Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

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

[]

[ "Well it seems to tell you that your tensorboard directory already exists? Try with a different directory?", "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" ]

[]

[]

[ "Confirming the failure is spurious. Thanks a lot for your fix!" ]

[ { "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" ]

[]

[]

[ "It's ready to merge" ]

[]

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

[ "Any solutions to this?\r\n\r\nThanks!", "Hey, how are you setting `mem_len=384`? I am under the impression that the default `run_generation` script doesn't allow you to change the model from the default configuration, which would explain the warning.", "Here is how I am setting it\r\n`model = model_class.from_pretrained(args.model_name_or_path,cache_dir=args.cache_dir,mem_len=args.mem_len)`\r\n\r\n", "Ok, I've understood the issue - when using this call, the config first gets created with `mem_len=0` (hence the warning) then `mem_len` gets changed. I'll open a PR to move the warning to the `forward` call. In the meantime, you can suppress the warnings; memories are actually getting used.", "Okay great! \r\nThanks a lot. Closing the issue." ]

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

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

[ "This 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" ]

[]

[]

[ "Fyi the example tests don't currently run on the multi-gpu setup, as pretty much none of the examples tests were made to run on a multi-gpu setup.\r\n\r\nWe can look into adding them to the CI.", "> Fyi the example tests don't currently run on the multi-gpu setup, as pretty much none of the examples tests were made to run on a multi-gpu setup.\r\n> \r\n> We can look into adding them to the CI.\r\n\r\nAbsolutely!\r\n\r\nDo you prefer to start with a CI job that explicitly lists tests that were converted to run on multi-gpu (3 at the moment) or just run them all? It's all that `require_torch_mutigpu` plus ` that `require_torch_gpu` as it's flexible.", "> Do you prefer to start with a CI job that explicitly lists tests that were converted to run on multi-gpu (3 at the moment) or just run them all? It's all that require_torch_mutigpu plus that `require_torch_gpu` as it's flexible.\r\n\r\nI would send a PR with whatever you think makes most sense. Or wait until the correct answer is clearer. At this point you know pretty much as much about CI tradeoffs as I do :)", "I will prepare a new PR for github CIs.\r\n\r\notherwise I think this PR is complete." ]

[]

[]

[ "@LysandreJik Not sure who I should ping to review this :) btw I am not sure what's the issue with the code quality here. ", "Hello, we're in the midst of deprecating that method and shifting the methods to `datasets` instead. Could you hold on until we have the updated approach, which should land sometimes within the next ~2 weeks?\r\n\r\nThis current method is not tested, so ensuring that a bug fix isn't creating all sorts of other bugs takes us a lot of time, so we're not going to do that here as this method will be deprecated very soon.", "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." ]

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

[]

[ "I even set batch_size to 1 and still the error is there. any suggestion? ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[]

[ { "id": 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 } ]

[]

[ "Maybe @patil-suraj has an idea!", "Hi\r\nI really appreciate assisting me with this question, I am still struggling with it, could you tell me some pointers, how to add the handling capability of iterative datasets to finetune_trainer.py codes. thanks @patil-suraj ", "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" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Note that the old script is still available [here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/run_language_modeling.py).\r\n\r\nThe new one expects slightly different arguments:\r\n- you should remove the `mlm` flag now, since the script has been split in several others\r\n- `train_data_file` is now `train_file`\r\n- there is no `model_type` argument anymore as this is not necessary when we have the model checkpoint\r\n- `block_size` has been renamed `max_seq_length`", "Thank you so much!!!\r\n\r\nYou saved my day ", "I am getting the following error while running the below code. Could anyone please help me with this?\r\n\r\n!python \"/content/transformers/examples/language-modeling/run_mlm.py\" \\\r\n--output_dir \"/content/drive/MyDrive/Bert_models/test-mlm\" \\\r\n--model_name_or_path \"/content/bert-base-uncased\" \\\r\n--do_train \\\r\n--do_eval \\\r\n--train_file \"/content/train.txt\" \\\r\n--validation_file \"/content/test.txt\" \\\r\n\r\n\r\nError:\r\n[INFO|modeling_utils.py:1022] 2020-12-26 10:18:28,002 >> loading weights file /content/bert-base-uncased/pytorch_model.bin\r\nTraceback (most recent call last):\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_utils.py\", line 1035, in from_pretrained\r\n state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/serialization.py\", line 595, in load\r\n return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/serialization.py\", line 764, in _legacy_load\r\n magic_number = pickle_module.load(f, **pickle_load_args)\r\n_pickle.UnpicklingError: invalid load key, 'v'.\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"/content/transformers/examples/language-modeling/run_mlm.py\", line 420, in <module>\r\n main()\r\n File \"/content/transformers/examples/language-modeling/run_mlm.py\", line 264, in main\r\n cache_dir=model_args.cache_dir,\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/models/auto/modeling_auto.py\", line 1092, in from_pretrained\r\n pretrained_model_name_or_path, *model_args, config=config, **kwargs\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_utils.py\", line 1038, in from_pretrained\r\n f\"Unable to load weights from pytorch checkpoint file for '{pretrained_model_name_or_path}' \"\r\nOSError: Unable to load weights from pytorch checkpoint file for '/content/bert-base-uncased' at '/content/bert-base-uncased/pytorch_model.bin'If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. ", "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." ]

[]

[]

[ "I cannot reproduce the error on master with\r\n```\r\ntransformers 4.0.0rc1\r\ntokenizers 0.9.4\r\n```", "@thomwolf I think it must be specific to the particular version of transformers (v3.1)", "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." ]

[ { "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" ]

[]

[]

[ "Hi! Usually this error is preceded by another message. The exception with exit code 17 is not the error in itself. Please paste the full error stack trace here, alongside the command you used to launch your script.", "Hi Lysandre, here is the full error I get on the google cloud it is\nterminated because of this exception\n\nException: process 0 terminated with exit code 17\n\nbut I do not see any other error messages, do you know what might be\ncausing this?\nthanks for your help.\nBest\nRabeeh\n\n\n\"Traceback (most recent call last): File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/runpy.py\", line 193, in\n_run_module_as_main \"__main__\", mod_spec) File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/runpy.py\", line 85, in\n_run_code exec(code, run_globals) File \"/workdir/seq2seq/xla_spawn.py\",\nline 79, in <module> app.run(main, flags_parser=parse_args) File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/absl/app.py\",\nline 300, in run _run_main(main, args) File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/absl/app.py\",\nline 251, in _run_main sys.exit(main(argv)) File\n\"/workdir/seq2seq/xla_spawn.py\", line 75, in main xmp.spawn(mod._mp_fn,\nargs=(), nprocs=xla_args.num_cores) File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\",\nline 395, in spawn start_method=start_method) File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/multiprocessing/spawn.py\",\nline 157, in start_processes while not context.join(): File\n\"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/multiprocessing/spawn.py\",\nline 112, in join (error_index, exitcode) Exception: process 0 terminated\nwith exit code 17\n\n\n\nOn Wed, Nov 4, 2020 at 7:45 PM Lysandre Debut <[email protected]>\nwrote:\n\n> Hi! Usually this error is preceded by another message. The exception with\n> exit code 17 is not the error in itself.\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/8304#issuecomment-721907293>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ARPXHH2HI6ICB6HB24S7CWDSOGOMNANCNFSM4TKLTIGA>\n> .\n>\n", "Hi\nthank you for the hint, now I got it, I can debug the rest, thanks for this.\nBest\nRabeeh\n\nOn Wed, Nov 4, 2020 at 7:50 PM Rabeeh Karimi Mahabadi <[email protected]>\nwrote:\n\n> Hi Lysandre, here is the full error I get on the google cloud it is\n> terminated because of this exception\n>\n> Exception: process 0 terminated with exit code 17\n>\n> but I do not see any other error messages, do you know what might be\n> causing this?\n> thanks for your help.\n> Best\n> Rabeeh\n>\n>\n> \"Traceback (most recent call last): File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/runpy.py\", line 193, in\n> _run_module_as_main \"__main__\", mod_spec) File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/runpy.py\", line 85, in\n> _run_code exec(code, run_globals) File \"/workdir/seq2seq/xla_spawn.py\",\n> line 79, in <module> app.run(main, flags_parser=parse_args) File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/absl/app.py\",\n> line 300, in run _run_main(main, args) File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/absl/app.py\",\n> line 251, in _run_main sys.exit(main(argv)) File\n> \"/workdir/seq2seq/xla_spawn.py\", line 75, in main xmp.spawn(mod._mp_fn,\n> args=(), nprocs=xla_args.num_cores) File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch_xla/distributed/xla_multiprocessing.py\",\n> line 395, in spawn start_method=start_method) File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/multiprocessing/spawn.py\",\n> line 157, in start_processes while not context.join(): File\n> \"/root/anaconda3/envs/pytorch/lib/python3.6/site-packages/torch/multiprocessing/spawn.py\",\n> line 112, in join (error_index, exitcode) Exception: process 0 terminated\n> with exit code 17\n>\n>\n>\n> On Wed, Nov 4, 2020 at 7:45 PM Lysandre Debut <[email protected]>\n> wrote:\n>\n>> Hi! Usually this error is preceded by another message. The exception with\n>> exit code 17 is not the error in itself.\n>>\n>> —\n>> You are receiving this because you authored the thread.\n>> Reply to this email directly, view it on GitHub\n>> <https://github.com/huggingface/transformers/issues/8304#issuecomment-721907293>,\n>> or unsubscribe\n>> <https://github.com/notifications/unsubscribe-auth/ARPXHH2HI6ICB6HB24S7CWDSOGOMNANCNFSM4TKLTIGA>\n>> .\n>>\n>\n" ]

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

[]

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

[]

[]

[]

[]

[]

[]

[ { "id": 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 never know which tags are auto-generated, so please correct me if I did something useless!\r\n\r\nlooks good to me in terms of the tags 👍 " ]

[ { "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" } ]

[]

[ "Pretty cool!", "Thank you so much, @julien-c :) More models are coming ;)" ]

[]

[]

[]

[ { "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" } ]

[ { "login": "stas00", "id": 10676103, "node_id": "MDQ6VXNlcjEwNjc2MTAz", "avatar_url": "https://avatars.githubusercontent.com/u/10676103?v=4", "gravatar_id": "", "url": "https://api.github.com/users/stas00", "html_url": "https://github.com/stas00", "followers_url": "https://api.github.com/users/stas00/followers", "following_url": "https://api.github.com/users/stas00/following{/other_user}", "gists_url": "https://api.github.com/users/stas00/gists{/gist_id}", "starred_url": "https://api.github.com/users/stas00/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/stas00/subscriptions", "organizations_url": "https://api.github.com/users/stas00/orgs", "repos_url": "https://api.github.com/users/stas00/repos", "events_url": "https://api.github.com/users/stas00/events{/privacy}", "received_events_url": "https://api.github.com/users/stas00/received_events", "type": "User", "site_admin": false } ]

[ "wdyt @stas00 ", "I will work on that, thank you.", "A minor correction to the command (corrected `data_dir`):\r\n```\r\npython -m torch.distributed.launch --nproc_per_node=2 run_distributed_eval.py --model_name Helsinki-NLP/opus-mt-en-ro --save_dir test_data/opus_wmt_en_ro_gens --data_dir test_data/wmt_en_ro\r\n```\r\n\r\nQuestion: why only 1 gpu? we currently don't have it tested at all.\r\n", "I thought 1 GPU test coverage would be runnable in current CI/by more users.\r\nBut if much easier to test 2 gpu/easy to add test for 2 GPU that is great!", "Bottom line - run with as many GPUs as available. \r\n\r\nThank you for clarifying.\r\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" } ]

[]

[ "I have a newer pull req", "I have a newer pull req" ]

[]

[]

[ "Very good catch! Thanks for pointing it out, the PR mentioned above should fix this." ]

[]

[]

[ "After thinking about it a bit, I don't think the `PegasusForConditionalGeneration` should be going there. The `MODEL_MAPPING` is a mapping to all the headless models, i.e., the models that output hidden states without processing them in a head.\r\n\r\nIntroducing `PegasusForConditionalGeneration` would result in a mismatch between every single other models defined in that mapping and this newly added model. \r\n\r\nAdding `BartModel` would fail because of the configuration, as you've said, so imo the best thing to do here is to create a `PegasusModel` that inherits from `BartModel`, and use this in the `AutoModel`.", "Why is summarization pipeline using AutoModel? Shouldn't it require a model with a head?", "The pipeline is using `AutoModel` to load the weights to see if they load. It follows the current installed platform (PT or TF), but if both are installed, it first tries to load the checkpoint in `AutoModel`, and if it fails (wrongly formatted weights), it tries to load it in `TFAutoModel`.\r\n\r\nThis does mean that the model is loaded twice (once in `AutoModel` and another time in the appropriate auto model), which may not be the best performance-wise.\r\n\r\nThe easy fix here is to add a base model for Pegasus (and all models should have base models imo), the somewhat more robust fix is to load the checkpoint directly in the appropriate auto model.", "> This does mean that the model is loaded twice (once in AutoModel and another time in the appropriate auto model), which may not be the best performance-wise.\r\n\r\nYes this is not ideal. If there was simpler way do determine appropriate framework from config that would be much better. Or attempt the AutoModel way but without going the full way (stopping at checking filenames).\r\n\r\n> The easy fix here is to add a base model for Pegasus (and all models should have base models imo), the somewhat more robust fix is to load the checkpoint directly in the appropriate auto model.\r\n\r\nThat seems probably like the best solution. (at least in the short term)", "> and all models should have base models imo\r\n\r\nMarianMT, Pegasus, Blenderbot are all only published/trained/used for one task, why should they have base models?\r\n\r\nWhat ever happened to `config.architectures`? Would that help?", "Some configs (old ones maybe) don't have `architectures` defined.", "> The pipeline is using `AutoModel` to load the weights to see if they load. It follows the current installed platform (PT or TF), but if both are installed, it first tries to load the checkpoint in `AutoModel`, and if it fails (wrongly formatted weights), it tries to load it in `TFAutoModel`.\r\n\r\nJust a note that I'm not 100% sure that our design goal with Pipelines is to be able to load a model automatically in PT/TF without any user input (e.g. in case the model is only TF)\r\n\r\nBesides, in most cases you would have access to the huggingface.co model list API so you would know if model has PT/TF files.", "> Some configs (old ones maybe) don't have `architectures` defined.\r\n\r\nJust a note that we can always backport architectures into the hosted config files (will be easier with the new model versioning system)", "Obsolete." ]

[]

[]

[ "Ping @LysandreJik " ]

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

[]

[ "What is your `metrics`?", "```py\r\noptimizer= AdamWeightDecay(\r\n learning_rate=5e-5,\r\n beta_1=0.9,\r\n beta_2=0.999,\r\n weight_decay_rate=0.01,\r\n epsilon=1e-6,\r\n exclude_from_weight_decay=['layer_norm', 'bias'])\r\noptimizer._HAS_AGGREGATE_GRAD = False\r\nloss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\r\nmetrics=[tf.keras.metrics.SparseCategoricalAccuracy(name=\"acc\")]\r\n```", "@jplu might know what's going on", "Hello @aks2193!\r\n\r\nSorry for this, but for now you cannot use `.compile()` + `.fit()` to train a Token Classification model. To make it short, this is because a layer is not used and then the gradients will be None, something that `.fit()` cannot handle.\r\n\r\nIf you want to train a NER I suggest you to use the [example](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_tf_ner.py).", "Hey @jplu \r\nThanks for the reply. I tried to follow the link and below is how I changed my code\r\n\r\nmodell = TFBertForTokenClassification.from_pretrained('bert-base-uncased',num_labels=len(tag2idx))\r\nmodell.layers[2].activation = tf.keras.activations.softmax\r\nmodell.layers[0].trainable = False\r\nmodell.compile(optimizer=optimizer, loss=loss, metrics=[metrics])\r\nmodell.fit(batch_train_data, epochs=epochs, validation_data=batch_val_data)\r\n\r\ndef compute_metrics(pred):\r\n labels = pred.label_ids\r\n preds = pred.predictions.argmax(-1)\r\n precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')\r\n acc = accuracy_score(labels, preds)\r\n return {\r\n 'accuracy': acc,\r\n 'f1': f1,\r\n 'precision': precision,\r\n 'recall': recall\r\n }\r\ntraining_args = TFTrainingArguments(\r\n output_dir='./bert_test', # output directory\r\n num_train_epochs=5, # total # of training epochs\r\n per_device_train_batch_size=32, # batch size per device during training\r\n per_device_eval_batch_size=32, # batch size for evaluation\r\n warmup_steps=500, # number of warmup steps for learning rate scheduler\r\n weight_decay=0.01, # strength of weight decay\r\n logging_dir='./logs', # directory for storing logs\r\n learning_rate=3e-5,\r\n )\r\n\r\ntrainer = TFTrainer(\r\n model = modell,\r\n args=training_args,\r\n train_dataset=train_dataset,\r\n eval_dataset=val_dataset,\r\n compute_metrics=compute_metrics\r\n )\r\n \r\ntrainer.train()\r\n\r\nBut I am still getting the below error\r\n ValueError: Trying to create optimizer slot variable under the scope for tf.distribute.Strategy (<tensorflow.python.distribute.one_device_strategy.OneDeviceStrategy object at 0x7fb46c2dec50>), which is different from the scope used for the original variable (<tf.Variable 'tf_bert_for_token_classification_14/classifier/kernel:0' shape=(768, 10) dtype=float32\r\nMake sure the slot variables are created under the same strategy scope. This may happen if you're restoring from a checkpoint outside the scope\r\n\r\n\r\nThen I tried defining a strategy scope and include all the above code inside that\r\n\r\nstrategy = tf.distribute.OneDeviceStrategy(device=\"/cpu:0\")\r\nwith strategy.scope():\r\n The ABOVE CODE\r\n\r\nOn doing this getting the below error\r\n\r\nMixing different tf.distribute.Strategy objects: <tensorflow.python.distribute.one_device_strategy.OneDeviceStrategy object at 0x7fb446d98eb8> is not <tensorflow.python.distribute.one_device_strategy.OneDeviceStrategy object at 0x7fb45fe2d7b8>\r\n\r\nHow does this exactly work?\r\nHow do I define the strategy scope for all this calculations?", "Hey @aks2193 \r\nI'm facing the same problem. Please let me know if you find a solution.", "@aks2193 \r\nTry replacing `strategy = tf.distribute.OneDeviceStrategy(device=\"/cpu:0\")` to `training_args.strategy.scope()`. Worked for me.", "@alibi123 has right, you are not properly instanciate your model. Please use the example as it is. You example won't work as well with the `TFTrainer` if you are setting the activation to `softmax` because we don't compute the loss from the logits.", "@jplu I have a question about restoring weights from the checkpoint. How to do it correctly?\r\n\r\nThis is how I try to load weights:\r\n```\r\n>>> model = TFBertForTokenClassification.from_pretrained(settings.BERT_NAME)\r\n>>> model.load_weights('/models/exp1/checkpoint/ckpt-55')\r\n```\r\nI get very long exception message starting with:\r\n```\r\nNothing except the root object matched a checkpointed value. Typically this means that the checkpoint does not match the Python program. The following objects have no matching checkpointed value: [<tf.Variable 't\r\n```\r\n\r\nHere's my training code:\r\n```\r\n dataset = get_dataset(in_fn, debug)\r\n args = TFTrainingArguments(\r\n os.path.join(os.path.join(settings.MODELS_DIR, save_name)),\r\n overwrite_output_dir=True,\r\n do_train=True,\r\n logging_dir=os.path.join(settings.DATA_DIR, 'exp1_logs'),\r\n save_total_limit=2,\r\n )\r\n with args.strategy.scope():\r\n model = TFBertForTokenClassification.from_pretrained(settings.BERT_NAME)\r\n trainer = TFTrainer(\r\n model=model,\r\n args=args,\r\n train_dataset=dataset,\r\n )\r\n trainer.train()\r\n```\r\n`BERT_NAME = 'bert-base-multilingual-cased'`\r\n\r\nI was also trying to use ckpt path in `.from_pretrained()` but also got errors regarding format.\r\n", "You cannot use Keras `load_weights` on a TF checkpoint. If you want to load your model you just have to use the path where you saved your model `model = TFBertForTokenClassification.from_pretrained(\"my_output_dir\")`", "@jplu Thank you for the quick response.\r\nI've tried that in the first place, but get an error as well. \r\n\r\nThis is my output dir:\r\n```\r\ncheckpoint$ ls\r\ncheckpoint ckpt-55.data-00000-of-00001 ckpt-55.index ckpt-56.data-00000-of-00001 ckpt-56.index\r\n```\r\nHere are my attempts with error messages:\r\n1:\r\n```\r\nconfig = BertConfig.from_pretrained(settings.BERT_NAME)\r\n>>> model = TFBertForTokenClassification.from_pretrained('/models/exp1/checkpoint', config=config)\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 653, in from_pretrained\r\n [WEIGHTS_NAME, TF2_WEIGHTS_NAME], pretrained_model_name_or_path\r\nOSError: Error no file named ['pytorch_model.bin', 'tf_model.h5'] found in directory /models/exp1/checkpoint or `from_pt` set to False\r\n```\r\n2:\r\n```\r\n>>> model = TFBertForTokenClassification.from_pretrained('/models/exp1/checkpoint/ckpt-55', config=config)\r\nTraceback (most recent call last):\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 711, in from_pretrained\r\n load_tf_weights(model, resolved_archive_file)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 268, in load_tf_weights\r\n with h5py.File(resolved_archive_file, \"r\") as f:\r\n File \"/usr/local/lib/python3.6/dist-packages/h5py/_hl/files.py\", line 408, in __init__\r\n swmr=swmr)\r\n File \"/usr/local/lib/python3.6/dist-packages/h5py/_hl/files.py\", line 173, in make_fid\r\n fid = h5f.open(name, flags, fapl=fapl)\r\n File \"h5py/_objects.pyx\", line 54, in h5py._objects.with_phil.wrapper\r\n File \"h5py/_objects.pyx\", line 55, in h5py._objects.with_phil.wrapper\r\n File \"h5py/h5f.pyx\", line 88, in h5py.h5f.open\r\nOSError: Unable to open file (file signature not found)\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 714, in from_pretrained\r\n \"Unable to load weights from h5 file. \"\r\nOSError: Unable to load weights from h5 file. If you tried to load a TF 2.0 model from a PyTorch checkpoint, please set from_pt=True.\r\n```\r\n3: I've even tried `from_pt=True` even though I used TFTrainer and TFBert\r\n```\r\n>>> model = TFBertForTokenClassification.from_pretrained('/models/exp1/checkpoint/ckpt-55', config=config, from_pt=True)\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 703, in from_pretrained\r\n return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file, allow_missing_keys=True)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_pytorch_utils.py\", line 89, in load_pytorch_checkpoint_in_tf2_model\r\n pt_state_dict = torch.load(pt_path, map_location=\"cpu\")\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/serialization.py\", line 595, in load\r\n return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/serialization.py\", line 764, in _legacy_load\r\n magic_number = pickle_module.load(f, **pickle_load_args)\r\n_pickle.UnpicklingError: invalid load key, '\\x00'.\r\n```\r\n4: I've also tried to add config.json into the output dir\r\n```\r\n>>> model = TFBertForTokenClassification.from_pretrained('/models/exp1/checkpoint')\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 653, in from_pretrained\r\n [WEIGHTS_NAME, TF2_WEIGHTS_NAME], pretrained_model_name_or_path\r\nOSError: Error no file named ['pytorch_model.bin', 'tf_model.h5'] found in directory /models/exp1/checkpoint or `from_pt` set to False\r\n```", "As I said, these are normal errors because: **You cannot use Keras load_weights on a TF checkpoint.** You have to use your output dir not the file or the checkpoint dir: `model = TFBertForTokenClassification.from_pretrained('/models/exp1')`.", "@jplu Sorry for bothering. But still doesn't work. It expects `'pytorch_model.bin', 'tf_model.h5'`.\r\n```\r\n>>> model = TFBertForTokenClassification.from_pretrained('/models/exp1')\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 653, in from_pretrained\r\n [WEIGHTS_NAME, TF2_WEIGHTS_NAME], pretrained_model_name_or_path\r\nOSError: Error no file named ['pytorch_model.bin', 'tf_model.h5'] found in directory /models/exp1 or `from_pt` set to False\r\n\r\n>>> model = TFBertForTokenClassification.from_pretrained('/models/exp1', from_pt=True)\r\nTraceback (most recent call last):\r\n File \"<stdin>\", line 1, in <module>\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_utils.py\", line 653, in from_pretrained\r\n [WEIGHTS_NAME, TF2_WEIGHTS_NAME], pretrained_model_name_or_path\r\nOSError: Error no file named ['pytorch_model.bin', 'tf_model.h5'] found in directory /models/exp1 or `from_pt` set to False\r\n\r\n```\r\n\r\nBut my output_dir only contains `checkpoint` dir", "This is because you are trying to load a PyTorch model into a TensorFlow one with `from_pt=True`, remove this parameter. If not working it means that your models have not been properly saved.\r\n\r\nDid you call the `save` method of the trainer?", "No, I haven't. Sorry, my bad. I thought that I can use checkpoints. \r\nThanks for your help!", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

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

[ "in the main page of examples you mention one can pass any script to run it on tpu, but inside the seq2seq it seems one needs to use finetune_trainer and not finetune.py for tpus, I am confused which one to use, thanks for your help", "@sshleifer @patil-suraj Maybe we could improve the documentation here", "Hi, thank you @LysandreJik, do you know which version of finetune.py to finetune_trainer.py are working with tpus? in the documentation it is written any example can be run with multiple tpus by using xla_spawn.py but I am not sure if this is true for finetune.py too. thanks ", "`finetune_trainer.py` works with TPU, here is [the wmt script](https://github.com/huggingface/transformers/blob/master/examples/seq2seq/builtin_trainer/finetune_tpu.sh)\r\n\r\n\r\n1) They were developed at different times. We are trying to get them both working well.\r\n2) finetune.py should not be used with TPU at all.\r\n3) yes, see script.\r\n4) see script\r\n5) see script\r\n6) see script\r\n7) I would guess that those samplers don't work on TPU.\r\n8) No it does not.\r\n9) No idea, maybe @patil-suraj knows.\r\n\r\n\r\n", "thanks Sam for this, so finetune.py does not work? In the documentation,\nthis is written one can run all examples with xla_spawn on TPU, I\nappreciate updating the README mentioning it.\nthank you.\n\nOn Wed, Nov 4, 2020 at 8:06 PM Sam Shleifer <[email protected]>\nwrote:\n\n> finetune_trainer.py works with TPU, cc @patil-suraj\n> <https://github.com/patil-suraj> .\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/8290#issuecomment-721917741>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ARPXHH345RVTXO3EA7ZUPUTSOGQ3XANCNFSM4TKBDLQQ>\n> .\n>\n", "Correct, updated docs.", "thank you Sam\n\nOn Wed, Nov 4, 2020 at 9:46 PM Sam Shleifer <[email protected]>\nwrote:\n\n> Correct, updated docs.\n>\n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/8290#issuecomment-721966350>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ARPXHH4QZWN6KEL2BKSF7BTSOG4TVANCNFSM4TKBDLQQ>\n> .\n>\n" ]

[]

[]

[ "I also have this question" ]

[ { "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 } ]

[ "This notebook should help: https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[]

[]

[]

[]

[]

[ { "login": "lhoestq", "id": 42851186, "node_id": "MDQ6VXNlcjQyODUxMTg2", "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhoestq", "html_url": "https://github.com/lhoestq", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "repos_url": "https://api.github.com/users/lhoestq/repos", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "type": "User", "site_admin": false } ]

[ "Maybe @lhoestq @patrickvonplaten have an idea", "Hey @gaobo1987,\r\n\r\nWe checked that the models match the performance as reported in the paper. \r\n\r\nDid you run the model as stated in https://github.com/huggingface/transformers/blob/master/examples/rag/README.md ? ", "Which index did you use exactly with wiki_dpr ? This EM value is expected if you used the `compressed` one. For the `exact` one you might need to increase the efSearch parameter of the index. I ran some indexing experiments recently and I'll update the default parameters of the wiki_dpr index with the optimized ones that reproduce RAG's paper results.\r\n\r\nEDIT: they've been updated a few weeks ago", "> Hey @gaobo1987,\r\n> \r\n> We checked that the models match the performance as reported in the paper.\r\n> \r\n> Did you run the model as stated in https://github.com/huggingface/transformers/blob/master/examples/rag/README.md ?\r\n\r\nThanks for your reply @patrickvonplaten ,\r\n\r\nwe did not use the example run script there, but followed the code snippets provided in the huggingface documentation:\r\n\r\n```python\r\nfrom transformers import RagTokenizer, RagRetriever, RagSequenceForGeneration\r\nimport torch\r\ntokenizer = RagTokenizer.from_pretrained(\"facebook/rag-sequence-nq\")\r\nretriever = RagRetriever.from_pretrained(\"facebook/rag-sequence-nq\", index_name=\"exact\", use_dummy_dataset=True)\r\n# initialize with RagRetriever to do everything in one forward call\r\nmodel = RagSequenceForGeneration.from_pretrained(\"facebook/rag-token-nq\", retriever=retriever)\r\ninput_dict = tokenizer.prepare_seq2seq_batch(\"How many people live in Paris?\", \"In Paris, there are 10 million people.\", return_tensors=\"pt\")\r\ninput_ids = input_dict[\"input_ids\"]\r\noutputs = model(input_ids=input_ids, labels=input_dict[\"labels\"])\r\n# or use retriever seperately\r\nmodel = RagSequenceForGeneration.from_pretrained(\"facebook/rag-sequence-nq\", use_dummy_dataset=True)\r\n# 1. Encode\r\nquestion_hidden_states = model.question_encoder(input_ids)[0]\r\n# 2. Retrieve\r\ndocs_dict = retriever(input_ids.numpy(), question_hidden_states.detach().numpy(), return_tensors=\"pt\")\r\ndoc_scores = torch.bmm(question_hidden_states.unsqueeze(1), docs_dict[\"retrieved_doc_embeds\"].float().transpose(1, 2)).squeeze(1)\r\n# 3. Forward to generator\r\noutputs = model(context_input_ids=docs_dict[\"context_input_ids\"], context_attention_mask=docs_dict[\"context_attention_mask\"], doc_scores=doc_scores, decoder_input_ids=input_dict[\"labels\"])\r\n```\r\nsee here: https://huggingface.co/transformers/model_doc/rag.html#ragsequenceforgeneration\r\n\r\nWe did use our own evaluation script for computing EM scores.\r\n\r\nIn general, we tried to follow the prescribed steps from official source as exactly as possible, as for the customized EM calculation, difference may arise there, but I believe the main source of performance difference lies somewhere else.", "> Which index did you use exactly with wiki_dpr ? This EM value is expected if you used the `compressed` one. For the `exact` one you might need to increase the efSearch parameter of the index. I ran some indexing experiments recently and I'll update the default parameters of the wiki_dpr index with the optimised ones that reproduce RAG's paper results.\r\n\r\nthanks for the reply @lhoestq , we used the \"exact\" mode of the wiki_dpr index, indeed, we haven't tried the \"compressed\" mode, nor did we tune the \"exact\" index. Thanks for the update, we will check the \"compressed\" alternative, and the parameter tuning of the \"exact\" index. Also great to know that you will update the default parameters!", "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, to provide an update on this issue. Recently I refactored my own RAG code based transformers-4.1.1, and obtained EM=40.7 performance on the open NQ dataset with rag-sequence-nq model (n_beams=4) and FAISS HNSW index with n_docs=5, efSearch=256 and efConstruction=200. Unfortunately it still didn't reach the expected 44.5 score. Are these sound parameters? Am I missing any? What is the best parameter combination used at Huggingface? Any advice is much appreciated, thanks! (Note that I couldn't use the original rag code as there is firewall restrictions on my server that prevented downloading the wiki_dpr.py script as well the arrow files for exact indexing, so I have to download these files on a much less powerful laptop and upload them to my server. Consequently, I am using a modified version of RagSequenceForGeneration along with a modified RagRetriever) @lhoestq ", "@gaobo1987 \r\nCan you please share how exactly you played around with the efSearch and efConstruction parameters?\r\n\r\nAs in where in the code did you make the changes??", "hello @krishanudb , thanks for your reply. What I did is merely manually downloading the wiki_dpr-train.arrow file, then use it to construct a faiss hnsw index with efSearch=256, efConstruction=200, then save this index to disk. I wrote a wrapper around RagRetriever and RagSequenceForGeneration respectively so that rag can run directly on the aforementioned faiss index, instead of relying on huggingFace.Datasets utilities and other caching sub-routines. I did not change the models in any way. Could you provide an answer to my question regarding the best combination of parameters from huggingFace to reach the performance as reported in the original paper? Thanks for your time", "@gaobo1987 \r\nThere are several versions of the DPR model (single-nq vs multiset) as well as the precomputed passages wiki_dpr\r\nI am not sure which one the authors used to get 44% EM but I think they have used the single-nq models for the tasks.\r\n\r\nMake sure that you are using the 'right; model. Maybe the authors can shed more light on this..\r\n\r\nEven I am facing the same issue... Not getting more than 40% EM no matter if I use the multiset or the nq-single models..", "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." ]

[ { "id": 2648621985, "node_id": "MDU6TGFiZWwyNjQ4NjIxOTg1", "url": "https://api.github.com/repos/huggingface/transformers/labels/Feature%20request", "name": "Feature request", "color": "FBCA04", "default": false, "description": "Request for a new feature" } ]

[ { "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 }, { "login": "lhoestq", "id": 42851186, "node_id": "MDQ6VXNlcjQyODUxMTg2", "avatar_url": "https://avatars.githubusercontent.com/u/42851186?v=4", "gravatar_id": "", "url": "https://api.github.com/users/lhoestq", "html_url": "https://github.com/lhoestq", "followers_url": "https://api.github.com/users/lhoestq/followers", "following_url": "https://api.github.com/users/lhoestq/following{/other_user}", "gists_url": "https://api.github.com/users/lhoestq/gists{/gist_id}", "starred_url": "https://api.github.com/users/lhoestq/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/lhoestq/subscriptions", "organizations_url": "https://api.github.com/users/lhoestq/orgs", "repos_url": "https://api.github.com/users/lhoestq/repos", "events_url": "https://api.github.com/users/lhoestq/events{/privacy}", "received_events_url": "https://api.github.com/users/lhoestq/received_events", "type": "User", "site_admin": false } ]

[ "@stas00 \r\n\r\nCan you please write a test code for **finetune.sh**. ", "As you can see I'm waiting for this ticket to be addressed before I'm able to write the tests. \r\n\r\nPerhaps you can address that, and then I will have all the info needed to write the tests.", "Until then please file a normal issue about it. I haven't done any rag work yet, so that's why I'm asking for support.", "@lhoestq is working on this at the moment :-) ", "Actually I'm working on the finetuning script example, not eval ;)\r\nBut maybe this can help with adding a test for the eval script example.", "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", "stale" ]

[]

[]

[ "ping", "Looks good to me. Thanks for handling this one @stas00 and sorry for the delay." ]

[]

[]

[]

[ { "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": "julien-c", "id": 326577, "node_id": "MDQ6VXNlcjMyNjU3Nw==", "avatar_url": "https://avatars.githubusercontent.com/u/326577?v=4", "gravatar_id": "", "url": "https://api.github.com/users/julien-c", "html_url": "https://github.com/julien-c", "followers_url": "https://api.github.com/users/julien-c/followers", "following_url": "https://api.github.com/users/julien-c/following{/other_user}", "gists_url": "https://api.github.com/users/julien-c/gists{/gist_id}", "starred_url": "https://api.github.com/users/julien-c/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/julien-c/subscriptions", "organizations_url": "https://api.github.com/users/julien-c/orgs", "repos_url": "https://api.github.com/users/julien-c/repos", "events_url": "https://api.github.com/users/julien-c/events{/privacy}", "received_events_url": "https://api.github.com/users/julien-c/received_events", "type": "User", "site_admin": false } ]

[ "Very cool. Is possible, can you add metadata as described in https://huggingface.co/docs#what-metadata-can-i-add-to-my-model-card?" ]

[]

[]

[ "The issue solved with setting --gpus 1 explicitly. thanks. " ]

[]

[]

[]

[]

[]

[ "Another error:\r\n```\r\n---------------------------------------------------------------------------\r\nImportError Traceback (most recent call last)\r\n<ipython-input-7-279c49635b32> in <module>\r\n----> 1 import transformers\r\n\r\n/scratch/a1204a01/.conda/envs/notebook/lib/python3.7/site-packages/transformers/__init__.py in <module>\r\n 133 \r\n 134 # Pipelines\r\n--> 135 from .pipelines import (\r\n 136 Conversation,\r\n 137 ConversationalPipeline,\r\n\r\n/scratch/a1204a01/.conda/envs/notebook/lib/python3.7/site-packages/transformers/pipelines.py in <module>\r\n 35 from .file_utils import add_end_docstrings, is_tf_available, is_torch_available\r\n 36 from .modelcard import ModelCard\r\n---> 37 from .tokenization_auto import AutoTokenizer\r\n 38 from .tokenization_bert import BasicTokenizer\r\n 39 from .tokenization_utils import PreTrainedTokenizer\r\n\r\n/scratch/a1204a01/.conda/envs/notebook/lib/python3.7/site-packages/transformers/tokenization_auto.py in <module>\r\n 117 \r\n 118 if is_tokenizers_available():\r\n--> 119 from .tokenization_albert_fast import AlbertTokenizerFast\r\n 120 from .tokenization_bart_fast import BartTokenizerFast\r\n 121 from .tokenization_bert_fast import BertTokenizerFast\r\n\r\n/scratch/a1204a01/.conda/envs/notebook/lib/python3.7/site-packages/transformers/tokenization_albert_fast.py in <module>\r\n 21 \r\n 22 from .file_utils import is_sentencepiece_available\r\n---> 23 from .tokenization_utils_fast import PreTrainedTokenizerFast\r\n 24 from .utils import logging\r\n 25 \r\n\r\n/scratch/a1204a01/.conda/envs/notebook/lib/python3.7/site-packages/transformers/tokenization_utils_fast.py in <module>\r\n 28 from tokenizers.decoders import Decoder as DecoderFast\r\n 29 \r\n---> 30 from .convert_slow_tokenizer import convert_slow_tokenizer\r\n 31 from .file_utils import add_end_docstrings\r\n 32 from .tokenization_utils import PreTrainedTokenizer\r\n\r\n/scratch/a1204a01/.conda/envs/notebook/lib/python3.7/site-packages/transformers/convert_slow_tokenizer.py in <module>\r\n 26 \r\n 27 # from transformers.tokenization_openai import OpenAIGPTTokenizer\r\n---> 28 from transformers.utils import sentencepiece_model_pb2 as model\r\n 29 \r\n 30 from .file_utils import requires_sentencepiece\r\n\r\nImportError: cannot import name 'sentencepiece_model_pb2' from 'transformers.utils' (/home01/a1204a01/.local/lib/python3.7/site-packages/transformers/utils/__init__.py)\r\n```", "Hi! could you let us know how you installed `transformers`?", "I built from source, by ```git clone``` and ```pip install .```\r\n\r\nEDIT: Huh, it's now giving error ```ImportError: cannot import name 'is_main_process' from 'transformers.trainer_utils'```", "Fixed by reinstalling python3 and reinstalling transformer with latest commit" ]

[]

[]

[ "Hello! First of all, you're using the `squeezebert-mnli` checkpoint, which is a checkpoint that was fine-tuned on the MNLI dataset. It cannot be used to do masked language modeling.\r\n\r\nI believe you should be using the `squeezebert-uncased` checkpoint instead.\r\n\r\nHowever, even when using that checkpoint with the MLM pipeline I cannot obtain sensible results. Maybe @forresti can chime in and let us know if something's up!\r\n\r\n", "Thanks @LysandreJik . I used both squeezebert-mnli and squeezebert-uncased (not shown). Same type of results. Thanks for checking. @forresti any thoughts? Is there something wrong with the squeezbert tokenizer? ", "@ontocord Sorry for the slow reply. I will dig into this on Thursday this week.", "@ontocord Thanks so much for bringing this to my attention! I was able to reproduce the issue. And, I think I was able to fix the issue in PR #8479.\r\n\r\nNow, let's try running your example code with...\r\n* PR #8479\r\n* the `squeezebert-uncased` checkpoint\r\n\r\n... this produces the following output:\r\n```\r\nSome weights of the model checkpoint at squeezebert/squeezebert-uncased were not used when initializing SqueezeBertForMaskedLM: ['cls.seq_relationship.weight', 'cls.seq_relationship.bias']\r\n- This IS expected if you are initializing SqueezeBertForMaskedLM from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\r\n- This IS NOT expected if you are initializing SqueezeBertForMaskedLM from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\r\nSome weights of SqueezeBertForMaskedLM were not initialized from the model checkpoint at squeezebert/squeezebert-uncased and are newly initialized: ['transformer.embeddings.position_ids']\r\nYou should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\r\n**\r\nhe was an american politician and lawyer who served as the 16th president of the united states from 1861 to 1865 . he led the nation through the american civil war , the country ' s greatest war , war , and economic crisis . , war , economic economic war\r\njohnson is a americans statesman & attorney and serve the interim 17th presidency in of confederate state in 1860 until 1866 \" white lead a country throughout a america black wars and a nation ’ largest largest economic and famine and , political crises \" and famine war and political crisis\r\n**\r\ngeorge washington , who served as the first president of the united states from 1796 to 1797 , was an american political leader , patriot patriot , statesman , and founding father . previously , he led patriot forces to victory in the nation ' s war for independence . ,\r\njames harrison s jr serve in inaugural inaugural presidency in s u united in 1789 until 1799 ) is a americans politician figure and military statesman and politician and , adoptive fathers \" historically was his lead revolutionary troops in fight during a country ’ the fight of freedom \" and\r\n**\r\njohnson , the first african - american president of the united states , is an american politician and attorney who served as the 44th president of the united states from 2016 to 2017 . he was a member of the republican party . , john the republican republican party . the\r\nwilliams is , second black – americans governor in this colored senate islander was a americans political , lawyer , serves the a 43rd governor for of union state in 2015 until 2016 , she is an part the house democratic assembly \" . james senate democratic democratic assembly party and\r\n```\r\n\r\nAlas, the model seems to think Obama's name is \"Johnson,\" but it does get George Washington correct.\r\n\r\nAnyway, does this output look a bit more like what you expected? :)", "Thsnks a lot @forresti! This works as well with the fill-mask pipeline:\r\n\r\n```py\r\n>>> from transformers import AutoModelForMaskedLM, AutoTokenizer\r\n\r\n>>> model = AutoModelForMaskedLM.from_pretrained('squeezebert/squeezebert-uncased')\r\n>>> tokenizer = AutoTokenizer.from_pretrained('squeezebert/squeezebert-uncased')\r\n>>> input_txt = [\r\n... \"George Washington, who served as the first [MASK] of the United States from 1789 to 1797, was an American political leader.\"\r\n... ]\r\n\r\n>>> from transformers import pipeline\r\n>>> nlp = pipeline(\"fill-mask\", model=model, tokenizer=tokenizer)\r\n>>> print(nlp(input_txt))\r\n[{'sequence': '[CLS] george washington, who served as the first president of the united states from 1789 to 1797, was an american political leader. [SEP]', 'score': 0.9644643664360046, 'token': 2343, 'token_str': 'president'}, {'sequence': '[CLS] george washington, who served as the first governor of the united states from 1789 to 1797, was an american political leader. [SEP]', 'score': 0.026940250769257545, 'token': 3099, 'token_str': 'governor'}, {'sequence': '[CLS] george washington, who served as the first king of the united states from 1789 to 1797, was an american political leader. [SEP]', 'score': 0.0013772461097687483, 'token': 2332, 'token_str': 'king'}, {'sequence': '[CLS] george washington, who served as the first lieutenant of the united states from 1789 to 1797, was an american political leader. [SEP]', 'score': 0.0012003666488453746, 'token': 3812, 'token_str': 'lieutenant'}, {'sequence': '[CLS] george washington, who served as the first secretary of the united states from 1789 to 1797, was an american political leader. [SEP]', 'score': 0.0008091009221971035, 'token': 3187, 'token_str': 'secretary'}]\r\n\r\n```", "Thank @forresti! Yes this fixes the problem! Thank you @LysandreJik as well! I noticed that different models have different capacities to store facts. Roughly based on the number of parameters, but not always. As a question, do you know of any models that are trained to identify a relationship and not a word in the mask:, leader($X, president,united_states,1789,1797) served as the first president of the united states from 1789 to 1797 ... in theory this should reduce the number of facts the model needs to learn as the relationships are already being learned by the attention mechanism, I belive. \r\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 } ]

[ "Hey @zhiheng-huang,\r\n\r\nit would be great if you could take a look at the failing tests :-) ", "Hey @patrickvonplaten, I fixed all failed tests except check_code_quality. Currently the relative embedding is implemented for BERT only. In check_code_quality, `utils/check_copies.py` tries to copy the relative embedding implementation from BERT model to other models including `albert`, `electra`, `roberta` etc. I understand this may make the relative embedding methods ready to be used in those models. However, we haven't pre-trained those type of models with relative embedding and thus cannot assess their effectiveness. Please advise if I should fix this failing test (by ensuring relative embedding implementation copied to those BERT variants) or leave it as is. ", "Hey @zhiheng-huang, \r\n\r\nSadly there is still a problem with the git commit history. As you can see 54 files are changed in this PR. Could you make sure to keep the commit tree clean. It is not really possible to review the PR otherwise :-/ \r\n\r\nTry to make use of `git rebase` to avoid appending the master's commit history to your branch maybe", "In the worst case, you can just make the changes to the files you intend to change without `rebasing` or `merging` and then I can review and merge/rebase for you. ", "Rebased and removed the unintended merge commit. @patrickvonplaten, can you comment on the `utils/check_copies.py` question so we can move forward?", "Hi @patrickvonplaten @LysandreJik, I see one approval already, is it ready to merge? If not, can you point to the embedding (for example absolute position embedding) unit tests so I can try to come up with similar tests?", "Regarding tests, I think adding integration tests in the `test_modeling_bert.py` would be nice. What do you think @patrickvonplaten?\r\n\r\nThe BERT model doesn't have any such tests right now, but you can take inspiration from the `RobertaModelIntegrationTest` class in `test_modeling_roberta.py`, which you can find [here](https://github.com/huggingface/transformers/blob/master/tests/test_modeling_roberta.py#L402).\r\n\r\nYou could add a couple of tests, each testing that you get the expected results: this will ensure the implementation will not diverge in the future. If you need a checkpoint, you can use `lysandre/tiny-bert-random`, which is a very small model (with random values), so it will be very light on the CI.\r\n\r\nLet me know if you need anything.", "@patrickvonplaten @LysandreJik \r\n1. Added forward test to ensure forward runs okay for `LayoutLM`, `Roberta`, `ELECTRA`, and `BERT` for three position embeddings: \"absolute\", \"relative_key\", \"relative_key_query\".\r\n2. Added integration test for `BERT` check points `bert-base-uncased`, `zhiheng-huang/bert-base-uncased-embedding-relative-key`, and `zhiheng-huang/bert-base-uncased-embedding-relative-key-query` to ensure that models predictions match expected outputs.", "@zhiheng-huang - Let me fix the CI later, don't worry about it :-) ", "> This looks good to me. Thanks a lot for your PR!\r\n> Any reason ALBERT and Longformer don't get this new functionality? (But RoBERTa and ELECTRA do?)\r\n\r\nGreat question! I ALBERT should get this functionality (I just added it - great catch!). Longformer has weird attention_scores which does not work with those embeddings.", "Good to merge! Thanks a mille @zhiheng-huang! ", "> Good to merge! Thanks a mille @zhiheng-huang!\r\n\r\nThanks! @patrickvonplaten @sgugger @LysandreJik " ]

[]

[]

[ "And and update on proposing this multiple report files feature to be a core feature in pytest https://github.com/pytest-dev/pytest/issues/7972 - one of the developers vetoed it, so I guess it will remain just here for now. It makes no sense for this not to be a core feature of pytest, as we are just splitting the huge mess of everything being dumped to the terminal to just one file per report, but there was no invitation to discuss that - just NO. If someone wants to make it into a pytest plugin it'd surely be useful to others.", "> I saw the discussion (or lack thereof) on pytest. Their loss! We don't mind having the post-processing in transformers.\r\n\r\nEventually we will either have to port it to pytest hooks or keep up with the pytest API changes, since currently the code uses `pytest` internals and could break should they change those. It's just so much simpler doing that than reinventing the wheel." ]

[]

[]

[ "I just tried it, and noticed that it doesn't work if `features` is `List[Dict[str, torch.Tensor]]`,\r\nbecause `tokenizer.pad()` will set `return_tensors` to `pt` if `input_ids` is `torch.Tensor` and `return_tensors` is `None`.\r\n\r\nFor example my dataset looked like this.\r\n```python\r\ndef __getitem__(self, i): \r\n return {k: torch.tensor(v, dtype=torch.long) for k,v in self.examples[i].items()} \r\n```\r\nChanging to this solves the problem.\r\n```python\r\ndef __getitem__(self, i): \r\n return self.examples[i] \r\n```\r\n\r\nMaybe I should have always used this and leave it to collator to tensorize the features.", "Will have look. In general, yes it's better to have your examples be the results of the tokenization (so `Dict[str, List[int]]`) and let the data collator handles the conversion to tensors." ]

[]

[]

[ "@sgugger changes are made...", "Thanks!" ]

[]

[]

[ "I'm encountering the same problem. Were you able to solve it?", "Could you add the following lines:\r\n\r\n```py\r\nfrom transformers import logging as hf_logging\r\n\r\nhf_logging.set_verbosity_info()\r\n```\r\nbefore reloading the model, and paste the results here?\r\n\r\ncc @sgugger ", "loading configuration file trained_models/checkpoint-8000/config.json\r\nModel config DistilBertConfig {\r\n \"_name_or_path\": \"distilbert-base-uncased\",\r\n \"activation\": \"gelu\",\r\n \"architectures\": [\r\n \"DistilBertForSequenceClassification\"\r\n ],\r\n \"attention_dropout\": 0.1,\r\n \"dim\": 768,\r\n \"dropout\": 0.1,\r\n \"hidden_dim\": 3072,\r\n \"id2label\": { ... },\r\n \"initializer_range\": 0.02,\r\n \"label2id\": { ... },\r\n \"max_position_embeddings\": 512,\r\n \"model_type\": \"distilbert\",\r\n \"n_heads\": 12,\r\n \"n_layers\": 6,\r\n \"pad_token_id\": 0,\r\n \"qa_dropout\": 0.1,\r\n \"seq_classif_dropout\": 0.2,\r\n \"sinusoidal_pos_embds\": false,\r\n \"tie_weights_\": true,\r\n \"vocab_size\": 30522\r\n}\r\n\r\nloading weights file trained_models/checkpoint-8000/pytorch_model.bin\r\nAll model checkpoint weights were used when initializing DistilBertForSequenceClassification.\r\n\r\nAll the weights of DistilBertForSequenceClassification were initialized from the model checkpoint at trained_models/checkpoint-8000.\r\nIf your task is similar to the task the model of the checkpoint was trained on, you can already use DistilBertForSequenceClassification for predictions without further training.", "When I go and evaluate the model from this point (either manually or by making a Trainer and using trainer.evaluate()) I get terrible scores. \r\n\r\nIf I make a Trainer and try to continue training, I get terrible loss scores _except_ if I provide the checkpoint directory as part of the input to trainer.train(). If I supply the checkpoint directory there, the training appears to continue from the checkpoint, and if I train for ~300 more iterations, trainer.evaluate() gives decent performance but still not what I was seeing during the initial run. ", "Okay, that's interesting. Do you mind sharing with us your environment? You can run `!transformers-cli env` and put the result here, we'll look into it.", "Thanks @LysandreJik.\r\n\r\n- `transformers` version: 3.5.0\r\n- Platform: Linux-5.4.0-1030-aws-x86_64-with-Ubuntu-18.04-bionic\r\n- Python version: 3.6.9\r\n- PyTorch version (GPU?): 1.6.0 (True)\r\n- Tensorflow version (GPU?): 2.3.1 (False)\r\n- Using GPU in script?: Yes, single K80 on AWS\r\n- Using distributed or parallel set-up in script?: No", "There is little we can do to debug without a reproducer, which we don't have as the initial code contains a `train_dataset` and an `eval_dataset` we don't have access to. I just tried the notebook on {GLUE](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb) and ran until the end of training (before hyperparameter-search), saved the model with `trainer.save_model(some_path)`, the restarted the notebook, ran all the cells up until the training then a new one with\r\n```\r\nmodel = AutoModelForSequenceClassification.from_pretrained(some_path, local_files_only=True)\r\ntrainer.model = model.cuda()\r\ntrainer.evaluate()\r\n```\r\nand it gave the exact same results as the end of training, so the `from_pretrained` method works well with the distilbert models.", "As an update, I find that it's not just Distilbert models which will not save/reload for me, but also an Albert model gives the same behavior. Evaluation at the end of training gives 68% accuracy on my problem, whereas save/reload/reevaluate gives <1% accuracy. Currently trying transformers 4.0 rather than 3.5.", "Thanks for the update. As mentioned above, it does not help us fix this problem. We need a reliable reproducer for that.", "@sgugger Turns out the problem was my fault. I was not keeping a consistent mapping of label names to integers across my runs. Once I corrected this my models performed identically after reload. Perhaps the OPs problem was similar. In any case, thanks for the help (the benchmark you linked helped me to debug) and sorry for the wild goose chase.", "Glad you found the reason to your issue!", "@mkreisel How are you \"keeping a consistent mapping of label names to integers\" across your runs now? Do you use a huggingface dataset ClassLabel ? I noticed in this original problem that it might be that the label mapping is somehow different after reloading the model when using num_labels in from_pretrained().", "> There is little we can do to debug without a reproducer, which we don't have as the initial code contains a `train_dataset` and an `eval_dataset` we don't have access to. I just tried the notebook on {GLUE](https://github.com/huggingface/notebooks/blob/master/examples/text_classification.ipynb) and ran until the end of training (before hyperparameter-search), saved the model with `trainer.save_model(some_path)`, the restarted the notebook, ran all the cells up until the training then a new one with\r\n> \r\n> ```\r\n> model = AutoModelForSequenceClassification.from_pretrained(some_path, local_files_only=True)\r\n> trainer.model = model.cuda()\r\n> trainer.evaluate()\r\n> ```\r\n> \r\n> and it gave the exact same results as the end of training, so the `from_pretrained` method works well with the distilbert models.\r\n\r\n@sgugger The issue is strictly with tokenclassification class. The index of the labels gets misaligned somehow when reloading a tokenclassification model. The problem happens across many model types: bert, distilbert, roberta, etc. If just giving num_labels = x when loading the model. I believe the issue has to do with the tokenizers and the fact that setting subwords equal to -100 creates another class when training the model, but that class is no longer available when you reload a pretrained tokenclassification model using from_pretrained(local_path). ", "@nhsmith85 I was doing index -> class mapping using my own dictionary, not using anything internal to HuggingFace. I created a dataset class as an extension of torch.utils.data.Dataset:\r\n\r\n```\r\nclass RecordsDataset(torch.utils.data.Dataset):\r\n def __init__(self, encodings, labels):\r\n self.encodings = encodings\r\n self.labels = labels\r\n\r\n def __getitem__(self, idx):\r\n item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}\r\n item['labels'] = torch.tensor(self.labels[idx])\r\n return item\r\n\r\n def __len__(self):\r\n return len(self.labels)\r\n```\r\n\r\nAt this point the text labels had already been mapped to integers. ", "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.", "Hello, Did anyone manage to get a solution for this? I am facing a very similar issue on `ViTForImageClassification` on using \"google/vit-base-patch16-224\". Upon training, I am getting an accuracy of 0.75 and a very low loss. However, once I save and reload it, say after a day, the loss is back to ~10 and accuracy is 0.\r\n\r\nPlease find the necessary parts of my code here: https://gist.github.com/thevishnupradeep/d5efc0b0510d8a30d997cadd836d2c61", "Also encountering the exact same problem with DistilBERT for QA.", "As of July 2023 facing same issue with Bert model. Some one suggest a fix", "Saved model performance is very bad compared to online model. Why???? " ]

[]

[]

[ "The problem is mine. As other user suggest me in Stackoverflow, I have to save the model this way\r\n```\r\nmodel.save_pretrained (\"my-model\")`\r\n```" ]

[]

[]

[ "Thanks a lot!" ]

[]

[]

[ "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." ]

[]

[]

[]

[]

[]

[]

[ { "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": 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" ]

[]

[]

[ "# 1. Input data\r\n\r\n> The PyTorch side of the library (and all the PyTorch scripts) have datasets that eventually yield dictionaries containing all inputs of the model as well as the labels. In an ideal world, it would be great if we could leverage that format easily for TF as well (so that the same script can be used for PT and TF by just changing a few lines, especially when using the datasets library). I don't know if that's possible or not but one thing to explore more I believe.\r\n\r\nCan you elaborate a bit more please? Do you mean that the input data given to the `.fit()` method should be a dictionary? If it is what you mean, it is already the case.\r\n\r\n# 2. Optimizer and scheduler\r\n> I like the new optimizer for gradient accumulation a lot. This feels like a very good design. Should we deprecate GradientAccumulator?\r\n\r\nYes, this should be deprecated because we won't use it anymore.\r\n\r\n> But where are the schedulers? Is this something you intend to control via callbacks? If that's the case I didn't see one with a sensible default (the linear + warmup used in PT for instance).\r\n\r\nThe schedulers are directly inside the optimizer, if you look at the `create_optimizer` method you can see that the scheduler is first created and then given to the Adam optimizer as input for the `learning_rate` parameter. In the previous Trainer the scheduler was returned only for being used in the logging, the scheduling is done automatically internally in the `tf.keras.optimizers.Optimizer` class.\r\n\r\n# 3. Callbacks\r\n> Leveraging Keras callbacks is definitely a good idea. My only remark here is that is should be more customizable. On the PT side we have some default callbacks and the init takes a list of additional callbacks the user can add.\r\n\r\nNo worries, this will be added in the next push :)\r\n\r\n# 4. Metrics\r\n> I'm not in favor of adding a new file of metrics we will have to maintain forever. We should provide an adapter for datasets Metric object and rely on the datasets library (or existing Keras metrics if users prefer them) but they shouldn't be in the transformers library (on the PT we will soon deprecate the ones in the metric folder).\r\n\r\nI fully agree, and this is why I asked @thomwolf and @lhoestq opinion's on this on the best way to integrate Keras metrics inside datasets :) For this will require, I think, a non negligeable amont of work in `datasets` that I would prefer to do not do alone.", "> Can you elaborate a bit more please? Do you mean that the input data given to the `.fit()` method should be a dictionary? If it is what you mean, it is already the case.\r\n\r\nIf you look at the new example scripts (like [run_glue](https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_glue.py)) the datasets are immediately sent to `Trainer` with no platform-specific processing needed. It would be really cool if we could just replace `Trainer` by `TFTrainer` in that script and have it work the same way. I'm not sure if the easiest for that is to change the input of `training_step` or do some internal processing of the dataset inside `TFTrainer`.", "> If you look at the new example scripts (like run_glue) the datasets are immediately sent to Trainer with no platform-specific processing needed. It would be really cool if we could just replace Trainer by TFTrainer in that script and have it work the same way. I'm not sure if the easiest for that is to change the input of training_step or do some internal processing of the dataset inside TFTrainer.\r\n\r\nHum I see. As a first glance, I would say it will requires much more changes than just replace Trainer by TFTrainer, at least all the metrics part won't be compliant, and the way we create the model is different (in TF we have to do that in a strategy).", "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." ]

[]

[]

[ "Hi! I believe those are warnings and not errors? Does it change the expected results when tracing the model?", "You are right, despite the warnings (and my - limited - understanding of what should work inside tracing), the output of the compiled model with different inputs are comparable to the base ones. \r\n\r\nThanks a lot!", "Glad it works!", "@gcompagnoni @LysandreJik \r\nHi, there\r\nWhy do I try the example below occurs error ? issue reported in #15598\r\n\r\n```\r\nimport torch\r\nfrom transformers import GPT2Tokenizer, GPT2LMHeadModel\r\n\r\ntokenizer = GPT2Tokenizer.from_pretrained('gpt2')\r\nmodel = GPT2LMHeadModel.from_pretrained('gpt2')\r\nmodel.eval()\r\n\r\ntokens=tokenizer('The cat is on the table.', return_tensors='pt')['input_ids']\r\n\r\nwith torch.jit.optimized_execution(True):\r\n traced_model = torch.jit.trace(model, tokens)\r\n```\r\n" ]

[]

[]

[]

[]

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

[ "Maybe I found out, is it:\r\n\r\n```\r\nfor i, sample_output in enumerate(generated):\r\n print(\"{}: {}\".format(i, tokenizer.decode(sample_output, skip_special_tokens=True)))\r\n```\r\n?", "You can also make use of `tokenizer.batch_decode(...)`" ]

[]

[]

[]

[ { "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 } ]

[ "Hi! Right now the sigmoid function is applied when the pipeline detects that there is a single label. You would like the option to disable the sigmoid function in that case?", "@LysandreJik \r\nSorry for my late reply.\r\nYes, cuz when people are doing regression tasks using the single-label SequenceClassification model, the output range depends on the specific task. For example, when predicting age from the text, [0,1] output after a sigmoid function is not a good fit here. ", "Indeed, I understand! I'm adding an option to return the raw outputs in #8328 ", "Thank you! I'm also wondering if this will be reflected by the Inference API? The inference API is using the sequence classification pipeline, therefore the API output on my model page is different from the original model output, which might confuse potential users. ", "@Jiaxin-Pei see discussion in #8328", "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" } ]

[]

[ "Add model card for pedropei/question-intimacy" ]

[]

[]

[]

[]

[]

[ "Ran all pipeline tests with @slow too to make sure:\r\n\r\n```\r\n==================================================================== warnings summary ====================================================================\r\n.venv/lib/python3.8/site-packages/tensorflow/python/autograph/utils/testing.py:21\r\n /home/nicolas/src/transformers/.venv/lib/python3.8/site-packages/tensorflow/python/autograph/utils/testing.py:21: DeprecationWarning: the imp module is \r\ndeprecated in favour of importlib; see the module's documentation for alternative uses\r\n import imp\r\n\r\ntests/test_pipelines_fill_mask.py::FillMaskPipelineTests::test_tf_fill_mask_results\r\n /home/nicolas/src/transformers/src/transformers/pipelines.py:1200: FutureWarning: The `topk` argument is deprecated and will be removed in a future vers\r\nion, use `top_k` instead.\r\n warnings.warn(\r\n\r\n-- Docs: https://docs.pytest.org/en/stable/warnings.html\r\n====================================================== 93 passed, 2 warnings in 2234.37s (0:37:14) =======================================================\r\n\r\n```" ]