/mfuntowicz/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/mfuntowicz/subscriptions", "organizations_url": "https://api.github.com/users/mfuntowicz/orgs", "repos_url": "https://api.github.com/users/mfuntowicz/repos", "events_url": "https://api.github.com/users/mfuntowicz/events{/privacy}", "received_events_url": "https://api.github.com/users/mfuntowicz/received_events", "type": "User", "site_admin": false } ]

[ "Are those `p3.16xlarge` is the same AZ? Even in the same AZ, the throughput and latency between machines might be the bottleneck here.", "i.e. you usually need infiniband between cluster machines. But @mfuntowicz knows about this stuff way better than I do...", "Thanks for your reply. The two hosts are indeed in the same AZ. \r\n\r\nHere is how i run them:\r\n```\r\n# Node 1:\r\npython3.6 -m torch.distributed.launch \\\r\n--nproc_per_node=8 \\\r\n--nnodes=2 \\\r\n--node_rank=0 \\\r\n--master_addr=\"10.0.0.83\" \\\r\n--master_port=12345 \\\r\nrun_lm_finetuning.py \\\r\n --output_dir=output \\\r\n --model_type=roberta \\\r\n --model_name_or_path=roberta-base \\\r\n --do_train \\\r\n --train_data_file=$TRAIN_FILE \\\r\n --do_eval \\\r\n --eval_data_file=$TEST_FILE \\\r\n --mlm\r\n\r\n# Node 2 \r\npython3.6 -m torch.distributed.launch \\\r\n--nproc_per_node=8 \\\r\n--nnodes=2 \\\r\n--node_rank=1 \\\r\n--master_addr=\"10.0.0.83\" \\\r\n--master_port=12345 \\\r\nrun_lm_finetuning.py \\\r\n --output_dir=output \\\r\n --model_type=roberta \\\r\n --model_name_or_path=roberta-base \\\r\n --do_train \\\r\n --train_data_file=$TRAIN_FILE \\\r\n --do_eval \\\r\n --eval_data_file=$TEST_FILE \\\r\n --mlm\r\n```\r\nI also have the nccl debug info here:\r\nThe first part is on multiple nodes, where the training is slow. The second part is on single node, and the training is fast. I can definitely see that on single node, there are many Channels, which can't be found on multiple node. \r\nhttps://gist.github.com/YingleiZhang/a8df48eb534ba20ff8f26b5309094b55\r\n\r\nI was also suspecting that i might need high speed connections (link infiniband) between cluster machines, but in this case, would MPI help? My PyTorch did not built with MPI yet. ", "I think the bandwidth between two different nodes are indeed the problem. Consider the amount of data (Our estimation is about 8G for each step) we need to move between different nodes, and the bandwidth for intra-gpu communication is much higher than that of inter-node communication. NCCL folks mentioned that this could be 120 GB/s vs 10 GB/s for all reduce operation. (See it here https://github.com/NVIDIA/nccl/issues/318)\r\n\r\nI am closing this issue here. Thanks for the help. ", "Thanks for investigating.\r\n\r\nAlso from your logs @mfuntowicz was saying that it looks like NCCL does not use [AWS EFA](https://aws.amazon.com/hpc/efa/) – maybe something to investigate there." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/3706?src=pr&el=h1) Report\n> Merging [#3706](https://codecov.io/gh/huggingface/transformers/pull/3706?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/f0c96fafd16d206b22a74fe76b251414f7314703&el=desc) will **decrease** coverage by `0.82%`.\n> The diff coverage is `90.64%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3706?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3706 +/- ##\n==========================================\n- Coverage 78.47% 77.65% -0.83% \n==========================================\n Files 106 106 \n Lines 17930 17904 -26 \n==========================================\n- Hits 14071 13903 -168 \n- Misses 3859 4001 +142 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3706?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYWxiZXJ0LnB5) | `89.42% <ø> (-0.20%)` | :arrow_down: |\n| [src/transformers/tokenization\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYmVydC5weQ==) | `95.29% <ø> (-0.04%)` | :arrow_down: |\n| [src/transformers/tokenization\\_bert\\_japanese.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYmVydF9qYXBhbmVzZS5weQ==) | `67.07% <ø> (-0.79%)` | :arrow_down: |\n| [src/transformers/tokenization\\_camembert.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fY2FtZW1iZXJ0LnB5) | `36.25% <ø> (+0.88%)` | :arrow_up: |\n| [src/transformers/tokenization\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fY3RybC5weQ==) | `96.11% <ø> (-0.08%)` | :arrow_down: |\n| [src/transformers/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fZ3B0Mi5weQ==) | `96.82% <ø> (-0.05%)` | :arrow_down: |\n| [src/transformers/tokenization\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdDUucHk=) | `95.71% <ø> (-0.12%)` | :arrow_down: |\n| [src/transformers/tokenization\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdHJhbnNmb194bC5weQ==) | `40.67% <ø> (-0.29%)` | :arrow_down: |\n| [src/transformers/tokenization\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25feGxtLnB5) | `83.33% <ø> (-0.14%)` | :arrow_down: |\n| [src/transformers/tokenization\\_xlm\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25feGxtX3JvYmVydGEucHk=) | `97.61% <ø> (-0.06%)` | :arrow_down: |\n| ... and [10 more](https://codecov.io/gh/huggingface/transformers/pull/3706/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3706?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/3706?src=pr&el=footer). Last update [f0c96fa...5b54450](https://codecov.io/gh/huggingface/transformers/pull/3706?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Ok here is a new version @mfuntowicz.\r\nI'll add some tests later before merging." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/3705?src=pr&el=h1) Report\n> Merging [#3705](https://codecov.io/gh/huggingface/transformers/pull/3705?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/bc65afc4dfac3badf3de3be395d4023b44c61bdd&el=desc) will **decrease** coverage by `0.11%`.\n> The diff coverage is `50.00%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3705?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3705 +/- ##\n==========================================\n- Coverage 78.14% 78.02% -0.12% \n==========================================\n Files 104 104 \n Lines 17723 17710 -13 \n==========================================\n- Hits 13849 13818 -31 \n- Misses 3874 3892 +18 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3705?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fb3BlbmFpLnB5) | `82.50% <0.00%> (ø)` | |\n| [src/transformers/tokenization\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdHJhbnNmb194bC5weQ==) | `40.67% <100.00%> (-0.29%)` | :arrow_down: |\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `85.76% <0.00%> (-1.77%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_flaubert.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9mbGF1YmVydC5weQ==) | `21.32% <0.00%> (-0.58%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9iZXJ0LnB5) | `96.07% <0.00%> (-0.47%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.96% <0.00%> (-0.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `91.84% <0.00%> (-0.13%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9hbGJlcnQucHk=) | `85.53% <0.00%> (-0.05%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl94bG0ucHk=) | `90.40% <0.00%> (-0.03%)` | :arrow_down: |\n| ... and [7 more](https://codecov.io/gh/huggingface/transformers/pull/3705/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3705?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/3705?src=pr&el=footer). Last update [bc65afc...e31554a](https://codecov.io/gh/huggingface/transformers/pull/3705?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 } ]

[]

[ "Hi! \r\n\r\n- 1 - casing is the difference between lowercasing and uppercasing. Uncased models do not handle uppercase letters, and therefore lowercase them:\r\n\r\n```py\r\nfrom transformers import AutoTokenizer\r\n\r\nuncased_tokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\r\ncased_tokenizer = AutoTokenizer.from_pretrained(\"bert-base-cased\")\r\n\r\nprint(uncased_tokenizer.tokenize(\"Hi, this is Lysandre\"))\r\n# ['hi', ',', 'this', 'is', 'l', '##ys', '##and', '##re'] <-- notice how uppercase letters are now lowercased\r\n\r\nprint(cased_tokenizer.tokenize(\"Hi, this is Lysandre\"))\r\n# ['Hi', ',', 'this', 'is', 'L', '##ys', '##and', '##re']\r\n```\r\n\r\n- 2 - These should be clarified with model cards on the [model hub](https://huggingface.co/models) but we haven't gotten to changing them yet. \r\n\r\nXLM models are usually multilingual, which is the case for those you mentioned: `ende` means english-german, `enfr`, english-french, `xnli15` means the 15 languages that are used in [XNLI](https://www.nyu.edu/projects/bowman/xnli/).\r\n\r\nThe following number is the hidden size, e.g. `1024` means that the hidden size of the model is 1024.\r\n\r\n- 3 - You may useT5 for sentiment classification, ELECTRA as well but with a bit more additional work. \r\n\r\nAs @craffel said in the issue you mentioned, T5 was trained with SST-2 so should work out-of-the-box if you follow what he mentioned in this issue.\r\n\r\nThere is no current `ElectraForSequenceClassification` as ELECTRA is so new, but it will certainly make its way in the library in the coming weeks! Once this head is here (feel free to add it yourself, it would be as easy as copying one head from one other modeling file and putting it for ELECTTRA), ELECTRA can be used for sentiment classification, but it would require you to fine-tune it first to a sentiment classification dataset (like the SST-2 dataset). \r\n\r\nIf you're looking at easy sentiment classification, please take a look at the pipelines and at the [already-finetuned sequence classification models](https://huggingface.co/models?filter=text-classification) and look for sentiment classification especially.", "@LysandreJik thanks, it was helpful 🙂", "Hi, it is easy to use the pre-trained T5 models for sentiment ID. You could do something like\r\n```Python\r\nMODEL_NAME = \"t5-base\"\r\nmodel = transformers.T5ForConditionalGeneration.from_pretrained(MODEL_NAME)\r\ntokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)\r\ninput_text = \"sst2 sentence: This movie was great! I loved the acting.\"\r\ninputs = tokenizer.encode_plus(input_text, return_token_type_ids=False, return_tensors=\"pt\")\r\nprint(tokenizer.decode(model.generate(**inputs)[0]))\r\ninput_text = \"sst2 sentence: The acting was so bad in this movie I left immediately.\"\r\ninputs = tokenizer.encode_plus(input_text, return_token_type_ids=False, return_tensors=\"pt\")\r\nprint(tokenizer.decode(model.generate(**inputs)[0]))\r\n```\r\nThe `\"sst2 sentence:\"` prefix is what we used for the SST-2 task. It is a sentiment ID task. The model needs to see this prefix to know what task you want it to undertake.", "Hi, @craffel Thank for your quick response and the intuitive code snippet. As I said, I am trying to implement **T5** for a `binary sentiment classification` task (label as `1` and `0`). So, if I want to use **T5**, I've to treat my task as a **text-to-text**, in other words, `positive` and `negative`. But I feel a bit confused, if I have the following scenario how should I approach. \r\n\r\n## Model loading\r\n```python\r\nMODEL_NAME = \"t5-base\"\r\ntransformer_layer = transformers.T5ForConditionalGeneration.from_pretrained(MODEL_NAME)\r\ntokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)\r\n```\r\n\r\n## A general encoder \r\n```python\r\ndef regular_encode(texts, tokenizer, maxlen=512):\r\n enc_di = tokenizer.batch_encode_plus(\r\n texts, \r\n return_attention_masks=False, \r\n return_token_type_ids=False,\r\n pad_to_max_length=True,\r\n max_length=maxlen\r\n )\r\n return np.array(enc_di['input_ids'])\r\n```\r\n\r\n## Build the model (as per my task)\r\n```python\r\ndef build_model(transformer, max_len=190):\r\n input_word_ids = Input(shape=(max_len,), dtype=tf.int32)\r\n sequence_output = transformer(input_word_ids)[0]\r\n cls_token = sequence_output[:, 0, :]\r\n out = Dense(1, activation='sigmoid')(cls_token)\r\n model = Model(inputs=input_word_ids, outputs=out)\r\n model.compile(Adam(lr=1e-5), loss='binary_crossentropy', metrics=['accuracy'])\r\n\r\n return model\r\n```\r\n\r\n## Tokenized the data and grab the targets int(1,0)\r\n```python\r\nx_train = regular_encode(data.text, tokenizer, maxlen=190)\r\ny_train = data.target.values # (0, 1)\r\nmodel = build_model(transformer_layer, max_len=190)\r\nmodel.fit...\r\nmodel.predict...\r\n```\r\nI sure I'm missing something crucial part that is not considering `text-to-text` manner. If I convert `1` and `0` of labels as `Positive` and `Negative`...I mean shouldn't the target need to be numeric! And about the prefix, `sst2 sentence:` so, this is, in other words, is a string indicator to inform the model about the goal or task. So, do I have to add this string at the beginning of every text sentence or (samples)?", "> I sure I'm missing something crucial part that is not considering text-to-text manner. If I convert 1 and 0 of labels as Positive and Negative...I mean shouldn't the target need to be numeric!\r\n\r\nNo, the target should *always* be text for T5. You should map your 0/1 labels to the words \"negative\" and \"positive\" and fine-tune T5 to predict those words, and then map them back to 0/1 after the model outputs the text if needed. This is the point of the text-to-text framework - all tasks take text as input and produce text as output. So, for example, your \"build model\" code should not include a dense layer with a sigmoid output, etc. There is no modification to the model structure necessary whatsoever.\r\n\r\n> And about the prefix, sst2 sentence: so, this is, in other words, is a string indicator to inform the model about the goal or task. So, do I have to add this string at the beginning of every text sentence or (samples)?\r\n\r\nYes, that is the intention.", "@LysandreJik @craffel \r\nPlease check this issue!\r\nAs per the discussion I have a similar approach on binary classification on the text. But it seems that I am doing something wrong. I have also converted the target 0 and 1 to \"0\" and \"1\". Don't know where I am getting wrong.\r\n```\r\nMODEL_NAME = \"t5-base\"\r\ntransformer_layer = transformers.TFT5ForConditionalGeneration.from_pretrained(MODEL_NAME)\r\ntokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME)\r\n```\r\n\r\n```\r\ndef regular_encode(texts, tokenizer, maxlen=512):\r\n enc_di = tokenizer.batch_encode_plus(\r\n texts, \r\n return_attention_masks=False, \r\n return_token_type_ids=False,\r\n pad_to_max_length=True,\r\n max_length=maxlen\r\n )\r\n return np.array(enc_di['input_ids'])\r\n```\r\n```\r\ndef build_model(transformer, max_len=190):\r\n input_word_ids = Input(shape=(max_len,), dtype=tf.int32)\r\n sequence_output = transformer(input_word_ids)[0]\r\n cls_token = sequence_output[:, 0, :]\r\n out = Dense(1, activation='sigmoid')(cls_token)\r\n model = Model(inputs=input_word_ids, outputs=out)\r\n model.compile(Adam(lr=1e-5), loss='binary_crossentropy', metrics=['accuracy'])\r\n\r\n return model\r\n\r\n```\r\n\r\n\r\n```\r\nx_train = regular_encode(train_df.new_text, tokenizer, maxlen=190)\r\ny_train = train_df.target.values # (0, 1) 0 and 1 convert to string\r\nmodel = build_model(transformer_layer, max_len=190)\r\n```\r\n```\r\nValueError: in converted code:\r\n\r\n /opt/conda/lib/python3.6/site-packages/transformers/modeling_tf_t5.py:854 call *\r\n encoder_outputs = self.encoder(\r\n /opt/conda/lib/python3.6/site-packages/tensorflow_core/python/keras/engine/base_layer.py:822 __call__\r\n outputs = self.call(cast_inputs, *args, **kwargs)\r\n /opt/conda/lib/python3.6/site-packages/transformers/modeling_tf_t5.py:445 call\r\n raise ValueError(\"You have to specify either input_ids or inputs_embeds\")\r\n\r\n ValueError: You have to specify either input_ids or inputs_embeds\r\n```\r\nAll inputs are converted to this format\r\n\"sst2 sentence: our deeds are the reason for this...\"\r\nI used the same things but having trouble with this error. I need to fine-tune the model on my custom dataset.", "Hi @vapyc, this seems to be an unrelated issue. Would you mind opening a new issue? When you do, would it be possible for you to show the entire stack trace, e.g. the line where it fails in your code, alongside all the information you've provided here? Thanks.", "@LysandreJik I'd be very interested in an `ElectraForSequenceClassification` head, as I'm not confident I could implement it myself since I'm quite new to Transformers and still learning how the library is organized. Any chance this is coming soon?", "i just posted a pull request ... was super simple to get it working\r\n\r\nhttps://github.com/huggingface/transformers/pull/4257", "@liuzzi awesome! I look forward to trying it out.", "@liuzzi wonderful, thanks a lot. Well done brother. Can you share a working notebook on this, please? Thank you.", "@innat i did not use a notebook to fine-tune, but for sentiment analysis you can just use the run_glue.py script with the SST-2 task which is a binary sentiment analysis task. You shouldn't even need to change any code, just make sure your dataset follows the format of SST-2.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/3703?src=pr&el=h1) Report\n> Merging [#3703](https://codecov.io/gh/huggingface/transformers/pull/3703?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/6435b9f908e7361330db89e263a65b0a58060d11&el=desc) will **decrease** coverage by `0.98%`.\n> The diff coverage is `59.37%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3703?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3703 +/- ##\n==========================================\n- Coverage 78.13% 77.14% -0.99% \n==========================================\n Files 104 104 \n Lines 17723 17752 +29 \n==========================================\n- Hits 13847 13695 -152 \n- Misses 3876 4057 +181 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3703?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19hbGJlcnQucHk=) | `74.50% <0.00%> (-0.75%)` | :arrow_down: |\n| [src/transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG5ldC5weQ==) | `73.33% <50.00%> (-2.45%)` | :arrow_down: |\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `87.88% <66.66%> (-0.28%)` | :arrow_down: |\n| [src/transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19yb2JlcnRhLnB5) | `85.04% <66.66%> (-10.67%)` | :arrow_down: |\n| [src/transformers/modeling\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG0ucHk=) | `88.78% <66.66%> (-0.38%)` | :arrow_down: |\n| [src/transformers/modeling\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19kaXN0aWxiZXJ0LnB5) | `97.02% <70.00%> (-0.57%)` | :arrow_down: |\n| [src/transformers/modeling\\_electra.py](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19lbGVjdHJhLnB5) | `72.38% <75.00%> (-0.27%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_pytorch\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3703/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/3703/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19jdHJsLnB5) | `95.61% <0.00%> (-2.64%)` | :arrow_down: |\n| ... and [3 more](https://codecov.io/gh/huggingface/transformers/pull/3703/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3703?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/3703?src=pr&el=footer). Last update [6435b9f...c6692e6](https://codecov.io/gh/huggingface/transformers/pull/3703?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "LGTM but I'll let others chime in.", "For ELECTRA, the discrepancy should have been fixed by 500aa12318ce5acd289d5edb6cb8266b3c3b162e, so can you propagate your changes there too?", "Related to naming, as it's for the `XXXForTokenClassification` when you have a single label wouldn't you expect to get a cross-entropy loss such as [binary cross-entropy](https://pytorch.org/docs/stable/nn.html#bceloss), rather than regression? Seeing as it's a classification model?", "I'd say the same would apply to `XXXForSentenceClassification`, right? It would probably be best to decouple regression and classification for those two classes instead of having regression as a special case, and make the behavior for `num_labels == 1` the same as `num_labels == 2` for labels that are not one-hot encoded.\r\n\r\nProbably worth a dedicated PR to decouple both instead of handling it here!", "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", "Any chance of blowing new life into this? Token classification is particularly familiar due to NER, but in many research fields (e.g. psycholinguistic studies) we are interested in a lot more than that. Continuous values for tokens are very common there. I'd love to see regression and multilabel classification for token classification models.", "Sad to see this die :(" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/3702?src=pr&el=h1) Report\n> Merging [#3702](https://codecov.io/gh/huggingface/transformers/pull/3702?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/f68d22850ced09bb194b30068ff94ca3409f0879&el=desc) will **decrease** coverage by `0.02%`.\n> The diff coverage is `41.66%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3702?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3702 +/- ##\n==========================================\n- Coverage 78.06% 78.03% -0.03% \n==========================================\n Files 100 100 \n Lines 17134 17144 +10 \n==========================================\n+ Hits 13375 13378 +3 \n- Misses 3759 3766 +7 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3702?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3702/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `91.30% <36.36%> (-0.81%)` | :arrow_down: |\n| [src/transformers/configuration\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3702/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX3V0aWxzLnB5) | `97.01% <100.00%> (+0.02%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3702/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.45% <0.00%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3702?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/3702?src=pr&el=footer). Last update [f68d228...6e959fd](https://codecov.io/gh/huggingface/transformers/pull/3702?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hmm don't worry about me, this shouldn't be a problem – feel free to merge if ready @LysandreJik !", "@jysohn23 I'm trying to run a variant of `run_glue_tpu.py` on TPUs and am stuck at an oom error. The first iteration of the below [for loop](https://github.com/jysohn23/transformers/blob/tpu/examples/run_tpu_glue.py#L150) runs fine, but it breaks on the second one. Any pointers on how to fix this?\r\n```\r\ntrain_dataloader = pl.ParallelLoader(dataloader, [args.device]).per_device_loader(args.device)\r\nepoch_iterator = tqdm(train_dataloader, desc=\"Iteration\", total=len(dataloader), disable=disable_logging)\r\nfor step, batch in enumerate(epoch_iterator):\r\n```\r\nI tried reducing the batch-size to 1 and running on a single core, both led to the same error. I'm using this `gcr.io/tpu-pytorch/xla:nightly_3.6` image for my experiments.\r\n\r\nfull log - shorturl.at/iswxR\r\nfew lines of the error log -\r\n```\r\n020-06-30 21:49:29.304998: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] >>> Dumping Computation 0 | 1/6136 [01:16<131:08:36, 76.95s/it]\r\n2020-06-30 21:49:29.305126: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] HloModule SyncTensorsGraph.33776, input_output_alias={ {0}: (250, {}), {1}: (249, {}), {2}: (265, {}), {3}: (248, {}), {4}: (247, {}), {5}: (246, {}), {6}: (245, {}), {7}: (244, {}), {8}: (269, {}), {9}: (243, {}), {10}: (242, {}), {11}: (241, {}), {12}: (240, {}), {13}: (239, {}), {14}: (271, {}), {15}: (238, {}), {16}: (237, {}), {17}: (236, {}), {18}: (235, {}), {19}: (234, {}), {20}: (273, {}), {21}: (233, {}), {22}: (232, {}), {23}: (231, {}), {24}: (230, {}), {25}: (229, {}), {26}: (274, {}), {27}: (228, {}), {28}: (227, {}), {29}: (226, {}), {30}: (225, {}), {31}: (224, {}), {32}: (276, {}), {33}: (223, {}), {34}: (222, {}), {35}: (221, {}), {36}: (220, {}), {37}: (219, {}), {38}: (277, {}), {39}: (218, {}), {40}: (217, {}), {41}: (216, {}), {42}: (215, {}), {43}: (214, {}), {44}: (279, {}), {45}: (213, {}), {46}: (212, {}), {47}: (211, {}), {48}: (210, {}), {49}: (209, {}), {50}: (280, {}), {51}: (208, {}), {52}: (207, {}), {53}: (206, {}), {54}: (205, {}), {55}: (204, {}), {56}: (282, {}), {57}: (203, {}), {58}: (202, {}), {59}: (201, {}), {60}: (200, {}), {61}: (199, {}), {62}: (283, {}), {63}: (198, {}), {64}: (197, {}), {65}: (196, {}), {66}: (195, {}), {67}: (194, {}), {68}: (285, {}), {69}: (193, {}), {70}: (192, {}), {71}: (191, {}), {72}: (190, {}), {73}: (189, {}), {74}: (286, {}), {75}: (188, {}), {76}: (187, {}), {77}: (186, {}), {78}: (185, {}), {79}: (184, {}), {80}: (288, {}), {81}: (183, {}), {82}: (182, {}), {83}: (181, {}), {84}: (180, {}), {85}: (179, {}), {86}: (289, {}), {87}: (178, {}), {88}: (177, {}), {89}: (176, {}), {90}: (175, {}), {91}: (174, {}), {92}: (291, {}), {93}: (173, {}), {94}: (172, {}), {95}: (171, {}), {96}: (170, {}), {97}: (169, {}), {98}: (292, {}), {99}: (168, {}), {100}: (167, {}), {101}: (166, {}), {102}: (165, {}), {103}: (164, {}), {104}: (294, {}), {105}: (163, {}), {106}: (162, {}), {107}: (161, {}), {108}: (160, {}), {109}: (159, {}), {110}: (295, {}), {111}: (158, {}), {112}: (157, {}), {113}: (156, {}), {114}: (155, {}), {115}: (154, {}), {116}: (297, {}), {117}: (153, {}), {118}: (152, {}), {119}: (151, {}), {120}: (150, {}), {121}: (149, {}), {122}: (298, {}), {123}: (148, {}), {124}: (147, {}), {125}: (146, {}), {126}: (145, {}), {127}: (144, {}), {128}: (300, {}), {129}: (143, {}), {130}: (142, {}), {131}: (141, {}), {132}: (140, {}), {133}: (139, {}), {134}: (301, {}), {135}: (138, {}), {136}: (137, {}), {137}: (136, {}), {138}: (135, {}), {139}: (134, {}), {140}: (303, {}), {141}: (133, {}), {142}: (132, {}), {143}: (131, {}), {144}: (130, {}), {145}: (129, {}), {146}: (304, {}), {147}: (128, {}), {148}: (127, {}), {149}: (126, {}), {150}: (125, {}), {151}: (124, {}), {152}: (306, {}), {153}: (123, {}), {154}: (122, {}), {155}: (121, {}), {156}: (120, {}), {157}: (119, {}), {158}: (307, {}), {159}: (118, {}), {160}: (117, {}), {161}: (116, {}), {162}: (115, {}), {163}: (114, {}), {164}: (309, {}), {165}: (113, {}), {166}: (112, {}), {167}: (111, {}), {168}: (110, {}), {169}: (109, {}), {170}: (310, {}), {171}: (108, {}), {172}: (107, {}), {173}: (106, {}), {174}: (105, {}), {175}: (104, {}), {176}: (312, {}), {177}: (103, {}), {178}: (102, {}), {179}: (101, {}), {180}: (100, {}), {181}: (99, {}), {182}: (313, {}), {183}: (98, {}), {184}: (97, {}), {185}: (96, {}), {186}: (95, {}), {187}: (94, {}), {188}: (315, {}), {189}: (93, {}), {190}: (92, {}), {191}: (91, {}), {192}: (90, {}), {193}: (89, {}), {194}: (316, {}), {195}: (88, {}), {196}: (87, {}), {197}: (86, {}), {198}: (85, {}), {199}: (84, {}), {200}: (318, {}), {201}: (83, {}), {202}: (82, {}), {203}: (81, {}), {204}: (80, {}), {205}: (79, {}), {206}: (319, {}), {207}: (78, {}), {208}: (77, {}), {209}: (76, {}), {210}: (75, {}), {211}: (74, {}), {212}: (321, {}), {213}: (73, {}), {214}: (72, {}), {215}: (71, {}), {216}: (70, {}), {217}: (69, {}), {218}: (322, {}), {219}: (68, {}), {220}: (67, {}), {221}: (66, {}), {222}: (65, {}), {223}: (64, {}), {224}: (324, {}), {225}: (63, {}), {226}: (62, {}), {227}: (61, {}), {228}: (60, {}), {229}: (59, {}), {230}: (325, {}), {231}: (58, {}), {232}: (57, {}), {233}: (56, {}), {234}: (55, {}), {235}: (54, {}), {236}: (327, {}), {237}: (53, {}), {238}: (52, {}), {239}: (51, {}), {240}: (50, {}), {241}: (49, {}), {242}: (328, {}), {243}: (48, {}), {244}: (47, {}), {245}: (46, {}), {246}: (45, {}), {247}: (44, {}), {248}: (330, {}), {249}: (43, {}), {250}: (42, {}), {251}: (41, {}), {252}: (40, {}), {253}: (39, {}), {254}: (331, {}), {255}: (38, {}), {256}: (37, {}), {257}: (36, {}), {258}: (35, {}), {259}: (34, {}), {260}: (333, {}), {261}: (33, {}), {262}: (32, {}), {263}: (31, {}), {264}: (30, {}), {265}: (29, {}), {266}: (334, {}), {267}: (28, {}), {268}: (27, {}), {269}: (26, {}), {270}: (25, {}), {271}: (24, {}), {272}: (336, {}), {273}: (23, {}), {274}: (22, {}), {275}: (21, {}), {276}: (20, {}), {277}: (19, {}), {278}: (337, {}), {279}: (18, {}), {280}: (17, {}), {281}: (16, {}), {282}: (15, {}), {283}: (14, {}), {284}: (339, {}), {285}: (13, {}), {286}: (12, {}), {287}: (8, {}), {288}: (7, {}), {289}: (5, {}), {290}: (340, {}), {291}: (346, {}), {292}: (4, {}), {377}: (342, {}) }\r\n2020-06-30 21:49:29.305162: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] \r\n2020-06-30 21:49:29.305173: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %MaxComputation.2092 (x.2093: f32[], y.2094: f32[]) -> f32[] {\r\n2020-06-30 21:49:29.305181: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %x.2093 = f32[] parameter(0)\r\n2020-06-30 21:49:29.305196: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %y.2094 = f32[] parameter(1)\r\n2020-06-30 21:49:29.305204: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] ROOT %maximum.2095 = f32[] maximum(f32[] %x.2093, f32[] %y.2094)\r\n2020-06-30 21:49:29.305212: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] }\r\n2020-06-30 21:49:29.305221: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] \r\n2020-06-30 21:49:29.305235: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %AddComputation.2101 (x.2102: f32[], y.2103: f32[]) -> f32[] {\r\n2020-06-30 21:49:29.305244: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %x.2102 = f32[] parameter(0)\r\n2020-06-30 21:49:29.305254: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %y.2103 = f32[] parameter(1)\r\n2020-06-30 21:49:29.305264: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] ROOT %add.2104 = f32[] add(f32[] %x.2102, f32[] %y.2103)\r\n2020-06-30 21:49:29.305273: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] }\r\n2020-06-30 21:49:29.305283: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] \r\n.\r\n.\r\n.\r\n.\r\n2020-06-30 21:49:29.568300: E 5603 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %subtract.5549 = f32[] subtract(f32[] %constant.5532, f32[] %constant.5533)\r\n2020-06-30 21:49:29.568320: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %constant.20745 = f32[] constant(0.125)\r\n2020-06-30 21:49:29.568321: E 5603 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %broadcast.5550 = f32[1,16,128,128]{3,2,1,0} broadcast(f32[] %subtract.5549), dimensions={}\r\n2020-06-30 21:49:29.568331: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %broadcast.20746 = f32[1024,4096]{1,0} broadcast(f32[] %constant.20745), dimensions={}\r\n2020-06-30 21:49:29.568332: E 5603 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %multiply.5551 = f32[1,16,128,128]{3,2,1,0} multiply(f32[1,16,128,128]{3,2,1,0} %multiply.5548, f32[1,16,128,128]{3,2,1,0} %broadcast.5550)\r\n2020-06-30 21:49:29.568342: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %multiply.20747 = f32[1024,4096]{1,0} multiply(f32[1024,4096]{1,0} %get-tuple-element.20744, f32[1024,4096]{1,0} %broadcast.20746)\r\n2020-06-30 21:49:29.568344: E 5603 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %broadcast.5552 = f32[1,16,128,128]{3,2,1,0} broadcast(f32[] %constant.5533), dimensions={}\r\n2020-06-30 21:49:29.568353: E 6014 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %reshape.25706 = f32[1,1]{1,0} reshape(f32[] %p263.1975)\r\n2020-06-30 21:49:29.568354: E 5603 tensorflow/compiler/xla/xla_client/xla_util.cc:76] %add.5553 = f32[1,16,128,128]{3,2,1,0} add(f32[1,16,128,128]{3,2,1,0} %multiply.5551, f32[1,16,128,128]{3,2,1,0} %broadcast.5552)\r\n.\r\n.\r\n.\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n (1) Resource exhausted: Ran out of memory in memory space vmem. It should not be possible to run out of vmem - please file a bug against XLA.\r\n\r\nLargest program allocations in vmem:\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n\t [[{{node XRTCompile}}]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n\t [[XRTCompile_G6]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n0 successful operations.\r\n0 derived errors ignored.\r\nTraceback (most recent call last):\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 235, in _mp_start_fn\r\n _start_fn(index, pf_cfg, fn, args)\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 229, in _start_fn\r\n fn(gindex, *args)\r\n File \"/export/share/akhilesh-gotmare/tpu_gedi/transformers/examples/run_tpu_glue.py\", line 797, in _mp_fn\r\n main(args)\r\n File \"/export/share/akhilesh-gotmare/tpu_gedi/transformers/examples/run_tpu_glue.py\", line 607, in main\r\n global_step, tr_loss = train(args, train_dataset, model, tokenizer, disable_logging=disable_logging)\r\n File \"/export/share/akhilesh-gotmare/tpu_gedi/transformers/examples/run_tpu_glue.py\", line 186, in train\r\n for step, batch in enumerate(epoch_iterator):\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/tqdm/std.py\", line 1107, in __iter__\r\n for obj in iterable:\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/parallel_loader.py\", line 31, in __next__\r\n return self.next()\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/parallel_loader.py\", line 37, in next\r\n xm.mark_step()\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/core/xla_model.py\", line 549, in mark_step\r\n wait=xu.getenv_as('XLA_SYNC_WAIT', bool, False))\r\nRuntimeError: Resource exhausted: From /job:tpu_worker/replica:0/task:0:\r\n2 root error(s) found.\r\n (0) Resource exhausted: Ran out of memory in memory space vmem. It should not be possible to run out of vmem - please file a bug against XLA.\r\n\r\nLargest program allocations in vmem:\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n\t [[{{node XRTCompile}}]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n (1) Resource exhausted: Ran out of memory in memory space vmem. It should not be possible to run out of vmem - please file a bug against XLA.\r\n\r\nLargest program allocations in vmem:\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n\t [[{{node XRTCompile}}]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n\t [[XRTCompile_G6]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n0 successful operations.\r\n0 derived errors ignored.\r\nTraceback (most recent call last):\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 235, in _mp_start_fn\r\n _start_fn(index, pf_cfg, fn, args)\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 229, in _start_fn\r\n fn(gindex, *args)\r\n File \"/export/share/akhilesh-gotmare/tpu_gedi/transformers/examples/run_tpu_glue.py\", line 797, in _mp_fn\r\n main(args)\r\n File \"/export/share/akhilesh-gotmare/tpu_gedi/transformers/examples/run_tpu_glue.py\", line 607, in main\r\n global_step, tr_loss = train(args, train_dataset, model, tokenizer, disable_logging=disable_logging)\r\n File \"/export/share/akhilesh-gotmare/tpu_gedi/transformers/examples/run_tpu_glue.py\", line 186, in train\r\n for step, batch in enumerate(epoch_iterator):\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/tqdm/std.py\", line 1107, in __iter__\r\n for obj in iterable:\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/parallel_loader.py\", line 31, in __next__\r\n return self.next()\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/parallel_loader.py\", line 37, in next\r\n xm.mark_step()\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/core/xla_model.py\", line 549, in mark_step\r\n wait=xu.getenv_as('XLA_SYNC_WAIT', bool, False))\r\nRuntimeError: Resource exhausted: From /job:tpu_worker/replica:0/task:0:\r\n2 root error(s) found.\r\n (0) Resource exhausted: Ran out of memory in memory space vmem. It should not be possible to run out of vmem - please file a bug against XLA.\r\n\r\nLargest program allocations in vmem:\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n\t [[{{node XRTCompile}}]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n (1) Resource exhausted: Ran out of memory in memory space vmem. It should not be possible to run out of vmem - please file a bug against XLA.\r\n\r\nLargest program allocations in vmem:\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n XLA label: %fusion.4431 = (f32[1024]{0:T(1024)}, f32[24,128]{1,0:T(8,128)}, f32[24,128]{1,0:T(8,128)}, f32[1024]{0:T(1024)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24,128,1024]{2,1,0:T(8,128)}, f32[24...\r\n Allocation type: scoped\r\n\r\n\t [[{{node XRTCompile}}]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n\t [[XRTCompile_G6]]\r\nHint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.\r\n\r\n0 successful operations.\r\n0 derived errors ignored.\r\n/root/anaconda3/envs/pytorch/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 1 leaked semaphores to clean up at shutdown\r\n len(cache))\r\n/root/anaconda3/envs/pytorch/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 1 leaked semaphores to clean up at shutdown\r\n len(cache))\r\n/root/anaconda3/envs/pytorch/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 1 leaked semaphores to clean up at shutdown\r\n len(cache))\r\n/root/anaconda3/envs/pytorch/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 1 leaked semaphores to clean up at shutdown\r\n len(cache))\r\n/root/anaconda3/envs/pytorch/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 1 leaked semaphores to clean up at shutdown\r\n len(cache))\r\n/root/anaconda3/envs/pytorch/lib/python3.6/multiprocessing/semaphore_tracker.py:143: UserWarning: semaphore_tracker: There appear to be 1 leaked semaphores to clean up at shutdown\r\n len(cache))\r\nTraceback (most recent call last):\r\n File \"run_tpu_glue.py\", line 806, in <module>\r\n main_cli()\r\n File \"run_tpu_glue.py\", line 802, in main_cli\r\n xmp.spawn(_mp_fn, args=(args,), nprocs=args.num_cores)\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 300, in spawn\r\n start_method=start_method)\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/multiprocessing/spawn.py\", line 158, in start_processes\r\n while not context.join():\r\n File \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/multiprocessing/spawn.py\", line 113, in join\r\n (error_index, exitcode)\r\nException: process 6 terminated with exit code 17\r\n```\r\n\r\n\r\n\r\n\r\n" ]

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

[]

[ "Duplicate of https://github.com/huggingface/transformers/issues/3452.\r\n\r\nWe should just remove the option cc @julien-c ", "You’re right, I’ll get on it", "Also best username ever, @MarxEngelsLeninStalin ", "Oh, I like that model as someone who has problems with spelling/grammar and fatigue it's useful to use xl one has more knowledge on it (yes I know it's not accurate but more accurate then large model). Is there any prospect of it being turned on again in the future?", "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", "@Marxist-Leninist, unfortunately adding it back isn't on our near future roadmap." ]

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

[]

[ "Hi, except if you freeze the transformer's (transformer being the base model) weights during the training, these will be modified during the fine-tuning. I believe none of our scripts in the library freeze the transformer's weight; so the weights would not be identical.", "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" ]

[]

[]

[ "Oh no, it should be `self.config.num_labels` instead of `self.num_labels` [here](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_electra.py#L665)!", "My bad, it should be fixed now!" ]

[]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[]

[]

[]

[ "Can't test this right now as I don't have access to a Windows machine - does this still work if the model doesn't already exist?", "> \r\n> \r\n> Can't test this right now as I don't have access to a Windows machine - does this still work if the model doesn't already exist?\r\n\r\nYes, I tested that 😄. There shouldn't be any further differences between `rename` and `replace`. Here's [what the documentation says](https://docs.python.org/3/library/os.html#os.replace)." ]

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

[]

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

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

[]

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

[]

[]

[ "I've worked around this problem now by not extending the huggingface framework directly but implementing XLMRobertaForQuestionAnswering externally (using huggingface classes etc.). ", "Would make sense to have it as a new feature though but that's a different issue type. ;)", "Could you share the code or make PR? ", "@djstrong The code corresponds to my solution presented above. The issue was related to me changing the library instead of just adding the file externally." ]

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

[]

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

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "This was implemented ~1 month ago so closing this issue." ]

[]

[]

[ "Created another environment from scratch and it got resolved.", "you need install transformers this way:\r\n\r\n```\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\npip install .\r\n```", "> you need install transformers this way:\r\n> \r\n> ```\r\n> git clone https://github.com/huggingface/transformers\r\n> cd transformers\r\n> pip install .\r\n> ```\r\n\r\nThis worked for me thanks!", "> you need install transformers this way:\r\n> \r\n> ```\r\n> git clone https://github.com/huggingface/transformers\r\n> cd transformers\r\n> pip install .\r\n> ```\r\n\r\nI was having a similar issue on colab:\r\n```\r\ncan't pickle AddedToken objects\r\n```\r\nThis solution also worked for me. Thanks!", "> you need install transformers this way:\r\n> \r\n> ```\r\n> git clone https://github.com/huggingface/transformers\r\n> cd transformers\r\n> pip install .\r\n> ```\r\n\r\nThanks much, this worked for me!", "> you need install transformers this way:\r\n> \r\n> ```\r\n> git clone https://github.com/huggingface/transformers\r\n> cd transformers\r\n> pip install .\r\n> ```\r\n\r\nThis solution doesn't work for me! I don't know maybe there is a conflict with `pip` an `conda`. I guess after I installed a package `bert_score` by conda, this error appeared and won't go" ]

[]

[]

[ "It's in [`BertSelfOutput`](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_bert.py#L261) which is called right after the `BertSelfAttention` in [`BertAttention`](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_bert.py#L316)." ]

[]

[]

[]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/3688?src=pr&el=h1) Report\n> Merging [#3688](https://codecov.io/gh/huggingface/transformers/pull/3688?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/715aa5b1356b878cbab7a7415a1c1b03a7777ae2&el=desc) will **increase** coverage by `0.03%`.\n> The diff coverage is `10.63%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3688?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3688 +/- ##\n==========================================\n+ Coverage 78.02% 78.06% +0.03% \n==========================================\n Files 104 104 \n Lines 17710 17708 -2 \n==========================================\n+ Hits 13819 13823 +4 \n+ Misses 3891 3885 -6 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3688?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/data/processors/utils.py](https://codecov.io/gh/huggingface/transformers/pull/3688/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvdXRpbHMucHk=) | `24.68% <0.00%> (ø)` | |\n| [src/transformers/data/processors/glue.py](https://codecov.io/gh/huggingface/transformers/pull/3688/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvZ2x1ZS5weQ==) | `29.79% <10.86%> (+2.26%)` | :arrow_up: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3688/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `91.84% <0.00%> (-0.13%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3688?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/3688?src=pr&el=footer). Last update [715aa5b...b867779](https://codecov.io/gh/huggingface/transformers/pull/3688?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "There is a combinatorial explosion between `tokenizer classes` x `glue datasets` so it's not super practical to test everything, but:\r\n- I've tested \"roberta-large\", \"bert-base-uncased\", \"xlnet-based\" tokenizers x \"sst2\" and \"mnli\" datasets and results are identical ✅ \r\n- encoding a batch of sentences or sentence pairs, while padding to a specific length, is really a native feature of a tokenizer at this point, so [those lines](https://github.com/huggingface/transformers/pull/3688/files#diff-8bc8284670454c05520b097dd51ad787R137-R139) in the current PR call the canonical API to do that. If there's a discrepancy with the historical way of tokenizing at this point it's probably outside the scope of this PR.\r\n", "**Note, however**, that following this PR the performance boost associated with using a fast tokenizer coupled with using `batch_encode` seems very variable, cc: @n1t0 @mfuntowicz \r\n\r\nOn my (mac OS) local machine it takes pretty much the same time using a fast and a non-fast tokenizer (even though the fast one burns all my CPU cores).\r\n\r\nOn a Colab notebook seems like perf varies a lot between executions (https://colab.research.google.com/drive/1DXOegSz7Tyr7MeSHYBBg40kiDhu4-JPr?authuser=1#scrollTo=NlygQfeyg-5b), with the fast tokenizer not always being faster than the other one.\r\n\r\nSee [notebook](https://colab.research.google.com/drive/1DXOegSz7Tyr7MeSHYBBg40kiDhu4-JPr)\r\n\r\nWould be interesting to dive in and do a more systematic benchmark than I did, considering that GLUE is a good benchmark for a real-world training workload. " ]

[ { "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", "I have the same the problem. When you close the issue does it mean that it is fixed? ", "I have also met this problem, how can I solve this? Thanks！！！！", "> I have also met this problem, how can I solve this? Thanks！！！！\r\n\r\nI switched to using `nlp.pipe`, which is the built-in function for multiprocessing, instead of doing it by hand. " ]

[]

[]

[ "I also see the same issue, has there been a fix for this?\r\nAttributeError: 'BertTokenizer' object has no attribute 'num_added_tokens'", "@minhtuev you can change `num_added_tokens` to `num_special_tokens_to_add`. This made the fix for me" ]

[]

[]

[ "Hi! We don't currently have a `run_tf_squad` script, but we would appreciate a contribution! The preprocessing of SQuAD shouldn't be different between pytorch and tensorflow. We haven't gotten to testing that as we haven't gotten to writing that script yet.\r\n\r\nIf we're to have a TF SQuAD script, we would have to align the pre-processing techniques as well!", "And we would definitely welcome a PR introducing `TFAlbertForQuestionAnswering`!", "Resolved in https://github.com/huggingface/transformers/commit/6d00033e97e1751a897f2317fdfd35dd853cee29 ." ]

[]

[]

[]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1771187924, "node_id": "MDU6TGFiZWwxNzcxMTg3OTI0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Core:%20Pipeline", "name": "Core: Pipeline", "color": "FF7066", "default": false, "description": "Internals of the library; Pipeline." } ]

[]

[ "hmmm, `mrm8488/bert-uncased-finetuned-qnli` is a sequence classification model, not a QA model.\r\n\r\nYou probably get warnings while loading it in a QA Pipeline.\r\n\r\nDoes this happen with other (QA) models?", "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/3682?src=pr&el=h1) Report\n> Merging [#3682](https://codecov.io/gh/huggingface/transformers/pull/3682?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/a594ee9c84dde933a3d0b4e07ff2994a1960574c&el=desc) will **increase** coverage by `0.07%`.\n> The diff coverage is `89.76%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3682?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3682 +/- ##\n==========================================\n+ Coverage 78.02% 78.09% +0.07% \n==========================================\n Files 104 104 \n Lines 17710 17786 +76 \n==========================================\n+ Hits 13818 13890 +72 \n- Misses 3892 3896 +4 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3682?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/3682/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190NS5weQ==) | `83.21% <89.68%> (+1.72%)` | :arrow_up: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3682/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `92.03% <100.00%> (+0.18%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3682?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/3682?src=pr&el=footer). Last update [a594ee9...67ae81f](https://codecov.io/gh/huggingface/transformers/pull/3682?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[]

[]

[ { "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 } ]

[ "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 see that the issue is closed? what is the answer?", "We are trying to move such questions more and more to the forum because they get more traction there and the library's issues should primarily be used for \"real\" issues. \r\n\r\nIt would be awesome if you guys could post the question on \r\nhttps://discuss.huggingface.co/", "I meet the same question. " ]

[]

[]

[ "@sshleifer It's not 100% needed to have model cards but definitely encouraged as they unlock new features e.g. here, discoverability of the models in the model hub, etc.\r\n\r\nYeah, adding an item to the checklist (I guess in `templates/adding_a_new_model`) would be nice, do you want to do it?" ]

[ { "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 @r0levrai, \r\n\r\nsorry for responding so late. Thanks for spotting the error. Once the linked PR is merged generation with an empty prompt should be fine! \r\n\r\nAlso note that there is a generation pipeline now which you could use as follows (once the PR is merged):\r\n\r\n```python\r\nfrom transformers import pipeline\r\ngenerator = pipeline(\"text-generation\")\r\ngenerator(\"\") # empty prompt\r\n```", "This is good news, thanks!" ]

[]

[]

[ "With ALBERT v1:\r\n\r\nhttps://github.com/huggingface/transformers/pull/1683#issuecomment-556001607\r\n\r\nI got better results using the recent integrated ELECTRA model :)\r\n\r\n", "@stefan-it j/w but do you know why bert-lg fine-tuned is listed as achieving 92.8 f1 on conll03 in [this paper?](https://paperswithcode.com/sota/named-entity-recognition-ner-on-conll-2003) Noticing it's over a pt higher f1 than Transformers's version. ", "@stefan-it \r\n\r\nYour albert result is very close to what I got. \r\n\r\nBy any chance, did you experiment with the XLNet?\r\n\r\nMy problem is:\r\nIn https://github.com/stevezheng23/xlnet_extension_tf, the author reported the ner performance as 0.9267. But I can only obtain a performance of 0.7626 with the same batch size and learning rate but longer training steps using the this library. I would like to confirm if the problem is my implementation but there is no baseline on XLNet.", "@stefan-it Thank you for your contribution to the Electra model finetuned on CoNLL03. I see you shared the weights in this repository. Could you please share the license for these weights? I could not find a model card for it.", "Hi @stefan-it !\r\nI know that this issue is not about electra, but I have the same question regarding electra too :sweat_smile: \r\nI ran NER CoNLL-2003 training with electra small like this:\r\n\r\n`python examples/ner/run_ner.py --model_type electra --model_name_or_path google/electra-small-discriminator --do_train --do_eval --do_predict --data_dir /home/corpora/ner/conll2003 --labels /home/corpora/ner/conll2003/labels.txt --num_train_epochs 6 --per_gpu_train_batch_size 256 --per_gpu_eval_batch_size 256 --max_seq_length 128 --output_dir /home/models/electra/conll2003 --evaluate_during_training --save_steps 1000 --logging_steps 1000 --overwrite_output_dir`\r\n\r\nBut got only 83.20% in F1. I know you ran electra small on the actual electra repository, but can you describe what you did in hugging face?", "Hi @petulla , the problem with the BERT paper is, that they've used document context for each token during evaluation. See e.g. this [discussion](https://github.com/allenai/allennlp/pull/2067#issuecomment-443961816) in the AllenNLP repo :)", "@guillaume-be I normally use MIT for all trained models (so should be fine for the fine-tuned ELECTRA model as well) :)", "@pvcastro Try to use a smaller batch size, for example with the configuration:\r\n\r\n```json\r\n{\r\n \"data_dir\": \"./data_en\",\r\n \"labels\": \"./data_en/labels.txt\",\r\n \"model_name_or_path\": \"google/electra-small-discriminator\",\r\n \"output_dir\": \"electra-small-en-1\",\r\n \"max_seq_length\": 128,\r\n \"num_train_epochs\": 5,\r\n \"per_gpu_train_batch_size\": 16,\r\n \"save_steps\": 878,\r\n \"seed\": 1,\r\n \"do_train\": true,\r\n \"do_eval\": true,\r\n \"do_predict\": true,\r\n \"--fp16\": true\r\n}\r\n```\r\n\r\nYou should be able to reach ~88.35% on the test set and 92.13% on development set.\r\n\r\nI just used the latest `master` version of Transformers + saved the JSON-based configuration as `config-electra-small.json`, then you can run training via `python3 run_ner.py config-electra-small.json` :)", "Hi @stefan-it , thanks for the input! I ran this config and got a pretty similar result. I had no idea that a larger batch size had this impact. I get 85% for bs 128. Is this for all transformer models, or for electra only? Or for sequence labeling tasks, perhaps? Do you know why this happens? ", "For Transformer-based model on sequence labeling tasks batch sizes of 8, 16 or 32 are a good choice for hyper-parameter search. So e.g. the [BERT paper](https://arxiv.org/abs/1810.04805) mentions [16. 32] for their experiments (see appendix A.3).\r\n\r\nAnd there's an interesting paper from Reimers and Gurevych about hyper-parameters for lstm-based networks for sequence labeling, but in my opinion their recommendations for batch sizes are also valid for Transformer-based models: [\"Optimal Hyperparameters for Deep LSTM-Networks for Sequence Labeling Tasks\"](https://arxiv.org/abs/1707.06799), see section 7.10 :)", "Thanks @stefan-it , I'll dig into these references! ", "Hi @stefan-it, it is quite strange that I got F1 0 for conll-03 dataset with **electra-large-discriminator**, while 0.88 and 0.91 with **small** and **base** models. The other settings are the same for the three models. Have you encountered this?", "Hi @lkluo , I can't remember fine-tuning instabilities with the ELECTRA models... but could you paste the fine-tuning configuration that you've used 🤔\r\n\r\n", "> Hi @lkluo , I can't remember fine-tuning instabilities with the ELECTRA models... but could you paste the fine-tuning configuration that you've used 🤔\r\n\r\nThanks @stefan-it. I think I may figure it out after I checked the loss, which converges slowly and which value remains 9.x after 5 epochs. Then I lower the learning rate from default **5e-5** to 10 times smaller, i.e. **5e-6**, then I can get 0.92 score. \r\n\r\nI also fine-tuned with **BERT-large** model using the default learning rate, and I am able to get a reasonable f1 score. Is there any special about **ELECTRA** large settings? Does batch size matter? It is limited to 12 due to GPU in my case. I saw somewhere people suggest larger batch size, smaller learning rate and longer training duration to reproduce good results. Could you share your configuration of **ELECTRA-LARGE**? Thanks a lot?\r\n\r\np.s., my configuration:\r\n\r\n> {\r\n \"data_dir\": \"\",\r\n \"labels\": \"\",\r\n \"model_name_or_path\": \"google/electra-large-discriminator\",\r\n \"output_dir\": \"\",\r\n \"max_seq_length\": 128,\r\n \"num_train_epochs\": 5,\r\n \"per_device_train_batch_size\": 12,\r\n \"save_steps\": 750,\r\n \"seed\": 1,\r\n \"do_train\": true,\r\n \"do_eval\": true,\r\n \"do_predict\": true\r\n}\r\n" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 2107543444, "node_id": "MDU6TGFiZWwyMTA3NTQzNDQ0", "url": "https://api.github.com/repos/huggingface/transformers/labels/fp16", "name": "fp16", "color": "d93f0b", "default": false, "description": "" } ]

[ { "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 } ]

[ "I can provide a google colab notebook, the fine-tuned model, the training data (or whatever needed) to show the issue. I hope it's a tokenizer issue rather than a training fault on my side, since re-training would cost a lot of cash (the training data is quite big - ~22 million lines)\r\n\r\n ", "I'm suspecting `fp16` to be the reason. Not sure whether this is supported for `generation` yet. @sshleifer - do you know more about this maybe?", "A colab notebook would be great though. Or even better would be if you could upload your model to the community models :-) \r\nThis would make it very easy for us to find the bug:\r\nhttps://huggingface.co/transformers/model_sharing.html", "Yeah some sort of sharing to diagnose. I don't think fp16 is the problem. What does `outputs[0]` look like? ", "hi, \r\n\r\nthanks for your reply, \r\n\r\n[https://huggingface.co/albertbn/gpt2-medium-finetuned-ads-fp16-blocksz512](url)\r\n\r\nthe above is the model, \r\n\r\nyou can re-create the error using the following:\r\n\r\n```\r\nimport torch\r\nfrom transformers import GPT2Tokenizer, GPT2LMHeadModel\r\n\r\ntokenizer = GPT2Tokenizer.from_pretrained(path)\r\nmodel = GPT2LMHeadModel.from_pretrained(path)\r\n\r\nif torch.cuda.is_available():\r\n model.to('cuda')\r\n\r\ninput_context = '''Find a plumber nearby!'''\r\ninput_ids = torch.tensor(tokenizer.encode(input_context)).unsqueeze(0) \r\nif torch.cuda.is_available():\r\n input_ids = input_ids.cuda()\r\n\r\nmax_length=150; temperature=.175; repetition_penalty=1.3; top_k=70; top_p=0.67\r\noutputs = model.generate(\r\n input_ids=input_ids, max_length=max_length, temperature=temperature, repetition_penalty=repetition_penalty,\r\n bos_token_id=tokenizer.bos_token_id,\r\n top_k=top_k,\r\n top_p=top_p\r\n )\r\n\r\nret = tokenizer.decode(outputs[0], skip_special_tokens=True)\r\n\r\nprint(ret)\r\n\r\n# Setting `pad_token_id` to 50256 (first `eos_token_id`) to generate sequence\r\n# Find a plumber nearby!\r\n# Plumbing Services in Wellington NZ.\r\n# 24/7 Emergency Plumbers Near You, Call Now For Fast Service or Repair of Your Plumbing!Need to Fix Leaking Pipes?‎NZ's #1 Gasfitter™️Call Us Today for Expert Advice & The Best Service!Get the Right Gasfitting Solution for your Home. Get It Installed Now - Free Quote Here !{KeyWord:Gas Fitting Installation}Quick And Efficient Installers\r\n\r\n```\r\n\r\nyou can see the issue (lines stuck without \\n) in the last line, starting with: 24/7 Emergency Plumbers...\r\n\r\nthank you in advance,\r\nAlbert", "> Yeah some sort of sharing to diagnose. I don't think fp16 is the problem. What does `outputs[0]` look like?\r\n\r\noutputs[0] for the example I've posted looks like this:\r\n\r\n```\r\ntensor([16742, 257, 458, 4494, 6716, 0, 198, 3646, 28149, 6168,\r\n 287, 30597, 26905, 13, 198, 1731, 14, 22, 18154, 1345,\r\n 17024, 20173, 921, 11, 4889, 2735, 1114, 12549, 4809, 393,\r\n 28912, 286, 3406, 1345, 28149, 0, 23037, 284, 13268, 1004,\r\n 868, 350, 18636, 30, 48261, 37371, 338, 1303, 16, 14345,\r\n 69, 1967, 8151, 37929, 14134, 4021, 6288, 329, 25516, 42708,\r\n 1222, 383, 6705, 4809, 0, 3855, 262, 6498, 14345, 32232,\r\n 28186, 329, 534, 5995, 13, 3497, 632, 2262, 4262, 2735,\r\n 532, 3232, 19879, 3423, 5145, 90, 9218, 26449, 25, 39699,\r\n 376, 2535, 32588, 92, 21063, 843, 412, 5632, 15545, 364],\r\n device='cuda:0')\r\n```\r\nthere is no white space separating 0, 23037 (23037 is the only index in the output: Plumbing**!Need** )\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", "Regarding the second question\r\n\r\n> 2. I get a 'warning':\r\n> Setting `pad_token_id` to 50256 (first `eos_token_id`) to generate sequence\r\n\r\nIt's explained [here](https://jaketae.github.io/study/gpt2/#setup): \"For open-end generation, HuggingFace will set the padding token ID to be equal to the end-of-sentence token ID\". Code is here: https://github.com/huggingface/transformers/blob/b880508440f43f80e35a78ccd2a32f3bde91cb23/src/transformers/generation_utils.py#L410-L414" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1771187924, "node_id": "MDU6TGFiZWwxNzcxMTg3OTI0", "url": "https://api.github.com/repos/huggingface/transformers/labels/Core:%20Pipeline", "name": "Core: Pipeline", "color": "FF7066", "default": false, "description": "Internals of the library; Pipeline." } ]

[]

[ "I'm having the same issue", "> I'm having the same issue\r\n\r\nI got this working by using the following code:\r\n```\r\n # Allocate a pipeline for sentiment-analysis\r\n nlp = pipeline(\"sentiment-analysis\")\r\n nlp.tokenizer = transformers.DistilBertTokenizer.from_pretrained(\"**distilbert-base-uncased**\")\r\n```\r\n\r\nThanks for pointing me in the right direction LysandreJik!", "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 } ]

[]

[ "looks like it's not able to find vocabulary file. Make sure there is a vocab.txt file for bert. Otherwise, you can simply load it by `tokenizer = BertTokenizer(vocab_file=\"path to vocab\", and configs)`.", "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" ]

[]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "You can try to move transformer_base file to the same location of run_bart_sum.py.", "Hi,\r\n1. As written in `README.md`\r\n> \"To use your own data, copy that files format. Each article to be summarized is on its own line.\"\r\n\r\n I think you should insert in cnn_dm folder your files renamed `train.source`, `train.target`, `test.source`, `test.target`, `val.source`, `val.target`, where in each file you have respectively a source text and a target text per line.\r\n\r\n2. You are not using the script `run_train.sh`, as suggested in the `README.md`. In the `run_train.sh` there are a series of export commands that you missed. \r\nThe last one should fix your issue.\r\nHope it helps.\r\n\r\n ```export OUTPUT_DIR_NAME=bart_sum\r\nexport CURRENT_DIR=${PWD}\r\nexport OUTPUT_DIR=${CURRENT_DIR}/${OUTPUT_DIR_NAME}\r\n\r\n# Make output directory if it doesn't exist\r\nmkdir -p $OUTPUT_DIR\r\n\r\n#Add parent directory to python path to access transformer_base.py\r\nexport PYTHONPATH=\"../../\":\"${PYTHONPATH}\"```\r\n\r\n", "Closing, @teelinsan 's answer is correct. " ]

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

[]

[ "Hi! I don't really understand how you obtained what you did, what script did you use, what arguments did you put in? The procedure to convert an ELECTRA checkpoint from the official implementation to our implementation is to do the following (feel free to skip the first steps if you already have your checkpoint):\r\n\r\n```bash\r\n# Get a checkpoint\r\nwget https://storage.googleapis.com/electra-data/electra_small.zip \r\n\r\n# Unzip it\r\nunzip electra_small.zip \r\n\r\n# Get an appropriate configuration file for your model (see below)\r\nvim electra_small/config.json\r\n\r\n# Run the script\r\npython $TRANSFORMERS/src/transformers/convert_electra_original_tf_checkpoint_to_pytorch.py \\ \r\n --tf_checkpoint_path=./electra_small/electra_small \\\r\n --config_file=./electra_small/config.json \\\r\n --pytorch_dump_path=pytorch_model.bin \\\r\n --discriminator_or_generator=discriminator\r\n\r\n```\r\n\r\nFrom this you should get the following output:\r\n\r\n```bash\r\nInitialize PyTorch weight ['discriminator_predictions', 'dense', 'bias'] discriminator_predictions/dense/bias\r\nInitialize PyTorch weight ['discriminator_predictions', 'dense', 'kernel'] discriminator_predictions/dense/kernel\r\nInitialize PyTorch weight ['discriminator_predictions', 'dense_prediction', 'bias'] discriminator_predictions/dense_1/bias\r\nInitialize PyTorch weight ['discriminator_predictions', 'dense_prediction', 'kernel'] discriminator_predictions/dense_1/kernel\r\nInitialize PyTorch weight ['electra', 'embeddings', 'LayerNorm', 'beta'] electra/embeddings/LayerNorm/beta\r\nInitialize PyTorch weight ['electra', 'embeddings', 'LayerNorm', 'gamma'] electra/embeddings/LayerNorm/gamma\r\n[...]\r\nSkipping generator_predictions/dense/bias ['generator_predictions', 'dense', 'bias'] 'ElectraForPreTraining' object has no attribute 'generator_predictions'\r\nSkipping generator_predictions/dense/kernel ['generator_predictions', 'dense', 'kernel'] 'ElectraForPreTraining' object has no attribute 'generator_predictions'\r\nSkipping generator_predictions/output_bias ['generator_lm_head', 'bias'] 'ElectraForPreTraining' object has no attribute 'generator_lm_head'\r\nINFO:transformers.modeling_electra:Skipping generator_predictions/temperature\r\nINFO:transformers.modeling_electra:Skipping global_step\r\nSave PyTorch model to pytorch_model.bin\r\n```\r\n\r\nWhich tells you that it ignored the generator layers, but saved the discriminator layers :).\r\n\r\nThe tricky part here is to craft a configuration file specific to the model. I want to obtain the small discriminator from this checkpoint, so the configuration file is the following:\r\n\r\n```json\r\n{\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"hidden_size\": 256,\r\n \"intermediate_size\": 1024,\r\n \"num_attention_heads\": 4,\r\n \"num_hidden_layers\": 12,\r\n \"embedding_size\": 128,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"initializer_range\": 0.02,\r\n \"layer_norm_eps\": 1e-12,\r\n \"max_position_embeddings\": 512,\r\n \"type_vocab_size\": 2,\r\n \"vocab_size\": 30522\r\n}\r\n```\r\n\r\nYou can either write it yourself or instantiate it from a `transformers.ElectraConfig` and save it as a JSON file.", "@LysandreJik I used your [script](https://github.com/huggingface/transformers/blob/master/src/transformers/convert_electra_original_tf_checkpoint_to_pytorch.py) to convert the trained model from the [origin repo](https://github.com/google-research/electra) (training on my own data) and it worked. I wonder whether an equal technique to convert this trained model to the Electra tf2 model that implemented in HuggingFace?", "@nguyenvulebinh glad the script worked! The script only outputs a PyTorch model but it's very simple to convert that model to TF2. Once you have the converted model, you can then load it in TensorFlow by specifying the `from_pt` option:\r\n\r\n```py\r\nfrom transformers import TFElectraForPreTraining\r\n\r\nmodel = TFElectraForPreTraining.from_pretrained(\"directory\", from_pt=True)\r\n```\r\n\r\nYou can then save that model in `.h5` format so that it gets natively loaded by TensorFlow in the future:\r\n\r\n```py\r\nmodel.save_pretrained(\"directory-tf\")\r\n\r\n# Can now load directly from TensorFlow without the `from_pt` option:\r\nmodel = TFElectraForPreTraining.from_pretrained(\"directory-tf\")\r\n```", "@LysandreJik It's really cool! Thank you! I did it 😍", "@LysandreJik \r\n\r\nHi,\r\n\r\nI have a question on pre-training Electra using the PyTorch base model.\r\n\r\nIf I want to continue pretraining the Electra model (HuggingFace implementation) on a domain-specific corpus, which model should I use to initialize - the generator or discriminator?\r\n\r\nThanks!", "When using the ELECTRA method what you're really interested in is the discriminator.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Hi! Maybe you can have a look at to the issue #1560 . ", "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/3669?src=pr&el=h1) Report\n> Merging [#3669](https://codecov.io/gh/huggingface/transformers/pull/3669?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0a9d09b42a9c7c1ccc00da48486a1188078e8594&el=desc) will **increase** coverage by `0.03%`.\n> The diff coverage is `83.54%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3669?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3669 +/- ##\n==========================================\n+ Coverage 78.03% 78.07% +0.03% \n==========================================\n Files 104 106 +2 \n Lines 17708 17787 +79 \n==========================================\n+ Hits 13819 13887 +68 \n- Misses 3889 3900 +11 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3669?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/hf\\_argparser.py](https://codecov.io/gh/huggingface/transformers/pull/3669/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9oZl9hcmdwYXJzZXIucHk=) | `74.00% <74.00%> (ø)` | |\n| [src/transformers/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/3669/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9fX2luaXRfXy5weQ==) | `98.98% <100.00%> (+0.02%)` | :arrow_up: |\n| [src/transformers/training\\_args.py](https://codecov.io/gh/huggingface/transformers/pull/3669/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90cmFpbmluZ19hcmdzLnB5) | `100.00% <100.00%> (ø)` | |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3669/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `92.23% <0.00%> (+0.12%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3669/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.28% <0.00%> (+0.16%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3669?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/3669?src=pr&el=footer). Last update [0a9d09b...b63747d](https://codecov.io/gh/huggingface/transformers/pull/3669?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Phew! Ok, I went through multiple rewrites today and I think it is pretty good now.\r\n\r\n**TL;DR:**\r\n\r\n- I only made changes to the `run_glue.py` example script to keep the diff smaller.\r\n- I have to pass `DataClassType`s (_types_, not instances) to `HfArgumentParser` because if they have required properties/arguments, we wouldn't be able to instantiate them before \"filling\" them\r\n- The class is designed to play well with the native `argparse`. In particular, you can get back any not-known args and parse them using a different argparse.ArgumentParser, to make adoption easier in complex scripts.\r\n- **read the unit tests for (a subset of) the supported arguments and how the properties translate into arguments.** " ]

[]

[]

[ "Hi @Colanim,\r\n\r\nThat's indeed a very good question! The only reason why we add these hacks here is because that's the way Fairseq implemented it and you get better results on summarization using Bart this way. We measured the differences in performance when leaving out those \"force token\" hacks and it was quite significant. \r\nPlease read through these PRs to better understand why we made this decision: \r\n\r\nhttps://github.com/huggingface/transformers/pull/3225\r\n\r\nand \r\n\r\nhttps://github.com/huggingface/transformers/pull/3140", "> I don't understand why it's necessary, because anyway the decoder input ids already contain BOS :\r\n\r\nRegarding this point, is the situation different? I went through the PRs and the code, but it seems that the default `decoder_start_token_id` is still the EOS_token.\r\n\r\nUltimately, the question I want to ask is \r\n**If I want to use BART for fine-tuning on another summarization task, do I set the `decoder_start_token_id` to EOS_token or BOS_token?**" ]

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

[]

[ "CTRL is a very large model so generating in bulk would require a lot of RAM. \r\nAlso generating with padded batches is not really supported yet, see: https://github.com/huggingface/transformers/issues/3021", "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/3666?src=pr&el=h1) Report\n> Merging [#3666](https://codecov.io/gh/huggingface/transformers/pull/3666?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0a9d09b42a9c7c1ccc00da48486a1188078e8594&el=desc) will **increase** coverage by `0.01%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3666?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3666 +/- ##\n==========================================\n+ Coverage 78.03% 78.04% +0.01% \n==========================================\n Files 104 104 \n Lines 17708 17708 \n==========================================\n+ Hits 13819 13821 +2 \n+ Misses 3889 3887 -2 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3666?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3666/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `92.23% <0.00%> (+0.12%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3666/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.28% <0.00%> (+0.16%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3666?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/3666?src=pr&el=footer). Last update [0a9d09b...7df3a3b](https://codecov.io/gh/huggingface/transformers/pull/3666?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "This is really cool, @seyonechithrananda. Can you add a \r\n\r\n```\r\n---\r\ntags: \r\n- chemistry\r\n---\r\n```\r\n\r\nmetadata block to the top of the file? Also cc'ing @mrm8488 who might be interested", "Thank you @julien-c. I uploaded two models from ChEMBL25/26 for drug structure learning (SMILES) using same technique. In fact, they have been used for COVID-19 drug discovery", "@mrm8488 Are you targeting ligand-protein modelling techniques with transformers? ", "@julien-c Made the changes. Let me know what you think!", "> @mrm8488 Are you targeting ligand-protein modelling techniques with transformers?\r\n\r\nAs told you via Twitter I am getting started into it. Getting chemical knowledge :)" ]

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

[]

[ "Couldn't we achieve the same result by specifying `add_special_tokens=False` instead of that? This isn't robust to different models, as GPT-2 (which doesn't have special tokens) would get some tokens removed.", "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 this as `run_language_modeling.py` is now based on the trainer. Thanks for your contribution!!" ]

[ { "id": 1834052129, "node_id": "MDU6TGFiZWwxODM0MDUyMTI5", "url": "https://api.github.com/repos/huggingface/transformers/labels/High-Level%20feature", "name": "High-Level feature", "color": "f7c9a3", "default": false, "description": "" }, { "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 } ]

[ "do you solve it ?", "> do you solve it ?\r\n\r\nNot yet. Let me know if you are able to find the fix.", "yes, I have the same issue. I inspected the tensorboard graph and there is no config operation or anything of that sort. I also tried to save it with a manually defined signature, which didn't work either.\r\n\r\nWorkaround for now is to use the `save_pretrained`. Is there a way to convert pretrained to TF2.0 saved_model?", "We also see a similar issue in transformers 2.9.1.\r\n\r\nAlso curious if people have a workaround or solution to use with TF model serving?", "FYI: The cause for this issue is documented in #4709. The current workaround/fix is to remove `config` from the call to this function:\r\n\r\nhttps://github.com/huggingface/transformers/blob/d6a677b14bcfd56b22fafeb212a27c6068886e07/src/transformers/modeling_tf_roberta.py#L331 \r\n\r\nThis prevents `trainable` from being set to `config` in the initialization function of `tf.keras.layers.Layer`. Then the model will be serialized correctly instead of failing to serialize the `trainable` value later.", "Hello!\r\n\r\nThe saving in saved model format is not implemented yet, but it is planned to work on it :) I will reply here once there will be something about this. Sorry for the inconvenience.", "I think this issue can be closed now due to PR #4884.\r\n\r\nThe sample code in the issue runs successfully in `master`.\r\n\r\n<img width=\"1010\" alt=\"Screenshot 2020-06-10 at 12 09 12\" src=\"https://user-images.githubusercontent.com/5602332/84255742-5363d000-ab13-11ea-822c-72da88399995.png\">\r\n", "Great, thanks for solving the issue @harkous ", "@harkous Thanks for the work.\r\nI have still the issue of the author even with your code and upgrade the last version of transformer with pip install transformers --upgrade , is it still working with you ?", "You have to install transformers from the master branch. The fix has not been released yet.", "Hello ! \r\n\r\nIt seems that I have a similar issue with a model based on Camembert when trying to save my model with : \r\n\r\n`model.save(\"model\",save_format='tf')`\r\n\r\nGive me : \r\n\r\n```\r\nTypeError: ('Not JSON Serializable:', CamembertConfig {\r\n \"architectures\": [\r\n \"CamembertForMaskedLM\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"bos_token_id\": 5,\r\n \"eos_token_id\": 6,\r\n \"gradient_checkpointing\": false,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 768,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"layer_norm_eps\": 1e-05,\r\n \"max_position_embeddings\": 514,\r\n \"model_type\": \"camembert\",\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 12,\r\n \"output_past\": true,\r\n \"pad_token_id\": 1,\r\n \"type_vocab_size\": 1,\r\n \"vocab_size\": 32005\r\n}\r\n)\r\n```\r\n\r\nAt first with transformers 2.11.0 but also after upgrading to 3.3 (with TF 2.3)\r\n\r\nI can give a code snippet to reproduce if necessary and Custom Model construction can be found here : https://github.com/MAIF/melusine/blob/master/melusine/models/neural_architectures.py#L312\r\n\r\n\r\n**Complete Stack Trace**\r\n```\r\n---------------------------------------------------------------------------\r\nTypeError Traceback (most recent call last)\r\n<ipython-input-32-ae508742561b> in <module>\r\n----> 1 model.model.save(\"test4\",save_format='tf')\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/engine/training.py in save(self, filepath, overwrite, include_optimizer, save_format, signatures, options)\r\n 1977 \"\"\"\r\n 1978 save.save_model(self, filepath, overwrite, include_optimizer, save_format,\r\n-> 1979 signatures, options)\r\n 1980 \r\n 1981 def save_weights(self,\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/saving/save.py in save_model(model, filepath, overwrite, include_optimizer, save_format, signatures, options)\r\n 132 else:\r\n 133 saved_model_save.save(model, filepath, overwrite, include_optimizer,\r\n--> 134 signatures, options)\r\n 135 \r\n 136 \r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/saving/saved_model/save.py in save(model, filepath, overwrite, include_optimizer, signatures, options)\r\n 78 # we use the default replica context here.\r\n 79 with distribution_strategy_context._get_default_replica_context(): # pylint: disable=protected-access\r\n---> 80 save_lib.save(model, filepath, signatures, options)\r\n 81 \r\n 82 if not include_optimizer:\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/saved_model/save.py in save(obj, export_dir, signatures, options)\r\n 974 \r\n 975 _, exported_graph, object_saver, asset_info = _build_meta_graph(\r\n--> 976 obj, export_dir, signatures, options, meta_graph_def)\r\n 977 saved_model.saved_model_schema_version = constants.SAVED_MODEL_SCHEMA_VERSION\r\n 978 \r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/saved_model/save.py in _build_meta_graph(obj, export_dir, signatures, options, meta_graph_def)\r\n 1074 \r\n 1075 object_graph_proto = _serialize_object_graph(saveable_view,\r\n-> 1076 asset_info.asset_index)\r\n 1077 meta_graph_def.object_graph_def.CopyFrom(object_graph_proto)\r\n 1078 \r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/saved_model/save.py in _serialize_object_graph(saveable_view, asset_file_def_index)\r\n 719 for obj, obj_proto in zip(saveable_view.nodes, proto.nodes):\r\n 720 _write_object_proto(obj, obj_proto, asset_file_def_index,\r\n--> 721 saveable_view.function_name_map)\r\n 722 return proto\r\n 723 \r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/saved_model/save.py in _write_object_proto(obj, proto, asset_file_def_index, function_name_map)\r\n 759 version=versions_pb2.VersionDef(\r\n 760 producer=1, min_consumer=1, bad_consumers=[]),\r\n--> 761 metadata=obj._tracking_metadata)\r\n 762 # pylint:enable=protected-access\r\n 763 proto.user_object.CopyFrom(registered_type_proto)\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py in _tracking_metadata(self)\r\n 3009 @property\r\n 3010 def _tracking_metadata(self):\r\n-> 3011 return self._trackable_saved_model_saver.tracking_metadata\r\n 3012 \r\n 3013 def _list_extra_dependencies_for_serialization(self, serialization_cache):\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/saving/saved_model/base_serialization.py in tracking_metadata(self)\r\n 52 # TODO(kathywu): check that serialized JSON can be loaded (e.g., if an\r\n 53 # object is in the python property)\r\n---> 54 return json_utils.Encoder().encode(self.python_properties)\r\n 55 \r\n 56 def list_extra_dependencies_for_serialization(self, serialization_cache):\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/saving/saved_model/json_utils.py in encode(self, obj)\r\n 42 \r\n 43 def encode(self, obj):\r\n---> 44 return super(Encoder, self).encode(_encode_tuple(obj))\r\n 45 \r\n 46 \r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/json/encoder.py in encode(self, o)\r\n 197 # exceptions aren't as detailed. The list call should be roughly\r\n 198 # equivalent to the PySequence_Fast that ''.join() would do.\r\n--> 199 chunks = self.iterencode(o, _one_shot=True)\r\n 200 if not isinstance(chunks, (list, tuple)):\r\n 201 chunks = list(chunks)\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/json/encoder.py in iterencode(self, o, _one_shot)\r\n 255 self.key_separator, self.item_separator, self.sort_keys,\r\n 256 self.skipkeys, _one_shot)\r\n--> 257 return _iterencode(o, 0)\r\n 258 \r\n 259 def _make_iterencode(markers, _default, _encoder, _indent, _floatstr,\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/keras/saving/saved_model/json_utils.py in default(self, obj)\r\n 39 items = obj.as_list() if obj.rank is not None else None\r\n 40 return {'class_name': 'TensorShape', 'items': items}\r\n---> 41 return serialization.get_json_type(obj)\r\n 42 \r\n 43 def encode(self, obj):\r\n\r\n~/.conda/envs/emails_maif_vie/lib/python3.6/site-packages/tensorflow/python/util/serialization.py in get_json_type(obj)\r\n 70 return obj.__wrapped__\r\n 71 \r\n---> 72 raise TypeError('Not JSON Serializable:', obj)\r\n\r\nTypeError: ('Not JSON Serializable:', CamembertConfig {\r\n \"architectures\": [\r\n \"CamembertForMaskedLM\"\r\n ],\r\n \"attention_probs_dropout_prob\": 0.1,\r\n \"bos_token_id\": 5,\r\n \"eos_token_id\": 6,\r\n \"gradient_checkpointing\": false,\r\n \"hidden_act\": \"gelu\",\r\n \"hidden_dropout_prob\": 0.1,\r\n \"hidden_size\": 768,\r\n \"initializer_range\": 0.02,\r\n \"intermediate_size\": 3072,\r\n \"layer_norm_eps\": 1e-05,\r\n \"max_position_embeddings\": 514,\r\n \"model_type\": \"camembert\",\r\n \"num_attention_heads\": 12,\r\n \"num_hidden_layers\": 12,\r\n \"output_past\": true,\r\n \"pad_token_id\": 1,\r\n \"type_vocab_size\": 1,\r\n \"vocab_size\": 32005\r\n}\r\n)\r\n```", "Please open another issue with a code snippet to make us able to reproduce your problem." ]

[]

[]

[ "Non slow test speed before change:\r\n\r\n", "There is a tiny TF T5 model now as well via: \r\n`model = TFAutoModelWithLMHead.from_pretrained(\"patrickvonplaten/t5-tiny-random\")`", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/3663?src=pr&el=h1) Report\n> Merging [#3663](https://codecov.io/gh/huggingface/transformers/pull/3663?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/0a9d09b42a9c7c1ccc00da48486a1188078e8594&el=desc) will **decrease** coverage by `0.01%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3663?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3663 +/- ##\n==========================================\n- Coverage 78.03% 78.02% -0.02% \n==========================================\n Files 104 104 \n Lines 17708 17708 \n==========================================\n- Hits 13819 13817 -2 \n- Misses 3889 3891 +2 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3663?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3663/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.96% <0.00%> (-0.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3663/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `91.97% <0.00%> (-0.13%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3663?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/3663?src=pr&el=footer). Last update [0a9d09b...4751188](https://codecov.io/gh/huggingface/transformers/pull/3663?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "LGTM, but FYI for context, I think the reason we were using the real models was that we intended to do integration testing: in `_test_mono_column_pipeline` we only test equality of keys but we could have tested equality (or closeness) of values.", "Makes sense @julien-c . I'd be happy to add some `@slow` integration tests and try to fulfill the original intent" ]

[ { "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/3662?src=pr&el=h1) Report\n> Merging [#3662](https://codecov.io/gh/huggingface/transformers/pull/3662?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/261c4ff4e297e919ba993e1214a805e988bc9e79&el=desc) will **decrease** coverage by `0.04%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3662?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3662 +/- ##\n==========================================\n- Coverage 78.29% 78.25% -0.05% \n==========================================\n Files 104 104 \n Lines 17628 17628 \n==========================================\n- Hits 13802 13794 -8 \n- Misses 3826 3834 +8 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3662?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3662/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `90.65% <0.00%> (-0.84%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3662/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.12% <0.00%> (-0.17%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3662/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `92.10% <0.00%> (-0.13%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3662?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/3662?src=pr&el=footer). Last update [261c4ff...fccd0c7](https://codecov.io/gh/huggingface/transformers/pull/3662?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "@MichalMalyska This is super interesting, thanks for sharing\r\n\r\n[**Model page**](https://huggingface.co/NLP4H/ms_bert)" ]

[]

[]

[]

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

[]

[ "Did you find the solution to the problem?", "@jhashekhar it seems like pytorch xla has some memory issue itself, xla team is working on it, tf tpu is much better at this moment so i am not using pytorch tpu anymore,probably later this year torch xla team will solve all the performance issues they are having at this moment,until then i recommend tf tpu or if you need gpu then pytorch gpu.just my recommendation", "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", "Had this come up when parallel training on GPUS with multiprocessing - thoughts on a solution?", "Take a look at this issue: https://github.com/pytorch/xla/issues/1870#issuecomment-612217012\r\n\r\nIt's pretty long, but they helped me solve this problem last year. I got the model working, but ended up using TF.", "> pytorch/xla#1870 (comment)\r\n\r\ndo you have a summary of what we need to do for solving this for pytorch?", "@brando90 please use bf16 and follow this simple tutorial of mine for better understanding : https://www.kaggle.com/mobassir/faster-pytorch-tpu-baseline-for-cld-cv-0-9\r\n\r\nDon't forget to reduce batch size,image size etc that fits in xla\r\n\r\nI think it will help you to solve your oom error,thanks", "> Had this come up when parallel training on GPUS with multiprocessing - thoughts on a solution?\r\n\r\nI got this problem. Did you solve it?" ]

[ { "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/3659?src=pr&el=h1) Report\n> Merging [#3659](https://codecov.io/gh/huggingface/transformers/pull/3659?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/39a34cc375ed79d18888464289b83713fc20f7d4&el=desc) will **not change** coverage by `%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3659?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3659 +/- ##\n=======================================\n Coverage 78.29% 78.29% \n=======================================\n Files 104 104 \n Lines 17628 17628 \n=======================================\n Hits 13801 13801 \n Misses 3827 3827 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3659?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3659/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.28% <0.00%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3659?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/3659?src=pr&el=footer). Last update [39a34cc...8d1de79](https://codecov.io/gh/huggingface/transformers/pull/3659?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/3658?src=pr&el=h1) Report\n> Merging [#3658](https://codecov.io/gh/huggingface/transformers/pull/3658?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/39a34cc375ed79d18888464289b83713fc20f7d4&el=desc) will **decrease** coverage by `0.05%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3658?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3658 +/- ##\n==========================================\n- Coverage 78.29% 78.23% -0.06% \n==========================================\n Files 104 104 \n Lines 17628 17628 \n==========================================\n- Hits 13801 13791 -10 \n- Misses 3827 3837 +10 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3658?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3658/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `90.65% <0.00%> (-0.84%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3658/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `92.63% <0.00%> (-0.66%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3658?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/3658?src=pr&el=footer). Last update [39a34cc...f390cb4](https://codecov.io/gh/huggingface/transformers/pull/3658?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 } ]

[ { "login": "sshleifer", "id": 6045025, "node_id": "MDQ6VXNlcjYwNDUwMjU=", "avatar_url": "https://avatars.githubusercontent.com/u/6045025?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sshleifer", "html_url": "https://github.com/sshleifer", "followers_url": "https://api.github.com/users/sshleifer/followers", "following_url": "https://api.github.com/users/sshleifer/following{/other_user}", "gists_url": "https://api.github.com/users/sshleifer/gists{/gist_id}", "starred_url": "https://api.github.com/users/sshleifer/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sshleifer/subscriptions", "organizations_url": "https://api.github.com/users/sshleifer/orgs", "repos_url": "https://api.github.com/users/sshleifer/repos", "events_url": "https://api.github.com/users/sshleifer/events{/privacy}", "received_events_url": "https://api.github.com/users/sshleifer/received_events", "type": "User", "site_admin": false } ]

[ "You can pass `summarizer(data, min_length=10, max_length=20)` to get a summary whose length is between 10 and 20 tokens. By default, summaries will be between 56 and 142 tokens. ", "Thanks @sshleifer, interestingly now by having a max_length the summary is just arbitrarily cut, which is not great either. Is there a way to constrain the summary length and actually preserve the sense? \r\n\r\n> [{'summary_text': '\"We have a telephony partner who is very interested in this program and may be'}]", "The logic of the program is \"generate the most likely summary\" of between `min_length` and `max_length`. So it's not programmed to cut the summary in a rules based way.\r\n\r\nWith that in mind, I've also seen poor results summarizing documents that are very different than the finetuning distribution (news articles of ~1024 tokens).\r\n\r\nYou *might* get better results with `summarizer = pipeline(task=\"summarization\", model='bart-large-xsum')` .", "> The logic of the program is \"generate the most likely summary\" of between min_length and max_length. So it's not programmed to cut the summary in a rules based way.\r\nThanks for confirming - seems to be the right approach :)! \r\n\r\n> You might get better results with summarizer = pipeline(task=\"summarization\", model='bart-large-xsum') .\r\nOk, will give it a try then! \r\n\r\n> With that in mind, I've also seen poor results summarizing documents that are very different than the finetuning distribution (news articles of ~1024 tokens).\r\nSo you want to keep it open as a bug or should we close? \r\n\r\nAs a side request, it would be awesome to have metrics associated with each models that are part of transformers to help users choose the right one for their job (cc: @julien-c ).\r\n", "Hi @sshleifer Can we increase token length beyond 1024 for generating a summary.\r\nI got the following message while generating a summary of the 20000-word document.\r\n\r\n`Your max_length is set to 1300, but you input_length is only 1024. You might consider decreasing max_length manually, e.g. summarizer('...', max_length=50)`\r\n", "Unfortunately, Bart can only process 1024 tokens at once, so your best best would be to split your doc into chunks, summarize each one, and concatenate the summaries.", "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" } ]

[]

[]

[ { "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/3655?src=pr&el=h1) Report\n> Merging [#3655](https://codecov.io/gh/huggingface/transformers/pull/3655?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/39a34cc375ed79d18888464289b83713fc20f7d4&el=desc) will **increase** coverage by `0.00%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3655?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3655 +/- ##\n=======================================\n Coverage 78.29% 78.29% \n=======================================\n Files 104 104 \n Lines 17628 17628 \n=======================================\n+ Hits 13801 13802 +1 \n+ Misses 3827 3826 -1 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3655?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3655/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.28% <0.00%> (ø)` | |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3655/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `92.23% <0.00%> (+0.12%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3655?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/3655?src=pr&el=footer). Last update [39a34cc...0ee8e8c](https://codecov.io/gh/huggingface/transformers/pull/3655?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" } ]

[]

[]

[ { "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" } ]

[]

[ "[**Model page**](https://huggingface.co/ktrapeznikov/biobert_v1.1_pubmed_squad_v2)" ]

[]

[]

[ "docs are pretty nice imho;", "If I had found what I looking for in the documentation then why would I open an issue and waste someone else's time? I know other approaches to this problem but one seemed to be more time saving so just checking for implementation available. But now it seems that people are more concerned about critical bugs only in the issue section. Closing the issue for good.", "There is no way to dynamically add unknown tokens to the vocabulary. The simplest way to do it would be to encode the sequence, detect unknowns, and then add these to the vocabulary, which seems to be what you did!\r\n\r\nPlease be aware that you will have to resize the model's embedding matrix according to the tokens you've added.", "Or you can map all such tokens a group of them maybe to an OOV kinda token as well. ", "So the only way to update tokenizer is to get all unknowns first then resize model embedding matrix. Thanks @LysandreJik and @AdityaSoni19031997 " ]

[ { "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": 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/3649?src=pr&el=h1) Report\n> Merging [#3649](https://codecov.io/gh/huggingface/transformers/pull/3649?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/2ee410560e45ae3c619dc1e0b0fc4d257c48e18a&el=desc) will **increase** coverage by `0.00%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/3649?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #3649 +/- ##\n=======================================\n Coverage 78.28% 78.29% \n=======================================\n Files 104 104 \n Lines 17628 17628 \n=======================================\n+ Hits 13800 13801 +1 \n+ Misses 3828 3827 -1 \n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/3649?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/3649/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.28% <0.00%> (+0.16%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/3649?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/3649?src=pr&el=footer). Last update [2ee4105...5b63e6a](https://codecov.io/gh/huggingface/transformers/pull/3649?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Awesome, thanks for sharing – also cc @joeddav \r\n\r\n[**Model page**](https://huggingface.co/ixa-ehu/berteus-base-cased)" ]

[]

[]

[ "You might want to check this, It has a very nice example. [https://huggingface.co/transformers/model_doc/bert.html#bertforquestionanswering](https://huggingface.co/transformers/model_doc/bert.html#bertforquestionanswering) ", "> You might want to check this, It has a very nice example. https://huggingface.co/transformers/model_doc/bert.html#bertforquestionanswering\r\n\r\nWell, I've known this solution for quite some time, this give outputs of start and end logits.\r\nBut i'd like to know if there is any implementation like if i just feed in just a corpus of text say a article, essay and so on and ask questions it would give some relevant outout" ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1990944155, "node_id": "MDU6TGFiZWwxOTkwOTQ0MTU1", "url": "https://api.github.com/repos/huggingface/transformers/labels/bertabs", "name": "bertabs", "color": "9ab22e", "default": false, "description": "" } ]

[]

[ "bump. i have the same question. paper says 5 BeamSize and better accuracy for bert-base. ", "same here", "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" ]