Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Great! \r\n\r\nDo you want for us to make this a \"model\" file first, merge it, see how it feels and then replicate to the rest? Or should I just proceed with the rest?\r\n", "I think you ca just proceed.", "Bummer! If I make it into a function and thus remove `if ...` - isort and flake now complain:\r\n```\r\n$ flake8 tests\r\ntests/test_modeling_bart.py:37:1: E402 module level import not at top of file\r\ntests/test_modeling_bart.py:39:1: E402 module level import not at top of file\r\ntests/test_modeling_bart.py:57:1: E402 module level import not at top of file\r\n$ isort --check-only tests\r\nERROR: /mnt/nvme1/code/huggingface/transformers-torch-req/tests/module_skip_pytorch.py Imports are incorrectly sorted and/or formatted.\r\n```\r\nso that would require adding #noqa to all the subsequent imports ;( which leaves the code ugly just in a different way.\r\n\r\nThese tools have so little flexibility. They are supposed to make things better but lead to a much uglier code :(\r\n", "meh! I call this experiment a failure thanks to `make quality` oppression.", "I will just leave the helper I wrote here, in case someone figures out a magical way to solve the ugliness. \r\n\r\n```\r\ndef test_module_skip_require_pytorch():\r\n \"\"\"\r\n Call this one on top of test module to skip the whole module if pytorch is not available:\r\n test_module_skip_require_pytorch()\r\n \"\"\"\r\n if not _torch_available:\r\n raise unittest.SkipTest(\"Skip the whole module as it requires pytorch\")\r\n```" ]

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

[]

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

[]

[]

[]

[]

[]

[]

[]

[]

[ "Should we start thinking about automating the creation of the metadata block for the model's model card?\r\n\r\nhere for instance we'd already have this info:\r\n```\r\n---\r\ndatasets:\r\n- mrpc\r\nmetrics:\r\n- f1\r\nfinetuned_from: bert-base-cased\r\n---\r\n```", "We could think of something like that and add a blank model card to be completed by the user in the final checkpoint. We could also include the results of the last evaluation if there is one." ]

[]

[]

[ "same issue", "Hello! I think this is due to a mismatch between your `transformers` and `tokenizers` versions. `transformers` version v3.3.1 expects `tokenizers == 0.8.1.rc2`.\r\n\r\nIf you want to use `tokenizers == 0.9.2` you should work on the current `master` branch or wait for version v3.4.0 which should be released sometimes today.", "Thank you! I upgraded both and it works." ]

[]

[]

[]

[]

[]

[ "I don't understand why you have a problem. `make quality` runs fine for me on master (and the CI is also happy).\r\n\r\n**Edit:** Not saying that this is useless, I just want to understand why it fails for you and not for me :-)", "Something is wrong on your side and CI. \r\n\r\n`check_repo.py` actually did its job correctly and reported problems that this PR fixes. Have a look at the fix and tell me if that checker shouldn't have caught it.\r\n\r\nSpecifically:\r\n\r\n* `test_modeling_flax_bert.py` and `test_modeling_flax_roberta.py` don't run common tests, and yet they weren't added to the ignore list\r\n* there is no `tests/test_modeling_flax_utils.py` so `modeling_flax_utils` has to be on the other ignore list\r\n\r\nI have no idea why you and CI don't reproduce these failures.", "I'm in the middle of something else right now, but will investigate once I'm done. I'm unsure of why the CI hasn't been more angry with the jax PR myself. But I'd like to understand why it's not failing for me and the CI before taking the fix if that makes sense.", "Absolutely, @sgugger. I will try to poke and see if the script behaves differently for some reason. I will report back if I find the culprit.\r\n\r\nIt should be safe for you to merge this as the lack of it may impacts other devs, I posted all the details why it's correct in here https://github.com/huggingface/transformers/pull/7914#issuecomment-712396143\r\n\r\nIt's your call.", "Found part of the culprit - my py37 env doesn't catch the problem, whereas py38 does - now need to figure out whether it's some package difference or python itself. ", "Oh, interesting! I'm indeed on Python 3.7.9", "and so is CI\r\n\r\nUsing my mightly https://github.com/stas00/conda-tools/blob/master/conda-env-compare.pl - I should get to the root of it in a few minutes", "I downgraded the the py38 env to py37 and it doesn't detect the problem anymore. Upgraded it back to py38 via conda and it fails now too! bummer - so it must have been some package. I need to dig more.\r\n\r\nI'd check `jax` and `flax` since the new flax code depends on it.", "I'm confused by your report: by \"it fails now too!\" do you mean you don't see the problem anymore?\r\n\r\n**Edit:** Think I've found the issue. It's because the presence/absence of jax/flax will change what's in the __init__. And neither me nor the CI have it installed.", "If my reasoning is correct, #7919 should be red. We can then add it to your fixes and merge all of this together.", "Yes, this is the culprit. \r\n\r\nIf you `pip install jax jaxlib flax` you should be able to reproduce the problem.\r\n\r\nSo basically the validation script is as good as the preinstalled pre-requisites allow it to be, therefore to move forward to do proper testing we need to have a prerequisites set that contains **all possible external packages** used by the core library.\r\n\r\nPerhaps we need to change `setup.py` to add:\r\n\r\n`extras[\"all\"] = list all groups here`\r\n\r\nand have the `check_code_quality` CI job installing `pip install -e .[all].\r\n\r\nBut specifically for this issue https://github.com/huggingface/transformers/pull/7919 will do the trick. I merged it here as you suggested.", "> I'm confused by your report: by \"it fails now too!\" do you mean you don't see the problem anymore?\r\n\r\nYeah, I was trying to figure out the difference and there were too many differences in installed modules, so I downgraded to py37, then upgraded to py38 and lost an environment that was good for the purpose of this issue. I eventually recovered it. I need to remember to back up conda envs before I try to mess with them :(" ]

[ { "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 } ]

[]

[ "Hey, I am getting the same error. \r\n\r\nI am using a three column CSV file which looks like this,\r\n\r\ndata.csv\r\nlabel,sent1,sent2\r\n0,he,so\r\n1,yes,why\r\n\r\nAny help would be appreciated. ", "The error is caused by the following function\r\n`\r\ntransformed_ds[k] = ds[k].map(\r\n lambda example: tokenizer.batch_encode_plus(\r\n (example[features_name[0]], example[features_name[1]]),\r\n truncation=True,\r\n max_length=max_seq_length,\r\n padding=\"max_length\",\r\n ),\r\n batched=True,\r\n )\r\n`\r\nWhen I set `batched=False`, it could pass; however, another error arises. Any idea? @jplu", "This issue has been automatically marked as stale and been closed because it has not had recent activity. Thank you for your contributions.\n\nIf you think this still needs to be addressed please comment on this thread." ]

[]

[]

[]

[]

[]

[ "Hey @ahoho - good question! \r\n\r\nI'm pretty confident that you should mask all sentinel tokens (with -100) and only compute the loss the \"real\" labels being \"cute dog\", \"the\" and \"</s>\". \r\n\r\nAlso they are definitely not automatically ignored as is done for the pad_token_id in `examples/seq2seq`\r\n\r\nI could not find a more detailed explanation in the paper - so maybe @craffel could take a quick look as well and confirm (hope it's fine to tag you here Colin)", "No need to treat the sentinel tokens specially (masking out their loss or otherwise). The model is trained to output both the sentinel tokens and the filled-in blanks." ]

[ { "id": 2368374212, "node_id": "MDU6TGFiZWwyMzY4Mzc0MjEy", "url": "https://api.github.com/repos/huggingface/transformers/labels/pegasus", "name": "pegasus", "color": "1f76a8", "default": false, "description": "" } ]

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

[ "I would also try without `--fp16`", "Sure @sshleifer I will try without `--fp16` and update the results here. Thanks for looking into this.", "Hi @sshleifer \r\n\r\nI ran the below command for distillation (without --fp16 as you suggested):\r\n`python finetune.py --learning_rate=3e-5 --do_train --do_predict --val_check_interval 0.25 --n_val 500 --num_train_epochs 6 --freeze_embeds --data_dir ./cnn_dm/ --max_target_length 142 --val_max_target_length=142 --train_batch_size=1 --eval_batch_size=1 --gradient_accumulation_steps=256 --model_name_or_path sshleifer/student_pegasus_cnn_12_2 --tokenizer_name google/pegasus-cnn_dailymail --warmup_steps 500 --output_dir distilpegasus-cnn-12-2 --gpus 1 --num_workers=0 --adafactor --freeze_encoder --task summarization --dropout 0.1 --attention_dropout 0.1 --label_smoothing 0.1 `\r\n\r\nHowever, the rouge scores are not improving even after 1 epoch\r\n`{\r\n \"val\": [\r\n {\r\n \"val_avg_loss\": 940.3131713867188,\r\n \"val_avg_rouge1\": 0.0,\r\n \"val_avg_rouge2\": 0.0,\r\n \"val_avg_rougeL\": 0.0,\r\n \"val_avg_rougeLsum\": 0.0,\r\n \"val_avg_gen_time\": 1.9830520153045654,\r\n \"val_avg_gen_len\": 128.0,\r\n \"step_count\": 1\r\n },\r\n {\r\n \"val_avg_loss\": 457.8860168457031,\r\n \"val_avg_rouge1\": 0.8307167999999999,\r\n \"val_avg_rouge2\": 0.0106524,\r\n \"val_avg_rougeL\": 0.8102172,\r\n \"val_avg_rougeLsum\": 0.8177266,\r\n \"val_avg_gen_time\": 1.9989106116294861,\r\n \"val_avg_gen_len\": 128.0,\r\n \"step_count\": 2\r\n },\r\n {\r\n \"val_avg_loss\": 297.9767761230469,\r\n \"val_avg_rouge1\": 2.7392655999999995,\r\n \"val_avg_rouge2\": 0.08615479999999999,\r\n \"val_avg_rougeL\": 2.4773216,\r\n \"val_avg_rougeLsum\": 2.6349664,\r\n \"val_avg_gen_time\": 1.7901806454658509,\r\n \"val_avg_gen_len\": 93.732,\r\n \"step_count\": 3\r\n },\r\n {\r\n \"val_avg_loss\": 272.0320129394531,\r\n \"val_avg_rouge1\": 4.0338778,\r\n \"val_avg_rouge2\": 0.2913826,\r\n \"val_avg_rougeL\": 3.4839722,\r\n \"val_avg_rougeLsum\": 3.7919970000000003,\r\n \"val_avg_gen_time\": 1.4304678964614868,\r\n \"val_avg_gen_len\": 47.67,\r\n \"step_count\": 4\r\n },\r\n {\r\n \"val_avg_loss\": 259.57611083984375,\r\n \"val_avg_rouge1\": 7.9237036000000005,\r\n \"val_avg_rouge2\": 0.7740864000000001,\r\n \"val_avg_rougeL\": 6.5176862,\r\n \"val_avg_rougeLsum\": 7.265688,\r\n \"val_avg_gen_time\": 1.3813148093223573,\r\n \"val_avg_gen_len\": 37.046,\r\n \"step_count\": 5\r\n }\r\n ]\r\n}\r\n`\r\n\r\nAfter 1 epoch, rogue2 score is 0.77. Could you please help if I am doing something wrong here?\r\n\r\nThanks in advance for your help.\r\n\r\nRegards,\r\nKarthik\r\n\r\n", "+ Note that scores are improving, just very slowly.\r\n+ I have not had good luck with `sshleifer/student_pegasus_cnn_12_2`, I'd try to make your own student with a full encoder and a 4+ layer decoder starting. Using, for example:\r\n\r\n```bash\r\npython make_student.py sshleifer/pegasus-cnn-ft-v2 -save_path student_peg_cnn_16_4 -e 16 -d 4\r\n```\r\n\r\nHere is the [wandb log](https://wandb.ai/sshleifer/pegasus_ft/runs/32ov7btf?workspace=user-) for a run that used `student_peg_cnn_16_4` \r\n\r\n\r\n\r\n\r\nI started at `--max_target_length 56` and then finetuned more with `--max_target_length 142`. That log is the first run. The second run is [here](https://wandb.ai/sshleifer/pegasus_ft/runs/2z1t4r0t?workspace=user-)\r\n\r\n\r\nFWIW, XSUM trains much faster!", "Thanks @sshleifer for your inputs.\r\n\r\nI am using this model (https://huggingface.co/sshleifer/distill-pegasus-cnn-16-4) which has 16 encoders and 4 decoders. I am trying to reduce the inference runtime of the model - for this reason, I am trying distillation with lesser encoders and decoders.\r\n\r\nCould you please suggest if I should try something different to reduce the inference runtime? \r\n\r\nRegards,\r\nKarthik\r\n", "Try generating with the 16/4 model and `num_beams=2`.\r\n", "Thanks @sshleifer for your suggestion. This improved the runtime. Please let me know if you have more such ideas.\r\n", "Besides that, all that's easy is to make your input documents shorter, or make your generations shorter (with min_length, max_length).\r\n" ]

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

[]

[ "This model is very big. Is there a good way to prune it?", "Moving to #8054 which is a duplicate (that I created)!", "> This model is very big. Is there a good way to prune it?\r\n\r\n@Bachstelze Did you find any ways to distill or prune such a large model?", "@robotsp \r\nThere is a smaller version: https://huggingface.co/alirezamsh/small100\r\n[SMaLL-100: Introducing Shallow Multilingual Machine Translation Model for Low-Resource Languages](https://aclanthology.org/2022.emnlp-main.571.pdf)" ]

[]

[]

[ "I don't know the answer to this question. Your numbers are close enough that all I can suggest is to either try fairseq's [command](https://github.com/pytorch/fairseq/blob/master/examples/bart/README.summarization.md#4-fine-tuning-on-cnn-dm-summarization-task) or look at differences between our command and fairseq.\r\n", "> Summarization: @sshleifer Bart: @sshleifer\r\n> \r\n> ## Information\r\n> I'm trying to finetune bart large for Xsum and unable to reproduce the results from the paper.\r\n> \r\n> When I try eval with facebook/bart-large-xsum, I get R1=45.3595, RLSum=37.1717 so I assume my eval script is working ok. For finetuning bart large, I use the same config as bart-large-xsum with vocab size=50265 to enable starting from bart-large. However, I am unable to reach the same scores. The best I have is R1=45.4188, RLSum=36.6986 with LR=1.2e-4, gbs=128 and --max_target_length=60 --max_source_length=1024 --val_check_interval 0.1 --val_max_target_length=60 --warmup_steps 50 --max_steps 5000.\r\n> \r\n> How can I reproduce the results?\r\n\r\nHi @swethmandava . I'm trying to reproduce the result by Transformres. Would you mind sharing your fine-tuning script?" ]

[]

[]

[]

[]

[]

[ "@patrickvonplaten can you please help", "Hey @lalitpagaria,\r\n\r\nUsing embedding, retrieval and generation separately for RagSequence is not yet available sadly.\r\nYou should take a look into the `generate()` function of `RagSequenceForGeneration` for more detail on how to run it separately yourself.", "Thanks @patrickvonplaten . I think we (haystack) will wait for implementation in transformers and use only RagToken for now.\r\nPlease let me know should I keep this open in case you plan to add functionality in the future? or close this.\r\n\r\ncc: @tholor ", "Closing" ]

[]

[]

[ "Hey Sam, it looks like the labels in the example you quoted are not shifted - can you be more specific about you think the labels are shifted?", "yes, I think the `labels` should be unshifted here (i.e `labels` should be same as `input_ids`) since `shift_right` takes care of preparing shifted `decoder_input_ids`.", "@craffel I assumed the labels were shifted because:\r\n+ Original: `The cute dog walks in the park`\r\n+ Input_ids: `The <extra_id_0> walks in <extra_id_1> park`\r\n+ Labels: `<extra_id_0> cute dog <extra_id_1> the <extra_id_2> </s>` \r\n\r\n`input_ids` starts with unmasked \"The\", whereas labels starts with a sentinel token. ", "I'm still not following - are you think the sentinel token `<extra_id_0>` is the same as the start-of-sequence token? They are different tokens.", "@sshleifer - I don't really understand the problem here either. In the example the `labels` are provided as:\r\n\r\n```python\r\n<extra_id_0> cute dog <extra_id_1> the <extra_id_2> </s>\r\n```\r\nwhich means that `decoder_input_ids` will be automatically created as:\r\n```python\r\n<s> <extra_id_0> cute dog <extra_id_1> the <extra_id_2>\r\n```\r\n\r\n=> This looks correct to me\r\n", "+1 to Patrick's take", "Aah, yes, for t5 we just predict the masked out spans, unlike BART. So this looks correct. ", "In the docs, the `</s>` is omitted from `input_ids`, but will be silently added due to #5866. Is this also the correct behavior?", "@ahoho => good point - I will update the docs to reflect this behavior", "@patrickvonplaten, thanks! Does this mean the docs were incorrect before? I guess my question is, for the denoising training, is it correct to append the `</s>` token to the `input_ids` (not `labels`) or isn't it?", "`</s>` should be appended IMO -> It's just that this is done automatically since #5866 as you mentioned above :-) " ]

[]

[]

[ "@patrickvonplaten @sgugger, please review", "Great catch @ayubSubhaniya !" ]

[]

[]

[ "But why doesn't this PR trigger a CI test? @LysandreJik ", "I have no idea, this is the second time it happens. I sent an empty commit on your branch to trigger the build @donchev7.", "Pinging @stefan-it for review." ]

[]

[]

[ "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 issue is still present.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "This issue is still present in `transformers==4.5.1`.", "Hey @jantrienes, sorry for getting back so late to you. The issue here is on our side - the added tokens work differently between the slow and fast tokenizers and we'll have to patch that. In the meantime, you can use `AddedToken`s instead of the strings to define the strategy w.r.t the stripping of whitespace. This should behave normally:\r\n\r\n```py\r\nimport os\r\nos.makedirs('models/', exist_ok=True)\r\n\r\nfrom transformers import GPT2Tokenizer\r\nfrom tokenizers import ByteLevelBPETokenizer, AddedToken\r\n\r\n\r\nopen('train.txt', 'w').write('Training data including a <special> token.')\r\n\r\nspecial_tokens = [AddedToken('<special>')]\r\n\r\nbpe_tokenizer = ByteLevelBPETokenizer()\r\nbpe_tokenizer.train(\r\n files=['train.txt'],\r\n special_tokens=special_tokens\r\n)\r\nbpe_tokenizer.save_model('models/')\r\n\r\ngpt2_tokenizer = GPT2Tokenizer.from_pretrained(\r\n 'models/',\r\n)\r\ngpt2_tokenizer.add_special_tokens({\"additional_special_tokens\": special_tokens})\r\n\r\ntext = 'A <special> token.'\r\n\r\nprint(bpe_tokenizer.encode(text).tokens)\r\nprint(gpt2_tokenizer.tokenize(text))\r\nassert bpe_tokenizer.encode(text).tokens == ['A', 'Ġ', '<special>', 'Ġ', 't', 'o', 'k', 'e', 'n', '.']\r\nassert gpt2_tokenizer.tokenize(text) == ['A', 'Ġ', '<special>', 'Ġ', 't', 'o', 'k', 'e', 'n', '.'\r\n```\r\n\r\nNote how the special tokens are defined using `AddedToken` instead of a string. Unfortunately, these cannot be passed during initialization as you've done, but I'm fixing this in https://github.com/huggingface/transformers/pull/11325.\r\n\r\nYou can control the `AddedToken`'s behavior relative to whitespace using the `rstrip` and `lstrip` keyword arguments.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "This issue is present for \"non-special\" tokens in transformer 4.12.3. The code explicitly strips spaces from the surrounding tokens. It does NOT look for the `AddedToken` class to get the behavior so there is no way to change this.\r\n\r\nSee `tokenization_utils.py::tokenizer()` line 517:\r\n```\r\n else:\r\n # We strip left and right by default\r\n if right:\r\n tokens[i + 1] = right.lstrip()\r\n if left:\r\n tokens[i - 1] = left.rstrip()\r\n```\r\nThis leads to incorrect behavior unless the added token is in the middle of a longer word.\r\n\r\nI should point out that the situation is more complicated than just changing the behavior above. Even if you comment out those lines (at least with the T5Tokenizer) there are still no spaces around the added tokens because, just below those lines, the logic to concatenate all the tokens is also stripping the spaces as well (see `tokenized_text.extend(self._tokenize(token))`).\r\n\r\nI'm not sure what the right solution is but this is in the base class so it's happening for a number of tokenizers including T5 and bart. It is also happening for T5Fast.", "@SaulLu, would you like to take a look at this?", "Glad to look at this issue in more detail! I'll dive in tomorrow :monocle_face: ", "\r\n> I'm not sure what the right solution is but this is in the base class so it's happening for a number of tokenizers including T5 and bart. It is also happening for T5Fast.\r\n\r\nYeah, I'm noticing spaces around my added tokens when creating a custom Bart tokenizer. Did this get resolved for you? Is there a way to work around it? " ]

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

[]

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

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

[]

[]

[]

[ { "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 } ]

[ "Confirmed by author that BOS and EOS have to be added -> fixed in https://github.com/huggingface/transformers/commit/0724c0f3a2d302246d0bd0b7d2f721fa902dee1b." ]

[ { "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 } ]

[ { "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 } ]

[ "Hey @Aakash12980,\r\n\r\nCould you please provide a complete code examples with which I can reproduce the results? Did you train the model yourself or did you use a pretrained model? (Which one in case you did?)", "> Hey @Aakash12980,\r\n> \r\n> Could you please provide a complete code examples with which I can reproduce the results? Did you train the model yourself or did you use a pretrained model? (Which one in case you did?)\r\n\r\n@patrickvonplaten I have used pretrained bert2bert model and I am fine-tuning it for sentence simplification. \r\n\r\nThe full code in run.py is below\r\n\r\n```\r\nlogging.basicConfig(filename=\"./drive/My Drive/Mini Project/log_file.log\", level=logging.INFO, \r\n format=\"%(asctime)s:%(levelname)s: %(message)s\")\r\nCONTEXT_SETTINGS = dict(help_option_names = ['-h', '--help'])\r\n\r\nTRAIN_BATCH_SIZE = 4 \r\nN_EPOCH = 5\r\nLOG_EVERY = 11000\r\n\r\nconfig_encoder = BertConfig()\r\nconfig_decoder = BertConfig()\r\nconfig_decoder.is_decoder = True\r\nconfig_decoder.add_cross_attention = True\r\nconfig = EncoderDecoderConfig.from_encoder_decoder_configs(config_encoder, config_decoder)\r\nmodel = EncoderDecoderModel.from_encoder_decoder_pretrained('bert-base-cased', 'bert-base-cased', config=config)\r\n\r\ndevice = \"cuda\" if torch.cuda.is_available() else \"cpu\"\r\nprint(f\"Using {device} as device\")\r\n\r\n\r\ndef collate_fn(batch):\r\n data_list, label_list = [], []\r\n for _data, _label in batch:\r\n data_list.append(_data)\r\n label_list.append(_label)\r\n return data_list, label_list\r\n\r\ndef save_model_checkpt(state, is_best, check_pt_path, best_model_path):\r\n f_path = check_pt_path\r\n torch.save(state, f_path)\r\n\r\n if is_best:\r\n best_fpath = best_model_path\r\n shutil.copyfile(f_path, best_fpath)\r\n\r\ndef load_checkpt(checkpt_path, optimizer=None):\r\n if device == \"cpu\":\r\n checkpoint = torch.load(checkpt_path, map_location=torch.device(\"cpu\"))\r\n model.load_state_dict(checkpoint[\"model_state_dict\"])\r\n if optimizer is not None:\r\n optimizer.load_state_dict(checkpoint[\"optimizer_state_dict\"], map_location=torch.device(\"cpu\"))\r\n \r\n else:\r\n model.load_state_dict(checkpoint[\"model_state_dict\"])\r\n if optimizer is not None:\r\n optimizer.load_state_dict(checkpoint[\"optimizer_state_dict\"])\r\n\r\n eval_loss = checkpoint[\"eval_loss\"]\r\n epoch = checkpoint[\"epoch\"]\r\n\r\n return optimizer, eval_loss, epoch\r\n\r\n\r\n\r\ndef evaluate(batch_iter, e_loss):\r\n was_training = model.training\r\n model.eval()\r\n eval_loss = e_loss\r\n\r\n with torch.no_grad():\r\n for step, batch in enumerate(batch_iter):\r\n src_tensors, src_attn_tensors, tgt_tensors, tgt_attn_tensors = generate_tokens(batch)\r\n loss, _ = model(input_ids = src_tensors.to(device), \r\n decoder_input_ids = tgt_tensors.to(device),\r\n attention_mask = src_attn_tensors.to(device),\r\n decoder_attention_mask = tgt_attn_tensors.to(device),\r\n labels=tgt_tensors.to(device))[:2]\r\n \r\n eval_loss += (1/(step+1)) * (loss.item() - eval_loss)\r\n\r\n if was_training:\r\n model.train()\r\n\r\n return eval_loss \r\n\r\[email protected](context_settings=CONTEXT_SETTINGS)\r\[email protected]_option(version = '1.0.0')\r\ndef task():\r\n ''' This is the documentation of the main file. This is the reference for executing this file.'''\r\n pass\r\n\r\n\r\[email protected]()\r\[email protected]('--src_train', default=\"./drive/My Drive/Mini Project/dataset/src_train.txt\", help=\"train source file path\")\r\[email protected]('--tgt_train', default=\"./drive/My Drive/Mini Project/dataset/tgt_train.txt\", help=\"train target file path\")\r\[email protected]('--src_valid', default=\"./drive/My Drive/Mini Project/dataset/src_valid.txt\", help=\"validation source file path\")\r\[email protected]('--tgt_valid', default=\"./drive/My Drive/Mini Project/dataset/tgt_valid.txt\", help=\"validation target file path\")\r\[email protected]('--best_model', default=\"./drive/My Drive/Mini Project/best_model/model.pt\", help=\"best model file path\")\r\[email protected]('--checkpoint_path', default=\"./drive/My Drive/Mini Project/checkpoint/model_ckpt.pt\", help=\" model check point files path\")\r\[email protected]('--seed', default=123, help=\"manual seed value (default=123)\")\r\ndef train(**kwargs):\r\n print(\"Training data module executing...\")\r\n logging.info(f\"Train module invoked.\")\r\n seed = kwargs[\"seed\"]\r\n torch.manual_seed(seed)\r\n if device == \"cuda\":\r\n torch.cuda.manual_seed(seed)\r\n \r\n np.random.seed(seed)\r\n print(\"Loading dataset...\")\r\n \r\n src_train = open_file(kwargs['src_train'])\r\n tgt_train = open_file(kwargs['tgt_train'])\r\n src_valid = open_file(kwargs['src_valid'])\r\n tgt_valid = open_file(kwargs['tgt_valid'])\r\n train_len = len(src_train)\r\n valid_len = len(src_valid)\r\n print(\"Dataset Loaded.\")\r\n\r\n train_dataset = WikiDataset(src_train, tgt_train)\r\n valid_dataset = WikiDataset(src_valid, tgt_valid)\r\n del src_valid, src_train, tgt_train, tgt_valid\r\n\r\n print(\"Creating Dataloader...\")\r\n train_dl = DataLoader(train_dataset, batch_size=TRAIN_BATCH_SIZE, collate_fn=collate_fn, shuffle=True)\r\n valid_dl = DataLoader(valid_dataset, batch_size=TRAIN_BATCH_SIZE, collate_fn=collate_fn, shuffle=True)\r\n print(\"Dataloader created.\")\r\n\r\n model.to(device)\r\n param_optimizer = list(model.named_parameters())\r\n no_decay = ['bias', 'gamma', 'beta']\r\n optimizer_grouped_parameters = [\r\n {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],\r\n 'weight_decay_rate': 0.01},\r\n {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)],\r\n 'weight_decay_rate': 0.0}\r\n ]\r\n optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=3e-3)\r\n eval_loss = float('inf')\r\n start_epoch = 0\r\n if os.path.exists(kwargs[\"checkpoint_path\"]):\r\n optimizer, eval_loss, start_epoch = load_checkpt(kwargs[\"checkpoint_path\"], optimizer)\r\n print(f\"Loading model from checkpoint with start epoch: {start_epoch} and loss: {eval_loss}\")\r\n logging.info(f\"Model loaded from saved checkpoint with start epoch: {start_epoch} and loss: {eval_loss}\")\r\n\r\n train_model(start_epoch, eval_loss, (train_dl, valid_dl), optimizer, kwargs[\"checkpoint_path\"], kwargs[\"best_model\"], (train_len, valid_len))\r\n print(\"Model Training Complete!\")\r\n \r\n \r\n\r\[email protected]()\r\[email protected]('--src_test', default=\"./drive/My Drive/Mini Project/dataset/src_test.txt\", help=\"test source file path\")\r\[email protected]('--tgt_test', default=\"./drive/My Drive/Mini Project/dataset/tgt_test.txt\", help=\"test target file path\")\r\[email protected]('--best_model', default=\"./drive/My Drive/Mini Project/best_model/model.pt\", help=\"best model file path\")\r\ndef test(**kwargs):\r\n print(\"Testing Model module executing...\")\r\n logging.info(f\"Test module invoked.\")\r\n\r\n src_test = open_file(kwargs['src_test'])\r\n tgt_test = open_file(kwargs['tgt_test'])\r\n len_data = len(src_test)\r\n\r\n test_dataset = WikiDataset(src_test, tgt_test)\r\n\r\n test_dl = DataLoader(test_dataset, batch_size=TRAIN_BATCH_SIZE, collate_fn=collate_fn, shuffle=True)\r\n\r\n _,_,_ = load_checkpt(kwargs[\"best_model\"])\r\n print(\"Model loaded...\")\r\n model.to(device)\r\n model.eval()\r\n\r\n test_start_time = time.time()\r\n epoch_test_loss = evaluate(test_dl, 0)\r\n epoch_test_loss = epoch_test_loss/len_data\r\n print(f'avg. test loss: {epoch_test_loss:.5f} | time elapsed: {time.time() - test_start_time}')\r\n logging.info(f'avg. test loss: {epoch_test_loss:.5f} | time elapsed: {time.time() - test_start_time}')\r\n print(\"Test Complete!\")\r\n \r\n\r\n#/drive/My Drive/Mini Project\r\[email protected]()\r\[email protected]('--src_file', default=\"./drive/My Drive/Mini Project/dataset/src_file.txt\", help=\"test source file path\")\r\[email protected]('--best_model', default=\"./drive/My Drive/Mini Project/checkpoint/model_ckpt.pt\", help=\"best model file path\")\r\[email protected]('--output', default=\"./drive/My Drive/Mini Project/outputs/decoded.txt\", help=\"file path to save predictions\")\r\ndef decode(**kwargs):\r\n print(\"Decoding sentences module executing...\")\r\n logging.info(f\"Decode module invoked.\")\r\n src_test = open_file(kwargs['src_file'])\r\n print(\"Saved model loading...\")\r\n _,_,_ = load_checkpt(kwargs[\"best_model\"])\r\n print(f\"Model loaded.\")\r\n model.to(device)\r\n model.eval()\r\n inp_tokens = create_sent_tokens(src_test)\r\n predicted_list = []\r\n print(\"Decoding Sentences...\")\r\n for tensor in inp_tokens:\r\n with torch.no_grad():\r\n predicted = model.generate(tensor.to(device), decoder_start_token_id=model.config.decoder.pad_token_id)\r\n print(f\"input: {tensor}\")\r\n print(f'output: {predicted.squeeze()}')\r\n predicted_list.append(predicted.squeeze())\r\n \r\n output = get_sent_from_tokens(predicted_list)\r\n with open(kwargs[\"output\"], \"w\") as f:\r\n for sent in output:\r\n f.write(sent + \"\\n\")\r\n print(\"Output file saved successfully.\")\r\n\r\ndef train_model(start_epoch, eval_loss, loaders, optimizer, check_pt_path, best_model_path, len_data):\r\n best_eval_loss = eval_loss\r\n print(\"Model training started...\")\r\n for epoch in range(start_epoch, N_EPOCH):\r\n print(f\"Epoch {epoch} running...\")\r\n epoch_start_time = time.time()\r\n epoch_train_loss = 0\r\n epoch_eval_loss = 0\r\n\r\n model.train()\r\n for step, batch in enumerate(loaders[0]):\r\n\r\n src_tensors, src_attn_tensors, tgt_tensors, tgt_attn_tensors = generate_tokens(batch)\r\n\r\n optimizer.zero_grad()\r\n model.zero_grad()\r\n loss = model(input_ids = src_tensors.to(device), \r\n decoder_input_ids = tgt_tensors.to(device),\r\n attention_mask = src_attn_tensors.to(device),\r\n decoder_attention_mask = tgt_attn_tensors.to(device),\r\n labels = tgt_tensors.to(device))[0]\r\n \r\n loss.backward()\r\n optimizer.step()\r\n epoch_train_loss += (1/(step+1))*(loss.item() - epoch_train_loss)\r\n\r\n if (step+1) % LOG_EVERY == 0:\r\n print(f'Epoch: {epoch} | iter: {step+1} | avg. loss: {epoch_train_loss/TRAIN_BATCH_SIZE} | time elapsed: {time.time() - epoch_start_time}')\r\n logging.info(f'Epoch: {epoch} | iter: {step+1} | avg. loss: {epoch_train_loss/TRAIN_BATCH_SIZE} | time elapsed: {time.time() - epoch_start_time}')\r\n eval_start_time = time.time()\r\n epoch_eval_loss = evaluate(loaders[1], epoch_eval_loss)\r\n epoch_eval_loss = epoch_eval_loss/TRAIN_BATCH_SIZE\r\n print(f'Completed Epoch: {epoch} | avg. eval loss: {epoch_eval_loss:.5f} | time elapsed: {time.time() - eval_start_time}')\r\n logging.info(f'Completed Epoch: {epoch} | avg. eval loss: {epoch_eval_loss:.5f} | time elapsed: {time.time() - eval_start_time}')\r\n \r\n check_pt = {\r\n 'epoch': epoch+1,\r\n 'model_state_dict': model.state_dict(),\r\n 'optimizer_state_dict': optimizer.state_dict(),\r\n 'eval_loss': epoch_eval_loss\r\n }\r\n check_pt_time = time.time()\r\n print(\"Saving Checkpoint.......\")\r\n if epoch_eval_loss < best_eval_loss:\r\n print(\"New best model found\")\r\n logging.info(f\"New best model found\")\r\n best_eval_loss = epoch_eval_loss\r\n save_model_checkpt(check_pt, True, check_pt_path, best_model_path)\r\n else:\r\n save_model_checkpt(check_pt, False, check_pt_path, best_model_path) \r\n print(f\"Checkpoint saved successfully with time: {time.time() - check_pt_time}\")\r\n logging.info(f\"Checkpoint saved successfully with time: {time.time() - check_pt_time}\")\r\n \r\n gc.collect()\r\n torch.cuda.empty_cache() \r\n\r\n\r\nif __name__ == \"__main__\":\r\n task()\r\n```\r\n\r\nThe GitHub link for my full project is here [https://github.com/Aakash12980/Sentence-Simplification-using-Transformer](url). ", "Hey @Aakash12980 - I think you forgot to replace the pad_token_ids with -100 in your code. You can checkout the map function here: https://huggingface.co/patrickvonplaten/bert2bert-cnn_dailymail-fp16#training-script to see how this should be done.", "> Hey @Aakash12980 - I think you forgot to replace the pad_token_ids with -100 in your code. You can checkout the map function here: https://huggingface.co/patrickvonplaten/bert2bert-cnn_dailymail-fp16#training-script to see how this should be done.\r\n\r\n@patrickvonplaten I have included -100 as well in the labels and again I am getting the same weird outputs from my model.generate() method. \r\n```\r\nSetting `pad_token_id` to 102 (first `eos_token_id`) to generate sequence\r\noutput: tensor([101, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119], device='cuda:0')\r\nSetting `pad_token_id` to 102 (first `eos_token_id`) to generate sequence\r\noutput: tensor([101, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119, 119,\r\n 119, 119], device='cuda:0')\r\n```\r\n\r\nMy updated code is this link: [https://github.com/Aakash12980/Sentence-Simplification-using-Transformer](url)", "Hey @Aakash12980, hmm - I'm not sure what to do here then either. I've done quite a lot of Bert2Bert fine-tune runs and they all worked for me. Also I will publish a more in-detail notebook in ~2 weeks on how to do Bert2Bert training. \r\n\r\nIt's quite time-consuming and difficult for me to dive deeper into your code, so the best I can do for you at the moment is this training code snippet: https://huggingface.co/patrickvonplaten/bert2bert-cnn_dailymail-fp16#training-script", "@patrickvonplaten Thank you so much for your time. ", "@patrickvonplaten Could you please explain to me why you have assigned -100 to PAD tokens?\r\n```\r\n# mask loss for padding\r\n batch[\"labels\"] = [\r\n [-100 if token == tokenizer.pad_token_id else token for token in labels] for labels in batch[\"labels\"]\r\n ]\r\n```\r\nI don't quite understand what mask loss for padding means. Also, Do I need to pass labels when I am not using Trainer() method?\r\n```\r\nloss, logits = self.model(input_ids = src_tensors.to(device), \r\n decoder_input_ids = tgt_tensors.to(device),\r\n attention_mask = src_attn_tensors.to(device),\r\n decoder_attention_mask = tgt_attn_tensors.to(device),\r\n labels = labels.to(device))[:2]\r\n```\r\nWhy do we pass labels since we have already provided decoder_iniput_ids?", "if label == -100, it means that on such tokens no loss is calculated and therefore does not affect the gradient. We don't want the model to learn to predict PAD tokens, so we make sure that no loss is calculated by putting a -100 on them. PyTorch's CE loss function uses -100 as the ignore token number.", "Thank you @patrickvonplaten I have trained my model but my model is generating tokens exactly similar to the input tokens. I tried to find out what have I done wrong but I couldn't. I don't know why it is generating the same tokens as inputs. Do you have any idea?\r\n```\r\nModel loaded.\r\nDecoding Sentences...\r\ninput tokens: tensor([[ 101, 140, 23339, 1706, 12788, 140, 22715, 8469, 1106, 2194,\r\n 1103, 5080, 1104, 172, 23339, 17458, 1107, 170, 172, 22715,\r\n 24759, 3536, 119, 102]])\r\nSetting `pad_token_id` to 102 (first `eos_token_id`) to generate sequence\r\noutput tokens: tensor([ 101, 140, 23339, 1706, 12788, 140, 22715, 8469, 1106, 2194,\r\n 1103, 5080, 1104, 172, 23339, 17458, 1107, 170, 172, 22715,\r\n 24759, 3536, 119, 102], device='cuda:0')\r\ninput tokens: tensor([[ 101, 18959, 18072, 1108, 1549, 1103, 13852, 6607, 1104, 156,\r\n 14046, 3263, 1670, 1118, 1624, 1600, 1107, 18563, 1571, 117,\r\n 1106, 9489, 1105, 2669, 1103, 2226, 112, 188, 16358, 20408,\r\n 119, 102]])\r\nSetting `pad_token_id` to 102 (first `eos_token_id`) to generate sequence\r\noutput tokens: tensor([ 101, 18959, 18072, 1108, 1549, 1103, 13852, 6607, 1104, 156,\r\n 14046, 3263, 1670, 1118, 1624, 1600, 1107, 18563, 1571, 117,\r\n 1106, 9489, 1105, 2669, 1103, 2226, 112, 188, 16358, 20408,\r\n 119, 102], device='cuda:0')\r\ninput tokens: tensor([[ 101, 153, 14272, 1633, 1110, 170, 5188, 1107, 1103, 1318,\r\n 26042, 2853, 1107, 1564, 118, 2466, 1699, 119, 102]])\r\nSetting `pad_token_id` to 102 (first `eos_token_id`) to generate sequence\r\noutput tokens: tensor([ 101, 153, 14272, 1633, 1110, 170, 5188, 1107, 1103, 1318,\r\n 26042, 2853, 1107, 1564, 118, 2466, 1699, 119, 102],\r\n device='cuda:0')\r\n```\r\nIt is driving me crazy. I ran the training module for 4 epochs and this was the best model I got so far.", "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" ]

[]

[]

[ "> For common tests to be marked slow, the slowest iteration of that common test must be > 15s.\r\n\r\nThis one probably needs to be more specific, if for a normal test we decide a threshold of 5s, and you get a common test runs like: 14s, 13, 13, 12s - this is still too slow. perhaps an average of all commont tests for this test should be calculated?\r\n\r\ne.g. this one:\r\n\r\n```\r\n$ grep test_model_outputs_equivalence stats.txt | perl -ne 's|^(.*?)s.|$x+=$1|e; END {print int $x}'\r\n230\r\n$ grep test_model_outputs_equivalence stats.txt | wc -l\r\n46\r\n$ perl -le 'print 230/46'\r\n5\r\n```\r\nHow interesting that it hits 5sec exactly, but it's on my machine so need to re-eval on CI.\r\n\r\nHere is a one liner to do all 3 lines at once: (`$.` contains the same data as `wc -l` - line counter in perl)\r\n```\r\n$ grep test_model_outputs_equivalence stats.txt | perl -ne 's|^(.*?)s.|$x+=$1|e; END {print int $x/$.}'\r\n5\r\n```\r\n\r\nThe full picture for this common test:\r\n```\r\n20.11s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_model_outputs_equivalence\r\n16.19s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_model_outputs_equivalence\r\n13.49s call tests/test_modeling_gpt2.py::GPT2ModelTest::test_model_outputs_equivalence\r\n9.94s call tests/test_modeling_bert.py::BertModelTest::test_model_outputs_equivalence\r\n9.56s call tests/test_modeling_albert.py::AlbertModelTest::test_model_outputs_equivalence\r\n8.81s call tests/test_modeling_flaubert.py::FlaubertModelTest::test_model_outputs_equivalence\r\n8.29s call tests/test_modeling_tf_xlm.py::TFXLMModelTest::test_model_outputs_equivalence\r\n7.98s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_model_outputs_equivalence\r\n7.87s call tests/test_modeling_xlnet.py::XLNetModelTest::test_model_outputs_equivalence\r\n6.85s call tests/test_modeling_tf_xlnet.py::TFXLNetModelTest::test_model_outputs_equivalence\r\n6.81s call tests/test_modeling_squeezebert.py::SqueezeBertModelTest::test_model_outputs_equivalence\r\n6.30s call tests/test_modeling_tf_roberta.py::TFRobertaModelTest::test_model_outputs_equivalence\r\n6.25s call tests/test_modeling_roberta.py::RobertaModelTest::test_model_outputs_equivalence\r\n5.90s call tests/test_modeling_reformer.py::ReformerLocalAttnModelTest::test_model_outputs_equivalence\r\n5.81s call tests/test_modeling_electra.py::ElectraModelTest::test_model_outputs_equivalence\r\n5.79s call tests/test_modeling_distilbert.py::DistilBertModelTest::test_model_outputs_equivalence\r\n5.69s call tests/test_modeling_xlm.py::XLMModelTest::test_model_outputs_equivalence\r\n5.35s call tests/test_modeling_tf_t5.py::TFT5ModelTest::test_model_outputs_equivalence\r\n4.64s call tests/test_modeling_tf_funnel.py::TFFunnelBaseModelTest::test_model_outputs_equivalence\r\n4.34s call tests/test_modeling_tf_bert.py::TFBertModelTest::test_model_outputs_equivalence\r\n3.79s call tests/test_modeling_dpr.py::DPRModelTest::test_model_outputs_equivalence\r\n3.71s call tests/test_modeling_tf_funnel.py::TFFunnelModelTest::test_model_outputs_equivalence\r\n3.61s call tests/test_modeling_bart.py::BARTModelTest::test_model_outputs_equivalence\r\n3.58s call tests/test_modeling_openai.py::OpenAIGPTModelTest::test_model_outputs_equivalence\r\n3.57s call tests/test_modeling_tf_transfo_xl.py::TFTransfoXLModelTest::test_model_outputs_equivalence\r\n3.53s call tests/test_modeling_ctrl.py::CTRLModelTest::test_model_outputs_equivalence\r\n3.40s call tests/test_modeling_tf_lxmert.py::TFLxmertModelTest::test_model_outputs_equivalence\r\n3.31s call tests/test_modeling_tf_gpt2.py::TFGPT2ModelTest::test_model_outputs_equivalence\r\n3.19s call tests/test_modeling_tf_electra.py::TFElectraModelTest::test_model_outputs_equivalence\r\n3.12s call tests/test_modeling_tf_openai.py::TFOpenAIGPTModelTest::test_model_outputs_equivalence\r\n2.98s call tests/test_modeling_tf_albert.py::TFAlbertModelTest::test_model_outputs_equivalence\r\n2.93s call tests/test_modeling_tf_flaubert.py::TFFlaubertModelTest::test_model_outputs_equivalence\r\n2.80s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_model_outputs_equivalence\r\n2.59s call tests/test_modeling_longformer.py::LongformerModelTest::test_model_outputs_equivalence\r\n2.37s call tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_model_outputs_equivalence\r\n2.17s call tests/test_modeling_funnel.py::FunnelModelTest::test_model_outputs_equivalence\r\n2.13s call tests/test_modeling_fsmt.py::FSMTModelTest::test_model_outputs_equivalence\r\n2.02s call tests/test_modeling_bert_generation.py::BertGenerationEncoderTest::test_model_outputs_equivalence\r\n1.94s call tests/test_modeling_funnel.py::FunnelBaseModelTest::test_model_outputs_equivalence\r\n1.70s call tests/test_modeling_t5.py::T5ModelTest::test_model_outputs_equivalence\r\n1.60s call tests/test_modeling_deberta.py::DebertaModelTest::test_model_outputs_equivalence\r\n1.44s call tests/test_modeling_lxmert.py::LxmertModelTest::test_model_outputs_equivalence\r\n1.22s call tests/test_modeling_layoutlm.py::LayoutLMModelTest::test_model_outputs_equivalence\r\n1.11s call tests/test_modeling_reformer.py::ReformerLSHAttnModelTest::test_model_outputs_equivalence\r\n0.86s call tests/test_modeling_tf_ctrl.py::TFCTRLModelTest::test_model_outputs_equivalence\r\n0.33s call tests/test_modeling_blenderbot.py::BlenderbotTesterMixin::test_model_outputs_equivalence\r\n```\r\n```", "Please have a look at this recent [report](https://github.com/huggingface/transformers/issues/7885#issuecomment-712287004) and let me know if anything else should be marked as slow.\r\n", "@LysandreJik, I think the 3 of us have had a good go at it. At your convenience please review the resulting prose and if there is something to modify please proceed to change it directly and then merge it. Much appreciated!" ]

[]

[]

[ "Hi, yes, this is a warning, nothing you should be afraid of." ]

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

[]

[ "Running `pip install transformers` installs the latest version for me. Which pip version are you running on?\r\n\r\nYou can install both cpu or gpu for tensorflow, if you want to run on a cpu or a gpu.", "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 } ]

[]

[ "Thanks for reporting, will investigate.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi @danyaljj and @LysandreJik I have the same issue. Are there any suggestions for running xlm-r (base/large) with multiple-choice qa?" ]

[]

[]

[ "@lhoestq would be great if you can review as well", "@patrickvonplaten Thanks for the review.\r\n\r\nwhile working on this PR I found that in `RagTokenForGeneration` we are computing `batch_size` as follows -\r\n```\r\nbatch_size = context_input_ids.shape[0] // n_docs\r\n```\r\nSo still issue can come when `((context_input_ids.shape[0] % n_docs) != 0)`, but I can't think of solution to address this.", "> @patrickvonplaten Thanks for the review.\r\n> \r\n> while working on this PR I found that in `RagTokenForGeneration` we are computing `batch_size` as follows -\r\n> \r\n> ```\r\n> batch_size = context_input_ids.shape[0] // n_docs\r\n> ```\r\n> \r\n> So still issue can come when `((context_input_ids.shape[0] % n_docs) != 0)`, but I can't think of solution to address this.\r\n\r\n`context_input_ids` is always supposed to have a size of `n_docs` times the number of input questions ", "> > @patrickvonplaten Thanks for the review.\r\n> > while working on this PR I found that in `RagTokenForGeneration` we are computing `batch_size` as follows -\r\n> > ```\r\n> > batch_size = context_input_ids.shape[0] // n_docs\r\n> > ```\r\n> > \r\n> > \r\n> > So still issue can come when `((context_input_ids.shape[0] % n_docs) != 0)`, but I can't think of solution to address this.\r\n> \r\n> `context_input_ids` is always supposed to have a size of `n_docs` times the number of input questions\r\n\r\nIt would be better if we mention it explicitly by assert. WDYT? \r\nIn one of my test case I used `n_docs=3` for retriever and `n_docs=2` for generator and it failed", "> It would be better if we mention it explicitly by assert. WDYT?\r\n> In one of my test case I used `n_docs=3` for retriever and `n_docs=2` for generator and it failed\r\n\r\nYes indeed. Also if `((context_input_ids.shape[0] % n_docs) != 0)` then we should raise an error otherwise some retrieved documents will be ignored for generation.", "Yes @lalitpagaria - it would be nice if you can add an asserte statement verifying that `n_docs` is correctly set. `n_docs` should be the same for both retriever and generator.", "@patrickvonplaten @lhoestq Added assert at two places please verify, along with supporting unit test. Pardon my naming convention for test function, and please suggest proper name :) \r\n\r\n> n_docs should be the same for both retriever and generator.\r\n\r\nThis can't be check if `generator` does not know about `retriever` hence using this `((context_input_ids.shape[0] % n_docs) != 0)`", "@patrickvonplaten and @lhoestq Thanks for the review. I liked the test coverage of this project. Initially I struggled but letter all worked nicely. You can merge when you want.", "Slow tests pass => ready to merge", "Good job @lalitpagaria !" ]

[]

[]

[ "@sshleifer, I don't think I can do anything here w/o repo permissions. I have no access to the runners.\r\n\r\nhttps://github.com/stas00/transformers/actions/runs/319165633\r\n\r\n> No runner matching the specified labels was found: self-hosted, multi-gpu", "github is so borked! It now won't let me delete the new workflow - and keeps running it automatically and emailing me on each commit I make elsewhere in this project - what horrors!", "ask on https://github.community/, they are pretty responsive." ]

[]

[]

[ "I let @LysandreJik answer about the model but I have a tangential question I love to have some information about @sebastianbujwid (we may be able to help):\r\nWhy is it not possible for you to update transformers past the 2.3.0 version?", "Hello! Indeed, there has been an issue with the TF ALBERT model that has been resolved in the following commit: https://github.com/huggingface/transformers/commit/1abd53b1aa2f15953bbbbbfefda885d1d9c9d94b#diff-143909fef69337378ffd5ccb1af28c3855ba498a11296fb62886b854d269342d\r\n\r\nThis commit is, unfortunately, only available on version v2.5.1. Is it possible for you to update your version to v2.5.1? If not, let me know specifically which models you would like and I'll give you URLs you can use to download these specific model weights, but please understand that *before v2.5.1*, the TF version of the ALBERT models will not correctly if using them with heads (e.g. SequenceClassification, TokenClassification, QuestionAnswering, etc.)", "Thank you so much for the quick response!\r\n\r\nI did more digging into the issue and realized that I actually still have the old model weights in my `.cache`, they are just not used since the model checks the URL and downloads the new weights anyway (expects the cached weights with the newest e-tag).\r\n\r\nAlthough I haven't verified it yet, I'm quite confident I should be able to either use the cached weights I have or maybe upgrade to v2.5.1 (assuming the API hasn't changed much, for practical reasons I'd strongly prefer not to change the code since would have run additional test and re-do at least some of my main experiments).\r\nOtherwise, I might reach out again for the specific URLs. Thanks :)\r\n\r\nMaybe the issue with incompatibility is not fixable for older versions anymore but for the future maybe it would be good to do some versioning of the weight files and check for the right version when downloading/reading from the cache, in case that's not done yet? It's a shame if some code suddenly stops working without being changed.\r\nRegarding the logging, after a quick check now I see that the recent code from the main branch uses `logging.warning` instead which should at least make the issue easier to notice (sadly, it took me a few days to realize that all my problems were due to the model not loading the pre-trained weights - apparently ALBERT with random weights still works much better than random on certain problems).", "You're absolutely right that these incompatibility issues are due to not having a way to version models. You'll be happy to hear that we're currently working on this internally :).\r\n\r\nThank you for opening such a detailed issue!" ]

[]

[]

[ "Merging as this just fixes tests." ]

[ { "id": 1108649053, "node_id": "MDU6TGFiZWwxMTA4NjQ5MDUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Help%20wanted", "name": "Help wanted", "color": "008672", "default": false, "description": "Extra attention is needed, help appreciated" }, { "id": 1834088753, "node_id": "MDU6TGFiZWwxODM0MDg4NzUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Tests", "name": "Tests", "color": "a6fcca", "default": false, "description": "Related to tests" } ]

[]

[ "I will surely give it a try, @sshleifer ", "So the idea of running my own gh action didn't quite work https://github.com/huggingface/transformers/pull/7890 as I don't have the right permissions to do so.\r\n\r\nBut that doesn't mean this issue can't be worked on.\r\n\r\n# Making information easier/quicker to find \r\n\r\nBasically we have the logs of the tests run and nothing else. Currently finding something in those logs is not trivial, so this part can definitely be improved. We could `tee` the logs into a file and then do a postprocessing on those (will need to figure out how to run a subsequent task despite the test suite task failure, but I'm sure in the worst case some bash magic could work around it). So we can make various clickable tasks that will show just specific outputs. I can also check whether there are some pytest extensions that might help here. Or we might write our own extension that will do just that - log different interesting things into different files.\r\n\r\nYou can tell me which info chunks you want and I can work on generating those from the log file. Currently you mentioned:\r\n- show slow tests (already in the log file, just need to make them a separate one-click accessible item, I think pytest has an option to log that into a separate report)\r\n- show just the errors\r\n\r\nanything else?\r\n\r\nThis part would be just as useful to normal CIs.\r\n\r\n# Communicating with specific user\r\n\r\na. the first problem is which user? a scheduled job may have dozens of merges in it - how do you find out which of them made the commit that caused a failure? \r\n\r\n Unless you make some kind of contact-table and look up the contact based on which test file failed. I was under the impression you wanted to contact not the specific maintainer, but the original PR creator.\r\n\r\nb. any push communications require some kind of auth, which would be very tricky with public repo. it could probably use some webhooks to ping some external server that could do something with that info. e.g. it could have the right auth and post on failure to a slack channel which devs can watch - easier than watching git hub actions (pull model).\r\n", "@sgugger suggested that CircleCI has an Artifacts tab, where we could add various bit of info.", "I want exactly the artifacts file of the circleci job (just uses `|tee output.txt`) available in github actions, but it sounds like that is hard.\r\n\r\nRe, communicating with specific user > can be just lysandre for now. I already added the SLACK_WEBHOOK_URL to secrets, but this should definitely not be your task given privs.\r\n", "To automate the finding of the specific user we will need to `git bisect` for the range of commits since the last check with a rerun of just the failing tests , which will give us the first hash, and thus the PR that caused it. \r\n\r\nThis efficient approach may not work 100% of the time, if the failing test happens due to some other test (test dependency) affecting it, so if this speedy reduction doesn't work, then the full test suite will need to be re-run in conjunction with `git bisect`.", "WIP https://github.com/huggingface/transformers/pull/7995" ]

[ { "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", "This was patched on `master`, cannot reproduce anymore. Can you reproduce on `master`?", "> This was patched on `master`, cannot reproduce anymore. Can you reproduce on `master`?\r\n\r\nI installed the newest version (3.4.0) by pip, but this problem still arises.", "Facing the same issue with 3.5.1", "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", "Fixed on version 4.0.1", "transformers 4.0.1??" ]

[]

[]

[ "Please note that this is runtime on my machine and not CIs - so make sure you're evaluating the report relative to itself and not CI. Once the PR is merged we will start getting CI reports.\r\n\r\nTotal run time 4248s.\r\n\r\nSo it looks like on the torch-side `test_tokenization_fast.py` accounts for the main culprit adding up to 1300 secs. ~1/3rd of all test run.\r\n\r\nAnd the bulk of slowdown is tf tests.\r\n\r\n|time| tests|\r\n|-------|---------------------------------|\r\n|1300 | test_tokenization_fast.py|\r\n|1089 | the rest of torch tests|\r\n|1859 | tf tests|\r\n|-------------|-----|\r\n|4248 | Total|\r\n\r\nAnother thing I noticed `tests/test_modeling_marian.py` spends almost 40 secs in setup (11 x 3.3secs) - that's very slow:\r\n\r\n```\r\n3.95s setup tests/test_modeling_marian.py::TestMarian_FR_EN::test_batch_generation_fr_en\r\n3.57s setup tests/test_modeling_marian.py::TestMarian_EN_DE_More::test_auto_config\r\n3.40s setup tests/test_modeling_marian.py::TestMarian_en_ROMANCE::test_batch_generation_en_ROMANCE_multi\r\n3.35s setup tests/test_modeling_marian.py::TestMarian_EN_DE_More::test_batch_generation_en_de\r\n3.31s setup tests/test_modeling_marian.py::TestMarian_en_ROMANCE::test_tokenizer_handles_empty\r\n3.23s setup tests/test_modeling_marian.py::TestMarian_en_zh::test_batch_generation_eng_zho\r\n3.23s setup tests/test_modeling_marian.py::TestMarian_EN_FR::test_batch_generation_en_fr\r\n3.20s setup tests/test_modeling_marian.py::TestMarian_RU_FR::test_batch_generation_ru_fr\r\n3.13s setup tests/test_modeling_marian.py::TestMarian_MT_EN::test_batch_generation_mt_en\r\n3.12s setup tests/test_modeling_marian.py::TestMarian_EN_DE_More::test_forward\r\n3.06s setup tests/test_modeling_marian.py::TestMarian_en_ROMANCE::test_pipeline\r\n```", "I fixed marian in https://github.com/huggingface/transformers/pull/7888. Do you want to try to fix `test_tokenization_fast` or let somebody else?", "> Do you want to try to fix test_tokenization_fast\r\n\r\nI will give it a go.\r\n\r\n**edit**: Except it was just removed by the merge that just happened. So I have to start from scratch.", "https://github.com/huggingface/transformers/pull/7659 may have fixed the slow tokenization tests. Checking the most recent run it's back to ~2min for the torch-only job.\r\nhttps://app.circleci.com/pipelines/github/huggingface/transformers/13951/workflows/244749ce-d1ee-488f-a59d-d891fbc38ed6/jobs/100800\r\nI will check a few more and close it if that was the culprit.\r\n\r\n", "this test for some reason has `@slow` commented out - it takes 20+ seconds - can we put it back on?\r\nThis is not a test that tests functionality that is going to change much, so should be safe to turn it off for normal CIs. \r\nhttps://github.com/huggingface/transformers/blob/master/tests/test_tokenization_auto.py#L42\r\n\r\n```\r\npytest --durations=0 tests/test_tokenization_auto.py &\r\n22.15s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_pretrained\r\n4.42s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_identifier_non_existent\r\n2.75s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_model_type\r\n2.57s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_tokenizer_class\r\n2.36s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_pretrained_identifier\r\n2.06s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_from_pretrained_use_fast_toggle\r\n2.05s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_identifier_with_correct_config\r\n```", "This one should probably also be `@slow` - all the other tests around it are `@slow`:\r\n```\r\n15.51s call tests/test_pipelines.py::MonoColumnInputTestCase::test_torch_translation\r\n```", "Wrote a one liner to calculate the sub-totals for whatever pattern in the output of `pytest --durations=0` stats, as in:\r\n```\r\n22.15s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_pretrained\r\n4.42s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_identifier_non_existent\r\n2.75s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_model_type\r\n2.57s call tests/test_tokenization_auto.py::AutoTokenizerTest::test_tokenizer_from_tokenizer_class\r\n\r\n```\r\nTotal runtime:\r\n```\r\n$ cat stats.txt | perl -ne 's|^(.*?)s.|$x+=$1|e; END {print int $x}'\r\n3308\r\n```\r\nTotal tf runtime:\r\n```\r\ngrep _tf_ stats.txt | perl -ne 's|^(.*?)s.|$x+=$1|e; END {print int $x}'\r\n1609\r\n```", "It this common test a good candidate for `@slow`?\r\n```\r\ngrep test_model_outputs_equivalence stats.txt | perl -ne 's|^(.*?)s.|$x+=$1|e; END {print int $x}'\r\n230\r\n```\r\nAt least a few of them are quite slow:\r\n```\r\n20.11s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_model_outputs_equivalence\r\n16.19s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_model_outputs_equivalence\r\n13.49s call tests/test_modeling_gpt2.py::GPT2ModelTest::test_model_outputs_equivalence\r\n9.94s call tests/test_modeling_bert.py::BertModelTest::test_model_outputs_equivalence\r\n9.56s call tests/test_modeling_albert.py::AlbertModelTest::test_model_outputs_equivalence\r\n8.81s call tests/test_modeling_flaubert.py::FlaubertModelTest::test_model_outputs_equivalence\r\n8.29s call tests/test_modeling_tf_xlm.py::TFXLMModelTest::test_model_outputs_equivalence\r\n7.98s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_model_outputs_equivalence\r\n7.87s call tests/test_modeling_xlnet.py::XLNetModelTest::test_model_outputs_equivalence\r\n6.85s call tests/test_modeling_tf_xlnet.py::TFXLNetModelTest::test_model_outputs_equivalence\r\n6.81s call tests/test_modeling_squeezebert.py::SqueezeBertModelTest::test_model_outputs_equivalence\r\n6.30s call tests/test_modeling_tf_roberta.py::TFRobertaModelTest::test_model_outputs_equivalence\r\n6.25s call tests/test_modeling_roberta.py::RobertaModelTest::test_model_outputs_equivalence\r\n5.90s call tests/test_modeling_reformer.py::ReformerLocalAttnModelTest::test_model_outputs_equivalence\r\n5.81s call tests/test_modeling_electra.py::ElectraModelTest::test_model_outputs_equivalence\r\n5.79s call tests/test_modeling_distilbert.py::DistilBertModelTest::test_model_outputs_equivalence\r\n5.69s call tests/test_modeling_xlm.py::XLMModelTest::test_model_outputs_equivalence\r\n5.35s call tests/test_modeling_tf_t5.py::TFT5ModelTest::test_model_outputs_equivalence\r\n4.64s call tests/test_modeling_tf_funnel.py::TFFunnelBaseModelTest::test_model_outputs_equivalence\r\n4.34s call tests/test_modeling_tf_bert.py::TFBertModelTest::test_model_outputs_equivalence\r\n3.79s call tests/test_modeling_dpr.py::DPRModelTest::test_model_outputs_equivalence\r\n3.71s call tests/test_modeling_tf_funnel.py::TFFunnelModelTest::test_model_outputs_equivalence\r\n3.61s call tests/test_modeling_bart.py::BARTModelTest::test_model_outputs_equivalence\r\n3.58s call tests/test_modeling_openai.py::OpenAIGPTModelTest::test_model_outputs_equivalence\r\n3.57s call tests/test_modeling_tf_transfo_xl.py::TFTransfoXLModelTest::test_model_outputs_equivalence\r\n3.53s call tests/test_modeling_ctrl.py::CTRLModelTest::test_model_outputs_equivalence\r\n3.40s call tests/test_modeling_tf_lxmert.py::TFLxmertModelTest::test_model_outputs_equivalence\r\n3.31s call tests/test_modeling_tf_gpt2.py::TFGPT2ModelTest::test_model_outputs_equivalence\r\n3.19s call tests/test_modeling_tf_electra.py::TFElectraModelTest::test_model_outputs_equivalence\r\n3.12s call tests/test_modeling_tf_openai.py::TFOpenAIGPTModelTest::test_model_outputs_equivalence\r\n2.98s call tests/test_modeling_tf_albert.py::TFAlbertModelTest::test_model_outputs_equivalence\r\n2.93s call tests/test_modeling_tf_flaubert.py::TFFlaubertModelTest::test_model_outputs_equivalence\r\n2.80s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_model_outputs_equivalence\r\n2.59s call tests/test_modeling_longformer.py::LongformerModelTest::test_model_outputs_equivalence\r\n2.37s call tests/test_modeling_transfo_xl.py::TransfoXLModelTest::test_model_outputs_equivalence\r\n2.17s call tests/test_modeling_funnel.py::FunnelModelTest::test_model_outputs_equivalence\r\n2.13s call tests/test_modeling_fsmt.py::FSMTModelTest::test_model_outputs_equivalence\r\n2.02s call tests/test_modeling_bert_generation.py::BertGenerationEncoderTest::test_model_outputs_equivalence\r\n1.94s call tests/test_modeling_funnel.py::FunnelBaseModelTest::test_model_outputs_equivalence\r\n1.70s call tests/test_modeling_t5.py::T5ModelTest::test_model_outputs_equivalence\r\n1.60s call tests/test_modeling_deberta.py::DebertaModelTest::test_model_outputs_equivalence\r\n1.44s call tests/test_modeling_lxmert.py::LxmertModelTest::test_model_outputs_equivalence\r\n1.22s call tests/test_modeling_layoutlm.py::LayoutLMModelTest::test_model_outputs_equivalence\r\n1.11s call tests/test_modeling_reformer.py::ReformerLSHAttnModelTest::test_model_outputs_equivalence\r\n0.86s call tests/test_modeling_tf_ctrl.py::TFCTRLModelTest::test_model_outputs_equivalence\r\n0.33s call tests/test_modeling_blenderbot.py::BlenderbotTesterMixin::test_model_outputs_equivalence\r\n```", "Here is another possible candidate for `@slow`:\r\n\r\n```\r\ngrep test_torchscript stats.txt | perl -ne 's|^(.*?)s.|$x+=$1|e; END {print int $x}'\r\n289\r\n```\r\n\r\n```\r\n18.89s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript_output_attentions\r\n18.65s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript\r\n11.37s call tests/test_modeling_bert.py::BertModelTest::test_torchscript_output_hidden_state\r\n11.02s call tests/test_modeling_bert.py::BertModelTest::test_torchscript_output_attentions\r\n9.69s call tests/test_modeling_electra.py::ElectraModelTest::test_torchscript\r\n8.90s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript_output_hidden_state\r\n8.35s call tests/test_modeling_albert.py::AlbertModelTest::test_torchscript\r\n7.82s call tests/test_modeling_roberta.py::RobertaModelTest::test_torchscript\r\n7.78s call tests/test_modeling_flaubert.py::FlaubertModelTest::test_torchscript_output_attentions\r\n7.71s call tests/test_modeling_funnel.py::FunnelModelTest::test_torchscript_output_hidden_state\r\n7.71s call tests/test_modeling_electra.py::ElectraModelTest::test_torchscript_output_attentions\r\n7.68s call tests/test_modeling_albert.py::AlbertModelTest::test_torchscript_output_attentions\r\n7.65s call tests/test_modeling_funnel.py::FunnelModelTest::test_torchscript\r\n7.37s call tests/test_modeling_flaubert.py::FlaubertModelTest::test_torchscript\r\n7.12s call tests/test_modeling_roberta.py::RobertaModelTest::test_torchscript_output_attentions\r\n7.01s call tests/test_modeling_squeezebert.py::SqueezeBertModelTest::test_torchscript_output_hidden_state\r\n6.61s call tests/test_modeling_roberta.py::RobertaModelTest::test_torchscript_output_hidden_state\r\n6.51s call tests/test_modeling_squeezebert.py::SqueezeBertModelTest::test_torchscript_output_attentions\r\n5.48s call tests/test_modeling_distilbert.py::DistilBertModelTest::test_torchscript_output_hidden_state\r\n5.06s call tests/test_modeling_xlm.py::XLMModelTest::test_torchscript\r\n4.78s call tests/test_modeling_xlnet.py::XLNetModelTest::test_torchscript_output_attentions\r\n4.71s call tests/test_modeling_xlm.py::XLMModelTest::test_torchscript_output_attentions\r\n4.71s call tests/test_modeling_xlnet.py::XLNetModelTest::test_torchscript_output_hidden_state\r\n4.66s call tests/test_modeling_xlm.py::XLMModelTest::test_torchscript_output_hidden_state\r\n4.59s call tests/test_modeling_xlnet.py::XLNetModelTest::test_torchscript\r\n4.44s call tests/test_modeling_funnel.py::FunnelModelTest::test_torchscript_output_attentions\r\n4.32s call tests/test_modeling_bart.py::BARTModelTest::test_torchscript_output_attentions\r\n3.98s call tests/test_modeling_electra.py::ElectraModelTest::test_torchscript_output_hidden_state\r\n3.73s call tests/test_modeling_funnel.py::FunnelBaseModelTest::test_torchscript_output_hidden_state\r\n3.61s call tests/test_modeling_gpt2.py::GPT2ModelTest::test_torchscript_output_attentions\r\n3.55s call tests/test_modeling_bart.py::BARTModelTest::test_torchscript_output_hidden_state\r\n3.55s call tests/test_modeling_dpr.py::DPRModelTest::test_torchscript\r\n3.54s call tests/test_modeling_bert.py::BertModelTest::test_torchscript\r\n3.50s call tests/test_modeling_albert.py::AlbertModelTest::test_torchscript_output_hidden_state\r\n3.46s call tests/test_modeling_openai.py::OpenAIGPTModelTest::test_torchscript\r\n3.36s call tests/test_modeling_openai.py::OpenAIGPTModelTest::test_torchscript_output_hidden_state\r\n3.33s call tests/test_modeling_gpt2.py::GPT2ModelTest::test_torchscript\r\n3.31s call tests/test_modeling_dpr.py::DPRModelTest::test_torchscript_output_attentions\r\n3.27s call tests/test_modeling_layoutlm.py::LayoutLMModelTest::test_torchscript_output_attentions\r\n3.25s call tests/test_modeling_layoutlm.py::LayoutLMModelTest::test_torchscript_output_hidden_state\r\n3.07s call tests/test_modeling_funnel.py::FunnelBaseModelTest::test_torchscript_output_attentions\r\n2.58s call tests/test_modeling_flaubert.py::FlaubertModelTest::test_torchscript_output_hidden_state\r\n2.56s call tests/test_modeling_t5.py::T5ModelTest::test_torchscript_output_hidden_state\r\n2.50s call tests/test_modeling_fsmt.py::FSMTModelTest::test_torchscript_output_attentions\r\n2.30s call tests/test_modeling_t5.py::T5ModelTest::test_torchscript_output_attentions\r\n2.28s call tests/test_modeling_openai.py::OpenAIGPTModelTest::test_torchscript_output_attentions\r\n2.19s call tests/test_modeling_bert_generation.py::BertGenerationEncoderTest::test_torchscript_output_hidden_state\r\n2.13s call tests/test_modeling_distilbert.py::DistilBertModelTest::test_torchscript_output_attentions\r\n2.02s call tests/test_modeling_gpt2.py::GPT2ModelTest::test_torchscript_output_hidden_state\r\n1.87s call tests/test_modeling_fsmt.py::FSMTModelTest::test_torchscript_output_hidden_state\r\n1.82s call tests/test_modeling_bert_generation.py::BertGenerationEncoderTest::test_torchscript_output_attentions\r\n1.78s call tests/test_modeling_funnel.py::FunnelBaseModelTest::test_torchscript\r\n1.69s call tests/test_modeling_bart.py::BARTModelTest::test_torchscript\r\n1.51s call tests/test_modeling_layoutlm.py::LayoutLMModelTest::test_torchscript\r\n1.34s call tests/test_modeling_dpr.py::DPRModelTest::test_torchscript_output_hidden_state\r\n0.89s call tests/test_modeling_fsmt.py::FSMTModelTest::test_torchscript\r\n0.79s call tests/test_modeling_bert_generation.py::BertGenerationEncoderTest::test_torchscript\r\n0.05s call tests/test_modeling_lxmert.py::LxmertModelTest::test_torchscript_output_attentions\r\n0.05s call tests/test_modeling_lxmert.py::LxmertModelTest::test_torchscript_output_hidden_state\r\n0.01s call tests/test_modeling_reformer.py::ReformerLocalAttnModelTest::test_torchscript_output_attentions\r\n0.01s call tests/test_modeling_ctrl.py::CTRLModelTest::test_torchscript_output_attentions\r\n0.01s call tests/test_modeling_reformer.py::ReformerLocalAttnModelTest::test_torchscript_output_hidden_state\r\n```", "I am fine with marking any of these @slow , but don't care as much now that the issue is resolved.\r\n\r\nI do think we should have a repo-wide rule about what should be slow, and you have to write a comment if you want to override it.\r\n\r\n#### Proposed Rule for testing.rst\r\n\r\n+ All tests that take longer than 5s should be marked slow, that way we can save a lot of back and forth.\r\n+ For common tests to be marked slow, the slowest iteration of that common test must be > 15s.\r\n\r\n5s/15s was arbitrary, don't care much what the value is. WDYT @LysandreJik @sgugger ?\r\n\r\n", "That's a fabulous suggestion!\r\n\r\nA few caveats for `testing.rst`:\r\n* < 5s should include model download overhead - my numbers above exclude this, so once we get the data from CI it'll be the true measurement.\r\n* 5s as measured on CI, since otherwise each hardware is different\r\n", "While we are at it - there is a ton of very slow tf tests - I suppose the same rule applies there, right?", "As much as I love moving quickly, we need to wait for others to agree to the rule before we apply it.\r\nMy proposed rule does not differentiate between tf and torch.", "My communication wasn't clear - I meant to do that after we agreed on the threshold and regardless this PR https://github.com/huggingface/transformers/pull/7884 needs to be merged first to perform the correct measurements. \r\n\r\nI just saw that there was a **lot** of tf tests that were very slow and which were not marked as such, so I thought perhaps there was a special reason for them not to be `@slow`.", "I'm not sure setting up a 5/15 or any specific time requirement on tests to classify them as slow would be best. Some tests, like the `test_model_outputs_equivalence` are important, and running them on contributors' PR when their changes affect the modeling internals is too.\r\n\r\nI think the following proposition would be more suited: \r\n\r\nif the test is focused on one of the library's internal components (e.g., modeling files, tokenization files, pipelines), then we should run that test in the non-slow test suite. If it's focused on an other aspect of the library, such as the documentation, the examples, then we should run these tests in the slow test suite. And then, to refine this approach we should have exceptions:\r\n\r\n- All tests that need a specific set of weights (e.g., model or tokenizer integration tests, pipeline integration tests) should be set to slow.\r\n- All tests that need to do a training (e.g, trainer integration tests) should be set to slow.\r\n- We can introduce exceptions if some of these should-be-non-slow tests are excruciatingly long, and set them to slow. Some examples are some auto modeling tests, which save and load large files to disk, which are set to slow.\r\n- Others?\r\n\r\nTo that end, we should aim for all the non-slow tests to cover entirely the different internals, while making sure that the tests keep a fast execution time. Having some very small models in the tests (e.g, 2 layers, 10 vocab size, etc.) helps in that regard, as does having dummy sets of weights like the `sshleifer/tiny-xxx-random` weights. On that front, there seems to be something fishy going on with the MobileBERT model, as it's supposed to be an efficient model but takes a while to be tested. There's probably something to do for this model.\r\n\r\nWilling to iterate on this wording, or specify/change some aspects if you think of something better.\r\n\r\nFollowing this approach:\r\n\r\nFor the [tokenization_auto tests](https://github.com/huggingface/transformers/issues/7885#issuecomment-711443513), we can definitely uncomment the `@slow`.\r\n\r\nFor the [MonoColumnInputTestCase](https://github.com/huggingface/transformers/issues/7885#issuecomment-711444043), we can also set it as a slow test.\r\n\r\n", "I would change all things that need to do a training to all thing that need to do a real training.\r\nI spent a lot of time making a mock training fast for the tests of the Trainer and I don't want that marked as slow ;-)", "OK, sounds like @stas00 can mark a few at slow.\r\nLongformer test also uses 5 layers for some reason, not sure if that matters.", "@LysandreJik, just a clarification - so you propose not to have a fixed speed threshold in any of the \"clauses\". i.e. for non-essential tests as defined by you they should be marked as slow regardless of their speed, correct? I suppose this is smart since even very fast tests still add up to a lot since there could be many of them.", "OK, so here is the full non-slow run's report on CI:\r\nhttps://pastebin.com/8pkaZKjH (quoted from [this report](https://circleci.com/api/v1.1/project/github/huggingface/transformers/101622/output/108/0?file=true&allocation-id=5f8db77a5d41f9372b3f92cf-0-build%2F6825DA7E))\r\n\r\nThe top slow ones are (cut off at 10sec):\r\n```\r\n131.69s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_train_pipeline_custom_model\r\n101.16s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_graph_mode\r\n79.24s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_compile_tf_model\r\n40.68s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_keras_save_load\r\n38.58s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_compile_tf_model\r\n35.05s call tests/test_modeling_marian.py::TestMarian_en_ROMANCE::test_pipeline\r\n32.99s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_train_pipeline_custom_model\r\n27.40s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_graph_mode\r\n26.53s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_compile_tf_model\r\n26.17s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_graph_mode\r\n25.97s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_model_outputs_equivalence\r\n20.57s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_train_pipeline_custom_model\r\n18.71s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_pt_tf_model_equivalence\r\n18.59s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript_output_attentions\r\n17.73s call tests/test_pipelines.py::MonoColumnInputTestCase::test_torch_translation\r\n17.72s call tests/test_modeling_tf_flaubert.py::TFFlaubertModelTest::test_train_pipeline_custom_model\r\n17.27s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript\r\n17.15s call tests/test_modeling_tf_bert.py::TFBertModelTest::test_compile_tf_model\r\n16.72s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript_output_hidden_state\r\n16.49s call tests/test_modeling_tf_funnel.py::TFFunnelBaseModelTest::test_train_pipeline_custom_model\r\n16.20s call tests/test_benchmark_tf.py::TFBenchmarkTest::test_train_no_configs\r\n15.91s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_pt_tf_model_equivalence\r\n15.64s call tests/test_modeling_tf_xlnet.py::TFXLNetModelTest::test_compile_tf_model\r\n15.52s call tests/test_modeling_tf_flaubert.py::TFFlaubertModelTest::test_compile_tf_model\r\n15.36s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_pretokenized_inputs\r\n15.32s call tests/test_benchmark_tf.py::TFBenchmarkTest::test_train_with_configs\r\n15.28s call tests/test_modeling_tf_funnel.py::TFFunnelModelTest::test_compile_tf_model\r\n15.28s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_model_outputs_equivalence\r\n15.24s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_pt_tf_model_equivalence\r\n15.14s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_maximum_encoding_length_pair_input\r\n14.94s call tests/test_modeling_tf_funnel.py::TFFunnelModelTest::test_train_pipeline_custom_model\r\n14.62s call tests/test_modeling_tf_electra.py::TFElectraModelTest::test_compile_tf_model\r\n14.34s call tests/test_modeling_tf_funnel.py::TFFunnelModelTest::test_graph_mode\r\n14.32s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_model_outputs_equivalence\r\n14.12s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_add_special_tokens\r\n13.75s call tests/test_modeling_tf_xlm.py::TFXLMModelTest::test_compile_tf_model\r\n13.72s call tests/test_modeling_tf_t5.py::TFT5ModelTest::test_compile_tf_model\r\n13.67s call tests/test_modeling_tf_bert.py::TFBertModelTest::test_graph_mode\r\n13.33s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_maximum_encoding_length_single_input\r\n13.12s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_attention_outputs\r\n11.84s call tests/test_modeling_tf_mobilebert.py::TFMobileBertModelTest::test_save_load\r\n11.78s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_attention_outputs\r\n11.69s call tests/test_modeling_tf_roberta.py::TFRobertaModelTest::test_compile_tf_model\r\n11.56s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_attention_outputs\r\n11.54s call tests/test_modeling_tf_transfo_xl.py::TFTransfoXLModelTest::test_train_pipeline_custom_model\r\n11.41s call tests/test_modeling_tf_lxmert.py::TFLxmertModelTest::test_train_pipeline_custom_model\r\n11.40s call tests/test_modeling_tf_electra.py::TFElectraModelTest::test_train_pipeline_custom_model\r\n11.35s call tests/test_modeling_tf_roberta.py::TFRobertaModelTest::test_train_pipeline_custom_model\r\n11.30s call tests/test_modeling_tf_xlm.py::TFXLMModelTest::test_train_pipeline_custom_model\r\n10.82s call tests/test_modeling_tf_electra.py::TFElectraModelTest::test_graph_mode\r\n10.77s call tests/test_modeling_tf_distilbert.py::TFDistilBertModelTest::test_save_load\r\n10.74s call tests/test_modeling_tf_bert.py::TFBertModelTest::test_train_pipeline_custom_model\r\n10.71s call tests/test_modeling_tf_xlm.py::TFXLMModelTest::test_graph_mode\r\n10.60s call tests/test_modeling_tf_albert.py::TFAlbertModelTest::test_compile_tf_model\r\n10.57s call tests/test_modeling_tf_longformer.py::TFLongformerModelTest::test_save_load\r\n10.55s call tests/test_modeling_blenderbot.py::Blenderbot90MIntegrationTests::test_90_generation_from_short_input\r\n10.39s call tests/test_modeling_tf_flaubert.py::TFFlaubertModelTest::test_graph_mode\r\n10.24s call tests/test_modeling_tf_xlnet.py::TFXLNetModelTest::test_graph_mode\r\n10.08s call tests/test_benchmark_tf.py::TFBenchmarkTest::test_inference_encoder_decoder_with_configs\r\n```\r\n\r\n@sshleifer, here is a highlight for you:\r\n```\r\n35.05s call tests/test_modeling_marian.py::TestMarian_en_ROMANCE::test_pipeline\r\n```\r\n\r\nOther slow torch tests by group:\r\n```\r\n18.59s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript_output_attentions\r\n17.27s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript\r\n16.72s call tests/test_modeling_mobilebert.py::MobileBertModelTest::test_torchscript_output_hidden_state\r\n\r\n17.73s call tests/test_pipelines.py::MonoColumnInputTestCase::test_torch_translation\r\n\r\n15.36s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_pretokenized_inputs\r\n15.14s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_maximum_encoding_length_pair_input\r\n14.12s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_add_special_tokens\r\n13.33s call tests/test_tokenization_deberta.py::DebertaTokenizationTest::test_maximum_encoding_length_single_input\r\n\r\n10.55s call tests/test_modeling_blenderbot.py::Blenderbot90MIntegrationTests::test_90_generation_from_short_input\r\n```", "> @LysandreJik, just a clarification - so you propose not to have a fixed speed threshold in any of the \"clauses\". i.e. for non-essential tests as defined by you they should be marked as slow regardless of their speed, correct? I suppose this is smart since even very fast tests still add up to a lot since there could be many of them.\r\n\r\nYes, I think that would be best! I don't think there are many non-essential tests that are not slow though. We'd still like to get full coverage of the library's internals using only non-@slow tests, so getting these tests below a certain time threshold would still be important so that every PR could get quick feedback on the CI's status.", "Thank you for that clarification, @LysandreJik\r\n\r\nPlease have a look at how your suggestions have been integrated into the testing doc:\r\nhttps://github.com/huggingface/transformers/pull/7895/files" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

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

[]

[ "Hello, I would say that there's a mismatch between the number of labels you've provided and the number of labels supported by your `model.pt`.\r\n\r\nHow did you obtain your `model.pt`?", "I downloaded it directly from the links provided for each bert models with `!wget`", "Which links? Why did you switch from using the `from_pretrained` utility to the `load_state_dict` torch util?", "I had to use the same model for two tasks. Since the downloaded model using `.from_pretrained` method go to root/ directory in colab, they aren't directly accessible (I actually opened an issue about that as well [here](https://github.com/huggingface/transformers/issues/7847)). Thus, to use the model for both tasks, I downloaded it manually in a model/ directory using `!wget https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-config.json`", "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 had the case when we tried to provide our own tokenizer with vocab.txt file.\r\nThe mistake we did was that we update the config.json file and set vocab_size as the total line number in the new vocab.txt file.\r\n\r\nThis was the wrong step we took. config.json has to be the one exactly for model.pt we want to use. If the model.pt is from the hub, then no need to change the config.json.\r\n", "This happened when i evaluating model on CPU while fine-tune model on GPU. It disappeared when i evaluate it on GPU.", "> This happened when i evaluating model on CPU while fine-tune model on GPU. It disappeared when i evaluate it on GPU.\r\n\r\nHey @aakejiang, I am encountering the same issue, did you manage to fix it ? Thanks. " ]

[]

[]

[ "Ooh, that's an ugly typo. Thanks for fixing! Will make sure to add a test of this soon so it doesn't break with me being stupid like that." ]

[]

[]

[ "When you add special tokens to the tokenizer, you should resize the model embedding matrix so that it contains the extra embeddings.\r\n\r\nYou should then do an additional training/fine-tuning on data containing those special tokens so that the model may tune this new column in the embedding matrix.", "thanks !" ]

[]

[]

[ "OK, so how would you recommend we run a specific common test for a model whose test module isn't running common tests at all? \r\n\r\nPlease have a look at the approach I used for `mbart` https://github.com/huggingface/transformers/pull/7878/files#diff-f3b9af7b4721bdf972a72e35a35e8742ddc986f12218e728cd2b6ad30a156abb\r\n\r\nI basically made a very stripped down `ModelTester` and then used an alias hack to run just a specific common test via:\r\n```\r\nclass SelectiveCommonTest(unittest.TestCase):\r\n test_save_load_keys_to_never_save = ModelTesterMixin.test_save_load_keys_to_never_save\r\n```\r\nIs this any good, or would you recommend a different approach - the idea is not to copy the same test to each test module.\r\n", "I like how you did it.\r\nAdding @LysandreJik and @sgugger to see if they agree before you do it 5 times.\r\n", "> I imagine you double-checked the test was indeed executed.\r\n\r\nYes. \r\n\r\nIndeed, my initial test that I wrote for fsmt wasn't executing, since I renamed the keys in the code and forgot the test and the test was no-op if it didn't find the right keys. So some print()s helped to verify it indeed worked in this new incarnation.", "I think `test_save_load_missing_keys` should be handled separately since it can add a significant overhead if run by all models, since I noticed tf models are **very** slow at save/load tests.. I reverted moving it from fsmt to common tests for now - back to just fsmt - will do a separate PR about it.\r\n", "Anybody with a good knowledge of XLM? @sshleifer suggested to add xlm to this PR, but I'm not sure whether it's safe to not save `self.position_embeddings`, since as you can see here:\r\nhttps://github.com/huggingface/transformers/blob/master/src/transformers/modeling_xlm.py#L439\r\nthe deterministic non-trained weights happen only if `config.sinusoidal_embeddings` is True. Otherwise it gets trained like any other embed layer.\r\n\r\nWhich leaves just: `position_ids` which is tiny a set of 512-2048 floats, which is not worth bothering for IMHO.\r\n\r\n", "wrt t5 - I don't see anything positional or deterministic for that matter, but this is quite a different model - are there any state_dict keys here that would be a good candidate for not saving?", "I implemented this for: mbart, pegasus, marian - all subclasses of `BartForConditionalGeneration` w/ `static_position_embeddings=True`. `fsmt` already had it.\r\n\r\nAnything else that I may have missed?\r\n\r\nIf not, this PR is good to go." ]

[]

[]

[ "I have some concerns defaulting this logic to `True`, given that some `sshleifer/` models like `sshleifer/pegasus-cnn-ft-v2/`, `distill-pegasus-cnn-16-4/` were trained without newlines in the dataset. I'm sorry I forgot to mention these until now.\r\n\r\nI would merge this if the logic to do the regex were defaulted to False and we checked that `PegasusFastTokenizer` did the same thing.\r\nI would also be fine abandoning, since you have already solved the Pegasus replication issue with your `build/` scripts, and we haven't had anybody external ask for this yet.", "Sure, we can just leave it unmerged and if someone reports an issue get back to sorting this out. It was a trivial change, so I have no problem with closing this. It's your call, @sshleifer.", "will reopen if needed, thx for being flexible @stas00 " ]

[]

[]

[ "Hi, you can find this data [here](https://github.com/facebookresearch/XLM#the-17-and-100-languages)." ]

[]

[]

[ "Solved by adding `num_labels` to the constructor: `BertForSequenceClassification.from_pretrained(\"...\", num_labels=z)`. Then the onehot-encoding wasn't necessary either. Another error would probably make it easier to spot this though." ]

[]

[]

[ "@lalitpagaria - good point! I think the cleanest solution here would actually be to add `n_docs` as a argument to all RagModel forward's function that defaults to `self.config.n_docs` (exactly as it is the case for `use_cache` or `output_attentions`) => it would be great if you could open a PR for this :-) ", "@patrickvonplaten I think this approach also will not solve a situation when retriever return less docs than `n_docs`. For example let's say a dataset only have `4` documents but `n_docs` value is `5`. \r\n\r\nHow about solving this as follows -\r\n1) When fetching docs from retriever use `n_docs` value (either as a argument or default `self.config.n_docs`)\r\n2) All other places use value from number of documents returned by retriever ie dimension `0` of `context_input_ids` can be used\r\n\r\nLet me know what do you think.", "I think this is an edge case - if a dataset only has 4 documents, then the user should not define `n_docs` as 5. I agree that it would be easier to just use the incoming tensor's dimension as `n_docs`, but it would be less readable and can also lead to confusion. `n_docs` should always be the number of docs actually given to Rag. So I'd prefer to not use the retriever's dimension here. We could add an assert that the dimension 0 is always equal `n_docs` which would be helpful for the user, but I think we should still rely on the `n_docs` parameter.", "Sure I will create PR (using n_docs with assert) with related tests." ]

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

[]

[ "This has been done already, see [here](https://github.com/huggingface/transformers/blob/master/src/transformers/training_args.py#L335). But you need to install from source for that, since the change is recent,", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[ "Hey @LittlePea13 - yes you are very much correct here. This PR linked to the issue should fix it. Thanks a lot!" ]

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

[]

[]

[]

[]

[ "> I think the postinit is a great place for this sanity check, so the error comes fast to the user, and I don't think our current FP16 implementations (Apex and Amp) work with TPUs. I can check the latter on Monday and report back here.\r\n\r\nIf you'd rather be more forgiving, let me know. In that case we can do a logging.warning(f\"AMP only supported when using CUDA. Turning off AMP because not using a CUDA device ({self.device.type})\") and subsequently automatically turning off --fp16. ", "Ok, just tested on TPUs and FP16 does not work on them indeed, though native amp fails graciously (testing on TPUs with PyTorch < 1.6 is suicide so I did not try APEX). So if you fix the styling issue, I think this is good to be merged!\r\n", "It has to be run by the user, I can force-push the styling on your branch if you don't have the setup to run it easily.", "> It has to be run by the user, I can force-push the styling on your branch if you don't have the setup to run it easily.\r\n\r\nThanks but it should work fine now. Updated local dependencies and all that and reformatted.", "Thanks for your contribution!" ]

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

[]

[]

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

[]

[]

[]

[]

[ "> # ❓ Questions & Help\r\n> ## Details\r\n> tensorflow 2.0.0\r\n> after \"pip install transformers\", i import transformers, meet this error.\r\n> \r\n> **A link to original question on the forum/Stack Overflow**:\r\n\r\n兄弟，升级tf到2.3", "Hi, you should install TensorFlow 2.3 as mentioned by @wmathor " ]

[]

[]

[ "I also get confused by the notation and would appreciate some clarifications.\r\nThank you.", "@AI678 I think I got it. Do not take my words as absolute truth since I am new to the subject.\r\nIf you are talking about a full Transformer architecture (e.g. BART, T5, PEGASUS), the labels are the token ids to which you compare the logits generated by the Decoder in order to compute the cross-entropy loss. This should be the only input necessary during training or fine tuning.\r\nOn the other hand, the decoder_input_ids have the exact same structure of the labels, but are shifted one position to the right in order to add the \\<start of sentence\\> token. The decoder_input_ids are then passed to the decoder (along with the mask, which masks all unseen tokens!!). The first input of the decoder will be the \\<start of sentence\\> token so that it starts generating the sentence.\r\n\r\nHope this help, waiting for somebody to acknowledge my answer.", "Hi, the glossary has been completed with these terms, you can check it out [here](https://huggingface.co/transformers/glossary.html), let me know if it's still not clear enough." ]

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

[]

[ "Hi @julien-c, thank you so much for the pointer. I have added the dataset identifier and license.", "Thanks! Your model is now linked from https://huggingface.co/datasets/anli (and the other datasets)" ]

[]

[]

[ "I resolved conflicts - it's good to go now." ]

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

[ { "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 } ]

[ "Hello! The layers are shared in ALBERT because even though there are `config.num_hidden_layers=n`, there's usually `config.num_hidden_groups = 1`. Change that value to `n` to have `n` different layers.", "Thanks for your reply, I have another question about that\r\nHow can I load the AlbertModel parameter(in transformer block) to unshared Albert, When I load just \r\n\r\n###_IncompatibleKeys(missing_keys=['embeddings.position_ids'], unexpected_keys=[])###\r\nwhether is it wrong? can you tell me how I can load parameters correctly?\r\nlooking forward to your reply", "Could you show me the full code used that generated this error? Or can you give an example so I can reproduce it?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[ "Hi Stas, thanks, #7659 will fix this (we will now require at least one example checkpoint for each tokenizer and we test it automatically).", "@thomwolf, I'm not certain why you closed this. This is a tokenizer test that is needed - the issue caught in examples was just a flag that there was a missing test in the normal tests.\r\n\r\nActually, not only it's needed, I will have to expand this test to verify that it **doesn't** get the hardcoded default values, but fetches the correct values from `tokenizer_config.json`. And to change the default values to be different from `TINY_FSMT`, since otherwise it won't be testing the right thing. \r\n\r\nFeel free to add it as part of https://github.com/huggingface/transformers/pull/7659 but please make sure you used different from `TINY_FSMT` hardcoded defaults.\r\n\r\nI hope this makes sense. ", "Hmm, but you copied `TINY_FSMT` in https://github.com/huggingface/transformers/commit/1885ca7f29ffa5373ea848edad8efab02809c268, how will then the tests check it can fetch that data if this is now hardcoded? I am not following this. Do I need to create `TINY_FSMT2` with different values?\r\n\r\nI haven't read the new code in depth, but my gut feeling is that the defaults may mask a problem.", "Hi stas, I'll let you read the new code and then we can have a look together.\r\n\r\nThe basic idea is that we now require a full and working checkpoint for the tokenizers to be fully tested in various conditions and the slow vs. fast compared.\r\n\r\nThe question of testing that tokenizers load and use `tokenizer_config.json` is another question and we should indeed address it in a subsequent PR if it's not addressed already indeed.", "That works. \r\n\r\nBut please re-open this PR, since we need it anyway. I will add more changes to it after your big PR merge to ensure that the loading of the tokenizer is properly tested.", "This PR is complete now." ]

[]

[]

[ "@sshleifer, one thing I wasn't sure about - since now the temp dir creation/destruction is automated and it's easy to hardwire those by just passing the explicit path `self.get_auto_remove_tmp_dir(\"/tmp/favoritedir\")` when debugging - I removed the prefices that were used in `tempfile.mkdtemp`. Do you still want to have the option of having a specific prefix? With prefix you still have to hunt for the unique ID which changes on every test rerun. This is a much simpler option, IMHO.\r\n\r\nTo me it sounds it'd be useful during a debug to have one common path prefix for all tmp dirs of the same test and then add suffices instead for the different specific tmp dir of the test. ", "Whatever you think is best! ", "I think the prefixes are no longer needed. So it's good to go. " ]

[]

[]

[ "Now investigating this FAILED test.\r\nSo far, it is PASSED on 3 different environments I have tried.\r\n`FAILED tests/test_tokenization_fast.py::SentencePieceFastTokenizerTest::test_add_special_tokens` from `run_tests_torch_and_tf`", "@sgugger \r\n\r\nCircleCI `run_tests_torch_and_tf` is FAILED, because \r\n```\r\nFAILED tests/test_tokenization_fast.py::SentencePieceFastTokenizerTest::test_add_special_tokens\r\n==== 1 failed, 2964 passed, 493 skipped, 386 warnings in 925.76s (0:15:25) =====\r\n```\r\nOn 2 different environments, I have a PASSED.\r\nI reproduced the CI run step by step, and it is still PASSED:\r\n```\r\nconda create -n trans_test_tf_torch python=3.6 \r\nconda activate trans_test_tf_torch\r\ngit clone [email protected]:huggingface/transformers.git\r\ncd transformers\r\ngit fetch origin pull/7858/head:pull/7858\r\ngit pull origin pull/7858 \r\ngit checkout pull/7858\r\npip install --upgrade pip \r\npip install git+https://github.com/huggingface/datasets \r\npip install .[sklearn,tf-cpu,torch,testing] \r\npip install codecov pytest-cov \r\n```\r\n\r\nResult:\r\n`python -m pytest -n auto --dist=loadfile -s -v tests/test_tokenization_fast.py --cov`\r\n`91 passed, 11 warnings in 1851.05s (0:30:51)`\r\n\r\nFor all the tests:\r\n`python -m pytest -n auto --dist=loadfile -s -v ./tests/ --cov `\r\n`2965 passed, 493 skipped, 385 warnings in 3124.66s (0:52:04)`", "Thanks for your contribution!" ]

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

[]

[]

[]

[]

[]