Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "All tests are passing on my side and the CI so it's something in your environment that triggers the failures. Could you add what you have inside it?", "This?\r\n> the test should skip and not fail if the user doesn't have `Comet.ml`\r\n", "I don't have `comet.ml` installed in my dev env. I think the problem might be having it installed without the key properly set.", "Right, so this comes from `transformers`:\r\n\r\n```\r\nsrc/transformers/integrations.py: experiment = comet_ml.Experiment(**args)\r\nsrc/transformers/trainer_tf.py: experiment = comet_ml.Experiment(**args)\r\n```\r\nand it fails in:\r\n```\r\n if self.api_key is None:\r\n> raise ValueError(\r\n \"Comet.ml requires an API key. Please provide as the \"\r\n \"first argument to Experiment(api_key) or as an environment\"\r\n \" variable named COMET_API_KEY \"\r\n )\r\n```\r\nand it has a check:\r\n```\r\nif is_comet_available():\r\n import comet_ml\r\n```\r\nso the tests need to do something about it.\r\n\r\nI have never explicitly installed `comet_ml`, but some other package installed it as a dependency. Yet, I don't have this service configured. It's like `wandb` - it needs to gracefully skip tests if `comet_ml` is installed but not configured. To configure you have to go to that service website, sign up, create API key, etc. etc.\r\n\r\nI suppose there needs to be a skip decorator that does something like:\r\n```\r\nfrom .integrations import is_comet_available\r\ndef working_comet_ml(test_case):\r\n # handle the case where comel_ml is installed but not configured\r\n try:\r\n if is_comet_available:\r\n import comet_ml\r\n comet_ml.Experiment()\r\n except:\r\n return unittest.skip(\"test is slow\")(test_case)\r\n return test_case\r\n```\r\nthis is untested. \r\n\r\nMost likely you should be able to reproduce the problem by just installing `comet_ml`\r\n\r\nOr perhaps those 2 places where it's used in `transformers` it should issue a warning, rather then raise an error. Which is probably a better way to handle that.\r\n\r\n**edit**: thinking more, it is the library that shouldn't crash, and not the tests that should skip. It's an optional feature not a requirement, so it should behave gracefully.", "Seems to come from: https://github.com/huggingface/transformers/commit/b923871bb78f538e3c2e4bf36776986c800da1ae\r\nSo I guess we need to ask @dsblank, who made the initial PR and I see you were in the reviewers too.", "Yes I think raising a warning should be better than a hard error since it does not mean the training has to stop, and it blocks the tests. If you can suggest a PR with those changes, I'd happily review it.", "Found a better solution https://github.com/huggingface/transformers/pull/7830\r\n\r\n`comet_ml` needs to have `comet_ml.ensure_configured()` ala `wandb.ensure_configured()` - I am not sure whether ` @dsblank could add/ask for it. Thank you!\r\n\r\nUntil then it'll be emulated by:\r\n```\r\ntry:\r\n # Comet needs to be imported before any ML frameworks\r\n import comet_ml # noqa: F401\r\n\r\n # XXX: there should be comet_ml.ensure_configured(), like `wandb`, for now emulate it\r\n comet_ml.Experiment(project_name=\"ensure_configured\")\r\n _has_comet = True\r\nexcept (ImportError, ValueError):\r\n _has_comet = False\r\n```\r\n", "and wandb needed the latest version - which fixed the `OSError: [Errno 29] Illegal seek` issue" ]

[]

[]

[ "@LysandreJik done, thanks for pointing that out!" ]

[ { "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" } ]

[]

[]

[]

[]

[]

[]

[]

[ "Great!" ]

[]

[ { "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 } ]

[ "I was able to get passed this error by using my own RagRetriever instead of RagPyTorchDistributedRetriever inside \"transformers/examples/rag/finetune.py\"\r\n\r\nI am also using my own custom dataset and index #7763 \r\n\r\nThe following changes got me past the missing index error. However, I have no idea if this is efficient or if I am doing something that I shouldn't be doing...\r\n\r\n```\r\nif self.is_rag_model:\r\n if args.prefix is not None:\r\n config.generator.prefix = args.prefix\r\n config.label_smoothing = hparams.label_smoothing\r\n hparams, config.generator = set_extra_model_params(extra_model_params, hparams, config.generator)\r\n\r\n # commented out this line\r\n # retriever = RagPyTorchDistributedRetriever.from_pretrained(hparams.model_name_or_path, config=config)\r\n \r\n ############### new stuff ###############\r\n dataset = load_from_disk(args.passages_path) # to reload the dataset\r\n dataset.load_faiss_index(\"embeddings\", args.index_path) # to reload the index\r\n retriever = RagRetriever.from_pretrained(\r\n hparams.model_name_or_path, index_name=\"custom\", indexed_dataset=dataset\r\n )\r\n ######################################\r\n\r\n model = self.model_class.from_pretrained(hparams.model_name_or_path, config=config, retriever=retriever)\r\n prefix = config.question_encoder.prefix\r\n\r\n```", "Won't have the time in the next 1,2 weeks to take a closer look sadly. Maybe @lhoestq this is interesting to you", "Could you paste the full stacktrace ?", "Thank you @lhoestq .\r\n\r\n```\r\nGPU available: True, used: True\r\nINFO:lightning:GPU available: True, used: True\r\nTPU available: False, using: 0 TPU cores\r\nINFO:lightning:TPU available: False, using: 0 TPU cores\r\nLOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]\r\nINFO:lightning:LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]\r\nUsing native 16bit precision.\r\nINFO:lightning:Using native 16bit precision.\r\nValidation sanity check: 0it [00:00, ?it/s]Traceback (most recent call last):\r\n File \"examples/rag/finetune.py\", line 499, in <module>\r\n main(args)\r\n File \"examples/rag/finetune.py\", line 471, in main\r\n logger=logger,\r\n File \"/home/ioannis/Desktop/transformers-lhoestq-2/transformers/examples/lightning_base.py\", line 384, in generic_train\r\n trainer.fit(model)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py\", line 440, in fit\r\n results = self.accelerator_backend.train()\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py\", line 54, in train\r\n results = self.train_or_test()\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/accelerators/accelerator.py\", line 66, in train_or_test\r\n results = self.trainer.train()\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py\", line 462, in train\r\n self.run_sanity_check(self.get_model())\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py\", line 648, in run_sanity_check\r\n _, eval_results = self.run_evaluation(test_mode=False, max_batches=self.num_sanity_val_batches)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py\", line 568, in run_evaluation\r\n output = self.evaluation_loop.evaluation_step(test_mode, batch, batch_idx, dataloader_idx)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/trainer/evaluation_loop.py\", line 171, in evaluation_step\r\n output = self.trainer.accelerator_backend.validation_step(args)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py\", line 76, in validation_step\r\n output = self.__validation_step(args)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/pytorch_lightning/accelerators/gpu_accelerator.py\", line 86, in __validation_step\r\n output = self.trainer.model.validation_step(*args)\r\n File \"examples/rag/finetune.py\", line 240, in validation_step\r\n return self._generative_step(batch)\r\n File \"examples/rag/finetune.py\", line 280, in _generative_step\r\n max_length=self.target_lens[\"val\"],\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/torch/autograd/grad_mode.py\", line 15, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/transformers/modeling_rag.py\", line 873, in generate\r\n return_tensors=\"pt\",\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/transformers/retrieval_rag.py\", line 600, in __call__\r\n retrieved_doc_embeds, doc_ids, docs = self.retrieve(question_hidden_states, n_docs)\r\n File \"/home/ioannis/Desktop/transformers-lhoestq-2/transformers/examples/rag/distributed_retriever.py\", line 115, in retrieve\r\n doc_ids, retrieved_doc_embeds = self._main_retrieve(question_hidden_states, n_docs)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/transformers/retrieval_rag.py\", line 521, in _main_retrieve\r\n ids, vectors = self.index.get_top_docs(question_hidden_states, n_docs)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/transformers/retrieval_rag.py\", line 226, in get_top_docs\r\n _, ids = self.dataset.search_batch(\"embeddings\", question_hidden_states, n_docs)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/datasets/search.py\", line 607, in search_batch\r\n self._check_index_is_initialized(index_name)\r\n File \"/home/ioannis/anaconda3/envs/transformers-custom-dataset-in-rag-retriever/lib/python3.7/site-packages/datasets/search.py\", line 358, in _check_index_is_initialized\r\n f\"Index with index_name '{index_name}' not initialized yet. Please make sure that you call `add_faiss_index` or `add_elasticsearch_index` first.\"\r\ndatasets.search.MissingIndex: Index with index_name 'embeddings' not initialized yet. Please make sure that you call `add_faiss_index` or `add_elasticsearch_index` first.\r\n```\r\n", "> I was able to get passed this error by using my own RagRetriever instead of RagPyTorchDistributedRetriever inside \"transformers/examples/rag/finetune.py\"\r\n> \r\n> I am also using my own custom dataset and index #7763\r\n> \r\n> The following changes got me past the missing index error. However, I have no idea if this is efficient or if I am doing something that I shouldn't be doing...\r\n> \r\n> ```\r\n> if self.is_rag_model:\r\n> if args.prefix is not None:\r\n> config.generator.prefix = args.prefix\r\n> config.label_smoothing = hparams.label_smoothing\r\n> hparams, config.generator = set_extra_model_params(extra_model_params, hparams, config.generator)\r\n> \r\n> # commented out this line\r\n> # retriever = RagPyTorchDistributedRetriever.from_pretrained(hparams.model_name_or_path, config=config)\r\n> \r\n> ############### new stuff ###############\r\n> dataset = load_from_disk(args.passages_path) # to reload the dataset\r\n> dataset.load_faiss_index(\"embeddings\", args.index_path) # to reload the index\r\n> retriever = RagRetriever.from_pretrained(\r\n> hparams.model_name_or_path, index_name=\"custom\", indexed_dataset=dataset\r\n> )\r\n> ######################################\r\n> \r\n> model = self.model_class.from_pretrained(hparams.model_name_or_path, config=config, retriever=retriever)\r\n> prefix = config.question_encoder.prefix\r\n> ```\r\n\r\nThe above code seems to work (runs out of GPU memory in my local machine, so I am in the process of testing it on a server - will keep you posted).\r\n\r\nI noticed that the retrieval step 4 in _/examples/rag/use_own_knowledge_dataset.py_ takes a few minutes for every question, so I tried passing in _device=0_ to faiss to move it from cpu to gpu. I got this:\r\n\r\n`Faiss assertion 'blasStatus == CUBLAS_STATUS_SUCCESS' failed in virtual void faiss::gpu::StandardGpuResources::initializeForDevice(int) at gpu/StandardGpuResources.cpp:248`\r\n\r\nThe idea was to speed it up because I don't see how the finetuning can take place with such a slow index, but I might have misunderstood.", "Seems like my attempt to replace RagPyTorchDistributedRetriever with RagRetriever (in an 8 GPU machine) fails too. Too good to be true :)\r\n\r\n```\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/question_encoder_tokenizer/tokenizer_config.json from cache at /home/ubuntu/.cache/torch/transformers/8ade9cf561f8c0a47d1c3785e850c57414d776b3795e21bd01e58483399d2de4.11f57497ee659e26f830788489816dbcb678d91ae48c06c50c9dc0e4438ec05b\r\nModel name 'facebook/rag-sequence-base/generator_tokenizer' not found in model shortcut name list (facebook/bart-base, facebook/bart-large, facebook/bart-large-mnli, facebook/bart-large-cnn, facebook/bart-large-xsum, yjernite/bart_eli5). Assuming 'facebook/rag-sequence-base/generator_tokenizer' is a path, a model identifier, or url to a directory containing tokenizer files.\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/generator_tokenizer/vocab.json from cache at /home/ubuntu/.cache/torch/transformers/3b9637b6eab4a48cf2bc596e5992aebb74de6e32c9ee660a27366a63a8020557.6a4061e8fc00057d21d80413635a86fdcf55b6e7594ad9e25257d2f99a02f4be\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/generator_tokenizer/merges.txt from cache at /home/ubuntu/.cache/torch/transformers/b2a6adcb3b8a4c39e056d80a133951b99a56010158602cf85dee775936690c6a.70bec105b4158ed9a1747fea67a43f5dee97855c64d62b6ec3742f4cfdb5feda\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/generator_tokenizer/added_tokens.json from cache at None\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/generator_tokenizer/special_tokens_map.json from cache at /home/ubuntu/.cache/torch/transformers/342599872fb2f45f954699d3c67790c33b574cc552a4b433fedddc97e6a3c58e.6e217123a3ada61145de1f20b1443a1ec9aac93492a4bd1ce6a695935f0fd97a\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/generator_tokenizer/tokenizer_config.json from cache at /home/ubuntu/.cache/torch/transformers/e5f72dc4c0b1ba585d7afb7fa5e3e52ff0e1f101e49572e2caaf38fab070d4d6.d596a549211eb890d3bb341f3a03307b199bc2d5ed81b3451618cbcb04d1f1bc\r\nloading file https://s3.amazonaws.com/models.huggingface.co/bert/facebook/rag-sequence-base/generator_tokenizer/tokenizer.json from cache at None\r\nUsing native 16bit precision.\r\nINFO:lightning:Using native 16bit precision.\r\nINFO:__main__:Custom init_ddp_connection.\r\ninitializing ddp: GLOBAL_RANK: 7, MEMBER: 8/8\r\nINFO:lightning:initializing ddp: GLOBAL_RANK: 7, MEMBER: 8/8\r\nTraceback (most recent call last):\r\n File \"/home/ubuntu/transformers/examples/rag/finetune.py\", line 519, in <module>\r\n main(args)\r\n File \"/home/ubuntu/transformers/examples/rag/finetune.py\", line 491, in main\r\n logger=logger,\r\n File \"/home/ubuntu/transformers/examples/lightning_base.py\", line 384, in generic_train\r\n trainer.fit(model)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/trainer/states.py\", line 48, in wrapped_fn\r\n result = fn(self, *args, **kwargs)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py\", line 1046, in fit\r\n self.accelerator_backend.train(model)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/accelerators/ddp_backend.py\", line 57, in train\r\n self.ddp_train(process_idx=self.task_idx, mp_queue=None, model=model)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/accelerators/ddp_backend.py\", line 164, in ddp_train\r\n self.trainer.is_slurm_managing_tasks\r\n File \"/home/ubuntu/transformers/examples/rag/finetune.py\", line 180, in init_ddp_connection\r\n self.model.retriever.init_retrieval(self.distributed_port)\r\nTypeError: init_retrieval() takes 1 positional argument but 2 were given\r\nTraceback (most recent call last):\r\n File \"examples/rag/finetune.py\", line 519, in <module>\r\n main(args)\r\n File \"examples/rag/finetune.py\", line 491, in main\r\n logger=logger,\r\n File \"/home/ubuntu/transformers/examples/lightning_base.py\", line 384, in generic_train\r\n trainer.fit(model)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/trainer/states.py\", line 48, in wrapped_fn\r\n result = fn(self, *args, **kwargs)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/trainer/trainer.py\", line 1058, in fit\r\n results = self.accelerator_backend.spawn_ddp_children(model)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/accelerators/ddp_backend.py\", line 123, in spawn_ddp_children\r\n results = self.ddp_train(local_rank, mp_queue=None, model=model, is_master=True)\r\n File \"/home/ubuntu/anaconda3/envs/tran/lib/python3.7/site-packages/pytorch_lightning/accelerators/ddp_backend.py\", line 164, in ddp_train\r\n self.trainer.is_slurm_managing_tasks\r\n File \"examples/rag/finetune.py\", line 180, in init_ddp_connection\r\n self.model.retriever.init_retrieval(self.distributed_port)\r\nTypeError: init_retrieval() takes 1 positional argument but 2 were given\r\n```\r\n", "There are differences between the regular retriever and the distributed retriever:\r\n- the distributed retriever's index is only initialized on rank 0\r\n- the different processes communicate with rank 0 to do retrieval\r\n- the init_retrieval method requires to specify a port on which retrieval communication happen\r\n\r\nLet me know if you figure out a way to make it work in your case", "Will do, though I guess it's easier to go back to trying to make it work with _RagPyTorchDistributedRetriever_.\r\n\r\nTried adding _dataset.load_faiss_index_ inside get_dataset in finetune.py, but... 'Seq2SeqDataset' object has no attribute 'load_faiss_index'\r\n", "The Seq2SeqDataset is the one the model is trained *on*. The knowledge dataset is stored inside the retriever.\r\nThe `MissingIndex` error must come from init_retrieval not being called on the retriever in the process 0, or that the index is not properly loaded.", "Hi @lhoestq @patrickvonplaten any update on this? I'm also running into this issue when running finetune.sh. Though I am able to get the legacy index to work.", "@amogkam I also get the same error when trying to run fine-tuning. I also got an error saying self.opt is not there, but I did solve it. \r\n\r\nWhat do you mean by legacy index?", "I'll investigate this error this week. I'll let you know how it goes", "@lhoestq \r\n\r\nI actually did change the initialization in [this line (retrieval_rag.py)](https://github.com/huggingface/transformers/blob/master/src/transformers/retrieval_rag.py#L276).\r\n\r\nself.dataset_name, with_index=True,index_name=exact, split=self.dataset_split, dummy=self.use_dummy_dataset", "That's good to know thanks !\r\nHowever for the RagPyTorchDistributedRetriever we need to load the index only on the process 0 and keep `with_index=False` for the other processes. Ideally we have `with_index=False` in the `__init__` and `with_index=True` in `init_index`", "Oh get it!\n\nOn Tue, Nov 10, 2020, 04:56 Quentin Lhoest <[email protected]> wrote:\n\n> That's good to know thanks !\n> However for the RagPyTorchDistributedRetriever we need to load the index\n> only on the process 0 and keep with_index=False for the other processes.\n> Ideally we have with_index=False in the __init__ and with_index=True in\n> init_index\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/7816#issuecomment-724102348>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGQAIXISSTMTQW6L6KDSPAGKLANCNFSM4SSCPP5Q>\n> .\n>\n", "Sorry for spamming. I find it hard to understand index_name and index_paths\nwhen loading the datasets with fairsis\n\nOn Tue, Nov 10, 2020, 04:56 Quentin Lhoest <[email protected]> wrote:\n\n> That's good to know thanks !\n> However for the RagPyTorchDistributedRetriever we need to load the index\n> only on the process 0 and keep with_index=False for the other processes.\n> Ideally we have with_index=False in the __init__ and with_index=True in\n> init_index\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/7816#issuecomment-724102348>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGQAIXISSTMTQW6L6KDSPAGKLANCNFSM4SSCPP5Q>\n> .\n>\n", "You can specify index_name if you want to use one the index that comes with the dataset (exact/compressed), OR you can use index_path to use your own local index file.", "So the index name is like a column right ? Which controls whether thah\ncolumn should get loaded in to memory or not ?\n\nOn Wed, Nov 11, 2020, 02:35 Quentin Lhoest <[email protected]> wrote:\n\n> You can specify index_name if you want to use one the index that comes\n> with the dataset (exact/compressed), OR you can use index_path to use your\n> own local index file.\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/7816#issuecomment-724705036>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGUF5V5GO2KCEGIBH7TSPE6RHANCNFSM4SSCPP5Q>\n> .\n>\n", "In the RAG configuration you can specify index_name=\"exact\" or index_name=\"compressed\" for the \"wiki_dpr\" dataset. Wiki_dpr has indeed those two types of index. For more info you can check the docs of the [RagConfig](https://huggingface.co/transformers/model_doc/rag.html#ragconfig)\r\n\r\nOn the other hand in the datasets library and in particular in `Dataset.add_faiss_index` you can also see an \"index_name\" parameter. However this one is different from the one used in the RAG configuration on transformers side. In the datasets library, each dataset can have several indexes that are identified by their names, and by default their names correspond to the column that was used to build the index. See the docs of [the add_faiss_index method](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.add_faiss_index)\r\n\r\nThis is unfortunately the same variable name but not for the same purpose...\r\nDoes that make sense to you ?", "Thanks a lot. I got the idea. \r\n\r\n@lhoestq \r\n*Btw I tried to run the rag fine-tuning script\r\nwith a lower PyTorch lightning (0.9) version and it worked. I think the issue comes\r\nwith version miss-match.*\r\n\r\nOn Wed, Nov 11, 2020, 02:46 Quentin Lhoest <[email protected]> wrote:\r\n\r\n> In the RAG configuration you can specify index_name=\"exact\" or\r\n> index_name=\"compressed\" for the \"wiki_dpr\" dataset. Wiki_dpr has indeed\r\n> those two types of index. For more info you can check the docs of the\r\n> RagConfig\r\n> <https://huggingface.co/transformers/model_doc/rag.html#ragconfig>\r\n>\r\n> On the other hand in the datasets library and in particular in\r\n> Dataset.add_faiss_index you can also see an \"index_name\" parameter.\r\n> However this one is different from the one used in the RAG configuration on\r\n> transformers side. In the datasets library, each dataset can have several\r\n> indexes that are identified by their names, and by default their names\r\n> correspond to the column that was used to build the index. See the docs of the\r\n> add_faiss_index method\r\n> <https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.add_faiss_index>\r\n>\r\n> This is unfortunately the same variable name but not for the same\r\n> purpose...\r\n> Does that make sense to you ?\r\n>\r\n> —\r\n> You are receiving this because you commented.\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/huggingface/transformers/issues/7816#issuecomment-724711118>,\r\n> or unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGUB6XY3LU3EBGEV2D3SPE7Z3ANCNFSM4SSCPP5Q>\r\n> .\r\n>\r\n", "I managed to reproduce the issue, I'm working on a fix", "Perfect.\n\nOn Fri, Nov 13, 2020, 00:12 Quentin Lhoest <[email protected]> wrote:\n\n> I managed to reproduce the issue, I'm working on a fix\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/7816#issuecomment-726012277>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGUUQ6UV4KCAZGRLCWTSPO7LDANCNFSM4SSCPP5Q>\n> .\n>\n", "Update: it looks like it's because pytorch lightning removed the init_ddp_connection hook of their LightningModule.\r\nThe hook was used to initialize the index on process 0.\r\nI'll use something else to initialize the index.", "Ok, that's why the code still works with PL 0.9. \r\n\r\nSo now the problem is the initialization of the index in this [line](https://github.com/huggingface/transformers/blob/master/src/transformers/retrieval_rag.py#L274) ?\r\n\r\n\r\nThanks a lot.", "@lhoestq any update with this, please? \r\n\r\np.s sorry for spamming :) ", "Yes I'm working on a fix ! I'll make a PR tomorrow", "Thanks a lot. :) " ]

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

[]

[ "Might be of interest to @patrickvonplaten ", "Facing same issue here.", "Try this.\r\nYou can check other parameters for PyTorchBenchmarkArguments at [here](https://github.com/huggingface/transformers/blob/acf56408d81fdc04a32af139a45ae7b76e0c5b0d/src/transformers/benchmark/benchmark_args_utils.py#L34-L123).\r\n\r\n```python\r\n# main.py\r\ndef main():\r\n from transformers import PyTorchBenchmark, PyTorchBenchmarkArguments\r\n\r\n args = PyTorchBenchmarkArguments(models=[\"bert-base-uncased\"],\r\n batch_sizes=[8],\r\n sequence_lengths=[8, 32, 128, 512],\r\n multi_process=False)\r\n print(args.do_multi_processing)\r\n benchmark = PyTorchBenchmark(args)\r\n results = benchmark.run()\r\n print(results)\r\n\r\n\r\nif __name__ == '__main__':\r\n main()\r\n```\r\n```\r\nCUDA_VISIBLE_DEVICES=0 python main.py\r\n```", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "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." ]

[]

[ { "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 } ]

[]

[]

[]

[]

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

[]

[ "Maybe I not use the argument `--version_2_with_negative` while training on Squad V2.0 ?", "Indeed, not using that clarg would result in such bad results as you're not taking into account impossible answers. Can you let us know if you obtain better results with `--version_2_with_negative`?", "> Indeed, not using that clarg would result in such bad results as you're not taking into account impossible answers. Can you let us know if you obtain better results with `--version_2_with_negative`?\r\n\r\nAfter using `--version_2_with_negative`, the result is also not satisfactory.\r\nBoth exact and f1 are not exceed **80%**.\r\n\r\nThe results are as follows:\r\n{\"exact_20000\": 73.88191695443443, \"f1_20000\": 76.9165287314905, \"total_20000\": 11873, \"HasAns_exact_20000\": 68.60661268556005, \"HasAns_f1_20000\": 74.68453873633385, \"HasAns_total_20000\": 5928, \"NoAns_exact_20000\": 79.1421362489487, \"NoAns_f1_20000\": 79.1421362489487, \"NoAns_total_20000\": 5945, \"best_exact_20000\": 73.88191695443443, \"best_exact_thresh_20000\": 0.0, \"best_f1_20000\": 76.91652873149043,}\r\n\r\nAfter the 25000th steps of training, the performance got worse.\r\n", "That's already a lot better! What is this clarg `--squad_model bert_qa_model_squad`? It's not in the `run_squad.py` script?", "\r\n\r\n\r\n> That's already a lot better! What is this clarg `--squad_model bert_qa_model_squad`? It's not in the `run_squad.py` script?\r\n\r\nOh ,I have adjusted some code from `run_squad.py` so that it could load and train my custom model from clarg.\r\nAnd `bert_qa_model_squad` is the model whose code is copied from `BertForQuestionAnswering` in `transformers`.\r\nSo that I can design my custom model and then transmit it to the `run_squad.py` script.", "Does your script differ in any other way to the `run_squad.py` script? Does your model differ in any way to the `BertForQuestionAnswering` model in `transformers`? Have you tried running the standard command on the `transformers` script and model?\r\n\r\n```\r\npython run_squad.py\r\n--model_name_or_path bert-base-uncased\r\n--do_train\r\n--do_eval\r\n--do_lower_case\r\n--train_file $SQUAD_DIR/train-v2.0.json\r\n--predict_file $SQUAD_DIR/dev-v2.0.json\r\n--version_2_with_negative\r\n--per_gpu_train_batch_size 12\r\n--learning_rate 3e-5\r\n--num_train_epochs 2.0\r\n--max_seq_length 384\r\n--doc_stride 128\r\n--output_dir ./models/debug_squad/\r\n```", "> Does your script differ in any other way to the `run_squad.py` script? Does your model differ in any way to the `BertForQuestionAnswering` model in `transformers`? Have you tried running the standard command on the `transformers` script and model?\r\n> \r\n> ```\r\n> python run_squad.py\r\n> --model_name_or_path bert-base-uncased\r\n> --do_train\r\n> --do_eval\r\n> --do_lower_case\r\n> --train_file $SQUAD_DIR/train-v2.0.json\r\n> --predict_file $SQUAD_DIR/dev-v2.0.json\r\n> --version_2_with_negative\r\n> --per_gpu_train_batch_size 12\r\n> --learning_rate 3e-5\r\n> --num_train_epochs 2.0\r\n> --max_seq_length 384\r\n> --doc_stride 128\r\n> --output_dir ./models/debug_squad/\r\n> ```\r\n\r\nsorry, I will try later and tell you the results.", "> Does your script differ in any other way to the `run_squad.py` script? Does your model differ in any way to the `BertForQuestionAnswering` model in `transformers`? Have you tried running the standard command on the `transformers` script and model?\r\n> \r\n> ```\r\n> python run_squad.py\r\n> --model_name_or_path bert-base-uncased\r\n> --do_train\r\n> --do_eval\r\n> --do_lower_case\r\n> --train_file $SQUAD_DIR/train-v2.0.json\r\n> --predict_file $SQUAD_DIR/dev-v2.0.json\r\n> --version_2_with_negative\r\n> --per_gpu_train_batch_size 12\r\n> --learning_rate 3e-5\r\n> --num_train_epochs 2.0\r\n> --max_seq_length 384\r\n> --doc_stride 128\r\n> --output_dir ./models/debug_squad/\r\n> ```\r\nI have run the official `run_squad.py` script，and the result is as follows:\r\n10/16/2020 19:34:00 - INFO - main - Results: {'exact': 72.60170133917292, 'f1': 75.79599520268259, 'total': 11873, 'HasAns_exact': 72.2165991902834, 'HasAns_f1': 78.61434734167507, 'HasAns_total': 5928, 'NoAns_exact': 72.9857022708158, 'NoAns_f1': 72.9857022708158, 'NoAns_total': 5945, 'best_exact': 72.60170133917292, 'best_exact_thresh': 0.0, 'best_f1': 75.79599520268256, 'best_f1_thresh': 0.0}\r\n", "i have ran the official run_squad.py script, but got a 'cached_train_bert-base-uncased_384' not a result. Did i something wrong?", "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 @jsilter - sorry this fix was already pushed yesterday in #7903. Thanks for spotting the bug!!!" ]

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

[]

[ "Great, this looks like the correct fix! Do you want to open a PR?", "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" ]

[]

[]

[ "LGTM, thanks for aligning it! We just need some way to pass `eval_beams` and `max_gen_length`.\r\n\r\n>We can have a `Seq2SeqTrainingArguments` that subclasses `TrainingArguments` if that helps.\r\n\r\n@sgugger we do have `Seq2SeqTrainingArguments` class\r\nhttps://github.com/huggingface/transformers/blob/d99ed7ad618037ae878f0758157ed0764bd7f935/examples/seq2seq/finetune_trainer.py#L37", "> @sgugger we do have `Seq2SeqTrainingArguments` class\r\n\r\nAh, had forgotten about that :-)", "`eval_beams`/`eval_max_gen_length` reasoning:\r\n@patil-suraj said exactly this LOL, but in my words:\r\nusers are not good at modifying configs locally. We want to have a way to run `num_beams=2` during the generation step, but then end up with a trained model with the default # beams. In general, we try not to manipulate config attributes that would only be desired during training.\r\n", "Also would <3 an encoder decoder test in `examples/seq2seq/test_finetune_trainer.py`. \r\n", "After discussion @sshleifer - changed the `Seq2SeqTrainer` to be fully backwards compatible and to work with EncoderDecoder.\r\n@sshleifer - cannot add EncDec test yet because the complete command line setup is too constrained (requires `prepare_seq2seq_batch` to be defined for all tokenizers, etc...) => will see how to add this in the future. \r\n\r\n@sshleifer , @patil-suraj - could you do another review please? :-) ", "Should be good - I don't really see how -100 would slow down the TPU, but let's wait for @LysandreJik opinion here. ", "Can't seem to reply to the comment, but yes, the line @sshleifer is pointing at will slow down on TPU since it's probably using a `torch.where` behind the scene which does not have an XLA operation AFAIK.", "> Can't seem to reply to the comment, but yes, the line @sshleifer is pointing at will slow down on TPU since it's probably using a `torch.where` behind the scene which does not have an XLA operation AFAIK.\r\n\r\nOkey, I see -> let's move back in the old CE loss function then to keep backward compatibility! \r\n\r\n@sshleifer - one last review please :-) " ]

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

[ { "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 guess that that traceback is either masking OOM or IndexError. You will get a better traceback if you try on CPU." ]

[]

[]

[ "I solved the issue by myself", "> I solved the issue by myself\r\n\r\nHow?", "> I solved the issue by myself\r\n\r\n@h56cho Would you mind sharing it? Thanks" ]

[]

[]

[ "Could you fill-in the issue template for bugs so that we may help you?", "Please checkout now.", "you need to upgrade tokenizers.\r\nIt should happen for you if you run `pip install -e \".[dev]\"` from the root of the repo.", "Thanks, it worked.", "> Thanks, it worked.\r\n\r\ncan you tell me how did you solve it?thanks\r\n\r\n", "> > Thanks, it worked.\r\n> \r\n> can you tell me how did you solve it?thanks\r\n\r\nJust run the command which @sshleifer suggested from the directory having setup.py file. ", "Hello,\r\nWhich directory are you talking about ? The transformer site package directory, the project directory ?" ]

[]

[]

[ "This seems to be an error when installing `regex`, can you install it on its own? If you cannot, please open an issue on [their repo](https://bitbucket.org/mrabarnett/mrab-regex/issues?status=new&status=open).", "Thank you, I was not aware of that even I looked into trace twice. 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 } ]

[]

[ "Hi, I believe there is an error in your pytorch/xla install. I would follow the docs shared on [pytorch/xla](https://github.com/pytorch/xla) for best results.", "Actually, I am using Google cloud for running the training job. And it comes already with torch-xla-1.6 environment and the imagenet example works file. \r\n\r\nAs fas as I understand the problem is here https://github.com/pytorch/xla/blob/53eb3fc5e701cb8cc79a45dd0f79c6e585a27a41/torch_xla/core/xla_model.py#L18. \r\nI do not understand what is the problem though. ", "I don't really understand how you're launching your training. Our `xla_spawn.py` already takes care of parallelizing. I would do the following:\r\n\r\n```\r\nexport TPU_IP_ADDRESS=xxx.xxx.xxx.xxx\r\nexport XRT_TPU_CONFIG=\"tpu_worker;0;$TPU_IP_ADDRESS:8470\"\r\n\r\npython xla_spawn.py --num_cores 8 /home/developer/XainBERT/train.py\r\n```", "This works when I am using 8 core TPU. However, when I try to use TPUv3-128, then I get the error about cluster resolver. So I follow the instructions from here https://cloud.google.com/tpu/docs/tutorials/pytorch-pod#configure-gcloud. Also, the PyTorch lightning website also refers to the same page https://pytorch-lightning.readthedocs.io/en/latest/tpu.html. ", "Ah, I see. Unfortunately, I don't have sufficient experience with TPU pods to help you debug this. Can you try to open an issue on pytorch/xla's github?", "That's sad. Am I right, that Huggingface does not have TF implementation of Language models? I could try to switch to TF and try. Anyways, I have opened an issue on PyTorch/xla.", "No, we do have TF implementations of language models. Most of our models are available in both PyTorch and TensorFlow.", "However, this link says that there is not TFtrainer support for language-modelling using raw data.\r\nhttps://huggingface.co/transformers/examples.html \r\n\r\nIs there any example available for TFtrainer?", "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 } ]

[]

[ "@LysandreJik yeah I think you're right, thanks! made the change", "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." ]

[ { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null }, { "id": 1897896961, "node_id": "MDU6TGFiZWwxODk3ODk2OTYx", "url": "https://api.github.com/repos/huggingface/transformers/labels/Migration", "name": "Migration", "color": "e99695", "default": false, "description": "" } ]

[]

[ "Hello! It seems you're running the BART conversion script for a BERT model?", "Yeah, but I got to fix this.\n-Anirban\n\nOn Thu, 15 Oct, 2020, 2:52 PM Lysandre Debut, <[email protected]>\nwrote:\n\n> Hello! It seems you're running the BART conversion script for a BERT model?\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/7801#issuecomment-709032258>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AMTD2W4GWGTXTL5DFR3DHPDSK25O3ANCNFSM4SRMD3AQ>\n> .\n>\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Also pinging @patrickvonplaten for info", "Hey @QuantumEntangledAndy, I updated the config parameters of all DialoGPT models to `max_length=1000`, see here: https://github.com/huggingface/transformers/issues/7764 -> this problem should now be solved for the DialoGPT models.\r\n\r\nI think this is the correct way to tackle this problem - we don't want to change the `generation` code logic here just for Conversational models.", "I do not believe this will resolve the issue with min_length. Surly the expected behavior of min_length in a conversation model would be the min length of the new utterence not the length of context + new utterence.\r\n\r\nLets say I set min length to 2, because I want my bot to always say something and I say, \"What shall we watch?\" This has a length of 4 (five with with eos) this is added into the context and counts towards the length for min_length. The bot therefore has every chance of saying nothing.\r\n\r\nWhat would be the ideal way to deal with this situation?", "Sorry I misread your issue a bit here. \r\n> Surly the expected behavior of min_length in a conversation model would be the min length of the new utterence not the length of context + new utterence\r\n\r\nI understand your reasoning here. Are you using the pipelines or the \"normal\" generate() function?\r\n\r\nI think we could tweak the `ConversationPipeline` a bit to better handle the `min_length` parameter. \r\n\r\n Note that because the `input_ids` get longer with every conversation you have with the bot, `min_length` only works for the very first conversation in pipelines.\r\n\r\nIf you directly use `generate()` you could just set `min_length=input_ids.shape[-1] + 2` to solve your problem.\r\n", "I use the ConversationPipeline. I think perhaps the ConversationPipeline should update the min_length on call like this:\r\n\r\n```python\r\nmin_length=input_ids.shape[-1] + self.min_length_for_response\r\n```\r\n\r\nSince we do have such a member in the class although it is set to 32 which seems a little high\r\n\r\nhttps://github.com/huggingface/transformers/blob/0911b6bd86b39d55ddeae42fbecef75a1244ea85/src/transformers/pipelines.py#L2375-L2382\r\n\r\nHowever I think it is high as this is the member that decides how many old responses need to be cleared to make room for new input.", "If you want to test this out for yourself and see the blank responses this should work\r\n\r\n```python\r\n#! /usr/bin/env python3\r\n\"\"\"Small example of conversational pipeline in python.\"\"\"\r\n\r\nfrom transformers.pipelines import (\r\n Conversation,\r\n ConversationalPipeline,\r\n)\r\nfrom transformers import (\r\n AutoConfig,\r\n AutoModelForCausalLM,\r\n AutoTokenizer,\r\n)\r\n\r\ncache_dir = \"cached\"\r\nmodel_name_or_path = \"microsoft/DialoGPT-medium\"\r\nconfig_name = \"microsoft/DialoGPT-medium\"\r\ntokenizer_name = \"microsoft/DialoGPT-medium\"\r\n\r\nconfig = AutoConfig.from_pretrained(\r\n config_name, cache_dir=cache_dir,\r\n)\r\ntokenizer = AutoTokenizer.from_pretrained(\r\n tokenizer_name, cache_dir=cache_dir,\r\n)\r\nmodel = AutoModelForCausalLM.from_pretrained(\r\n model_name_or_path,\r\n from_tf=False,\r\n config=config,\r\n cache_dir=cache_dir,\r\n)\r\n\r\nconfig.min_length = 2\r\nconfig.max_length = 1000\r\n\r\nprint(f\"min_length: {config.min_length}\")\r\nprint(f\"max_length: {config.max_length}\")\r\n\r\nconversation = Conversation()\r\nconversation_manager = ConversationalPipeline(model=model,\r\n tokenizer=tokenizer)\r\n\r\nconversation.add_user_input(\"Is it an action movie?\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"Is it a love movie?\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"What is it about?\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"Would you recommend it?\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"If not what would you recommend?\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"I think you need to think about it more.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"After all action is the best.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"But maybe not.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"What really matters is quality.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"Quality over all other things.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"But not at the expense of tradition.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"For advancement for advancments sake must\"\r\n \" be curtailed.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"Unethical practises must be trimmed.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"In truth nothing is of any good.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"Unless it is traditional.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n\r\nconversation.add_user_input(\"And sometimes not even then.\")\r\nconversation_manager([conversation])\r\nprint(f\"Response: {conversation.generated_responses[-1]}\")\r\n```", "> I use the ConversationPipeline. I think perhaps the ConversationPipeline should update the min_length on call like this:\r\n> \r\n> ```python\r\n> min_length=input_ids.shape[-1] + self.min_length_for_response\r\n> ```\r\n> \r\n> Since we do have such a member in the class although it is set to 32 which seems a little high\r\n> \r\n> https://github.com/huggingface/transformers/blob/0911b6bd86b39d55ddeae42fbecef75a1244ea85/src/transformers/pipelines.py#L2375-L2382\r\n> \r\n> However I think it is high as this is the member that decides how many old responses need to be cleared to make room for new input.\r\n\r\nI think I would be fine to add this to the Conversation Pipeline. Do you want to open a PR and we see how to integrate it? ", "Before I do any PR Id like some input on design choices.\n\n- Should I also set max_length?\n \n - If I do set it too I believe it will no longer be necessary to remove old conversations to have room for new content.\n \n - However for very long chats perhaps a chat bot that saves and reloads it's memory this may become computational expensive\n\n - In light of this would a convenience function that trims memory down to n last inputs be acceptable?\n\n- I am thinking to make min_length an optional parameter to init. It defaults to None, when None is given as input it sets min_length to that of the model at init time. With similar behaviour for max_length", "Thank you @QuantumEntangledAndy for sharing the issue here, as I believe it affects both implementation.\r\n\r\nIf I may, I'd like to add my view on the issue, which I believe is not tied to the `ConversationPipeline`, but rather on how `min_length` and `max_length` and handled for non-\"encoder-decoder\" architectures.\r\n\r\nI would like to question the validity of setting the `cur_len` to the input sequence for pure decoder architectures:\r\nhttps://github.com/huggingface/transformers/blob/15a189049e01fbd3bef902848a09f58a5f006c37/src/transformers/generation_utils.py#L451.\r\n\r\nI would argue for setting the `cur_len` to 1 (or 0?) for pure-decoder architectures as well for a few reasons:\r\n- I would believe that user looking to generate a sequence would be generally looking to set how many tokens they would like to generate, and not the length of the context + generated inputs. It would be great if you could share some use-cases where that is typically not the case.\r\n- This definition of `cur_len` leads to somewhat \"hacky\" workarounds when the context needs to be extended by a prefix (for example, XLNet). Setting `min_length` and `max_length` to refer to generated content would make it independent of any context pre-processing\r\n- The current behaviour differs between \"encoder-decoder\" and \"decoders\", and I am not entirely sure why. A definition based on the length generated would bring the behaviour of both together.\r\n- Lastly, and in my opinion more importantly, the current solution does not work for batched generation. Let's say the input is made of 2 sentences, with initial length of 3 and 8. For batched generation, the input prompt will be padded as follows (`0` indicates a padded token, `x` an input token):\r\n```\r\n[ x x x x x x x x]\r\n[ 0 0 0 0 0 x x x] \r\n```\r\nthe `cur_len` will be set to 8 (`input_ids.shape[-1]`). Let's assume the `min_len` is 12. The model could generate the following sequence (`g` indicates a generated token):\r\n```\r\n[ x x x x x x x x g g g g]\r\n[ 0 0 0 0 0 x x x g g g g] \r\n```\r\nThis shows that while the first sequence respects the `min_len` of 12, the effective length of the second sequence is below the minimum value. Using `min_length` and `max_length` to refer to generated content would lead to valid constraints on all of the sequences in the batch, regardless of padding. For example, with the previous example if `min_length` is 6, both sequence would have at a minimum 6 generated token. In summary, I believe the current handling of `min_length` and `max_length` make them misleading as soon as inputs are passed as batches - but working on generated sequences would prevent that.\r\n\r\nI would be more in favor of initializing`cur_len` to 1 (0) for decoders as well - I would be interested in your thoughts on that.", "Hmmm I think I see what you mean with batch generation.\n\nI am wondering what the use case of treating non encoder-decoders differently is. If for some reason we cannot reasonably change the definition perhaps we should consider adding different types of new lengths.\n\n- minmax document length\n - This is current definition of min/max\n- minmax generated length\n - This is the effective length the model generates.\n - This of course would mean new config variables. But we could default them to none and then ignore that condition in this case.", "Hey @guillaume-be,\r\n\r\nI fully understand your point of view and I also tend to agree that the handling of `min_length` and `max_length` as described by you and I tend to agree with you (actually @sshleifer suggested this change a while back as well).\r\n\r\nI guess the disadvantages of changing `max_length's` logic (making it \"max added tokens\" vs. \"max total tokens\") is the following:\r\n\r\n- For models like GPT2, users might want to generate multiple articles < 100 words including the prefix. The output of decoder models is always prefix + generated text so it makes sense to me that `max_length` is the maximum length of the output. It's also safer regarding the `max_embedding_positions` provided by the model - by setting `max_length=512`, there will never be an error independent of the prefix. I guess all arguments are based on the advantages one might have from knowing that `max_length` is independent of the input.\r\n\r\nThat's the main argument.\r\n- Changing the logic now would break backward compatibility quite heavily. Beam search is highly dependent on max_length *e.g.* - not sure whether we want that.\r\n\r\nLet's put @LysandreJik and @yjernite and @sshleifer in cc as well to see their opinions on changing the `max_length` logic.", "We could make it an option like `use_relative` lengths. It defaults to `None` meaning auto (the current behavour) but can be set to `True` of `False` in order to explicitly override the current auto logic. Ultimately I think the auto logic is not perfect and there will always be situations where it choices incorrectly, making it configurable will at least allow the library user to choose the appropiate choice.", "My opinion is that I have never seen \"users might want to generate multiple articles < 100 words including the prefix\", and I have seen, \"Why does prefix count?\" a number of times and had the same misunderstanding the first few times I used it. So I think we should change the behavior to not count `decoder_start_token_id` or tokens in `input_ids`. \r\nAnother argument that might land: you would expect a function called `generate` to `generate` N tokens. Not generate N tokens where N varies based on your inputs.\r\n\r\nThe `use_relative` compromise is my second favorite option.\r\nThe status quo is my least favorite.", "Was there any progress on making these changes?", "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." ]

[ { "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" } ]

[]

[ "Great point! I've already simplified the code and left only the tokenizer. Ideally, `XLMTokenizer` should have a converter to the appropriate `Fast` class. On second thought, I can see the problem would be to program the Moses pretokenization in the `tokenizers` library.", "Hi @rmroczkowski \r\n\r\nAs far as I know, you employed fastBPE to train a tokenizer: https://github.com/huggingface/transformers/tree/b23d3a5ad4aa08decd10671f85be5950767dd052/model_cards/allegro/herbert-klej-cased-tokenizer-v1\r\n\r\nI also employed fastBPE for the Vietnamese BERT-based tokenizer (i.e. PhoBERTTokenizer https://github.com/huggingface/transformers/pull/6129 ), but I am still struggling to implement a fast tokenizer based on fastBPE, e.g. handling the suffix \"@@\" of subword tokens. In particular, given https://huggingface.co/vinai/phobert-base/tree/main I can convert \"bpe.codes\" into a \"merge.txt\"-style file, but I am not sure about how to convert our \"vocab.txt\" into your \"vocab.json\"-style file.\r\n\r\nHow can you convert your fastBPE's code and vocab outputs into HuggingFace's tokenizers? So that you can call the tokenizer with the use_fast=True option. \r\n\r\ncc: @LysandreJik is there any idea for implementing a fast version of a fastBPE-based slow one?\r\n\r\nThank you both.\r\n\r\n\r\n\r\n" ]

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

[]

[ "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, I'm still getting this warning on version 4.3.2." ]

[]

[]

[ "`some other parameter to change?`: BINGO\r\n\r\n\r\nthere is a `min_length`/`max_length` parameter you can pass to beam search (in many ways) that is affecting your generations.\r\nIf you eval offline with min_length=0, max_length=3 it should work.\r\n", "Cool! Sorry for the n00biness. \r\n1. Is there somewhere I can read about when / why this happens? (or in brief, why does it happen?) \r\n2. min_length and max_length will just limit how long the output sequence can be? Where's the best place to input them? Just directly from finetune.py?\r\n3. Is there a different way to have the model learn when to stop outputting? (i.e to learn by itself that it should only be outputting one \"word\" since that's what all the train examples show)", "1) you can read the [docstring for `generate`](https://github.com/huggingface/transformers/blob/master/src/transformers/generation_utils.py#L135)\r\n2) I would edit `finetune.py ` around [here](https://github.com/huggingface/transformers/blob/master/examples/seq2seq/finetune.py#L215)\r\n3) It should learn good lengths within the hardcoded range. It's simply not allowed to go out of the hardcoded range.\r\nIf you set `min_length=0`, `max_length=10` I would guess it will learn to always generate word followed by `</s>` (This \"eos\" symbol is automatically added to input sequences by the `T5Tokenizer`.)\r\n\r\n\r\n\r\n\r\n", "Thanks! I am rerunning with the max length (I didn't see a spot for min length). \r\n\r\nI'm still a little confused as to why this happens though. For example, \r\n* why doesn't it get penalized for the gibberish? (is padding somehow affecting what it gets penalized for?)\r\n* why isn't the gibberish at all linguistic, even? I would expect it at least to add mostly english-like tokens? These strings seem entirely non-lingustic.\r\n\r\nRelated: is there an easy flag to change so that I could view part of the validation outputs at each epoch to keep track of when it learns to truncate? Right now I'm just waiting until end of training to look at the test generations.\r\n", "+ You need the min_length, just pass min_length=0 to `model.generate`\r\n+ re padding, yes. There is no loss for pad tokens.\r\n+ no flag to see intermediate generations, but https://github.com/huggingface/transformers/blob/master/examples/seq2seq/callbacks.py#L83 should maybe work.", "Okay thanks, I will work on these.\r\n\r\nI realize these are unrelated T5 issues, but before I file other feature requests /bugs I just wanted to run them by you:\r\n* auto_lr_find and auto_scale_batch_size (pytorch lightning flags) when used from the finetune.sh script throw errors. Should these be usable? (I can debug and figure out why they're not working; but I want to know if they should be working)\r\n* I am unable to get the finetune.sh script to resume from a checkpoint (I played around with this for ~2 hours last night) and was unable to make it resume. Should this be supported?", "auto*: Would be nice if they worked!\r\nit should work with `--resume_from_checkpoint`, but that part of lightning has been very flaky.\r\n\r\n\r\nI probably won't fix either of these but would definitely accept a PR that allow clargs that currently don't work. If you can't fix, you could also make separate issues for clargs that don't work, label them \"Help Wanted\" and see what happens.\r\nIf you make issues, make sure to include your PL version.\r\n\r\n", "@jsrozner did you `finetune.py` work for fine-tuning T5? \r\n\r\nWe're also having [some difficulties](https://discuss.huggingface.co/t/issue-with-finetuning-a-seq-to-seq-model/1680/2). Wanted to make sure if it has worked for someone else, at least. ", "@danyaljj will be fixed by #8435", "Thanks, @jsrozner for the update! \r\nDoes this address the issue [here](https://discuss.huggingface.co/t/issue-with-finetuning-a-seq-to-seq-model/1680/27?u=danyaljj)? Mainly your observation that: \r\n\r\n> But even after setting eval_beams=1, eval_max_gen_length=40, it still continues to generate many more tokens than it should ", "Did you pass `min_length=0` to generate?", "See issue #5142 for resolution" ]

[]

[]

[]

[]

[]

[ "Why not `>= 0.9.1` or `>=0.8.1.rc2`?", "Hi, we have a strict requirement on `tokenizers==0.8.1rc2`. We're updating it in https://github.com/huggingface/transformers/pull/7659 but the current `transformers` `master` branch will stay pinned until that PR is merged.\r\n\r\nBoth libraries evolve quickly and generally evolve together, so having a strict `==` dependency is necessary until tokenizers version 1.0.0 is released." ]

[]

[]

[]

[]

[]

[ "Here is part of the problem\r\n\r\n\r\n\r\nall one line (221 in vim)\r\n", "Oh, I know - it's \\cM characters. Let me take care of it.\r\n\r\n```\r\nGoogle is still the best company to work for, according to Fortune\r\n.^M<n>^M<n>The Mountain View-based tech giant earned the top \r\n^^^^^^^^^^^^^\r\n```", "Easiest way to clarify will be to say how I fixed (in vim)\r\n\r\n(1) `%s/^M//g` # This is not ctrl-m it must be typed by following these [instructions](https://stackoverflow.com/questions/5843495/what-does-m-character-mean-in-vim#:~:text=Windows%20uses%20a%20combination%20of,letter%20in%20the%20English%20alphabet).&text=Where%20%5EM%20is%20entered%20by,m%20%2C%20and%20then%20releasing%20Ctrl%20.)\r\n \r\n(2)`%s/<n><n>/<n>/g` (probably not necesarry, but did it anyway).\r\n", "```\r\ndos2unix filename\r\n```", "I will fix that in the build script", "```\r\n src = re.sub(r'[\\r\\n]+', '<n>', src)\r\n tgt = re.sub(r'[\\r\\n]+', '<n>', tgt)\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 } ]

[ "Hey @agemagician - did you train your model using the \"newer\" T5 model (see here https://github.com/huggingface/transformers/issues/6285) for reference or is it the \"original\" T5 model?", "No, this is the original T5 model.\r\n\r\nI just doubled checked the training script as well as the operative_config :\r\nhttps://storage.googleapis.com/t5_convert_tranformers/model/operative_config.gin ", "Ok! From a first check of your google colab it looks like the model was correctly converted to PT (the `\"Weights not copied to PyTorch model:` message is empty meaning that all PT weights are initialiazed). \r\n\r\nDo you think you could check if it might be the tokenizer that does not work correctly? Could you maybe run an integration test for some `input_ids` to check if original t5 implementation yields same output as the PT version?", "I have loaded the original T5 tokenizer then encoded the data and performed generation using Pytorch to make sure the input is the same for both original T5 script and Pytorch script, and the results is still rubbish.\r\n\r\nI have checked the original T5 tokenizer and Pytorch tokenizer and they produce the same encoding/decoding. The only difference is that Pytorch tokenizer doesn't append Eos.\r\n\r\nI have added a new section on the Colab \"Part IIII: Check tokenizers\" which perform these tests.", "Since the input is the same to both original T5 script and Pytorch script, I think the issue should be in one of the following:\r\n1. The conversion process.\r\n2. The generation process.\r\n3. The loading process.", "Thanks, I hope to be able to take a look at this soon!", "@patrickvonplaten Any update for fixing this issue ?\r\n\r\nWe started to release our models for the following tasks:\r\n\r\n1. api generation\r\n2. code comment generation\r\n3. commit generation\r\n4. function documentation generation\r\n5. program synthesis\r\n6. source code summarization\r\n7. Code generation\r\n\r\nfor the following languages:\r\n\r\n1. go\r\n2. java\r\n3. javascript\r\n4. php\r\n5. python\r\n6. ruby\r\n7. c#\r\n8. SQL\r\n9. LISP\r\n\r\nhttps://github.com/agemagician/CodeTrans\r\n\r\nHowever, we are using T5 original library for now, as huggingface transformers is still producing rubbish text after conversion.\r\n\r\nIt will be really useful if we can integrate and use huggingface transformers for this project too.\r\n", "Will take a look today!", "@agemagician - I looked into it. It's quite nightmarish to debug in mesh tensorflow ... :-/ I couldn't find the bug sadly and it's getting very time-consuming. I'll gonna spend some time now to integrate mt5 and T5v1.1, so I'll still be working with the mesh tensorflow library. I hope to be able to come back to this problem! A couple of things I found out:\r\n\r\n1) The `input_ids` passed to the Encoder for \r\n```\r\n\"Code: function isStandardBrowserEnv ( ) { if ( typeof navigator !== 'undefined' && ( navigator . product === 'ReactNative' || navigator . product === 'NativeScript' || navigator . product === 'NS' ) ) { return false ; } return ( typeof window !== 'undefined' && typeof document !== 'undefined' ) ; }\r\nDocumentation: Returns true if the browser is a native element .\"\r\n```\r\nis actually not the same for Hugging Face T5 and Mesh TF T5. => I suspect the tokenizers to behave differently here or mesh tf to do something under the hood with the input text\r\n\r\n2) Sadly even if I pass the exact same `input_ids` to the encoder of both models, the encoder outputs are still different => this means that there is a different in the architecture. I suspect that mesh TensorFlow handles the `relative_attention_bias` different for the `EncoderDecoderSelfAttention`. In the mesh tensorflow's gin it's set no `None`, but in our code its definitely used. Did not manage to check it here in more detail. \r\n\r\n=> Overall the problem is that `mesh_tensorflow` is constantly adding new features that are configurable with the gin config, but some of these new features are not implemented in HF and are therefore not used. So what is probably happening is that a mesh tensorflow trained model has the exact same weights as the HF implementation but has a slightly different architecture that cannot be configured with the HF T5 model...it's very hard for us to make sure that mesh tensorflow is kept constantly compatible with HF and we probably won't have the time to make sure it is. The only real solution is to use a HF pre-trained and train it within our environment or make sure that before mesh tensorflow training that the model is compatible with HF (checking the output of the pretrained models).\r\n\r\nIn case you want to take a deeper look here are my simplified scripts I used for debugging:\r\n\r\nfor mesh tf model:\r\n\r\n```python\r\nimport t5\r\nfrom t5.data.sentencepiece_vocabulary import SentencePieceVocabulary\r\n\r\nt5_model = t5.models.MtfModel(\r\n model_dir=\"./checkpoint\",\r\n batch_size=16,\r\n sequence_length={\"inputs\": 128, \"targets\": 32},\r\n learning_rate_schedule=0.003,\r\n save_checkpoints_steps=5000,\r\n keep_checkpoint_max=None,\r\n iterations_per_loop=100,\r\n tpu=None\r\n)\r\n\r\nvocab_model_path = 'gs://t5_convert_tranformers/spm/code_spm_unigram_40M.model'\r\nvocab = SentencePieceVocabulary(vocab_model_path, extra_ids=100)\r\n\r\nt5_model.predict(\r\n input_file=\"input.txt\",\r\n output_file=\"output.txt\",\r\n vocabulary=vocab,\r\n temperature=0\r\n)\r\n```\r\n\r\nand HF:\r\n\r\n```python\r\nfrom transformers import T5ForConditionalGeneration, T5Tokenizer\r\nimport torch\r\n\r\ninput_text = \"javascript documentation generation: function isStandardBrowserEnv ( ) { if ( typeof navigator !== 'undefined' && ( navigator . product === 'ReactNative' || navigator . product === 'NativeScript' || navigator . product === 'NS' ) ) { return false ; } return ( typeof window !== 'undefined' && typeof document !== 'undefined' ) ; }\"\r\n\r\nmodel = T5ForConditionalGeneration.from_pretrained(\"./pytorch_model\").to(\"cuda\")\r\ntok = T5Tokenizer.from_pretrained(\"./pytorch_model\")\r\n\r\n#input_ids = tok(input_text, return_tensors=\"pt\").input_ids.to(\"cuda\")\r\ninput_ids = torch.tensor([[69, 8316, 3952, 12059, 171, 69, 34, 11451, 7798,\r\n 6614, 5, 6, 12, 29, 5, 644, 16747, 494,\r\n 20, 3910, 36, 129, 5, 16747, 4, 1668, 232,\r\n 20, 23435, 6462, 36, 194, 16747, 4, 1668, 232,\r\n 20, 6462, 2769, 36, 194, 16747, 4, 1668, 232,\r\n 20, 4759, 36, 6, 6, 12, 30, 181, 9,\r\n 16, 30, 5, 644, 1066, 494, 20, 3910, 36,\r\n 129, 644, 722, 494, 20, 3910, 36, 6, 9,\r\n 16, 1]], dtype=torch.long, device=\"cuda\")\r\n\r\n\r\noutput = model.generate(input_ids, num_beams=4)\r\n\r\nprint(tok.batch_decode(output))\r\n```\r\n\r\nThen my folders had the following files (same as in your notebook). \r\n```bash\r\nls checkpoint\r\ncheckpoint code_spm_unigram_40M.model graph.pbtxt model.ckpt-16000.data-00000-of-00002 model.ckpt-16000.data-00001-of-00002 model.ckpt-16000.index model.ckpt-16000.meta operative_config.gin\r\n```\r\nand\r\n```bash\r\nls pytorch_model\r\nconfig.json pytorch_model.bin special_tokens_map.json spiece.model tokenizer_config.json\r\n```\r\n\r\nwith all the pytorch models converted from the mesh tf spm and mesh tf checkpoint (as you've done in the colab).\r\n\r\n\r\nAnd then one has to put a lot of `mtf.print(x, [x], \"output: \", summarize=-1)` statements in the mesh tensorflow code - here e.g.: https://github.com/tensorflow/mesh/blob/165d3dc7b4186ee5b6d31c9b17b3df4f7571cf42/mesh_tensorflow/transformer/transformer_layers.py#L729, but that's very painful ;-) \r\n\r\nAlso, see here for debugging advice: https://github.com/tensorflow/mesh/issues/235\r\n\r\nMaybe by some miracle I find the problem over the next two weeks while further looking into mesh tensorflow.\r\n\r\nSorry, to be not too much of help here.\r\n\r\n", "Hi @patrickvonplaten ,\r\n\r\nThanks a lot for looking into this issue.\r\nWe highly appreciate your effort and sorry if it wasted your time.\r\n\r\nI have also tested our protein model \"prot_t5_xl_bfd\" for protein sequence generation and it has the same issue. Also our next 11B model for protein sequences \"prot_t5_xxl_bfd\" will have the same issue.\r\nThis means the current results that we have from all our T5 models are not correct.\r\n\r\nDo you know if this issue exist in only the decoder or both the encoder and the decoder ?\r\nbecause currently we are only using the encoder on \"prot_t5_xl_bfd\" for feature extraction.\r\n\r\nI have also checked MT5 and T5v1.1 and they seem to have the same issue as our current models, so if you will work on T5v1.1, you will highly likely find the issue and the solution for path ProtTrans models and ProtCode models.\r\n\r\nThanks again for your time, and I will leave this issue open, until you finish T5v1.1 implementation. ", "It's both encoder and decoder. Even the same encoder input yielded a different encoder output", "This is really bad for the ProtTrans project.\r\nThanks a lot Patrick for your clear reply.\r\nI will try to debug it from my side, and I will update you if I found the issue.", " I got T5v1.1 working now I think: https://github.com/huggingface/transformers/pull/8488. But this code will certainly not work with your example since the Feed-Forward layer has different weights...\r\n\r\nLet me take a look again at this issue in a bit. Could you maybe provide me with a code example where I just need to download 1) of your pretrained checkpoints\r\n2) run a code snippet of the following format:\r\n\r\n```python\r\n#!/usr/bin/env python3\r\nimport os\r\n\r\nos.environ[\"TF_CPP_MIN_LOG_LEVEL\"] = \"3\" # or any {'0', '1', '2'}\r\n\r\nimport t5 # noqa: E402\r\nfrom t5.data.sentencepiece_vocabulary import SentencePieceVocabulary # noqa: E402\r\nfrom transformers import T5Tokenizer # noqa: E402\r\nfrom transformers.convert_t5_v1_1_original_tf_checkpoint_to_pytorch import ( # noqa: E402\r\n convert_tf_checkpoint_to_pytorch,\r\n)\r\nfrom transformers.modeling_t5v2 import T5Config, T5v2ForConditionalGeneration # noqa: E402\r\n\r\n\r\npath_to_tf_checkpoint = \"/home/patrick/hugging_face/t5v1.1/t5_mesh_checkpoints\"\r\n\r\n\r\ntok = T5Tokenizer.from_pretrained(\"t5-small\")\r\ntok.save_pretrained(path_to_tf_checkpoint)\r\nconfig = T5Config.from_pretrained(\"t5-small\")\r\nconfig.d_ff = 1024\r\nconfig.num_decoder_layers = 8\r\nconfig.num_layers = 8\r\nconfig.num_heads = 6\r\n\r\nconfig.save_pretrained(path_to_tf_checkpoint)\r\n\r\nconvert_tf_checkpoint_to_pytorch(path_to_tf_checkpoint, path_to_tf_checkpoint + \"/config.json\", path_to_tf_checkpoint)\r\n\r\nt5_model = t5.models.MtfModel(\r\n model_dir=path_to_tf_checkpoint,\r\n batch_size=1,\r\n tpu=None,\r\n sequence_length={\"inputs\": 4, \"targets\": 4},\r\n)\r\n\r\nvocab_model_path = path_to_tf_checkpoint + \"/sentencepiece.model\"\r\nvocab = SentencePieceVocabulary(vocab_model_path, extra_ids=100)\r\n\r\nscore = t5_model.score(\r\n inputs=[\"Hello there\"],\r\n targets=[\"Hi I am\"],\r\n vocabulary=vocab,\r\n)\r\n\r\nmodel = T5v2ForConditionalGeneration.from_pretrained(path_to_tf_checkpoint, return_dict=True)\r\n\r\ninput_ids = tok(\"Hello there\", return_tensors=\"pt\").input_ids\r\nlabels = tok(\"Hi I am\", return_tensors=\"pt\").input_ids\r\n\r\n# input_ids and labels are ok!\r\nloss = model(input_ids, labels=labels).loss\r\n\r\nassert -(labels.shape[-1] * loss.item()) - score[0][0] < 1e-4\r\n```\r\n\r\nIf all the code would be in one file -> this would really help me save time in debugging. Otherwise, maybe we can have a quick call early next week (Monday maybe?) to discuss how to best tackle the error. I got a bit lost in all the colab notebook. I'm sure it's not that hard to fix actually.", "Great @patrickvonplaten , \"du bist der Beste\" :\r\n\r\nI have created a Colab that runs your code and download one of the CodeTrans models:\r\nhttps://colab.research.google.com/drive/149F64wSOjm5O-HdLWpdWJE4dAMUA-Waa?usp=sharing\r\n\r\nImportant notes:\r\n1. This model is using the original T5 model not v1.1. ie (word embedding is tied, uses dropout, uses RELU)\r\n2. It is the base model.\r\n\r\nLet me know if anything else is required.", "should be fixed now. Everything is explained in the PR.", "Woohoo, thanks a lot @patrickvonplaten, you are the best 😄 " ]

[ { "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": "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 } ]

[ "\r\nI think the doc should recommend\r\n```python\r\nAdafactor(model.parameters(), relative_step=True, warmup_init=True, lr=None)\r\n```\r\nwant to fix it?", "I think what corresponds to the original T5 training code is `Adafactor(model.parameters(), lr=1e-3, relative_step=False, warmup_init=False)`, however that didn't work great for me so far (much slower than Adam, and giving me NaN's even in FP32).", "Hello @OyvindTafjord, have you been able to fine-tune T5 with Adafactor? Thanks, Sonali", "No, I haven't investigated further regarding the slowness and NaN's I was getting.", "This issue persists (i.e. the suggested defaults still produce the error).\r\n\r\nI can confirm that `Adafactor(lr=1e-3, relative_step=False, warmup_init=False)` seems to break training (i.e. I observe no learning over 4 epochs, whereas `Adafactor(model.parameters(), relative_step=True, warmup_init=True, lr=None)` works well (much better than adam)" ]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Sorry it was my mistake, I need to get generated ids calling `model.generate` instead of `model`.\r\n\r\nAdding fix here if anyone search for the same issue -\r\n\r\n```\r\ngenerated_ids = model.generate(input_ids=input_ids, labels=input_dict[\"labels\"])\r\n\r\ngenerated_string = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)\r\nprint(generated_string)\r\n``` \r\n" ]

[]

[]

[ "The failed tests are not from this pull request but from RAG. Can someone review this, please?\r\n@patrickvonplaten , @TevenLeScao", "Hey @nicola-decao - thanks a lot for your PR here! The failing tests are actually due to this pull request. See a part of the error message here:\r\n\r\n```\r\n )\r\nE TypeError: _generate_beam_search() missing 1 required positional argument: 'prefix_allowed_tokens_fn'\r\n\r\nsrc/transformers/modeling_rag.py:1400: TypeError\r\n_______________________ RagDPRT5Test.test_model_generate _______________________\r\n[gw6] linux -- Python 3.7.9 /usr/local/bin/python\r\n\r\nself = <tests.test_modeling_rag.RagDPRT5Test testMethod=test_model_generate>\r\n\r\n def test_model_generate(self):\r\n inputs_dict = self.config_and_inputs\r\n> self.check_model_generate(**inputs_dict)\r\n```\r\n\r\nCould you fix the errors by slightly adapting the `generate` method in RAG to make it pass with your example? \r\n\r\nIn general I'm fine with this PR :-) ", "@patrickvonplaten now It should be ready for the merge :)", "Awesome! There is probably going to be a merge conflict with the big `generate()` refactor PR that will be merged today: https://github.com/huggingface/transformers/pull/6949 . \r\n\r\nWe changed the design for these kinds of \"logits processing\" methods so that we'll probalby have to change the PR here a bit (and should also add a test). But let's put the PR on hold for a day and then I can help you merge it! ", "@nicola-decao, also what would be a typical use case of this function? *E.g.* could you give a quick example of a function that one would use as this function? \r\n\r\nAlso @sshleifer could you check if this is useful for Fairseq's Blenderbot? And @yjernite this might be interesting to you as well :-) ", "@patrickvonplaten we do have a few use cases in mind internally already :) \r\n- prefix-triggered multi task *a la* T5: in many cases having the prefix in the output sequence makes more sense than in the input\r\n- seq2seq model evaluation: some metrics (e.g. ROUGE-20 in the [ELI5 paper](https://arxiv.org/abs/1907.09190)) measure the model's ability to \"continue\" a generation, which can correlate better with human judgments of quality that full generation ROUGE\r\n- seq2seq diagnostics: being able to measure the the effect of the input vs local context\r\n\r\n@nicola-decao did you have something along those lines in mind?", "> @nicola-decao, also what would be a typical use case of this function? _E.g._ could you give a quick example of a function that one would use as this function?\r\n> \r\n> Also @sshleifer could you check if this is useful for Fairseq's Blenderbot? And @yjernite this might be interesting to you as well :-)\r\n\r\n@patrickvonplaten an example would be using a seq2seq model to predict a wikipedia title as in **Autoregressive Entity Retrieval** (https://arxiv.org/abs/2010.00904) (I am the fist author and I want to release my models 😊 - that is why I did this PR). In this case the possible outputs are all the 6M wikipedia titles so one can create a prefix tree and constrain the generation only on these 6M strings. Here an example of what I have:\r\n```python\r\n# custom class that creates a prefix tree where only \"London\" and \"Rome\" are possible outputs\r\ntrie = PrefixTree([\"London\", \"Rome\"])\r\n\r\n# from a batch of tokens (torch.Tensor) returns a list of lists of allowed tokens (batches/ beams)\r\n# `trie.get` returns the possible next tokens (leaves if any) given a prefix\r\ndef prefix_allowed_tokens_fn(batch_tokens):\r\n return [\r\n [\r\n trie.get(tokens.tolist())\r\n for tokens in beam_tokens\r\n ]\r\n for beam_tokens in batch_tokens\r\n ]\r\n\r\n# encoding inputs\r\ninput_args = {\r\n k: v.to(model.device) for k, v in tokenizer.batch_encode_plus(\r\n [\"[START_ENT] London [END_ENT] is the capital of the UK.\"],\r\n return_tensors=\"pt\"\r\n ).items()\r\n}\r\n\r\n# generating and decoding\r\ntokenizer.batch_decode(\r\n model.generate(\r\n **input_args,\r\n min_length=0,\r\n num_beams=2,\r\n num_return_sequences=2,\r\n prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,\r\n ),\r\n skip_special_tokens=True\r\n)\r\n>> ['London', 'Rome']\r\n```\r\n\r\n", "> Awesome! There is probably going to be a merge conflict with the big `generate()` refactor PR that will be merged today: #6949 .\r\n> \r\n> We changed the design for these kinds of \"logits processing\" methods so that we'll probalby have to change the PR here a bit (and should also add a test). But let's put the PR on hold for a day and then I can help you merge it!\r\n\r\n@patrickvonplaten I guess one way to do it is to implement a `logits_processor`. But what I want to touch the log-probabilities directly instead of the logits (I observed that this works better in practice). Maybe we can also add this logic too to the general method? ", "@nicola-decao - I think you should be able to directly touch the log-probs with a `LogitsProcessor` since there are applied after the `log_softmax` here: https://github.com/huggingface/transformers/blob/7abc1d96d114873d9c3c2f1bc81343fb1407cec4/src/transformers/generation_utils.py#L967 and from looking at this version of the PR it seems to work with a `LogitsProcessor`\r\n \r\nOr do you need to apply it to `log_probs + beam_score` ? so after this line: https://github.com/huggingface/transformers/blob/7abc1d96d114873d9c3c2f1bc81343fb1407cec4/src/transformers/generation_utils.py#L968? his would be more difficult then and we would have to see how to deal with it ... maybe introduce `logits_warper as well for `beam_search`... not sure yet!\r\n\r\nIt would be great if you could add a `LogitProcessor` - I kinda did the whole refactor to keep the functions clean :sweat_smile: . \r\n\r\nI'm sorry that the big generate refactor means that we have to change this PR now. Do you want to give it a shot with the new `generate()` design? Otherwise I'm happy to help :-) ", "@patrickvonplaten I can do it :) I'll make another PR today.\r\n", "@patrickvonplaten here the new PR: https://github.com/huggingface/transformers/pull/8529" ]

[]

[]

[]

[]

[]

[ "If you don't need early stopping just comment out `early_stop_callback=early_stopping_callback` on line 379 of `/home/ioannis/Desktop/transformers/examples/lightning_base.py`. You should be able to run your script.\r\n\r\nI think lightning may have changed their api\r\nYou can also just uninstall your pytorch lightning and do `pip install pytorch_lightning==0.9.0` and script should work", "Awesome! Installing 0.9.0 worked.\r\n\r\nI manually installed pytorch_lightning and gitpython as they were not included in the transformers installation and the rag requirements file." ]

[]

[ { "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 } ]

[ "@jsilter - great catch! I agree 100% with your suggestion! Do you want to open a PR to fix it? :-) ", "https://github.com/huggingface/transformers/pull/7903 this is already fixed yesterday @jsilter " ]

[]

[]

[ "The answer you got on the forum is pretty much the only one we have: copy paste the model code and customize it to your needs. If you're not a programmer, then you will need to learn a bit of Python/PyTorch to do this, but apart from making sure each model file contains the full code of each model, there is little more we can do to help you. There is an API to use the models in the common cases, users are expected to make the customizations they want if that does not suit their needs.", "Hello, thank you for your reply.\r\nSorry if my question sounds a bit silly, but I just have a question:\r\n\r\nWhen I do `my_Longformer_model.encoder.layer[layer_index].forward(hidden_output,my_attention_mask,output_attention=False)`, this error is generated:\r\n```python\r\n\r\n File \"/Users/hyunjindominiquecho/opt/anaconda3/lib/python3.7/site-packages/transformers/modeling_longformer.py\", line 384, in _sliding_chunks_query_key_matmul\r\n batch_size, seq_len, num_heads, head_dim = query.size()\r\n\r\nValueError: too many values to unpack (expected 4)\r\n```\r\n\r\nHere, `my_attention_mask` is the same attention mask that I would specify under the regular\r\n```python\r\nlongformer_output= my_longformer_model(input_ids=input_ids,....,attention_mask=my_attention_mask)\r\n```\r\n\r\nwhy exactly is the above error generated, and how can I remedy it?\r\n\r\n\r\nThank you," ]

[ { "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", "@sunnyville01 \r\nhi, facing the same issue, did you manage to solve this?" ]

[]

[]

[]

[ { "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 }, { "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 } ]

[ "Hey @lalitpagaria - RAG is quite different to other generation models so we don't have it on the short-term roadmap to add it to pipelines. We are still thinking about how to integrate retrieval augmented models to the pipelines.", "Thanks @patrickvonplaten \r\nYeah I totally agree with you.\r\nPlease let me know whether I close this issue or keep it open for future reference.", "Leave it open - I'll put it under projects so that I don't forget it :-) \r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "@patrickvonplaten Great work! I noticed that `transformers` included the implementation for `DPR`. But for `RAG`, I only find a [demo](https://huggingface.co/rag/). Is there a source code for `RAG`? Or do you know where is Facebook's source code for `RAG`? ", "transformers does include RAG.\r\nYou can even find the documentation here: https://huggingface.co/transformers/model_doc/rag.html" ]

[]

[]

[]

[ { "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" } ]

[]

[ "This is great, thanks for uploading and sharing.\r\n\r\nWill merge this now, but had two questions:\r\n- should we add a link back to your repo at https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/TinyBERT\r\n- Is this checkpoint version 1 or version 2 from https://github.com/huawei-noah/Pretrained-Language-Model/blame/master/TinyBERT/README.md#L62-L72\r\n\r\nThank you!" ]

[ { "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", "hello, this command \" --model_type xlm-roberta \\\" not work for me,\r\ncan you help me? please" ]

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

[]

[ "Update - this is not a bug in run_ner.py, but sort-of a documentation bug. The page that describes how to do NER does not document that you first need to copy \"tasks.py\" and other scripts, into your local current directory.\r\n\r\nhttps://github.com/huggingface/transformers/tree/master/examples/token-classification/README.md\r\n\r\nFor instance, you could provide a list of \"wget\" commands. This may seem obvious to experienced developers, but it really helps to spell it out for the occasional clueless people like me :-)", "Hello! In order to run the script, you should generally clone the repository and run the scripts from there. Running a single script will very rarely work. You'd be way safer doing the following (showing you the full setup + venv setup):\r\n\r\n```py\r\ngit clone https://github.com/huggingface/transformers\r\ncd transformers\r\n\r\npython -m venv .env\r\nsource .env/bin/activate\r\npip install -e .\r\n\r\ncd examples\r\npip install -r requirements.txt\r\n\r\n# You can run your scripts now :)\r\n```", "Thank you; for future users, that would be super helpful to add that to the README.\r\n\r\nAlso, I think you need to cd to https://github.com/huggingface/transformers/tree/master/examples because the scripts assume that tasks.py etc are in the current directory. But the “activate” may take care of that.\r\n\r\nThanks for everything – great software!\r\n\r\nDana\r\n\r\nFrom: Lysandre Debut <[email protected]>\r\nSent: Thursday, October 15, 2020 2:00 AM\r\nTo: huggingface/transformers <[email protected]>\r\nCc: Ludwig, Dana <[email protected]>; Author <[email protected]>\r\nSubject: Re: [huggingface/transformers] Error in run_ner.py - ModuleNotFoundError: No module named 'tasks' (#7773)\r\n\r\n\r\nHello! In order to run the script, you should generally clone the repository and run the scripts from there. Running a single script will very rarely work. You'd be way safer doing the following (showing you the full setup + venv setup):\r\n\r\ngit clone https://github.com/huggingface/transformers\r\n\r\ncd transformers\r\n\r\n\r\n\r\npython -m venv .env\r\n\r\nsource .env/bin/activate\r\n\r\npip install -e .\r\n\r\n\r\n\r\ncd examples\r\n\r\npip install -r requirements.txt\r\n\r\n\r\n\r\n# You can run your scripts now :)\r\n\r\n—\r\nYou are receiving this because you authored the thread.\r\nReply to this email directly, view it on GitHub<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_huggingface_transformers_issues_7773-23issuecomment-2D709013098&d=DwMCaQ&c=iORugZls2LlYyCAZRB3XLg&r=A2YbHreGE4p0vzAywzM_Uctk-D3fPuXcmLPnjKJ7Gqc&m=QMIf0fCBI0DJve9mdTNWpHYcZNw6G6lbgNMtwt27-LI&s=_PAC-teAXIZpD6bvXaxJJooVSpQXCW-g1E0P6Xx0nzE&e=>, or unsubscribe<https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_notifications_unsubscribe-2Dauth_ABUXNRIKGTLERRFCOBWSGRTSK22XTANCNFSM4SQGNKXQ&d=DwMCaQ&c=iORugZls2LlYyCAZRB3XLg&r=A2YbHreGE4p0vzAywzM_Uctk-D3fPuXcmLPnjKJ7Gqc&m=QMIf0fCBI0DJve9mdTNWpHYcZNw6G6lbgNMtwt27-LI&s=JROY2gIxImDPq_udIaCyPb85pHW64zCpy7-aa4ggSQY&e=>.\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[ { "id": 2627272588, "node_id": "MDU6TGFiZWwyNjI3MjcyNTg4", "url": "https://api.github.com/repos/huggingface/transformers/labels/Model%20Parallel", "name": "Model Parallel", "color": "8B66A5", "default": false, "description": "Model Parallelilsm Implementations" } ]

[]

[ "Very cool implementation that would interest a few of our team, pinging them here :)", "Indeed, adding a test would be nice! Will trigger the multi-gpu slow tests once the test is added.", "@LysandreJik @patrickvonplaten @sgugger Glad you like it! Yes, the basic form of the implementation works on T5. I've been working on that too. Testing solutions to the items you brought up locally and should make the improvements to the PR in the next day or two.", "Hi @alexorona! Before we merge this, having a test would really be necessary. Have you had any luck in implementing such a test?\r\n\r\nIf you're lacking time, we can also take over from here.", "Hi, I've successfully been able to start training gpt2-xl on multiple gpu's using the model parallelism code from this pull request, but I'm running into an issue when restarting the training process from a checkpoint model. It seems that in this case when reloading the model, only the memory of the first GPU keeps increasing until it reaches an Out Of Memory error, instead of spreading over all GPUs like it did when training from scratch. Is it possible that reloading from checkpoint triggers a different code path that has been overlooked until now?\r\n\r\n(edit)\r\nI seem to have found an issue and potential fix. Line 609 in Trainer.train() loads the optimizer saved in the checkpoint. This function has a _map_location_ parameter which seems to force the optimizer to load fully onto _self.args.device_ which I'm guessing would be the first GPU. \r\n`self.optimizer.load_state_dict(torch.load(os.path.join(model_path, \"optimizer.pt\"), map_location=self.args.device))`\r\nRemoving the _map_location_ parameter makes the function properly put back all the loaded parameters onto the correct devices. Maybe there is a better way of handling this but at least it is an indication.", "@LysandreJik Just haven't had the time to do that and it might be awhile before I can get around to writing tests. However, I have been able to implement the same thing on T5 and it's working. Fine-tuned a 3B model this weekend. Maybe I can add that to this branch tonight and the team can handle the tests? \r\n\r\n@MichielRuelens Did you try loading to CPU and then calling model.parallelize() ? ", "@LysandreJik I worked on t5 last weekend. Added that. Happy to explain things like why `get_device_map `should probably take a list of devices rather than a number. Can you help get this across the finish line?", "Very cool, thanks! Will take a look at it and open a PR on your branch.", "Great, with the tests PR merged this seems close to merge! Could you rebase & run the code quality tools (`make fixup` or `make style && make quality`) so that the test suite passes?\r\n\r\nAlso @patrickvonplaten and could you check the tests?", "Awesome, LGTM!", "oof that's a tough rebase. I don't think we'll be able to merge that. Could you close the PR and open a new one, so that we can see the diff? I don't think you need to do anything on the branch, just close the PR and open a new one.", "I'm researching doing the same for BART (https://github.com/huggingface/transformers/issues/8344) and stumbled upon the open source library DeepSpeed/ZeRO:\r\nhttps://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/\r\n\r\nI'm yet to experiment with deepspeed and fairscale, but I thought I'd ask whether you have already done this and decided that implementing from scratch is better. \r\n\r\nWhat are the pros/cons for using an in-house implementation as compared to using an external library, other than the obvious potential bugs and issues of using external libs and not having control over those?\r\n\r\nIf this has been discussed already please kindly send me to that issue/page? Thank you!\r\n\r\nWhatever the outcome, we can also qualititatively compare the results of this PR once it's done with doing the same via deepspeed and/or fairscale.\r\n", "Rebased with transformers 4.0.0 and moved PR to [here](https://github.com/huggingface/transformers/pull/8696).\r\n\r\n> I'm researching doing the same for BART (#8344) and stumbled upon the open source library DeepSpeed/ZeRO:\r\n> https://www.microsoft.com/en-us/research/blog/zero-deepspeed-new-system-optimizations-enable-training-models-with-over-100-billion-parameters/\r\n\r\nGreat point, @stas00 ! From the description, DeepSpeed is an optimization on top of data parallelism and model parallelism. I think you've identified the next step! My reading of DeepSpeed this that one still has to implement data parallelism or model parallelism, but DeepSpeed will reduce the GPU memory footprint. Practically speaking, this could open the door to even larger models. During the development of the [final model parallel PR](https://github.com/huggingface/transformers/pull/8696), I ran into hard limits on AWS GPU instance with t5-11b and t5-11b in terms of the number of tokens you can train. You'll need the largest AWS instances to train 512 tokens on t5-3b. For t5-11b, you're restricted even more. Note that I haven't tried this with apex, so it might be possible to squeeze out a little more from the current implementation.", "Thanks for the follow up, @alexorona. From what I read deepspeed implements parallelism too, but I'm still in the process of studying - didn't get to play with it yet. I've started with fairscale first.\r\n\r\nWhat about [`fairscale`](https://github.com/facebookresearch/fairscale) then? It implements the model parallelism w/o needing to make any changes to the model. All that code is abstracted into the trainer-level calls. \r\n\r\nNow that you added T5+gpt2 model parallelism `transformers` needs to add a separate code for each model architecture. Why not follow `fairscale`-style and do it outside the model?", "Similarly to what we've done with the `tie_weights` methods which ties weights thanks to the `get_input_embeddings()` and `get_output_embeddings()` methods, we could probably have a model-agnostic way of enabling parallelization using a `get_layers()` method, like [this](https://github.com/huggingface/transformers/blob/gradient-checkpointing-v2/src/transformers/modeling_utils.py#L669-L682) for example. \r\n\r\nThis would allow to have the same code for each model architecture, provided the `get_layers()` utility is correctly implemented.", "Would need to look into how `deepspeed` is implemented. My reading is that it supports parallelism, but still requires you to go through the process of defining how parallelism will work in the model definition (which modules on which devices). \r\n\r\nNot sure on `fairscale`. It may or may not be a simplification. There were quirks in implementing model parallelism on gpt2 and t5 that have to do with the pytorch graph:\r\n- The LM head has to be loaded on the same device as the embedding layer\r\n- Tensors (not just layers) have to be shifted to the appropriate device during training\r\n\r\nInitially, I went to `eisen` to see if it could handle model parallelism, but it couldn't deal with transformer models. If `fairscale` can abstract that process, that would be great.", "I hear you, @alexorona, that you're saying that each model may have its special needs. So we need to see if perhaps this can be done via some sort of callbacks and add the callbacks points as we discover new needs.\r\n\r\nAs I was planning to do the same for BART, perhaps I should replicate your implementation for Bart and then based on 3 different models we can then see how we can make this model-agnostic. What do you think?\r\n\r\nPerhaps we should continue this discussion in a dedicated issue where we discuss model_parallelism for all of `transformers`, including considering using or forking existing implemenations.", "I think those are both great ideas, @stas00. Should give us a better understanding of how to move forward with this. Created [this issue](https://github.com/huggingface/transformers/issues/8771) to continue conversation. " ]

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

[]

[ "`generate` currently does not support `torch.jit.trace`. This is sadly also not on the short-term roadmap.", "@patrickvonplaten, thanks for the response. In that case is there no way to trace the inference process of generative models provided here ? So any kind of inference of the form text -> text (for eg: summarization) cannot be exported to torchscript ? ", "Yes, if you want to make it faster, reduce `num_beams`.", "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, are summarization kinds of models still not traceable ??\r\n\r\nI am trying to deploy this onto AWS inferentia, whose prerequisite is that the model should be traceable !!\r\n\r\nfor example this sshleifer/distilbart-cnn-12-6\r\n\r\n@sshleifer, @patrickvonplaten can you guys please help ??", "Hey @DevBey,\r\n\r\nCould you maybe open a new issue that states exactly what doesn't work in your case? This issue is quite old now and it would be nice to have a reproducible code snippet with the current `transformers` version." ]

[]

[]

[ "Normally there's no difference. You just provide `[CLS] question [SEP] answer [SEP]` examples to the model, and the `start_positions` and `end_positions` can be indexes of the question. ", "Yes,\r\nCan you point me to such an implementation?\r\nhuggingface Squad Question answering code handles a lot of edge cases like start token should be before the end token, find n best predictions etc. \r\n@NielsRogge , Is there a simpler implementation for the question answering task?\r\n\r\nI found one in the documentation but that doesn't handle the predictions, sanity checks, n_best predictions etc.\r\n\r\n\r\n" ]

[]

[]

[ "You don't have PyTorch installed in your environment. This is a PyTorch model.", "Indeed , that was the problem.\r\nThe code in https://huggingface.co/transformers/model_doc/rag.html has import torch after the \"from transformers import RagSequenceForGeneration,...\" statement, so I incorrectly concluded that torch is not needed for the import. Also, that RagTokenizer got imported without errors.\r\nThanks!\r\n" ]

[]

[]

[]

[]

[]

[ "Thanks so much @sgugger - you're a legend! 🙇 ", "> # What does this PR do?\r\n> Currently, the Trainer accumulates all predictions on the host (GPU or TPU) before gathering across all hosts (in case of distributed training) and moving back to the CPU. This can result in OOM errors when users have a big dataset (the model is already taking up a lot of space on the host) as highlighted in #7232. However moving the predictions to the CPU at each prediction step is also inefficient (particularly on TPU).\r\n> \r\n> This PR aims at fixing the OOM problem while retaining efficiency by introducing a new training argument called `eval_accumulation_step`. If left untouched, the behavior is the same as right now (all predictions accumulated on the host and moved at the end of the prediction loop). If set to an int, the predictions are gathered and moved every `eval_accumulation_step`. This required some clever reorganization of the predictions (see the docstring of `DistributedTensorGatherer` for more details).\r\n> \r\n> In passing I cleaned up the code related to gathering tensors across multiple hosts and fixed the issue of the `loss.item()` (big slow down to do that at every step on TPUs) and accumulated the losses the same way predictions and labels are. This still works for any number of outputs/labels of the model.\r\n> \r\n> To check those changes did not break anything, I ran `test_trainer_distributed.py` on my local setup and created an equivalent for TPUs that I also ran (they both pass).\r\n> \r\n> This slightly change Seq2SeqTrainer (since we don't want the `loss.item()`) so cc @patil-suraj I don't think this should break anything in it.\r\n> \r\n> Fixes #7232\r\n\r\nThanks so much @sgugger \r\n`eval_accumulation_steps` for the argument name and not `eval_accumulation_step` 😉" ]

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

[]

[ "Would love a contribution that implemented it. \r\nCan you paste a working fairseq command to try to emulate?\r\nAlternatively you could just send a PR.", "Hi @sshleifer \r\n\r\nMarian NMT seems to have such functionality as well. Please check their script [here](https://github.com/marian-nmt/marian-examples/blob/master/wmt2017-transformer/run-me.sh#L189).\r\n\r\nI modify the [translation code](https://fairseq.readthedocs.io/en/latest/command_line_tools.html#fairseq-generate) in fairseq named as *validate_lm.py*, and run the bash script to get the per-word log-probability.\r\n```bash\r\npython validate_lm.py \\\r\n $binarized_data_dir \\\r\n --source-lang en --target-lang fr \\\r\n --path $path_to_checkpoint \\\r\n --task translation \\\r\n --valid-subset train \\\r\n --max-sentences 16 \\\r\n --nll-file $out_file\r\n```\r\n\r\nvalidate_lm.py\r\n```python\r\nimport logging\r\nimport sys\r\n\r\nimport torch\r\n\r\nfrom fairseq import checkpoint_utils, distributed_utils, options, utils, tasks\r\nfrom fairseq.logging import metrics, progress_bar\r\nfrom fairseq.options import add_distributed_training_args\r\nfrom fairseq.criterions.cross_entropy import CrossEntropyCriterion\r\n\r\nlogging.basicConfig(\r\n format='%(asctime)s | %(levelname)s | %(name)s | %(message)s',\r\n datefmt='%Y-%m-%d %H:%M:%S',\r\n level=logging.INFO,\r\n stream=sys.stdout,\r\n)\r\nlogger = logging.getLogger('fairseq_cli.validate')\r\n\r\n\r\ndef main(args, override_args=None):\r\n utils.import_user_module(args)\r\n\r\n assert args.max_tokens is not None or args.max_sentences is not None, \\\r\n 'Must specify batch size either with --max-tokens or --max-sentences'\r\n\r\n use_fp16 = args.fp16\r\n use_cuda = torch.cuda.is_available() and not args.cpu\r\n\r\n if override_args is not None:\r\n if isinstance(override_args, dict):\r\n overrides = override_args\r\n else:\r\n overrides = vars(override_args)\r\n print('override_args')\r\n overrides.update(eval(getattr(override_args, 'model_overrides', '{}')))\r\n else:\r\n overrides = None\r\n\r\n # Load ensemble\r\n logger.info('loading model(s) from {}'.format(args.path))\r\n task = tasks.setup_task(args)\r\n models, _model_args = checkpoint_utils.load_model_ensemble(\r\n [args.path],\r\n arg_overrides=overrides,\r\n task=task,\r\n )\r\n model = models[0]\r\n\r\n # Move models to GPU\r\n for model in models:\r\n if use_fp16:\r\n model.half()\r\n if use_cuda:\r\n model.cuda()\r\n\r\n # Print args\r\n logger.info('args', args)\r\n logger.info('overrides', overrides)\r\n\r\n # Build criterion\r\n # criterion = task.build_criterion(args)\r\n criterion = CrossEntropyCriterion(task, False)\r\n criterion.eval()\r\n\r\n for subset in args.valid_subset.split(','):\r\n try:\r\n task.load_dataset(subset, combine=False, epoch=1)\r\n dataset = task.dataset(subset)\r\n except KeyError:\r\n raise Exception('Cannot find dataset: ' + subset)\r\n\r\n # Initialize data iterator\r\n itr = task.get_batch_iterator(\r\n dataset=dataset,\r\n max_tokens=args.max_tokens,\r\n max_sentences=args.max_sentences,\r\n max_positions=utils.resolve_max_positions(\r\n task.max_positions(),\r\n *[m.max_positions() for m in models],\r\n ),\r\n ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,\r\n required_batch_size_multiple=args.required_batch_size_multiple,\r\n seed=args.seed,\r\n num_workers=args.num_workers,\r\n ).next_epoch_itr(shuffle=False)\r\n progress = progress_bar.progress_bar(\r\n itr,\r\n log_format=args.log_format,\r\n log_interval=args.log_interval,\r\n prefix=f\"valid on '{subset}' subset\",\r\n default_log_format=('tqdm' if not args.no_progress_bar else 'simple'),\r\n )\r\n\r\n fout = open(args.nll_file, 'w')\r\n nll = []\r\n for i, sample in enumerate(progress):\r\n sample = utils.move_to_cuda(sample) if use_cuda else sample\r\n model.eval()\r\n with torch.no_grad():\r\n loss, sample_size, log_output = criterion(model, sample, reduce=False)\r\n nsentences = log_output['nsentences']\r\n loss = loss.view(nsentences, -1).tolist()\r\n for j, sample_id in enumerate(sample['id'].tolist()):\r\n loss_j = [ll for ll in loss[j] if ll > 0]\r\n avg_nll = sum(loss_j) / len(loss_j)\r\n lstr = '\\t'.join([f'{ll:.4f}' for ll in loss_j])\r\n fout.write(f'{sample_id}\\t{avg_nll}\\t{lstr}\\n')\r\n nll.append((sample_id, avg_nll, lstr))\r\n fout.close()\r\n\r\ndef cli_main():\r\n parser = options.get_validation_parser()\r\n add_distributed_training_args(parser)\r\n args = options.parse_args_and_arch(parser)\r\n\r\n # only override args that are explicitly given on the command line\r\n override_parser = options.get_validation_parser()\r\n group = override_parser.add_argument_group(\"Valid BW\")\r\n group.add_argument('--nll-file', type=str, default=None)\r\n add_distributed_training_args(override_parser)\r\n override_args = options.parse_args_and_arch(override_parser, suppress_defaults=True)\r\n distributed_utils.call_main(args, main, override_args=override_args)\r\n\r\n\r\nif __name__ == '__main__':\r\n cli_main()\r\n```\r\n\r\n\r\n\r\n", "I'd tinker with that - perhaps in a few days.", "I started looking into this but to be able to reproduce any of these examples in order to replicate this in transformers I need to spend hours just to set things up :( It'd have been much easier if such requests came with ready data that we could use right away.\r\n\r\nCurrently waiting for very sloooooooow downloads of data from statmt.org - will try to setup fairseq for en-fr...\r\n\r\nwill keep you posted on the progress.\r\n\r\np.s. marianmt I couldn't even build on ubuntu-20.04, so hoping I could sort it out with fairseq", "OK, I found an easy way to set up something that your script would run on:\r\n\r\n```\r\nmkdir -p data-bin\r\ncurl https://dl.fbaipublicfiles.com/fairseq/models/wmt14.v2.en-fr.fconv-py.tar.bz2 | tar xvjf - -C data-bin\r\ncurl https://dl.fbaipublicfiles.com/fairseq/data/wmt14.v2.en-fr.newstest2014.tar.bz2 | tar xvjf - -C data-bin\r\npython validate_lm.py data-bin/wmt14.en-fr.joined-dict.newstest2014 --source-lang en --target-lang fr --path data-bin/wmt14.en-fr.joined-dict.transformer/model.pt --task translation --valid-subset test --max-tokens 128\r\n```\r\n\r\n(in the future requests please provide something similar so that the devs could quickly reproduce your example. And of course test that it works. Thank you)\r\n\r\nYour script won't run as is, had to do some tweaks to it. But it doesn't matter for now.\r\n\r\nWhat it produces is:\r\n```\r\n1596 0.7027008545895418 1.2075 1.4464 1.1499 0.1134 0.1958 0.1032\r\n1615 3.4015886902809145 5.8378 9.5874 0.6482 0.7548 0.1797\r\n2837 1.7323936596512794 0.8643 0.1775 7.3694 0.1477 0.1032\r\n256 1.5152931759754817 0.8425 7.1255 0.8411 0.0599 0.1208 0.1020\r\n556 2.3045735328147807 6.7926 0.0528 1.0420 2.4513 3.3704 0.1184\r\n2922 1.7860032822936773 0.4388 3.8867 1.9373 0.2531 6.5018 0.1011 1.0688 0.1005\r\n612 1.598365530371666 3.2449 2.8535 0.1290 0.1661\r\n605 1.7411062990625699 4.9944 0.1142 0.1147\r\n1481 0.4938086899263518 0.3987 0.5716 1.2258 0.5801 0.4607 0.1157 0.1041\r\n2013 0.7291228858133157 2.0326 0.1758 0.1322 1.7975 0.1314 0.1051\r\n75 1.2818062901496887 0.6200 0.2217 0.0730 0.1697 7.2470 0.4181 0.2230\r\n279 1.4058668922100748 8.1397 1.1370 0.1082 0.1427 0.0871 0.1210 0.1053\r\n2641 0.1703926378062793 0.3318 0.1146 0.2372 0.1261 0.1727 0.1082 0.1022\r\n1120 2.3883045655157833 2.9336 0.8035 1.6985 4.0525 8.6964 3.0033 0.0963 0.1081 0.1024\r\n1031 0.6803936168551445 4.6826 0.7309 0.0864 0.1111 0.1247 0.0863 0.5777 0.0856 0.2151 0.1035\r\n2484 0.17211276292800903 0.3391 0.2407 0.0678 0.4290 0.1203 0.0990 0.1382 0.1095 0.0758 0.1018\r\n2600 1.3488108797797136 4.0854 0.1176 1.0209 0.1188 0.0373 3.7863 0.2754\r\n2814 1.5876335703900881 1.6473 2.3662 0.2776 0.0731 4.8990 1.7478 0.1024\r\n2822 0.6652300975152424 0.6418 0.1277 0.6620 2.6037 0.4055 0.1121 0.1038\r\n2854 0.396458768418857 0.5300 0.1097 1.1606 0.4882 0.2649 0.1183 0.1035\r\n169 1.5903700785711408 4.9675 0.0435 0.1813 1.8350 0.1336 5.2166 0.2192 0.1262\r\n234 2.0860743997618556 2.8854 0.2894 0.3069 0.1739 0.0880 12.6388 0.1717 0.1345\r\n368 0.9114988259971142 0.7726 0.7938 1.8604 3.4764 0.0779 0.0802 0.1271 0.1035\r\n387 0.24928769376128912 0.4698 0.0763 0.2203 0.1345 0.3810 0.4948 0.1152 0.1024\r\n596 0.7321822415105999 5.0991 0.1796 0.1138 0.1113 0.1246 0.0427 0.0759 0.1104\r\n1200 0.587307695299387 1.3917 1.1277 1.4277 0.0914 0.1485 0.1969 0.2117 0.1029\r\n2015 1.1963263098150492 5.4510 1.0897 0.4743 0.0972 0.2240 1.2175 0.8575 0.1595\r\n2216 1.3388446755707264 4.0211 4.2651 0.4503 0.1335 0.0482 0.3432 1.3127 0.1366\r\n2994 0.3156363896559924 0.1788 0.1143 0.1824 1.7381 0.0294 0.0742 0.1021 0.1059\r\n```\r\nI don't suppose this is what you are after, do you?\r\n\r\n", "i think the original poster wants a way to get those log probs for models that are not in fair aew, like MarianMT.", "I think I understood that, I'm referring to:\r\n\r\n> I modify the translation code in fairseq named as validate_lm.py, and run the bash script to get the per-word log-probability.\r\n\r\nwhich is what I run and displayed the top of the results above. But in the OP it was requested:\r\n\r\n> For example:\r\n> Source input (x): I would like to run this experiment.\r\n> Translated output (y) in Chinese: 我 希望 跑 这个 实验 。\r\n> The word-level log probability: -0.3 -0.4 -0.23 -0.43 -0.23 -0.8\r\n\r\nSo the example code doesn't match the original request and I'm asking to clarify to what output is desired.\r\n\r\nIn other words - do you want the word and the average log probability of its tokens printed? If yes, I guess we gather the log probs for each token and then somehow keep track of the de-tokenizer and then average the ones that end up comprising each word. I don't have any experience with Kanji - do tokenizers there split the characters into sub-characters, or is it the case where each token is a word, in which case you're asking to just print the log probabilities of each token in the translation output?\r\n\r\nCould you redefine your request in terms of latin 2 latin language to remove this possible ambiguity and then once working adapt it to Chinese? I hope it makes sense where I am not sure what exactly you're after.\r\n", "@JunjieHu Is this roughly what you are looking for? \r\n```python\r\nbatch = tokenizer.prepare_seq2seq_batch(\": I would like to run this experiment.\")\r\nmodel = AutoModelForSeq2SeqLM.from_pretrained('opus-mt/marian-en-zh')\r\ngenerated_ids = model.generate(batch)\r\noutputs = model(batch.input_ids, labels=generated_ids)\r\nlog_probas == outputs.logits[:, generated_ids]\r\n```\r\n\r\n", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Hi @sshleifer Sorry for the late response! Yes! This is exactly what I want! Thanks!" ]

[]

[]

[ "I'll leave it to @sshleifer - haven't really used the seq2seq fine-tuning too much.\r\n\r\nMaybe @patil-suraj has also an idea here :-) ", "You can safely ignore the `lr_scheduler` warning. `optimzer` is saved, but torch `lr_scheduler` warns you anyway just so you don't forget.\r\n\r\nNot sure about the first warning, but you can also safely ignore that\r\n", "@patil-suraj Thank you! Would there be an easy way to somehow silence the warning, or print a logging statement from huggingface library that the warning can be safely ignored? ", "I believe this warning has been hidden on the `master` branch, and will be hidden in the next release. See [this](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_pt_utils.py#L114-L119).", "Cool! Looks like originally in #7401. I pulled master and confirmed that this is fixed. \r\n\r\nAny notes on the computational graph warning that also pops up?\r\n../python3.8/site-packages/pytorch_lightning/utilities/distributed.py:37: UserWarning: Could not log computational graph since the model.example_input_array attribute is not set or input_array was not given" ]

[]

[]

[ "Very true! Thanks for the notification - will upload all DialoGPT configs.", "Done - note however that the \"task\" name was `conversational` not `dialogue`" ]

[]

[]

[ "I took your comments into account, let me know if you have other things to improve.\r\nAlso I had to change the DPR encoder from the one trained on Natural Questions to the one trained ont the multiset/hybrid dataset to match the embeddings used by the Rag team.", "can't wait to try this out. \r\n@lhoestq , can this functionality be adapted to address #6399 ?", "Glad this got implemented! Many thanks @lhoestq . I checked out a copy, added a custom 25MB CSV file, and gave it a run:\r\npython examples/rag/use_own_knowledge_dataset.py\r\n\r\nGot this.\r\n```\r\nTraceback (most recent call last):\r\n File \"examples/rag/use_own_knowledge_dataset.py\", line 195, in <module>\r\n main(tmp_dir, rag_example_args, processing_args, index_hnsw_args)\r\n File \"examples/rag/use_own_knowledge_dataset.py\", line 84, in main\r\n ctx_tokenizer = DPRContextEncoderTokenizerFast.from_pretrained(rag_example_args.dpr_ctx_encoder_model_name)\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/tokenization_utils_base.py\", line 1544, in from_pretrained\r\n list(cls.vocab_files_names.values()),\r\nOSError: Model name 'facebook/dpr-ctx_encoder-multiset-base' was not found in tokenizers model name list (facebook/dpr-ctx_encoder-single-nq-base). We assumed 'facebook/dpr-ctx_encoder-multiset-base' was a path, a model identifier, or url to a directory containing vocabulary files named ['vocab.txt'] but couldn't find such vocabulary files at this path or url.\r\n```\r\n\r\nI then switched to **dpr-ctx_encoder-single-nq-base**. In this case, indexing was successful. However, after Step 3 - Load RAG, the script started downloading the 74G wiki dataset (which should not be necessary) and then errored out on line 195.\r\n\r\n```\r\n$ python examples/rag/use_own_knowledge_dataset.py\r\nINFO:__main__:Step 1 - Create the dataset\r\nUsing custom data configuration default\r\nReusing dataset csv (/home/ioannis/.cache/huggingface/datasets/csv/default-49c04e2dbd1cfa6f/0.0.0/49187751790fa4d820300fd4d0707896e5b941f1a9c644652645b866716a4ac4)\r\n100%|█████████████████████████████████████████| 155/155 [00:01<00:00, 82.27ba/s]\r\n100%|█████████████████████████████████████| 18531/18531 [28:23<00:00, 10.88ba/s]\r\nINFO:__main__:Step 2 - Index the dataset\r\n100%|█████████████████████████████████████████| 297/297 [30:22<00:00, 6.14s/it]\r\nINFO:__main__:Step 3 - Load RAG\r\nUsing custom data configuration psgs_w100.nq.no_index\r\nDownloading and preparing dataset wiki_dpr/psgs_w100.nq.no_index (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/ioannis/.cache/huggingface/datasets/wiki_dpr/psgs_w100.nq.no_index/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2...\r\nDownloading: 100%|█████████████████████████| 11.2k/11.2k [00:00<00:00, 2.65MB/s]\r\nDownloading: 100%|███████████████████████| 78.4G/78.4G [1:54:59<00:00, 11.4MB/s]\r\nDataset wiki_dpr downloaded and prepared to /home/ioannis/.cache/huggingface/datasets/wiki_dpr/psgs_w100.nq.no_index/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2. Subsequent calls will reuse this data.\r\nUsing custom data configuration psgs_w100.nq.custom\r\nDownloading and preparing dataset wiki_dpr/psgs_w100.nq.custom (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/ioannis/.cache/huggingface/datasets/wiki_dpr/psgs_w100.nq.custom/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2...\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\nTo disable this warning, you can either:\r\n\t- Avoid using `tokenizers` before the fork if possible\r\n\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\nhuggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\n...\r\n...\r\nDownloading: 100%|██████████████████████████| 1.33G/1.33G [29:21<00:00, 753kB/s]\r\nDownloading: 100%|██████████████████████████| 1.33G/1.33G [29:30<00:00, 749kB/s]\r\n...\r\n...\r\nTraceback (most recent call last): \r\n File \"examples/rag/use_own_knowledge_dataset.py\", line 195, in <module>\r\n main(tmp_dir, rag_example_args, processing_args, index_hnsw_args)\r\n File \"examples/rag/use_own_knowledge_dataset.py\", line 116, in main\r\n rag_example_args.rag_model_name, index_name=\"custom\", indexed_dataset=dataset\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 321, in from_pretrained\r\n config, question_encoder_tokenizer=question_encoder_tokenizer, generator_tokenizer=generator_tokenizer\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 310, in __init__\r\n self.init_retrieval()\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 338, in init_retrieval\r\n self.index.init_index()\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 248, in init_index\r\n dummy=self.use_dummy_dataset,\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/load.py\", line 611, in load_dataset\r\n ignore_verifications=ignore_verifications,\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/builder.py\", line 476, in download_and_prepare\r\n dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/builder.py\", line 553, in _download_and_prepare\r\n self._prepare_split(split_generator, **prepare_split_kwargs)\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/builder.py\", line 841, in _prepare_split\r\n generator, unit=\" examples\", total=split_info.num_examples, leave=False, disable=not_verbose\r\n File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/tqdm/std.py\", line 1133, in __iter__\r\n for obj in iterable:\r\n File \"/home/ioannis/.cache/huggingface/modules/datasets_modules/datasets/wiki_dpr/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2/wiki_dpr.py\", line 124, in _generate_examples\r\n id, text, title = line.strip().split(\"\\t\")\r\nValueError: not enough values to unpack (expected 3, got 2)\r\n\r\n```\r\n", "> Glad this got implemented! Many thanks @lhoestq . I checked out a copy, added a custom 25MB CSV file, and gave it a run:\r\n> python examples/rag/use_own_knowledge_dataset.py\r\n> \r\n> Got this.\r\n> \r\n> ```\r\n> Traceback (most recent call last):\r\n> File \"examples/rag/use_own_knowledge_dataset.py\", line 195, in <module>\r\n> main(tmp_dir, rag_example_args, processing_args, index_hnsw_args)\r\n> File \"examples/rag/use_own_knowledge_dataset.py\", line 84, in main\r\n> ctx_tokenizer = DPRContextEncoderTokenizerFast.from_pretrained(rag_example_args.dpr_ctx_encoder_model_name)\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/tokenization_utils_base.py\", line 1544, in from_pretrained\r\n> list(cls.vocab_files_names.values()),\r\n> OSError: Model name 'facebook/dpr-ctx_encoder-multiset-base' was not found in tokenizers model name list (facebook/dpr-ctx_encoder-single-nq-base). We assumed 'facebook/dpr-ctx_encoder-multiset-base' was a path, a model identifier, or url to a directory containing vocabulary files named ['vocab.txt'] but couldn't find such vocabulary files at this path or url.\r\n> ```\r\n> \r\n> I then switched to **dpr-ctx_encoder-single-nq-base**. In this case, indexing was successful. However, after Step 3 - Load RAG, the script started downloading the 74G wiki dataset (which should not be necessary) and then errored out on line 195.\r\n> \r\n> ```\r\n> $ python examples/rag/use_own_knowledge_dataset.py\r\n> INFO:__main__:Step 1 - Create the dataset\r\n> Using custom data configuration default\r\n> Reusing dataset csv (/home/ioannis/.cache/huggingface/datasets/csv/default-49c04e2dbd1cfa6f/0.0.0/49187751790fa4d820300fd4d0707896e5b941f1a9c644652645b866716a4ac4)\r\n> 100%|█████████████████████████████████████████| 155/155 [00:01<00:00, 82.27ba/s]\r\n> 100%|█████████████████████████████████████| 18531/18531 [28:23<00:00, 10.88ba/s]\r\n> INFO:__main__:Step 2 - Index the dataset\r\n> 100%|█████████████████████████████████████████| 297/297 [30:22<00:00, 6.14s/it]\r\n> INFO:__main__:Step 3 - Load RAG\r\n> Using custom data configuration psgs_w100.nq.no_index\r\n> Downloading and preparing dataset wiki_dpr/psgs_w100.nq.no_index (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/ioannis/.cache/huggingface/datasets/wiki_dpr/psgs_w100.nq.no_index/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2...\r\n> Downloading: 100%|█████████████████████████| 11.2k/11.2k [00:00<00:00, 2.65MB/s]\r\n> Downloading: 100%|███████████████████████| 78.4G/78.4G [1:54:59<00:00, 11.4MB/s]\r\n> Dataset wiki_dpr downloaded and prepared to /home/ioannis/.cache/huggingface/datasets/wiki_dpr/psgs_w100.nq.no_index/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2. Subsequent calls will reuse this data.\r\n> Using custom data configuration psgs_w100.nq.custom\r\n> Downloading and preparing dataset wiki_dpr/psgs_w100.nq.custom (download: Unknown size, generated: Unknown size, post-processed: Unknown size, total: Unknown size) to /home/ioannis/.cache/huggingface/datasets/wiki_dpr/psgs_w100.nq.custom/0.0.0/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2...\r\n> huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\n> To disable this warning, you can either:\r\n> \t- Avoid using `tokenizers` before the fork if possible\r\n> \t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\r\n> huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\r\n> ...\r\n> ...\r\n> Downloading: 100%|██████████████████████████| 1.33G/1.33G [29:21<00:00, 753kB/s]\r\n> Downloading: 100%|██████████████████████████| 1.33G/1.33G [29:30<00:00, 749kB/s]\r\n> ...\r\n> ...\r\n> Traceback (most recent call last): \r\n> File \"examples/rag/use_own_knowledge_dataset.py\", line 195, in <module>\r\n> main(tmp_dir, rag_example_args, processing_args, index_hnsw_args)\r\n> File \"examples/rag/use_own_knowledge_dataset.py\", line 116, in main\r\n> rag_example_args.rag_model_name, index_name=\"custom\", indexed_dataset=dataset\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 321, in from_pretrained\r\n> config, question_encoder_tokenizer=question_encoder_tokenizer, generator_tokenizer=generator_tokenizer\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 310, in __init__\r\n> self.init_retrieval()\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 338, in init_retrieval\r\n> self.index.init_index()\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/transformers/retrieval_rag.py\", line 248, in init_index\r\n> dummy=self.use_dummy_dataset,\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/load.py\", line 611, in load_dataset\r\n> ignore_verifications=ignore_verifications,\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/builder.py\", line 476, in download_and_prepare\r\n> dl_manager=dl_manager, verify_infos=verify_infos, **download_and_prepare_kwargs\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/builder.py\", line 553, in _download_and_prepare\r\n> self._prepare_split(split_generator, **prepare_split_kwargs)\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/datasets/builder.py\", line 841, in _prepare_split\r\n> generator, unit=\" examples\", total=split_info.num_examples, leave=False, disable=not_verbose\r\n> File \"/home/ioannis/anaconda3/envs/transformers/lib/python3.6/site-packages/tqdm/std.py\", line 1133, in __iter__\r\n> for obj in iterable:\r\n> File \"/home/ioannis/.cache/huggingface/modules/datasets_modules/datasets/wiki_dpr/14b973bf2a456087ff69c0fd34526684eed22e48e0dfce4338f9a22b965ce7c2/wiki_dpr.py\", line 124, in _generate_examples\r\n> id, text, title = line.strip().split(\"\\t\")\r\n> ValueError: not enough values to unpack (expected 3, got 2)\r\n> ```\r\n\r\nI am facing the same issue.", "Thanks for reporting I'm looking into it", "> Glad this got implemented! Many thanks @lhoestq . I checked out a copy, added a custom 25MB CSV file, and gave it a run:\r\n> python examples/rag/use_own_knowledge_dataset.py\r\n> \r\n> Got this.\r\n> \r\n> ```\r\n> OSError: Model name 'facebook/dpr-ctx_encoder-multiset-base' was not found in tokenizers model name list (facebook/dpr-ctx_encoder-single-nq-base). We assumed 'facebook/dpr-ctx_encoder-multiset-base' was a path, a model identifier, or url to a directory containing vocabulary files named ['vocab.txt'] but couldn't find such vocabulary files at this path or url.\r\n> ```\r\n> \r\n> I then switched to **dpr-ctx_encoder-single-nq-base**. In this case, indexing was successful. However, after Step 3 - Load RAG, the script started downloading the 74G wiki dataset (which should not be necessary) and then errored out on line 195.\r\n> \r\n> ```\r\n> $ python examples/rag/use_own_knowledge_dataset.py\r\n> ...\r\n> ValueError: not enough values to unpack (expected 3, got 2)\r\n> ```\r\n\r\nYou're having this issue because you are running the script with a version of transformers that doesn't include the changes I had to make in this PR to support custom datasets. This PR not only adds an example script, but there are also changes to make it possible in the core code of the RAG retriever.\r\n\r\nEverything works fine if you have all the changes of this PR", "> > Glad this got implemented! Many thanks @lhoestq . I checked out a copy, added a custom 25MB CSV file, and gave it a run:\r\n> > python examples/rag/use_own_knowledge_dataset.py\r\n> > Got this.\r\n> > ```\r\n> > OSError: Model name 'facebook/dpr-ctx_encoder-multiset-base' was not found in tokenizers model name list (facebook/dpr-ctx_encoder-single-nq-base). We assumed 'facebook/dpr-ctx_encoder-multiset-base' was a path, a model identifier, or url to a directory containing vocabulary files named ['vocab.txt'] but couldn't find such vocabulary files at this path or url.\r\n> > ```\r\n> > \r\n> > \r\n> > I then switched to **dpr-ctx_encoder-single-nq-base**. In this case, indexing was successful. However, after Step 3 - Load RAG, the script started downloading the 74G wiki dataset (which should not be necessary) and then errored out on line 195.\r\n> > ```\r\n> > $ python examples/rag/use_own_knowledge_dataset.py\r\n> > ...\r\n> > ValueError: not enough values to unpack (expected 3, got 2)\r\n> > ```\r\n> \r\n> You're having this issue because you are running the script with a version of transformers that doesn't include the changes I had to make in this PR to support custom datasets. This PR not only adds an example script, but there are also changes to make it possible in the core code of the RAG retriever.\r\n> \r\n> Everything works fine if you have all the changes of this PR\r\n\r\nHaving a separate script to fine-tune with custom datasets would be super useful!", "> Having a separate script to fine-tune with custom datasets would be super useful!\r\n\r\nI am adding flags to the fine-tuning scripts to make it work with a custom retriever ;)", "Ah, I think my mistake was that I was using the previous conda environment of transformers (with the new branch). Trying this again with a new env now :)", "> > I just checkd out [14420dc](https://github.com/huggingface/transformers/commit/14420dcbfa0f06012152cc66a7af84d7165a2a17) (committed one hour ago) and I am still getting the same error. Also, it was this branch that was giving the previous errors, not the main transformers branch. Perhaps some of your files have not been updated in the repository? Confused :)\r\n> > `OSError: Model name 'facebook/dpr-ctx_encoder-multiset-base' was not found in tokenizers model name list (facebook/dpr-ctx_encoder-single-nq-base). We assumed 'facebook/dpr-ctx_encoder-multiset-base' was a path, a model identifier, or url to a directory containing vocabulary files named ['vocab.txt'] but couldn't find such vocabulary files at this path or url.`\r\n> \r\n> Could you try to run\r\n> \r\n> ```python\r\n> from transformers import DPRContextEncoderTokenizerFast\r\n> tokenizer = DPRContextEncoderTokenizerFast.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n> ```\r\n> \r\n> and let me know if you're having the OSError on [14420dc](https://github.com/huggingface/transformers/commit/14420dcbfa0f06012152cc66a7af84d7165a2a17) ?\r\n> \r\n> Also just to help me fix this issue, could you also tell me the input of this code please:\r\n> \r\n> ```python\r\n> from transformers.tokenization_dpr_fast import CONTEXT_ENCODER_PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES\r\n> print(CONTEXT_ENCODER_PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES)\r\n> # it should print {'facebook/dpr-ctx_encoder-single-nq-base': 512, 'facebook/dpr-ctx_encoder-multiset-base': 512}\r\n> ```\r\n\r\nI was using the old transformers conda env. Doing things over again now and will report back asap!", "Ok thanks !\r\nI deleted my comment since I just noticed yours about the wrong env.\r\nIt should work for you now in a new env", "I added some tests for the distributed retriever for fine-tuning.\r\nUnless you still have an issue @ioannist this PR should be ready.\r\nCc @patrickvonplaten if you want to take a look at the new changes", "> Ok thanks !\r\n> I deleted my comment since I just noticed yours about the wrong env.\r\n> It should work for you now in a new env\r\n\r\nI created a new conda env from 14420dcb, installed pytorch and tf 2.2 (2.0 did not work). I ran again into the OS error. Here is the output on the tests you asked me to run +1 more.\r\n\r\n(HEAD detached at 14420dcb)\r\n\r\n```\r\nfrom transformers import DPRContextEncoderTokenizerFast\r\ntokenizer = DPRContextEncoderTokenizerFast.from_pretrained(\"facebook/dpr-ctx_encoder-single-nq-base\")\r\n```\r\nNo error\r\n\r\n```\r\nfrom transformers.tokenization_dpr_fast import CONTEXT_ENCODER_PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES\r\nprint(CONTEXT_ENCODER_PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES)\r\n```\r\nModuleNotFoundError: No module named 'transformers.tokenization_dpr_fast'\r\nI checked if the tokenization_dpr_fast file is under src/transformers. It's there.\r\n\r\n```\r\nfrom transformers.tokenization_dpr import CONTEXT_ENCODER_PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES\r\nprint(CONTEXT_ENCODER_PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES)\r\n```\r\n{'facebook/dpr-ctx_encoder-single-nq-base': 512}\r\n", "Give me an hour to test this more. I may be messing up somewhere and don't wanna waste your time..", "If you can't import `transformers.tokenization_dpr_fast` that must be an environment issue. This file was added recently on the master branch. It looks like your env imports a version of transformers that is not up-to-date. That's why you're having an OSError.\r\n\r\nMaybe you can run `pip show transformers` to check the location of your transformers installation ?", "> If you can't import `transformers.tokenization_dpr_fast` that must be an environment issue. This file was added recently on the master branch. It looks like your env imports a version of transformers that is not up-to-date. That's why you're having an OSError.\r\n> \r\n> Maybe you can run `pip show transformers` to check the location of your transformers installation ?\r\n\r\nOk, I just did the env again for the 3rd time and it works! No idea what i messed up before. \r\n\r\nStep -3 Load Rag in progress :)\r\n", "Slow tests pass. This one is ready to merge.\r\n", "@lhoestq - feel free to merge whenever!", "When I do \"from transformers import DPRContextEncoder\", I get an error :\r\nFile \"convert_slow_tokenizer.py\", line 24, in <module>\r\n from tokenizers.models import BPE, Unigram, WordPiece\r\nImportError: cannot import name 'Unigram'\r\n\r\nUnigram is missing in tokenizers 0.8.1.rc2. Needed to update to 0.9.0", "> I took your comments into account, let me know if you have other things to improve.\r\n> Also I had to change the DPR encoder from the one trained on Natural Questions to the one trained ont the multiset/hybrid dataset to match the embeddings used by the Rag team.\r\n\r\n@lhoestq \r\n\r\nHi, can you elaborate on the change you made in the DPR bit more? My understanding is, you have pretrained the DPR with a hybrid dataset to improve the performance when encoding a custom knowledge base. \r\n\r\nIf you have pretrained the DPR, can you please publish the code? \r\n\r\nCan you please refer to this issue also. \r\n\r\nhttps://github.com/huggingface/transformers/issues/8037\r\n\r\nThanks a lot.", "There are two versions of DPR in the paper. One trained on NQ and one trained on various datasets. The authors released the code and the weights in this [this repo](https://github.com/facebookresearch/DPR).\r\n\r\nThe change I did was just to use the weight of the second one, since it was the one used for RAG.", "Thanks. So as mentioned in the RAG paper, we can use the doc encoder to get\nembeddings for any custom dataset. Later we can only fine tune the BART\nmodel and the question encoder.\n\nOn Tue, Oct 27, 2020, 01:42 Quentin Lhoest <[email protected]> wrote:\n\n> There are two versions of DPR in the paper. One trained on NQ and one\n> trained on various datasets. The authors released the code and the weights\n> in this this repo <https://github.com/facebookresearch/DPR>.\n>\n> The change I did was just to use the weight of the second one, since it\n> was the one used for RAG.\n>\n> —\n> You are receiving this because you commented.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/pull/7763#issuecomment-716521222>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AEA4FGUJ5EC6J2ROO56RXB3SMVVDPANCNFSM4SPHCGQQ>\n> .\n>\n", "Yes exactly :) ", "@lhoestq , I tried use_own_knowledge_dataset.py to try and retrieve passages from a custom dataset in response to queries. It works, though the relevance of the results isn't great. \r\nI wanted to try and use finetune.sh. Is it possible to finetune just the retriever ? Is there a sample format for the training data ? Looks like it will need the question and passages as input. Need positive and negative passages ? \r\nThanks!", "@mchari He previously said (https://github.com/huggingface/transformers/issues/8037), it is possible to retrain DPR with Facebook code and then convert the checkpoint to the huge face compatible.\r\n\r\nWhat if you fine-tune the RAG system with your own data letting the question encoder to get fine-tuned? It seems like the pre-training of DPR can be a hard task since the results can depend on the selection procedure of the negative samples as mentioned in the paper. " ]

[]

[]

[]

[]

[]

[ "Do you want to post the full error message? (and the information asked in the issue template)", "These models have been added last months, so they shouldn't have changed much. The full issue template filled would be very helpful here!", "This is the error: \r\nValueError: Unrecognized model identifier: bert-generation. Should contain one of retribert, t5, mobilebert, distilbert, albert, camembert, xlm-roberta, pegasus, marian, mbart, bart, reformer, longformer, roberta, flaubert, bert, openai-gpt, gpt2, transfo-xl, xlnet, xlm, ctrl, electra, encoder-decoder\r\n\r\nUsing exactly the code appearing in the link I passed:\r\n```python\r\nfrom transformers import AutoTokenizer, AutoModelForSeq2SeqLM\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(\"google/bert2bert_L-24_wmt_de_en\", pad_token=\"<pad>\", eos_token=\"</s>\", bos_token=\"<s>\")\r\nmodel = AutoModelForSeq2SeqLM.from_pretrained(\"google/bert2bert_L-24_wmt_de_en\")\r\n\r\nsentence = \"Willst du einen Kaffee trinken gehen mit mir?\"\r\n\r\ninput_ids = tokenizer(sentence, return_tensors=\"pt\", add_special_tokens=False).input_ids\r\noutput_ids = model.generate(input_ids)[0]\r\nprint(tokenizer.decode(output_ids, skip_special_tokens=True))\r\n```\r\n\r\nTransformers version: 3.1.0\r\n", "@patrickvonplaten what should be done here? The `BertGeneration` model cannot be loaded directly through the `AutoModelForSeq2SeqLM` auto-model, can it?", "Then how could I load it?\r\n", "Hey @alexvaca0 - the google/encoder-decoder models were released in transformers 3.2.0 => so you will have to update your transformers version for it :-) It should then work as expected.", "Ohhh so sorry, my bad :( Thanks a lot for the quick response! :) ", "I think this may not have been fully resolved? I'm getting a simmilar error:\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/transformers/modeling_utils.py\", line 926, in from_pretrained\r\n state_dict = torch.load(resolved_archive_file, map_location=\"cpu\")\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/serialization.py\", line 527, in load\r\n with _open_zipfile_reader(f) as opened_zipfile:\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/serialization.py\", line 224, in __init__\r\n super(_open_zipfile_reader, self).__init__(torch._C.PyTorchFileReader(name_or_buffer))\r\nRuntimeError: version_ <= kMaxSupportedFileFormatVersion INTERNAL ASSERT FAILED at /opt/conda/conda-bld/pytorch_1579022060824/work/caffe2/serialize/inline_container.cc:132, please report a bug to PyTorch. Attempted to read a PyTorch file with version 3, but the maximum supported version for reading is 2. Your PyTorch installation may be too old. (init at /opt/conda/conda-bld/pytorch_1579022060824/work/caffe2/serialize/inline_container.cc:132)\r\nframe #0: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x47 (0x7fd86f9d2627 in /home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/lib/libc10.so)\r\nframe #1: caffe2::serialize::PyTorchStreamReader::init() + 0x1f5b (0x7fd82fbbb9ab in /home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/lib/libtorch.so)\r\nframe #2: caffe2::serialize::PyTorchStreamReader::PyTorchStreamReader(std::string const&) + 0x64 (0x7fd82fbbcbc4 in /home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/lib/libtorch.so)\r\nframe #3: <unknown function> + 0x6d2146 (0x7fd87067e146 in /home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/lib/libtorch_python.so)\r\nframe #4: <unknown function> + 0x28ba06 (0x7fd870237a06 in /home/ubuntu/anaconda3/lib/python3.7/site-packages/torch/lib/libtorch_python.so)\r\n<omitting python frames>\r\nframe #37: __libc_start_main + 0xe7 (0x7fd87474cb97 in /lib/x86_64-linux-gnu/libc.so.6)\r\n\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"wmt_test.py\", line 26, in <module>\r\n model = AutoModelForSeq2SeqLM.from_pretrained(\"google/bert2bert_L-24_wmt_de_en\").to(device)\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/transformers/modeling_auto.py\", line 1073, in from_pretrained\r\n return model_class.from_pretrained(pretrained_model_name_or_path, *model_args, config=config, **kwargs)\r\n File \"/home/ubuntu/anaconda3/lib/python3.7/site-packages/transformers/modeling_utils.py\", line 929, in from_pretrained\r\n \"Unable to load weights from pytorch checkpoint file. \"\r\nOSError: Unable to load weights from pytorch checkpoint file. If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True.\r\n```\r\n\r\npython = 3.76, torch==1.4.0, transformers==3.2.0", "Hi @ezekielbarnett could you open a new issue and fill the issue template? A reproducible code example would be particularly helpful here." ]

[]

[]

[ "As the error says, you've applied a `.detach()` method to a model output, which are *always* tuples. You can check the [documentation here](https://huggingface.co/transformers/main_classes/output.html).\r\n\r\nYou probably want the first output of your model so change this line:\r\n```py\r\nbatch_output = model(batch_data, token_type_ids=None, attention_mask=batch_masks) \r\n```\r\nto\r\n```py\r\nbatch_output = model(batch_data, token_type_ids=None, attention_mask=batch_masks)[0]\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 } ]

[]

[ "Hey @laibamehnaz , that's a great question for forum\r\nhttps://discuss.huggingface.co/.\r\n\r\nCould you post it there, someone might have tried it and the forum is better to discuss such questions :) ", "Great. I will post it there. Thanks a lot.", "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" ]

[]

[]

[ "Got 2 offline yes! from @thomwolf and @LysandreJik => merging." ]