Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2312?src=pr&el=h1) Report\n> Merging [#2312](https://codecov.io/gh/huggingface/transformers/pull/2312?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/cea04a244351a7c5bce44e1cfc01abde0ceb60fd?src=pr&el=desc) will **increase** coverage by `<.01%`.\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2312?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2312 +/- ##\n==========================================\n+ Coverage 73.54% 73.54% +<.01% \n==========================================\n Files 87 87 \n Lines 14789 14791 +2 \n==========================================\n+ Hits 10876 10878 +2 \n Misses 3913 3913\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2312?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2312/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `92.11% <100%> (+0.03%)` | :arrow_up: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2312/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `63.91% <0%> (ø)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2312?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/2312?src=pr&el=footer). Last update [cea04a2...b262577](https://codecov.io/gh/huggingface/transformers/pull/2312?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "This is awesome, thanks a lot @vitaliyradchenko " ]

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

[]

[ "You could do something like this when using gpt2\r\n\r\n```\r\nfrom transformers import GPT2LMHeadModel, GPT2Tokenizer\r\nfrom torch.nn import functional as F\r\nimport torch\r\n\r\nmodel = GPT2LMHeadModel.from_pretrained('gpt2-medium')\r\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2-medium')\r\n\r\n# encode input context\r\ninput_ids = torch.tensor(tokenizer.encode('I put the glass of the')).unsqueeze(0)\r\n# get logits of last predicted token\r\nnext_word_logits = model(input_ids)[0][0, -1].detach()\r\nnext_word_probs = F.softmax(next_word_logits, dim=0)\r\n\r\nnext_words = ['desk', 'table', 'car', 'shirt']\r\nnext_words_probs = []\r\n# encode tokens for which prob is to be estimated\r\nnext_word_ids = [tokenizer.encode(next_word) for next_word in next_words]\r\n\r\nfor next_word_id in next_word_ids:\r\n next_word_input_ids = input_ids.clone()\r\n next_word_prob = next_word_probs[next_word_id[0]].item()\r\n # We need a while loop here because a single word can be composed of multiple tokens\r\n # 'desk' is encoded to 2 tokens so that we have to call the model another time\r\n while(len(next_word_id) > 1):\r\n next_word_input_ids = torch.cat((next_word_input_ids, torch.tensor([next_word_id[0]]).unsqueeze(0)), dim=1)\r\n # get logits of last predicted token\r\n next_word_logits = model(next_word_input_ids)[0][0, -1].detach()\r\n # multiply prob of next token to prob of previous tokens\r\n next_word_prob *= F.softmax(next_word_logits, dim=0)[next_word_id[1]].item()\r\n # remove first token since already used\r\n next_word_id = next_word_id[1:]\r\n next_words_probs.append(next_word_prob)\r\n\r\n# print result\r\nfor next_word, next_word_prob in zip(next_words, next_words_probs):\r\n print('{} = {}'.format(next_word, next_word_prob))\r\n```\r\n \r\n", "Yes it is possible u need to take the topk of lm_logits (it will be output[0] in case of gpt)which essentially gives to 50257 probabilities (highest to lowest) which is the vocab size then you need to take top k which gives indices and values, values are nothing but urs scores(0.8, 0.1) and the indices which correspond to the 50257 vocabulary words which u can decode using tokenize decode.", "@patrickvonplaten Amazing thanks!\r\nAnd if I want the rank of these words from all the word in the vocab?\r\ne.g. desk is the most probable word , table in #12 , etc. ?", "Since GPT-2's output is based on byte-pair-encoding tokens and not on words you would have to define your own vocabulary. Having defined your vocabulary, I would simply calculate the probability for each word using the above procedure and then sort the tensor. \r\nTo better understand how byte-pair-encoding works [this](https://leimao.github.io/blog/Byte-Pair-Encoding/) might help. \r\nTo sort the tensor [this](https://stackoverflow.com/questions/56176439/pytorch-argsort-ordered-with-duplicate-elements-in-the-tensor) might help.", "@patrickvonplaten Thanks, you think it will be possible to do it for all (or at least most) of the words in English in my personal MAC?", "Yeah, I think that should definitely be feasible. \r\nMany words will consists of two tokens or less and will therefore need at most one additional forward pass (because the first forward pass is the same for all words and need to be calculated only once). \r\n\r\nSo if you have a vocabulary of say 300.000 words, I'd estimate that you would have to compute around 200.000 forward passes. You can calculate how much time a forward pass would take by averaging the computation time for 100 times calculating the probability for the word 'desk'. \r\n\r\nConcerning memory, there should not be a problem.", "And the final vector giving the probabilities over your defined vocabulary should be normalized to make a prob distribution.", "@patrickvonplaten You mean using softmax?", "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 was thinking to just normalize like this:\r\nhttps://stackoverflow.com/questions/26785354/normalizing-a-list-of-numbers-in-python\r\n\r\nbut you could also use softmax again - depends on what you want and what works better for you! ", "@patrickvonplaten is it possible with BERT pre-trained model?\r\nThanks!", "You might take a look at masked language modeling :-) https://huggingface.co/transformers/usage.html#masked-language-modeling", "@patrickvonplaten Nice! Thanks for the pointer!\r\nAnd let's say I want to check a specific word in a masked location (What is the probability of the word \"`package` \" in the sequence \"`HuggingFace is creating a { } that the community uses to`\"? Is this possible?" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2310?src=pr&el=h1) Report\n> Merging [#2310](https://codecov.io/gh/huggingface/transformers/pull/2310?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/81db12c3ba0c2067f43c4a63edf5e45f54161042?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `0%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2310?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2310 +/- ##\n=======================================\n Coverage 73.54% 73.54% \n=======================================\n Files 87 87 \n Lines 14789 14789 \n=======================================\n Hits 10876 10876 \n Misses 3913 3913\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2310?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2310/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `63.91% <0%> (ø)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2310?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/2310?src=pr&el=footer). Last update [81db12c...e1844d9](https://codecov.io/gh/huggingface/transformers/pull/2310?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hi, probably best to just use positional arguments here (instead of keywords) then, don't you think?", "Great, thanks!" ]

[]

[]

[]

[]

[]

[]

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

[]

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

[]

[]

[ "As an update I've been able to reproduce this problem using a python script", "I cannot reproduce this on Windows, PT1.4, latest transformers master.", "I think I was using an old version of transformers :( \r\n\r\nThis seems to have been fixed in v2.2.2 - After upgrading to latest I haven't observed this anymore so I'll close this issue.\r\n\r\nThanks!" ]

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

[]

[ "[CLS], [SEP] are \"special token\" of bert. Its included in the vocabulary of size 30522. Its starts with like 101, 103 something.", "of course these two tokens are special tokens. \r\n\r\n **is used as the aggregate sequence representation for classification tasks.**\r\n\r\nin which way does happen this aggregations?\r\n\r\n>>The first token of every sequence is always a special classification token ([CLS]). The final hidden state corresponding to this token is used as the aggregate sequence representation for classification tasks.\r\n\r\nfrom papert https://arxiv.org/pdf/1810.04805.pdf", "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" ]

[]

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

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2303?src=pr&el=h1) Report\n> Merging [#2303](https://codecov.io/gh/huggingface/transformers/pull/2303?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/81db12c3ba0c2067f43c4a63edf5e45f54161042?src=pr&el=desc) will **decrease** coverage by `0.01%`.\n> The diff coverage is `0%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2303?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2303 +/- ##\n==========================================\n- Coverage 73.54% 73.52% -0.02% \n==========================================\n Files 87 87 \n Lines 14789 14793 +4 \n==========================================\n Hits 10876 10876 \n- Misses 3913 3917 +4\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2303?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2303/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `63.45% <0%> (-0.46%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2303?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/2303?src=pr&el=footer). Last update [81db12c...18e5bdb](https://codecov.io/gh/huggingface/transformers/pull/2303?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Good catch.\r\nBut this is actually the technique mentioned in http://arxiv.org/abs/1909.05858.\r\nSo to fix it we should check the code of Nitish (https://github.com/salesforce/ctrl) and apply the same behavior here.", "I checked the code in https://github.com/salesforce/ctrl/blob/0f30306a8947ce0ede62e79c7e1f05a585cc56c9/generation.py#L217: \r\n`prompt_logits[_token][generated_token] /= penalty`\r\n\r\nSo in the original code division is always used no matter what sign the `prompt_logit` of the previously generated tokens. \r\n\r\nWhen going a bit deeper and looking at the actual values of the logit in \r\nhttps://github.com/huggingface/transformers/blob/81db12c3ba0c2067f43c4a63edf5e45f54161042/src/transformers/modeling_utils.py#L731\r\nfor different models the following can be observed:\r\n\r\nFor the models: **ctrl**, **xlm** the logit values tend to be positive, which explains why division by the `repetition penalty` is used. BUT, the values don't have to be positive, there were also very rare cases when using **ctrl** where the logit was actually negative in which case a division increases the probability of that word to be sampled. \r\n\r\nFor the models: **gpt2**, **openai-gpt**, **xlnet** the logit values tend to be negative, in which case dividing by a `repetition penalty` increases the probability of previously generated tokens to be sampled. \r\n\r\nIn the proposed PR, both cases would be correctly handled from a logical point of view. \r\nIf we want to stick to the original code on the other hand (only using division) we could add a warning that the `repetition penalty` should only be used in combination with **ctrl**.", "Ok, I see, thanks for documenting this. Let's go for this solution for now.", "Is this fix added to the pip package? So if we use pip install package this will be covered or not yet I have to install from source? ", "Reading this after it was mentioned in the PPLM example PR.\r\nThe fix makes total sense, but I have a concern: the amount by which a negative number is diminished is greater than the amount a positive number is diminished.\r\nIf we have two values, say -2 and 2 this happens:\r\n```\r\nx = np.array([-2, 2])\r\nsx = np.exp(x)/sum(np.exp(x))\r\nprint(sx) # array([0.01798621, 0.98201379])\r\n```\r\nif we apply the same penalty to both, we would want the probabilities to stay the same, but this is what happens:\r\n```\r\np = [1/1.2, 1.2]\r\nspx = np.exp(x/p)/sum(np.exp(x/p))\r\nprint(spx) # array([0.01684577, 0.98315423])\r\n```\r\nOn the other hand, if we apply the penalty to the probabilities after the softmax (and we renormalize) this is what happens:\r\n```\r\np2 = [1.2, 1.2]\r\nsp2x = (sx/p2)/sum(sx/p2)\r\nprint(sp2x) # array([0.01798621, 0.98201379])\r\n```\r\nThe probabilities are intact, as we want, because we don't want to penalize negative values more than we penalize positive values.\r\nSo my proposal is to perform the penalty after the softmax, on probability values, always dividing, rather than on the logits.\r\nWhat do you think?\r\n\r\nEdit:\r\nIn math i propose to move from:\r\n\r\nto:\r\n\r\n", "Sorry for the late response @w4nderlust ! \r\n\r\nI think you it makes a lot of sense what you are saying! \r\n\r\nTo implement your solution with minimal code change one could simply change Eq. (1): \r\n\r\n\r\n\r\nto the equivalent Eq. (2)\r\n\r\n\r\n\r\nOne question that remains is how the new repetition penalties  in Eq. (1) & (2) will have to differ from the old repetition penalties  in Eq. (3):\r\n\r\n\r\n\r\nto have a similar effect on the softmax. It is quite obvious that  reduces the prob of its token much more than \r\n\r\nFor the different LMHead models, I calculated  for different values of  . I simply generated randomly sampled sentences from the pretrained models and averaged the effect of the tokens for 5 runs with `max_length=100` so that the averaged is formed of ca.  tokens. \r\n\r\nThe following values show by how much  scales down the prob after the softmax which is equivalent of what  would have been set to:\r\n\r\n```\r\nGenerate repetition penalty comparison for ctrl\r\nPenalty factor: 1.1 - Without penalty / penalty ratio avg: 4e0\r\nPenalty factor: 1.2 - Without penalty / penalty ratio avg: 31e0\r\nPenalty factor: 1.3 - Without penalty / penalty ratio avg: 149e0\r\nPenalty factor: 1.4 - Without penalty / penalty ratio avg: 25e3\r\nPenalty factor: 1.5 - Without penalty / penalty ratio avg: 286e3\r\nGenerate repetition penalty comparison for distilgpt2\r\nPenalty factor: 1.1 - Without penalty / penalty ratio avg: 23e3\r\nPenalty factor: 1.2 - Without penalty / penalty ratio avg: 2e9\r\nPenalty factor: 1.3 - Without penalty / penalty ratio avg: 223e9\r\nPenalty factor: 1.4 - Without penalty / penalty ratio avg: 3e24\r\nGenerate repetition penalty comparison for gpt2\r\nPenalty factor: 1.1 - Without penalty / penalty ratio avg: 1e9\r\nPenalty factor: 1.2 - Without penalty / penalty ratio avg: 742e18\r\nGenerate repetition penalty comparison for xlm-clm-enfr-1024\r\nPenalty factor: 1.1 - Without penalty / penalty ratio avg: 2e0\r\nPenalty factor: 1.2 - Without penalty / penalty ratio avg: 3e0\r\nPenalty factor: 1.3 - Without penalty / penalty ratio avg: 5e0\r\nPenalty factor: 1.4 - Without penalty / penalty ratio avg: 9e0\r\nPenalty factor: 1.5 - Without penalty / penalty ratio avg: 13e0\r\nGenerate repetition penalty comparison for openai-gpt\r\nPenalty factor: 1.1 - Without penalty / penalty ratio avg: 1e0\r\nPenalty factor: 1.2 - Without penalty / penalty ratio avg: 2e0\r\nPenalty factor: 1.3 - Without penalty / penalty ratio avg: 4e0\r\nPenalty factor: 1.4 - Without penalty / penalty ratio avg: 15e0\r\nPenalty factor: 1.5 - Without penalty / penalty ratio avg: 19e0\r\nGenerate repetition penalty comparison for xlnet-base-cased\r\nPenalty factor: 1.1 - Without penalty / penalty ratio avg: 5e0\r\nPenalty factor: 1.2 - Without penalty / penalty ratio avg: 34e0\r\nPenalty factor: 1.3 - Without penalty / penalty ratio avg: 2e3\r\nPenalty factor: 1.4 - Without penalty / penalty ratio avg: 47e3\r\nPenalty factor: 1.5 - Without penalty / penalty ratio avg: 8e6\r\n```\r\n\r\nIt can be seen that `gpt2` for example produces much larger logit values which lead to much more drastic reductions in the prob after softmax. The repetition penalty was originally introduced for `ctrl` so it's probably best to look at its behaviour.\r\n\r\n\r\n\r\n", "So I think there are three possibilities:\r\n\r\n1) Follow the proposed solution from @w4nderlust implementing Eq.(1). \r\nThis would mean though that the proposed repetition penalty of 1.3 in the ctrl paper would have to be changed to something around 150 which is quite a large value. \r\n\r\n2) Instead of using substracting by the log(rep_penalty) as in: \r\n, \r\none could only substract by the rep_penalty to give the equation: \r\n,\r\nThis way the values for  would equal  and thus be much smaller. The repetition penalty in `ctlr` would thus only have to be around 5 to equal the behavior of the old penalty of 1.3. One disadvantage would be that the neutral element in this case is 0 instead of 1 which might be a bit confusing. \r\n\r\n3) Just leave as it is now since from what I seen most logits almost always all either positive or either all negative, so that the current behavior is not very prone to lead to errors. \r\n\r\nI would tend to solution 2, giving a clear explanation of the variable in the argument section of the language generation function. \r\n\r\nWhat do you think @w4nderlust and @thomwolf ?\r\n\r\n\r\n\r\n", "Thank you for the thorough analysis @patrickvonplaten ! I believe 2 would be fine. The nog just scales things differently, but there's no specific reason to have it, as it is a user tunable parameter anyway. The fact that the default would be 0 instead of one I think could be explained and one could point to this conversation in a comment to give the full picture. Although I understand this is not a huge issue (because of what you say in 3), I kinda believe 2 is better as the could potentially be in the future a different model that actually outputs both positive and negative logits and it that case this could make a substantial difference in the quality of the sampling. " ]

[]

[]

[ "Propsed fix in PR #2303 " ]

[]

[]

[ "You can now leave `--model_name_or_path` to None in `run_language_modeling.py` to train a model from scratch.\r\n\r\nSee also https://huggingface.co/blog/how-to-train" ]

[]

[]

[]

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

[]

[ "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 } ]

[]

[ "BERT was not designed to produce useful word / sentence embeddings that can be used with cosine similarities. Cosine-similarity treats all dimensions equally which puts high requirements for the created embeddings.\r\n\r\nBERT as not intended for this. See this post by Jacob Devlin:\r\nhttps://github.com/UKPLab/sentence-transformers/issues/80#issuecomment-565388257\r\n\r\nIf you want to use BERT with cosine similarities, you need to fine-tune it on suitable data. You can find data, code and examples in our repository:\r\nhttps://github.com/UKPLab/sentence-transformers", "@nreimers I have read your paper, it's great and thanks for the answer!", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "Someone suggested to use Pytorch v1.1.0 instead of 1.2.0. But not sure if it is ok. ", "it's a version inconsistent issue. \r\nin ver 1.1.0, torch.gt outputs:\r\ntorch.gt(torch.tensor([[1, 2], [3, 4]]), torch.tensor([[1, 1], [4, 4]]))\r\ntensor([[ 0, 1],\r\n [ 0, 0]], dtype=torch.uint8)\r\nwhile in ver 1.2.0, it outputs:\r\n>>> torch.ge(torch.tensor([[1, 2], [3, 4]]), torch.tensor([[1, 1], [4, 4]]))\r\ntensor([[True, True], [False, True]])\r\n", "Please see the pull request #2369" ]

[]

[]

[ "The sentence \"Today is a nice day\" will already be padded to 512 because of the tokenization process, so the input will remain 512 in the embedding layer", "> The sentence \"Today is a nice day\" will already be padded to 512 because of the tokenization process, so the input will remain 512 in the embedding layer\r\n\r\nThanks a lot! " ]

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

[]

[ "apparently the answer may be to feed smaller sequences of tokens and use the past input keyword itn pytorch models or hidden states in tensorflow. models both this past input and the stateful nature of models aren't documented. it would be interesting to have methods to manage big inputs", "Recent models like Transformers-XL and XLNet already support longer sequences. Although, the available pretrained models are imho only using 512 tokens. \r\n\r\nSome additional pointers:\r\n- Long-form document classification with BERT. [Blogpost](https://andriymulyar.com/blog/bert-document-classification), [Code](https://github.com/AndriyMulyar/bert_document_classification)\r\n- See ICLR 2020 reviews: \r\n - [BERT-AL: BERT for Arbitrarily Long Document Understanding](https://openreview.net/forum?id=SklnVAEFDB)\r\n - [Blockwise Self-Attention for Long Document Understanding](https://openreview.net/forum?id=H1gpET4YDB)\r\n- [Easy-to-use interface to fine-tuned BERT models for computing semantic similarity](https://github.com/AndriyMulyar/semantic-text-similarity)\r\n- Ye, Z. et al. 2019. BP-Transformer: Modelling Long-Range Context via Binary Partitioning. (2019). [Paper](https://arxiv.org/pdf/1911.04070.pdf) [Code](https://github.com/yzh119/BPT)\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "There are two main methods:\r\n- Concatenating 'short' BERT altogether (which consists of 512 tokens max)\r\n- Constructing a real long BERT (CogLTX, Blockwise BERT, Longformer, Big Bird)\r\n\r\nI resumed some typical papers of BERT for long text in this post : [Paper Dissected and Recap #4 : which BERT for long text ?](https://lethienhoablog.wordpress.com/2020/11/19/paper-dissected-and-recap-4-which-bert-for-long-text/)\r\nYou can have an overview of all methods there.", "@lethienhoa does `all long BERT` is capable of any length text?" ]

[]

[]

[ "i am also curious about it. ;;\r\nhttps://github.com/dsindex/etagger/blob/master/feed.py\r\nthis is not efficient way. just pooling.", "> i am also curious about it. ;;\r\n> https://github.com/dsindex/etagger/blob/master/feed.py\r\n> this is not efficient way. just pooling.\r\n\r\nHi, I think I get the solution. The problem can be solved via simple matrix computation.\r\nFor example, let the BERT representation of the above example be B, which has a size of (5, 100). \r\nWe should first construct a matrix according to the sub-words:\r\nm = \r\n[ 1, 1, 1, 1, 0\r\n 0, 0, 0, 0, 1] \r\nThen we can simply compute m.dot(B), which is exactly the result.\r\n\r\n", "I have solved this issue." ]

[ { "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." } ]

[]

[ "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" ]

[]

[]

[ "That a great idea, @patrickvonplaten \r\n\r\nI'll let you finish this PR when you have time and ping me for review or questions.", "True, I will implement this tomorrow!", "tested for transfo_xl, gpt2, openai-gpt and xlnet in combination with PR #2289 ", "This looks great!\r\nTo pass the code quality test, you can use `make style`.\r\nPlease read this section of the (new) CONTRIBUTING guidelines: https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#start-contributing-pull-requests", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2292?src=pr&el=h1) Report\n> Merging [#2292](https://codecov.io/gh/huggingface/transformers/pull/2292?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/aeef4823ab6099249679756182700e6800024c36?src=pr&el=desc) will **decrease** coverage by `0.17%`.\n> The diff coverage is `8.51%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2292?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2292 +/- ##\n==========================================\n- Coverage 73.49% 73.32% -0.18% \n==========================================\n Files 87 87 \n Lines 14793 14833 +40 \n==========================================\n+ Hits 10872 10876 +4 \n- Misses 3921 3957 +36\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2292?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2292/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG5ldC5weQ==) | `72.9% <0%> (-0.37%)` | :arrow_down: |\n| [src/transformers/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/transformers/pull/2292/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19ncHQyLnB5) | `83.17% <16.66%> (-1.3%)` | :arrow_down: |\n| [src/transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/2292/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19jdHJsLnB5) | `94.22% <16.66%> (-2.13%)` | :arrow_down: |\n| [src/transformers/modeling\\_transfo\\_xl.py](https://codecov.io/gh/huggingface/transformers/pull/2292/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190cmFuc2ZvX3hsLnB5) | `74.68% <20%> (-0.59%)` | :arrow_down: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2292/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `61.44% <3.84%> (-2.02%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2292?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/2292?src=pr&el=footer). Last update [aeef482...fc84bd5](https://codecov.io/gh/huggingface/transformers/pull/2292?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Ok good for now, merging" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2291?src=pr&el=h1) Report\n> Merging [#2291](https://codecov.io/gh/huggingface/transformers/pull/2291?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/072750f4dc4f586cb53f0face4b4a448bb0cdcac?src=pr&el=desc) will **decrease** coverage by `1.18%`.\n> The diff coverage is `50%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2291?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2291 +/- ##\n==========================================\n- Coverage 74.45% 73.26% -1.19% \n==========================================\n Files 85 85 \n Lines 14608 14603 -5 \n==========================================\n- Hits 10876 10699 -177 \n- Misses 3732 3904 +172\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2291?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [...c/transformers/modeling\\_tf\\_transfo\\_xl\\_utilities.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl90cmFuc2ZvX3hsX3V0aWxpdGllcy5weQ==) | `86.13% <ø> (+0.84%)` | :arrow_up: |\n| [src/transformers/modeling\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19hbGJlcnQucHk=) | `78.86% <ø> (-0.18%)` | :arrow_down: |\n| [src/transformers/data/metrics/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL21ldHJpY3MvX19pbml0X18ucHk=) | `27.9% <ø> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_pytorch\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9weXRvcmNoX3V0aWxzLnB5) | `8.72% <0%> (-81.28%)` | :arrow_down: |\n| [src/transformers/modeling\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190NS5weQ==) | `81.05% <100%> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl90NS5weQ==) | `96.54% <100%> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.12% <66.66%> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19yb2JlcnRhLnB5) | `54.1% <0%> (-10.15%)` | :arrow_down: |\n| [src/transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG5ldC5weQ==) | `71.19% <0%> (-2.32%)` | :arrow_down: |\n| ... and [5 more](https://codecov.io/gh/huggingface/transformers/pull/2291/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2291?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/2291?src=pr&el=footer). Last update [072750f...3e0cf49](https://codecov.io/gh/huggingface/transformers/pull/2291?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2290?src=pr&el=h1) Report\n> Merging [#2290](https://codecov.io/gh/huggingface/transformers/pull/2290?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/aeef4823ab6099249679756182700e6800024c36?src=pr&el=desc) will **decrease** coverage by `<.01%`.\n> The diff coverage is `0%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2290?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2290 +/- ##\n==========================================\n- Coverage 73.49% 73.48% -0.01% \n==========================================\n Files 87 87 \n Lines 14793 14794 +1 \n==========================================\n Hits 10872 10872 \n- Misses 3921 3922 +1\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2290?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2290/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `63.34% <0%> (-0.12%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2290?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/2290?src=pr&el=footer). Last update [aeef482...0f6017b](https://codecov.io/gh/huggingface/transformers/pull/2290?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Yes, actually the doc is still not complete for this new feature.\r\n\r\nWe should add some examples and double-check all. Feel free to clean this up if you feel like it.", "clean up documentation, add examples for documentation and rename some variables", "checked example generation for openai-gpt, gpt2, xlnet and xlm in combination with #2289.\r\n", "Also checked for ctrl", "also checked for transfo-xl", "Awesome, merging!" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2289?src=pr&el=h1) Report\n> Merging [#2289](https://codecov.io/gh/huggingface/transformers/pull/2289?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/81db12c3ba0c2067f43c4a63edf5e45f54161042?src=pr&el=desc) will **decrease** coverage by `0.01%`.\n> The diff coverage is `0%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2289?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2289 +/- ##\n==========================================\n- Coverage 73.54% 73.52% -0.02% \n==========================================\n Files 87 87 \n Lines 14789 14792 +3 \n==========================================\n Hits 10876 10876 \n- Misses 3913 3916 +3\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2289?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/2289/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG0ucHk=) | `86.23% <0%> (-0.23%)` | :arrow_down: |\n| [src/transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2289/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ194bG5ldC5weQ==) | `73.26% <0%> (-0.25%)` | :arrow_down: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2289?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/2289?src=pr&el=footer). Last update [81db12c...f18ac4c](https://codecov.io/gh/huggingface/transformers/pull/2289?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Indeed, thanks @patrickvonplaten " ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2288?src=pr&el=h1) Report\n> Merging [#2288](https://codecov.io/gh/huggingface/transformers/pull/2288?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/23dad8447c8db53682abc3c53d1b90f85d222e4b?src=pr&el=desc) will **increase** coverage by `0.2%`.\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2288?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2288 +/- ##\n=========================================\n+ Coverage 74.27% 74.47% +0.2% \n=========================================\n Files 85 85 \n Lines 14610 14608 -2 \n=========================================\n+ Hits 10851 10879 +28 \n+ Misses 3759 3729 -30\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2288?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdDUucHk=) | `93.93% <ø> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `63.91% <ø> (ø)` | :arrow_up: |\n| [src/transformers/commands/train.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb21tYW5kcy90cmFpbi5weQ==) | `0% <ø> (ø)` | :arrow_up: |\n| [src/transformers/tokenization\\_albert.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYWxiZXJ0LnB5) | `89.1% <ø> (ø)` | :arrow_up: |\n| [src/transformers/pipelines.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9waXBlbGluZXMucHk=) | `68.31% <ø> (+2.32%)` | :arrow_up: |\n| [src/transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3F1YWQucHk=) | `29.37% <ø> (ø)` | :arrow_up: |\n| [src/transformers/data/processors/utils.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvdXRpbHMucHk=) | `19.6% <ø> (ø)` | :arrow_up: |\n| [src/transformers/tokenization\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25feGxtLnB5) | `83.26% <ø> (ø)` | :arrow_up: |\n| [src/transformers/data/metrics/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL21ldHJpY3MvX19pbml0X18ucHk=) | `27.9% <ø> (-3.21%)` | :arrow_down: |\n| [src/transformers/tokenization\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25feGxuZXQucHk=) | `89.9% <ø> (ø)` | :arrow_up: |\n| ... and [7 more](https://codecov.io/gh/huggingface/transformers/pull/2288/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2288?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/2288?src=pr&el=footer). Last update [23dad84...4621ad6](https://codecov.io/gh/huggingface/transformers/pull/2288?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "Indeed, as you can see from the source code [here](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_gpt2.py#L414-L417), when no position ids are passed, they are created as absolute position embeddings.\r\n\r\nYou could have trained a model with a GPT-2 architecture that was using an other type of position embeddings, in which case passing your specific embeddings would be necessary. I'm sure several other use-cases would make sure of specific position embeddings.", "Ooohh, ok,\r\n\r\nso to clarify, absolute position embedding _**is automatically done**_ by the ```model( )``` statement, but if we want to use our custom position embedding (i.e. other than the absolute position embedding), we can use the ```position_ids``` option inside the ```model( )``` statement......is what I said above correct?\r\n\r\nThank you,", "Yes, that is correct!", "Thank you :) !" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2286?src=pr&el=h1) Report\n> Merging [#2286](https://codecov.io/gh/huggingface/transformers/pull/2286?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/23dad8447c8db53682abc3c53d1b90f85d222e4b?src=pr&el=desc) will **increase** coverage by `0.18%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2286?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2286 +/- ##\n==========================================\n+ Coverage 74.27% 74.45% +0.18% \n==========================================\n Files 85 85 \n Lines 14610 14610 \n==========================================\n+ Hits 10851 10878 +27 \n+ Misses 3759 3732 -27\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2286?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2286/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `92.08% <ø> (+0.77%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2286/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.12% <0%> (+1.58%)` | :arrow_up: |\n| [src/transformers/pipelines.py](https://codecov.io/gh/huggingface/transformers/pull/2286/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9waXBlbGluZXMucHk=) | `68.31% <0%> (+2.32%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/2286/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9hdXRvLnB5) | `32.14% <0%> (+7.14%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2286?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/2286?src=pr&el=footer). Last update [23dad84...7cef764](https://codecov.io/gh/huggingface/transformers/pull/2286?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 } ]

[]

[ "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 actually have the same problem with GPT-2 tokenizer. Is this the expected behavior?" ]

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

[]

[ "i have similar situation.\r\n\r\nhttps://github.com/dsindex/iclassifier#emb_classalbert\r\n\r\nin the paper( https://arxiv.org/pdf/1909.11942.pdf ), ALBERT xlarge has just 60M parameters which is much less than BERT large(334M)'s. \r\nbut, we are unable to load albert-xlarge-v2 on 32G GPU memory.\r\n(no problem on bert-large-uncased, bert-large-cased)", "A similar situation happened to me too.\r\nWhile fine-tuning Albert base on SQuAD 2.0, I had to lower the train batch size to manage to fit the model on 2x NVIDIA 1080 Ti, for a total of about 19 GB used.\r\nI find it quite interesting and weird as the same time, as I managed to fine-tune BERT base on the same dataset and the same GPUs using less memory...", "Same for the pytorch version of ALBERT, where my 8/11GB GPU could run BERT_base and RoBERTa.", "Interesting, I started to hesitate on using this ALBERT implementation but hope it will be fixed soon.", "Indeed, I can reproduce for the TensorFlow version. I'm looking into it, thanks for raising this issue.", "@jonanem, if you do this at the beginning of your script, does it change the amount of memory used?\r\n\r\n```py\r\ngpus = tf.config.experimental.list_physical_devices('GPU')\r\ntf.config.experimental.set_virtual_device_configuration(\r\n gpus[0],\r\n [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)]\r\n)\r\n```\r\n\r\nThis should keep the amount of memory allocated to the model to 1024MB, with possibility to grow if need be. Initializing the model after this only uses 1.3GB of VRAM on my side. Can you reproduce?\r\n\r\nSee this for more information: [limiting gpu memory growth](https://www.tensorflow.org/guide/gpu#limiting_gpu_memory_growth)", "@LysandreJik I just did some investigation and I found a similar problem with the Pytorch implementation.\r\nModel: ALBERT base v2, fine tuning on SQuAD v2 task\r\n\r\nI used the official code from Google Tensorflow repository and I managed to fine tune it on a single GTX 1080 Ti, with batch size 16 and memory consumption of about 10 GB.\r\nThen, I used transformers Pytorch implementation and did the same task on 4x V100 on AWS, with total batch size 48 and memory consumption of 52 GB (about 13 GB per GPU).\r\n\r\nNow, putting it in perspective, I guess the memory consumption of the Pytorch implementation is 10/15 GB above what I was expecting. Is this normal?\r\nIn particular, where in the code is there the Embedding Factorization technique proposed in the official paper?", "Hi @matteodelv, I ran a fine-tuning task on ALBERT (base-v2) with the parameters you mentioned: batch size of 16. I end up with a VRAM usage of 11.4GB, which is slightly more than the official Google Tensorflow implementation you mention. The usage is lower than when using BERT, which has a total usage of 14GB.\r\n\r\nHowever, when loading the model on its own without any other tensors, taking into account the pytorch memory overhead, it only takes about 66MB of VRAM.\r\n\r\nConcerning your second question, here is the definition of the Embedding Factorization technique proposed in the official paper: `[...] The first one is a factorized embedding parameterization. By decomposing\r\nthe large vocabulary embedding matrix into two small matrices, we separate the size of the hidden\r\nlayers from the size of vocabulary embedding.`\r\n\r\nIn this PyTorch implementation, there are indeed two smaller matrices so that the two sizes may be separate. The first embedding layer is visible in [the `AlbertEmbeddings` class](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_albert.py#L172), and is of size `(vocab_size, embedding_size)`, whereas the second layer is visible in [the `AlbertTransformer` class](https://github.com/huggingface/transformers/blob/master/src/transformers/modeling_albert.py#L317), with size `(embedding_size, hidden_size)`.", "Thanks for your comment @LysandreJik... I haven't looked in the `AlbertTransformer` class for the embedding factorization.\r\n\r\nHowever, regarding the VRAM consumption, I'm still a bit confused about it.\r\nI don't get why the same model with a batch size 16 consumes about 10/11 GB on a single GPU while the same training, on 4 GPUs (total batch size 48, so it's 12 per GPUs) requires more memory.\r\n\r\nCould you please check this? May it be related to Pytorch's `DataParallel`?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Did @matteodelv @LysandreJik find any issue or solution for this? The memory consumption given the parameter is insane", "Unfortunately not. I had to tune hyperparameters or use other hardware with more memory. But I was using an older version... I haven't checked if the situation has changed since then.", "Hey,\r\nI tried running on GTX 1080 (10GB) bert-base-uncased with **sucess** on IMDB dataset with a batch-size equal to 16 and sequence length equal to 128.\r\nRunning albert-base-v2 with the same sequence length and same batch size is giving me Out-of-memory issues.\r\n\r\nI am using pytorch, so I guess I have the same problem as you guys here.", "Same issue. ALBERT raises OOM requiring 32G. ", "ALBERT repeats the same parameters for each layer but increases each layer size, so even though it have fewer parameters than BERT, the memory needs are greater due to the much larger activations in each layer.", "> ALBERT repeats the same parameters for each layer but increases each layer size, so even though it have fewer parameters than BERT, the memory needs are greater due to the much larger activations in each layer.\r\n\r\nThat is true, still there is need for more computation, but BERT can fit into 16G memory. I had my albert reimplemented differently and I could fit its weights on a 24G gpu.", "> ALBERT repeats the same parameters for each layer but increases each layer size, so even though it have fewer parameters than BERT, the memory needs are greater due to the much larger activations in each layer.\r\n\r\nThanks for this explanation, which saves my life." ]

[]

[]

[ "Try to use the following commands:\r\n\r\n```bash\r\n$ wget \"https://s3-us-west-2.amazonaws.com/ai2-s2-research/scibert/huggingface_pytorch/scibert_scivocab_uncased.tar\"\r\n$ tar -xf scibert_scivocab_uncased.tar\r\n\r\nThe sciBERT model is now extracted and located under: `./scibert_scivocab_uncased`.\r\n\r\nTo load it:\r\n\r\n```python\r\nfrom transformers import BertModel\r\n\r\nmodel = BertModel.from_pretrained(\"./scibert_scivocab_uncased\")\r\nmodel.eval()\r\n```\r\n\r\nThis should work 🤗", "> Try to use the following commands:\r\n> \r\n> ```shell\r\n> $ wget \"https://s3-us-west-2.amazonaws.com/ai2-s2-research/scibert/huggingface_pytorch/scibert_scivocab_uncased.tar\"\r\n> $ tar -xf scibert_scivocab_uncased.tar\r\n> \r\n> The sciBERT model is now extracted and located under: `./scibert_scivocab_uncased`.\r\n> \r\n> To load it:\r\n> \r\n> ```python\r\n> from transformers import BertModel\r\n> \r\n> model = BertModel.from_pretrained(\"./scibert_scivocab_uncased\")\r\n> model.eval()\r\n> ```\r\n> \r\n> This should work 🤗\r\n\r\nIt works! Thank you very much!" ]

[]

[]

[]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2281?src=pr&el=h1) Report\n> Merging [#2281](https://codecov.io/gh/huggingface/transformers/pull/2281?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/ba2378ced560c12f8ee97ca7998fd28b93fcfb47?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2281?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2281 +/- ##\n=======================================\n Coverage 74.45% 74.45% \n=======================================\n Files 85 85 \n Lines 14610 14610 \n=======================================\n Hits 10878 10878 \n Misses 3732 3732\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2281?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/tokenization\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/2281/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fYmVydC5weQ==) | `96.34% <ø> (ø)` | :arrow_up: |\n| [src/transformers/configuration\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/2281/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2JlcnQucHk=) | `100% <ø> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/2281/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9iZXJ0LnB5) | `96.26% <ø> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_bert.py](https://codecov.io/gh/huggingface/transformers/pull/2281/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19iZXJ0LnB5) | `87.7% <ø> (ø)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2281?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/2281?src=pr&el=footer). Last update [ba2378c...5eb71e6](https://codecov.io/gh/huggingface/transformers/pull/2281?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hi @wietsedv! Did you try loading your tokenizer/model directly, cf. https://huggingface.co/wietsedv/bert-base-dutch-cased\r\n\r\ni.e. It should work out-of-the-box using:\r\n```python\r\ntokenizer = AutoTokenizer.from_pretrained(\"wietsedv/bert-base-dutch-cased\")\r\n\r\nmodel = AutoModel.from_pretrained(\"wietsedv/bert-base-dutch-cased\")\r\ntf_model = TFAutoModel.from_pretrained(\"wietsedv/bert-base-dutch-cased\")\r\n```\r\n\r\nLet us know if it's not the case (we can still merge this PR to have a nice shortcut inside the code but feature-wise it should be equivalent)", "Hi! Thanks for your response. I did notice that I can use that snippet and I can confirm it works. I am however not sure whether cased tokenization works correctly that way. Correct me if I am wrong, but it seems that Transformers always lowercases unless that is explicitly disabled. I did disable it in this PR, which makes it work correctly out of the box. But I think lowercasing is enabled by default if people use the snippet above.\r\n\r\nPlease correct my if I am wrong.", "We'll check, thanks for the report.", "Just a note: I absolutely love seeing a Dutch version of BERT but this isn't the only BERT model out there. As you mention in your paper, there's also [BERT-NL](http://textdata.nl/). You seem to claim that it performs a lot worse than your version and that it is even outperformed by multilingual BERT. At first glance I don't see any written-down experiments confirming that claim - a comparison between BERTje, BERT-NL, and multilingual BERT on down-stream tasks would've been much more informative. (BERT-NL will be presented at the largest computational linguistics conference on Dutch (Computational Linguistics in the Netherlands; CLIN) at the end of the month, so presumably it does carry some weight.)\r\n\r\nAll this to say: why does your version deserve to be \"the\" `bert-base-dutch-cased` model if there is an alternative? Don't get me wrong, I really value your research, but a fair and full comparison is missing.", "It is correct that there is another Dutch model and we do full fine-tuning results of this model. These numbers were included in an earlier draft of our paper, but we removed it since they did not add any value. For instance for named entity recognition (conll2002), multilingual BERT achieves about 80% accuracy, our BERT model about 88% and their BERT model just 41%.\r\n\r\nMore detailed comparison with their model would therefore not add any value since the scores are too low. The authors have not made any claims about model performance yet, so it would have been unfair to be too negative about their model before they have even released any paper.\r\n\r\nSomeone at Leiden has confirmed that their experiments also showed that they were outperformed by multilingual BERT. Therefore I think that our BERT model is the only _effective_ Dutch BERT model.\r\n\r\nPS: The entry barrier for CLIN is not extremely high. The reason we are not presenting at CLIN is that we missed the deadline. ", "I think it is in fact very useful information to see such huge difference. This way readers are not confronted with the same question that I posed before: there are two Dutch BERT models - but which one should I use/which one is better? You now clarify it to me, for which I am grateful, but a broader audience won't know. I think the added value is great. However, I do agree that is hard and perhaps unfair to compare to a model that hasn't been published/discussed yet. (Then again, their model is available online so they should be open to criticism.)\r\n\r\nThe CLIN acceptance rate is indeed high, and the follow-up CLIN Journal is also not bad. Still, what I was aiming for is peer review. If I were to review the BERT-NL (assuming they will submit after conference and assuming I was a reviewer this year), then I would also mention your model and ask for a comparison. To be honest, I put more faith in a peer reviewed journal/conference than arXiv papers.\r\n\r\nI really don't want to come off as arrogant and I very much value your work, but I am trying to approach this from a person who is just getting started with this kind of stuff and doesn't follow the trends or what is going on in the field. They might find this model easily available in the Transformers hub, but then they might read in a journal (possibly) about BERT-NL - which then is apparently different from the version in Transformers. On top of that, neither paper (presumably) refers or compares to the other! Those people _must be confused_ by that.\r\n\r\nThe above wall of text just to say that I have no problem with your model being \"the\" Dutch BERT model because it seems to clearly be the best one, but that I would very much like to see reference/comparison to the other model in your paper so that it is clear to the community what is going on with these two models. I hope that the authors of BERT-NL do the same. Do you have any plans to submit a paper somewhere?", "Thanks for your feedback and clear explanation. We do indeed intend to submit a long paper somewhere. The short paper on arxiv is mainly intended for reference and to demonstrate that the model is effective with some benchmarks. Further evaluation would be included in a longer paper.", "The authors of BERT-NL have reported results which can be compared to Bertje, see https://twitter.com/suzan/status/1200361620398125056\r\nAlso see the results in this thread https://twitter.com/danieldekok/status/1213378688563253249 and https://twitter.com/danieldekok/status/1213741132863156224\r\nIn both cases, Bertje does outperform BERT-NL. On the other hand, the results about Bertje vs multilingual BERT are different, so this needs to be investigated further.", "I think @BramVanroy raises some good points about naming convention.\r\n\r\nIn my opinion the organization name or author name should come after the \"bert-base-<language>\" skeleton.\r\n\r\nSo it seems that BERTje is current SOTA now. But: on the next conference maybe another BERT model for Dutch is better... I'm not a fan of the \"First come, first served\" principle here 😅\r\n\r\n/cc @thomwolf , @julien-c ", "> On the other hand, the results about Bertje vs multilingual BERT are different, so this needs to be investigated further.\r\n\r\nAgreed. Tweeting results of doing \"tests\" is one thing, but actual thorough investigating and reporting is something else. It has happened to all of us that you quickly wanted to check something and only later realized that you made a silly mistake. (My last one was forgetting the `-` in my learning rate and not noticing it, oh boy what a day.) As I said before I would really like to a see a thorough, reproducible comparison of BERTje, BERT-NL, and multilingual BERT, and I believe that that should be the basis of any new model. Many new models sprout from the community grounds - and that's great! - but without having at least _some_ reference and comparison, it is guess-work trying to figure out which one is best or which one you should use.\r\n\r\n> I think @BramVanroy raises some good points about naming convention.\r\n> \r\n> In my opinion the organization name or author name should come after the \"bert-base-\" skeleton.\r\n> \r\n> So it seems that BERTje is current SOTA now. But: on the next conference maybe another BERT model for Dutch is better... I'm not a fan of the \"First come, first served\" principle here 😅\r\n> \r\n> /cc @thomwolf , @julien-c\r\n\r\nPerhaps it's better to just make the model available through the user and that's all? In this case, only make it available through `wietsedv/bert-base-dutch-cased` and not `bert-base-dutch-cased`? That being said, where do you draw the line of course. Hypothetical question: why does _a Google_ get the rights to make its weights available without a `google/` prefix, acting as a \"standard\"? I don't know how to answer that question, so ultimately it's up to the HuggingFace people.\r\n\r\nI'm also not sure how diverging models would then work. If for instance you bring out a German BERT-derivative that has a slightly different architecture, or e.g. a different tokenizer, how would that then get integrated in Transformers? (For example, IIRC BERTje uses SOP instead of NSP, so that may lead to more structural changing in the available heads than just different weights.)\r\n", "I inherently completely agree with your points. I think the people at Huggingface are trying to figure out how to do this, but they have not been really consistent. Initially, the \"official\" models within Transformers were only original models (Google/Facebook) and it is unlikely that there would be competetition for better models with exactly the same architecture in English. But for other monolingual models this is different.\r\n\r\nI prefer a curated list with pre-trained general models that has more structure than the long community models list. But existing shortcuts should be renamed to be consistent. German for instance has a regular named german and there is one with the `dbmdz` infix. And Finnish has for some reason the `v1` suffix?\r\n\r\nMy preference would be to always use `institution/` or `institution-` prefixes in the curated list. In Transformers 2.x, the current shortcuts could be kept for backward compatibility but a structured format should be used in the documentation. I think this may prevent many frustrations if even more non-english models are trained and even more people are wanting to use and trust (!) these models.", "200% agree with that. That would be the fairest and probably clearest way of doing this. Curating the list might not be easy, though, unless it is curated by the community (like a wiki)? Perhaps requiring a description, website, paper, any other meta information might help to distinguish models as well, giving the authors a chance to explain, e.g., which data their model was trained on, which hyperparameters were used, and how their model differs from others.\r\n\r\nI really like http://nlpprogress.com/ which \"tracks\" the SOTA across different NLP tasks. It is an open source list and anyone can contribute through github. Some kind of lists like this might be useful, but instead discussing the models. ", "You all raise excellent questions, many (most?) of which we don’t have a definitive answer to right now 🤗\r\n\r\nSome kind of structured evaluation results (declarative or automated) could be a part of the solution. In addition to nlpprogress, sotabench/paperswithcode is also a good source of inspiration.\r\n\r\n", "On the previous point of being able to load the tokenizer's (remote) config correctly:\r\n- I've added a `tokenizer_config.json` to your user namespace on S3: https://s3.amazonaws.com/models.huggingface.co/bert/wietsedv/bert-base-dutch-cased/tokenizer_config.json\r\n- We're fixing the support for those remote tokenizer configs in https://github.com/huggingface/transformers/pull/2535 (you'll see that the unit test uses your model). Feedback welcome.", "Merging this as we haven't seen other \"better\" BERT models for Dutch (coincidentally, [`RobBERT`](https://people.cs.kuleuven.be/~pieter.delobelle/robbert/) from @iPieter looks like a great RoBERTa-like model)\r\n\r\nPlease see [this discussion on model descriptions/README.md](https://github.com/huggingface/transformers/issues/2520#issuecomment-579009439). If you can upload a README.md with eval results/training methods, that'd be awesome.\r\n\r\nThanks!" ]

[]

[]

[]

[]

[]

[ "I suggest you follow the [run_lm_finetuning.py](https://github.com/huggingface/transformers/blob/master/examples/run_lm_finetuning.py) script. Instead of downloading a pretrained model, simply start with a fresh one. Here's an example:\r\n\r\n```\r\nfrom transformers import BertModel, BertConfig, BertTokenizer\r\nmodel = BertModel(BertConfig())\r\ntokenizer = BertTokenizer.from_pretrained('bert-base-cased')\r\n```\r\n\r\nI personally don't have a reason to write my own tokenizer, but if you do feel free to do that as well. All you need to do is generate a vocab.txt file from your corpus.", "You can now leave `--model_name_or_path` to None in `run_language_modeling.py` to train a model from scratch.\r\n\r\nSee also https://huggingface.co/blog/how-to-train" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Check here: https://github.com/huggingface/transformers/blob/master/examples/distillation/run_squad_w_distillation.py" ]

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

[]

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

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2276?src=pr&el=h1) Report\n> Merging [#2276](https://codecov.io/gh/huggingface/transformers/pull/2276?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/ce50305e5b8c8748b81b0c8f5539a337b6a995b9?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `0%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2276?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2276 +/- ##\n=======================================\n Coverage 74.45% 74.45% \n=======================================\n Files 85 85 \n Lines 14610 14610 \n=======================================\n Hits 10878 10878 \n Misses 3732 3732\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2276?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2276/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ191dGlscy5weQ==) | `63.91% <0%> (ø)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2276?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/2276?src=pr&el=footer). Last update [ce50305...398bb03](https://codecov.io/gh/huggingface/transformers/pull/2276?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks a lot for catching that @ShnitzelKiller!", "I just realized that despite the official (current) documentation saying that the argument name is \"source\", after upgrading my pytorch version, this code is what throws an error saying the argument name is \"src\"! I should probably notify you to revert this pull request then." ]

[]

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

[ "Same problem. And it seems not fixed yet.", "It should be fixed now. Thank you for raising this issue." ]

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

[]

[ "Same problem", "I have found the reason.\r\n\r\nSo it turns out that the `generate()` method of the `PreTrainedModel` class is newly added, even newer than the latest release (2.3.0). Quite understandable since this library is iterating very fast.\r\n\r\nSo to make `run_generation.py` work, you can install this library like this:\r\n\r\n- Clone the repo to your computer\r\n- cd into the repo\r\n- Run `pip install -e .` (don't forget the dot)\r\n- Re-run `run_generation.py`\r\n\r\nI'll leave this ticket open until the `generate()` method is incorporated into the latest release.", "@jsh9's solution worked for me!\r\n\r\nAlso, if you want to avoid doing the manual steps, you can just `pip install` directly from the `master` branch by running:\r\n\r\n```bash\r\npip install git+https://github.com/huggingface/transformers.git@master#egg=transformers\r\n```\r\n", "i was getting the same error then i used repository before 7days which is working fine for me \r\n`!wget https://github.com/huggingface/transformers/archive/f09d9996413f2b265f1c672d7a4b438e4c5099c4.zip`\r\n\r\nthen unzip with\r\n\r\n`!unzip file_name.zip`\r\n\r\nthere is some bugs in recent update, hope they fix it soon", "@Weenkus's way worked for me. In `requirements.txt` you can use;\r\n```\r\n-e git+https://github.com/huggingface/transformers.git@master#egg=transformers\r\n```\r\n\r\n(all on one line)", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

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

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

[]

[ "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 issue on Colab TPU with tf-nightly 2.2.0. @dlauc did you solve the problem?", "Hi @BOUALILILila, I've switched to the TF/Keras - it works well with TPU-s", "Hi @dlauc. Could you elaborate how you fixed this? I am having the same problem. " ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2271?src=pr&el=h1) Report\n> Merging [#2271](https://codecov.io/gh/huggingface/transformers/pull/2271?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/23dad8447c8db53682abc3c53d1b90f85d222e4b?src=pr&el=desc) will **decrease** coverage by `0.58%`.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2271?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2271 +/- ##\n==========================================\n- Coverage 74.27% 73.68% -0.59% \n==========================================\n Files 85 87 +2 \n Lines 14610 14791 +181 \n==========================================\n+ Hits 10851 10899 +48 \n- Misses 3759 3892 +133\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2271?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/commands/user.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb21tYW5kcy91c2VyLnB5) | `0% <0%> (ø)` | |\n| [src/transformers/commands/serving.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb21tYW5kcy9zZXJ2aW5nLnB5) | `0% <0%> (ø)` | |\n| [src/transformers/tokenization\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fdXRpbHMucHk=) | `92.27% <0%> (+0.96%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.12% <0%> (+1.58%)` | :arrow_up: |\n| [src/transformers/pipelines.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9waXBlbGluZXMucHk=) | `68.31% <0%> (+2.32%)` | :arrow_up: |\n| [src/transformers/data/processors/squad.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9kYXRhL3Byb2Nlc3NvcnMvc3F1YWQucHk=) | `35.97% <0%> (+6.6%)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/2271/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9hdXRvLnB5) | `32.14% <0%> (+7.14%)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2271?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/2271?src=pr&el=footer). Last update [23dad84...10724a8](https://codecov.io/gh/huggingface/transformers/pull/2271?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2270?src=pr&el=h1) Report\n> Merging [#2270](https://codecov.io/gh/huggingface/transformers/pull/2270?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/b6ea0f43aeb7ff1dcb03658e38bacae1130abd91?src=pr&el=desc) will **increase** coverage by `1.2%`.\n> The diff coverage is `86.66%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2270?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2270 +/- ##\n=========================================\n+ Coverage 73.25% 74.45% +1.2% \n=========================================\n Files 85 85 \n Lines 14779 14610 -169 \n=========================================\n+ Hits 10826 10878 +52 \n+ Misses 3953 3732 -221\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2270?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [src/transformers/modeling\\_camembert.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ19jYW1lbWJlcnQucHk=) | `100% <ø> (ø)` | :arrow_up: |\n| [src/transformers/configuration\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX3JvYmVydGEucHk=) | `100% <ø> (ø)` | :arrow_up: |\n| [src/transformers/tokenization\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fZGlzdGlsYmVydC5weQ==) | `100% <ø> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_auto.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl9hdXRvLnB5) | `32.14% <ø> (-0.41%)` | :arrow_down: |\n| [src/transformers/configuration\\_distilbert.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2Rpc3RpbGJlcnQucHk=) | `100% <ø> (ø)` | :arrow_up: |\n| [src/transformers/modeling\\_tf\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl91dGlscy5weQ==) | `93.12% <ø> (-0.04%)` | :arrow_down: |\n| [src/transformers/configuration\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9jb25maWd1cmF0aW9uX2N0cmwucHk=) | `97.05% <ø> (-0.09%)` | :arrow_down: |\n| [src/transformers/tokenization\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fcm9iZXJ0YS5weQ==) | `100% <ø> (ø)` | :arrow_up: |\n| [src/transformers/tokenization\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy90b2tlbml6YXRpb25fY3RybC5weQ==) | `96.11% <ø> (-0.08%)` | :arrow_down: |\n| [src/transformers/modeling\\_tf\\_t5.py](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree#diff-c3JjL3RyYW5zZm9ybWVycy9tb2RlbGluZ190Zl90NS5weQ==) | `96.54% <ø> (-0.01%)` | :arrow_down: |\n| ... and [66 more](https://codecov.io/gh/huggingface/transformers/pull/2270/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2270?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/2270?src=pr&el=footer). Last update [b6ea0f4...1a948d7](https://codecov.io/gh/huggingface/transformers/pull/2270?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

[]

[]

[]

[]

[]

[ "Last test run failed only because of a flaky test — this is #2240." ]

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

[]

[ "`BertPredictionHeadTransform` is never used in itself AFAIK, and is only part of the full scale models (e.g. `BertModel`). As such, the weights for the prediction head are loaded when you run BertModel.from_pretrained(), since the `BertPredictionHeadTransform` is only a module in the whole model.", "> `BertPredictionHeadTransform` is never used in itself AFAIK, and is only part of the full scale models (e.g. `BertModel`). As such, the weights for the prediction head are loaded when you run BertModel.from_pretrained(), since the `BertPredictionHeadTransform` is only a module in the whole model.\r\n\r\nThanks, Bram. As you said `BertModel` does take `BertPreTrainedModel` in `super`. I did notice it, But It's just that my mind doesn't get around how/where and when those weights are getting used exactly.", "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 } ]

[]

[ "@aaugustin it seems like that i met the same problem when i use convert_bertabs_original_pytorch_checkpoint.py ,have you ever fixed it or find any way to make it work.\r\nAppriciate it if you can tell me!", "I didn't attempt to fix this issue. I merely noticed it while I was working on the overall quality of the `transformers` code base.\r\n\r\nI suspect these modules used to exist in `transformers` and were removed in a refactoring, but I don't know for sure.", "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" ]

[]

[]

[ "You need to give more information. What are you trying to do, what is the code that you use for it, what does not work the way you intended to?", "> You need to give more information. What are you trying to do, what is the code that you use for it, what does not work the way you intended to?\r\n\r\nI am learning how to use this repo.\r\nI use the example of the official website model2model, the link is as follows\r\nhttps://huggingface.co/transformers/quickstart.html\r\nAnd I use google colab to install the transformer and copy and run the official website code", "I mean, which error are you getting or what is not working as expected? ", "> I mean, which error are you getting or what is not working as expected?\r\n\r\nthis is my issue topic\r\nOnly the Bert model is currently supported\r\n```\r\n---------------------------------------------------------------------------\r\nValueError Traceback (most recent call last)\r\n<ipython-input-18-58aac9aa944f> in <module>()\r\n 9 \r\n 10 # Load pre-trained model (weights)\r\n---> 11 model = Model2Model.from_pretrained('fine-tuned-weights')\r\n 12 model.eval()\r\n 13 \r\n\r\n/usr/local/lib/python3.6/dist-packages/transformers/modeling_encoder_decoder.py in from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs)\r\n 281 or \"distilbert\" in pretrained_model_name_or_path\r\n 282 ):\r\n--> 283 raise ValueError(\"Only the Bert model is currently supported.\")\r\n 284 \r\n 285 model = super(Model2Model, cls).from_pretrained(\r\n\r\nValueError: Only the Bert model is currently supported.\r\n```", "That wasn't clear. In the future, please post the trace so it is clear what your error is, like so:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"C:/Users/bramv/.PyCharm2019.2/config/scratches/scratch_16.py\", line 62, in <module>\r\n model = Model2Model.from_pretrained('fine-tuned-weights')\r\n File \"C:\\Users\\bramv\\.virtualenvs\\semeval-task7-Z5pypsxD\\lib\\site-packages\\transformers\\modeling_encoder_decoder.py\", line 315, in from_pretrained\r\n raise ValueError(\"Only the Bert model is currently supported.\")\r\nValueError: Only the Bert model is currently supported.\r\n```\r\n\r\nThis is not a bug, then of course. In the example, where \"fine-tuned-weights\" is used, you can load your own fine-tuned model. So if you tuned a model and saved it as \"checkpoint.pth\" you can use that.", "> This is not a bug, then of course. In the example, where \"fine-tuned-weights\" is used, you can load your own fine-tuned model. So if you tuned a model and saved it as \"checkpoint.pth\" you can use that.\r\n\r\nthanks", "Please close this question." ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2264?src=pr&el=h1) Report\n> Merging [#2264](https://codecov.io/gh/huggingface/transformers/pull/2264?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/645713e2cb8307e41febb2b7c9f6036f6645efce?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2264?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2264 +/- ##\n=======================================\n Coverage 78.35% 78.35% \n=======================================\n Files 133 133 \n Lines 19878 19878 \n=======================================\n Hits 15576 15576 \n Misses 4302 4302\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2264?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/2264?src=pr&el=footer). Last update [645713e...9d00f78](https://codecov.io/gh/huggingface/transformers/pull/2264?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks @upura!" ]

[ { "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 set a fixed seed? If you want deterministic results, you should set a fixed seed. ", "> Did you set a fixed seed? If you want deterministic results, you should set a fixed seed.\r\n\r\nNo, I didn't. My problem seems that the model finetuned is totally wrong sometimes. Is it a problem related to random seed?", "What do you mean by \"bad\"? Post some code to better help you.", "> What do you mean by \"bad\"? Post some code to better help you.\r\n\r\nFor example, the normal performance is 140.1, but the bad performance is 1.0. As I mentioned before, this situation happened sometimes and I found the word embeddings produced by BertModel were all the same when the performance was bad.\r\n\r\nThe text encoding part of my code is as follows:\r\n \r\n\r\n", "I think you are accessing the CLS token embeddings and It is constant as the model you are using is trained on MLM objective. ", "> I think you are accessing the CLS token embeddings and It is constant as the model you are using is trained on MLM objective.\r\n\r\nIf I change the out to self.bert(text)[0][:,1,:], the output is the same as self.bert(text)[0][:,0,:]. It seems I get the same output no matter the input that I put in.", "I saved the checkpoint when this situation happened. I multiplied all the parameter values of bert by 10 and found the outputs were different, while I divided all parameter values by 10 and found the outputs were almost same. So I think the reason is that the parameter values are too small.", "@xuesong0309 https://github.com/huggingface/transformers/issues/1465，Could you look at this problem?", "> @xuesong0309 https://github.com/huggingface/transformers/issues/1465，Could you look at this problem?\r\n\r\nDid you always get same output? I suggest outputing the parameter values of your model as I mentioned above.", "> > @xuesong0309 [https://github.com/huggingface/transformers/issues/1465，Could](https://github.com/huggingface/transformers/issues/1465%EF%BC%8CCould) you look at this problem?\r\n> \r\n> Did you always get same output? I suggest outputing the parameter values of your model as I mentioned above.\r\n\r\nIt may not be the problem you mentioned, because the model is normal for multi-class classification, and this happens only for multi-label classification.", "> @xuesong0309 https://github.com/huggingface/transformers/issues/1465，Could you look at this problem?\r\n> \r\n> Did you always get same output? I suggest outputing the parameter values of your model as I mentioned above.\r\n> \r\n> It may not be the problem you mentioned, because the model is normal for multi-class classification, and this happens only for multi-label classification.\r\n\r\nYou could open a new issue to describe your problem in detail.", "Have you solved this issue? I have facing the same issue. Output same results during evaluation.", "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 } ]

[]

[ "#2198 Same here. A predict script can be really helpful to researchers.", "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" ]

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

[]

[ "The encoder itself would automatically add the [CLS] and [SEP] tokens so if you've done that during preprocessing, you would need to change the `add_special_tokens` parameter to `False`. So your code should probably be like this:\r\n\r\n```\r\ninput_ids = tokenizer.encode(input_text, add_special_tokens=False)\r\n```", "Thanks, but that does not fix the error. The problem is there is no [102] token in the list. Maybe because we're using AlbertTokenizer?\r\n\r\n```\r\ninput_ids = tokenizer.encode(input_text, add_special_tokens=False)\r\nprint(input_ids)\r\n[2, 72, 23, 2170, 27674, 60, 3, 2170, 27674, 23, 21, 2210, 10956, 3]\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" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2260?src=pr&el=h1) Report\n> Merging [#2260](https://codecov.io/gh/huggingface/transformers/pull/2260?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/645713e2cb8307e41febb2b7c9f6036f6645efce?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2260?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2260 +/- ##\n=======================================\n Coverage 78.35% 78.35% \n=======================================\n Files 133 133 \n Lines 19878 19878 \n=======================================\n Hits 15576 15576 \n Misses 4302 4302\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2260?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_mmbt.py](https://codecov.io/gh/huggingface/transformers/pull/2260/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX21tYnQucHk=) | `18.25% <ø> (ø)` | :arrow_up: |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2260?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/2260?src=pr&el=footer). Last update [645713e...b668a74](https://codecov.io/gh/huggingface/transformers/pull/2260?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Thanks" ]

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

[]

[ "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 } ]

[]

[ "Can confirm this issue -> a temporary workaround would be to change the line to:\r\n\r\n```python\r\nif os.path.exists(args.model_name_or_path) and \"checkpoint\" in args.model_name_or_path:\r\n```", "See also related (recent) fix on master: https://github.com/huggingface/transformers/commit/4d36472b96d144887cbe95b083f0d2091fd5ff03", "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" ]

[]

[]

[ "Having the same issue.", "May be related to this:\r\nhttps://twitter.com/Thom_Wolf/status/1208365367493636096\r\n", "Are you still having an issue or is it fixed (if it is, please close the issue)?", "Hello,\r\nI am still having the same issue.\r\n", "Ok, we are in the process of fixing it. Thanks for the report", "It works now, thank you for the help" ]

[]

[]

[ "After looking at \r\n\r\nhttps://github.com/huggingface/transformers/issues/1727\r\n\r\nI figured out we're getting the warning because we're not using the pooler, therefore it gets no updates. ", "@Santosh-Gupta can you please update your colab notebook with correct version. i am unbale to get this resolved", "Hi, I have exactly the same issue with TFXLMRobertaForSequenceClassification. How did you solve the issue ?", "> Hi, I have exactly the same issue with TFXLMRobertaForSequenceClassification. How did you solve the issue ?\r\n\r\nFrom what understand, this is not a bug. You won't get gradients calculated for variables (kernel and bias) of the layer tf_bert_model/bert/pooler/dense if you don't use it. As such, if you indeed don't use the pooler, you can simply ignore this warning. ", "> @Santosh-Gupta can you please update your colab notebook with correct version. i am unbale to get this resolved\r\n\r\nIt's not a bug, the forward pass doesn't go through the pooler, so the backwards pass doesn't go through it either. ", "I'm having the same issue, although I am using the pooling layer. \r\nmy model is like this one \r\n```\r\nclass TFAlbertForNaturalQuestionAnswering(TFAlbertPreTrainedModel):\r\n def __init__(self, config, *inputs, **kwargs):\r\n super().__init__(config, *inputs, **kwargs)\r\n \r\n self.albert = TFAlbertMainLayer(config)\r\n\r\n self.initializer = get_initializer(config.initializer_range)\r\n self.start = tf.keras.layers.Dense(1,\r\n kernel_initializer=self.initializer, name='start')\r\n self.end = tf.keras.layers.Dense(1,\r\n kernel_initializer=self.initializer, name='end')\r\n self.long_outputs = tf.keras.layers.Dense(1, kernel_initializer=self.initializer,\r\n name='long')\r\n\r\n self.answerable = tf.keras.layers.Dense(1, kernel_initializer=self.initializer,\r\n name='answerable', activation = \"sigmoid\")\r\n def call(self, inputs, **kwargs):\r\n outputs = self.albert(inputs, **kwargs)\r\n sequence_output = outputs[0]\r\n \r\n # tf.print(outputs[0].shape) (batch, len->0, hidden) 1->0\r\n # tf.print(outputs[1].shape) (batch, hidden_size)\r\n\r\n start_logits = tf.squeeze(self.start(sequence_output), -1)\r\n end_logits = tf.squeeze(self.end(sequence_output), -1)\r\n long_logits = tf.squeeze(self.long_outputs(sequence_output), -1)\r\n\r\n answerable = tf.squeeze(self.answerable(outputs[1]), -1)\r\n```", "Hey ! i'm getting the same issue with attention model and embedding layer, the weights of both layers are not updating .\r\n\r\n```\r\n\r\nembd=Embedding(input_dim=len(vocab),output_dim=100,name=\"embd\")\r\nlstm1=Bidirectional(LSTM(units=100,return_sequences=True,name=\"lstm1\"),name=\"bd1\")\r\nlstm2=Bidirectional(LSTM(units=100,return_sequences=True,name=\"lstm2\"),name=\"bd2\")\r\nattention_layer=Attention_Model(21,200)\r\ndense1=Dense(units=80,name=\"dense1\",kernel_regularizer=\"l2\")\r\ndropout1=Dropout(0.5)\r\nact1=Activation('sigmoid')\r\n\r\ndense2=Dense(units=50,name=\"dense2\",kernel_regularizer=\"l2\")\r\ndropout2=Dropout(0.4)\r\nact2=Activation('sigmoid')\r\n\r\ndense3=Dense(units=30,name=\"dense3\",kernel_regularizer=\"l2\")\r\ndropout3=Dropout(0.3)\r\nact3=Activation('sigmoid')\r\n\r\ndense4=Dense(units=len(classes),name=\"dense4\")\r\ndropout4=Dropout(0.2)\r\noutput=Activation('softmax')\r\n\r\n```\r\nForward Pass : \r\n\r\n```\r\ndef forward_pass(X):\r\n t=embd(X)\r\n \r\n t=lstm1(t)\r\n \r\n t=lstm2(t)\r\n \r\n\r\n \r\n t=attention_layer(t)\r\n \r\n \r\n t=dense1(t)\r\n t=dropout1(t)\r\n t=act1(t)\r\n\r\n t=dense2(t)\r\n t=dropout2(t)\r\n t=act2(t)\r\n\r\n t=dense3(t)\r\n t=dropout3(t)\r\n t=act3(t)\r\n \r\n t=dense4(t)\r\n t=dropout4(t)\r\n t=output(t)\r\n\r\n return t\r\n\r\n\r\n```\r\n\r\nAttention Model : \r\n\r\n```\r\n\r\nclass Attention_Model():\r\n def __init__(self,seq_length,units):\r\n self.seq_length=seq_length\r\n self.units=units\r\n self.lstm=LSTM(units=units,return_sequences=True,return_state=True)\r\n \r\n\r\n def get_lstm_s(self,seq_no):\r\n input_lstm=tf.expand_dims(tf.reduce_sum(self.X*(self.alphas[:,:,seq_no:seq_no+1]),axis=1),axis=1)\r\n a,b,c=self.lstm(input_lstm)\r\n self.output[:,seq_no,:]=a[:,0,:]\r\n\r\n return b\r\n\r\n def __call__(self,X):\r\n self.X=X\r\n self.output=np.zeros(shape=(self.X.shape[0],self.seq_length,self.units))\r\n self.dense=Dense(units=self.seq_length)\r\n self.softmax=Softmax(axis=1)\r\n \r\n\r\n for i in range(self.seq_length+1):\r\n if i==0 :\r\n s=np.zeros(shape=(self.X.shape[0],self.units))\r\n else :\r\n s=self.get_lstm_s(i-1)\r\n if(i==self.seq_length):\r\n break \r\n \r\n s=RepeatVector(self.X.shape[1])(s)\r\n concate_X=np.concatenate([self.X,s],axis=-1)\r\n \r\n self.alphas=self.softmax(self.dense(concate_X))\r\n\r\n return self.output\r\n \r\n```\r\n\r\n\r\nis anything wrong with implementation or something else ?", "@MarioBonse , your forward pass isn't going through the pooling layer\r\n\r\n'''sequence_output = outputs[0]'''\r\n\r\n@gajeshladhar \r\n\r\nIs that code from the hf library? Where are the classes defined? \r\n", "pooler_output of transformers TFRobertaModel have tf_roberta_model/roberta/pooler/dense/kernel:0. if you have not use pooler_output,tf_roberta_model/roberta/pooler/dense/kernel:0 do not update Gradients" ]

[]

[]

[ "This is in reasonably good shape. There are two tasks left:\r\n\r\n1. Figure out why isort doesn't behave the same locally and on Circle CI. Most likely this has to do with how it classifies first-party / third-party / unknown libraries. Then enable it on Circle CI. **EDIT - fixed** - this was a matter of installing optional dependencies, not listed in setup.py, on Circle CI so that isort can classify them correctly.\r\n2. Fix flake8 F841 warnings and stop ignoring them.\r\n\r\nAssuming tests pass, I think it would be best to merge this PR and deal with these two items later. I'd like to do the repository structure changes first, so we're done with the large changes.", "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2255?src=pr&el=h1) Report\n> Merging [#2255](https://codecov.io/gh/huggingface/transformers/pull/2255?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/645713e2cb8307e41febb2b7c9f6036f6645efce?src=pr&el=desc) will **decrease** coverage by `0.24%`.\n> The diff coverage is `44.23%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2255?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2255 +/- ##\n==========================================\n- Coverage 78.35% 78.11% -0.25% \n==========================================\n Files 133 133 \n Lines 19878 19655 -223 \n==========================================\n- Hits 15576 15354 -222 \n+ Misses 4302 4301 -1\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2255?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_tf\\_xlm.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3hsbS5weQ==) | `90.44% <ø> (-0.04%)` | :arrow_down: |\n| [transformers/configuration\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25feGxuZXQucHk=) | `93.47% <ø> (-0.4%)` | :arrow_down: |\n| [transformers/modeling\\_mmbt.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX21tYnQucHk=) | `18.25% <ø> (ø)` | :arrow_up: |\n| [transformers/modelcard.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsY2FyZC5weQ==) | `87.8% <ø> (ø)` | :arrow_up: |\n| [transformers/tests/modeling\\_bert\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX2JlcnRfdGVzdC5weQ==) | `98% <ø> (-0.02%)` | :arrow_down: |\n| [transformers/modeling\\_tf\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3JvYmVydGEucHk=) | `90.9% <ø> (ø)` | :arrow_up: |\n| [transformers/tests/modeling\\_tf\\_ctrl\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3RmX2N0cmxfdGVzdC5weQ==) | `95.74% <ø> (-0.18%)` | :arrow_down: |\n| [transformers/configuration\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fb3BlbmFpLnB5) | `97.22% <ø> (-0.22%)` | :arrow_down: |\n| [transformers/tests/tokenization\\_xlm\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL3Rva2VuaXphdGlvbl94bG1fdGVzdC5weQ==) | `82.22% <ø> (ø)` | :arrow_up: |\n| [transformers/commands/\\_\\_init\\_\\_.py](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbW1hbmRzL19faW5pdF9fLnB5) | `0% <ø> (ø)` | :arrow_up: |\n| ... and [237 more](https://codecov.io/gh/huggingface/transformers/pull/2255/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2255?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/2255?src=pr&el=footer). Last update [645713e...c11b3e2](https://codecov.io/gh/huggingface/transformers/pull/2255?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "`pyproject.toml` is supposed to be The Future (tm). `setup.cfg`was supposed to be The Future as well. That didn't quite work, but it gained some support.\r\n\r\nUnfortunately, in the Python packaging ecosystem, some futures never became the present. That's why I'm conservative, at least until this changes: https://packaging.python.org/specifications/distribution-formats/\r\n\r\nI'm happy to try converting setup.py to pyproject.toml if you're feeling adventurous. Let me know.\r\n\r\nWe can move the isort configuration there, but not the flake8 configuration until [this PR](https://gitlab.com/pycqa/flake8/issues/428) is merged. I like setup.cfg because we can put both in the same file.", "Alright thanks for the context!" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/transformers/pull/2254?src=pr&el=h1) Report\n> Merging [#2254](https://codecov.io/gh/huggingface/transformers/pull/2254?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/73f6e9817c744caae0b73fa343ceaf95ba76f9f8?src=pr&el=desc) will **increase** coverage by `1.46%`.\n> The diff coverage is `100%`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2254?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2254 +/- ##\n==========================================\n+ Coverage 77.28% 78.75% +1.46% \n==========================================\n Files 133 131 -2 \n Lines 19872 19742 -130 \n==========================================\n+ Hits 15358 15547 +189 \n+ Misses 4514 4195 -319\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/transformers/pull/2254?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [transformers/modeling\\_tf\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3RmX3JvYmVydGEucHk=) | `90.9% <100%> (+0.47%)` | :arrow_up: |\n| [transformers/modeling\\_mmbt.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX21tYnQucHk=) | | |\n| [transformers/configuration\\_mmbt.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL2NvbmZpZ3VyYXRpb25fbW1idC5weQ==) | | |\n| [transformers/pipelines.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3BpcGVsaW5lcy5weQ==) | `67.94% <0%> (+0.58%)` | :arrow_up: |\n| [transformers/modeling\\_utils.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3V0aWxzLnB5) | `64.29% <0%> (+0.71%)` | :arrow_up: |\n| [transformers/modeling\\_openai.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX29wZW5haS5weQ==) | `81.51% <0%> (+1.32%)` | :arrow_up: |\n| [transformers/modeling\\_ctrl.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX2N0cmwucHk=) | `96.47% <0%> (+2.2%)` | :arrow_up: |\n| [transformers/modeling\\_xlnet.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3hsbmV0LnB5) | `73.72% <0%> (+2.29%)` | :arrow_up: |\n| [transformers/modeling\\_roberta.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL21vZGVsaW5nX3JvYmVydGEucHk=) | `64.42% <0%> (+10.09%)` | :arrow_up: |\n| [transformers/tests/modeling\\_tf\\_common\\_test.py](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree#diff-dHJhbnNmb3JtZXJzL3Rlc3RzL21vZGVsaW5nX3RmX2NvbW1vbl90ZXN0LnB5) | `94.39% <0%> (+17.24%)` | :arrow_up: |\n| ... and [1 more](https://codecov.io/gh/huggingface/transformers/pull/2254/diff?src=pr&el=tree-more) | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2254?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/2254?src=pr&el=footer). Last update [73f6e98...77676c2](https://codecov.io/gh/huggingface/transformers/pull/2254?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Yeepee!" ]

[]

[]

[ "Yes, all the layer norms in T5 have no bias (so we keep the default value of 0)" ]

[]

[]

[]

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

[]

[ "@abb4s A workaround would be downgrading `transformers` to `2.2.0`. It worked for me that way.", "Indeed, this was an error introduced by #1548. It was patched by 1efc208. Thank you for raising this issue! ", "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" ]

[]

[]

[ "The first two are easy fixes. I put fixes in the test parallelization PR.\r\n\r\nThe last two are likely the same bug, but I'm out of my depth there.", "I guess this has been fixed by now" ]

[ { "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/2249?src=pr&el=h1) Report\n> Merging [#2249](https://codecov.io/gh/huggingface/transformers/pull/2249?src=pr&el=desc) into [master](https://codecov.io/gh/huggingface/transformers/commit/ac1b449cc938bb34bc9021feff599cfd3b2376ae?src=pr&el=desc) will **not change** coverage.\n> The diff coverage is `n/a`.\n\n[](https://codecov.io/gh/huggingface/transformers/pull/2249?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #2249 +/- ##\n=======================================\n Coverage 79.82% 79.82% \n=======================================\n Files 131 131 \n Lines 19496 19496 \n=======================================\n Hits 15562 15562 \n Misses 3934 3934\n```\n\n\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/transformers/pull/2249?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/2249?src=pr&el=footer). Last update [ac1b449...ea25498](https://codecov.io/gh/huggingface/transformers/pull/2249?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "@michael-wzhu did this increase f1 on CoNLL?", "> add crf layer for better performance in NER tasks\r\n\r\nHi. How do you solve the tokens with \"##...\" when they are fed into the crf layer?\r\ne.g. De ##duct ##ive reasoning\r\nDo \"##duct\" and \"\"##ive\" are fed into the crf layer? If they are, do they have chance to be transfered in the transition matrix?", "> e\" are fed into the crf layer? If they are, do they have chance to be transfered in the transition matrix?\r\n\r\nI have used this to develop my own version. To answer your quesiton, ## sub word tokens are treated as padding, as suggested by the original BERT authors. This code relies on The padding label token being \"X\", at the first position (to get0th index) from the output of get_labels function in crf_utils_ner.py \r\n\r\nThe pad token label id might need to be 0 for calculations in CRF, but you should be careful in declaring your mask so your model does not confuse padding with one of the tokens. ", "If ## sub word tokens are treated as padding, it will break the tag-tag dependencies, so definitely not ideal.", "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", "@srush @mezig351 What did you find after your work on this issue? Why didn't you merge it?", "I think you should add it, it's not trivial, and I myself spent 2 days to make it work...\r\nand then I just found this thread... " ]

[]

[]

[ ">My question is how to get feature embedding aligned to each token from a pretrained BertForQuestionAnswering\r\n\r\nCould you elaborate on what this means? I am currently working with BertForQuestionAnswering but haven't encountered this area before. " ]

[ { "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." } ]

[]

[ "That would be a great feature, +1.", "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" ]