Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "ok" ]

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

[]

[ "You could fix it with something like this\r\n```\r\nbatch_size = torch.tensor(data.shape[1]).to(device)\r\ndist.all_reduce(batch_size, op=dist.ReduceOp.SUM)\r\ndist.all_reduce(loss, op=dist.ReduceOp.SUM)\r\nmean_loss = loss/batch_size\r\n```", "Thanks for your solution. I don't think it could fix it, because the `loss` returned by each GPU is already averaged over its `batch_size`.", "@Lvzhh the examples (like `run_squad.py`) only work in two settings for multi-gpu:\r\n\r\n- using `DataParallel` (not distributed, triggered when you simply run the script on a multi-gpu server): in this case the batch size should be a multiple of the number of gpu or an error is thrown. This is the case in which this `loss = loss.mean()` will be triggered. No risk of differing batch sizes here.\r\n- using `DistributedDataParallel` with a *single* GPU assigned to each process (triggered when you run the example using PyTorch `distributed.launch` script for instance): this only works when 1 GPU is assigned to each process (see [here](https://github.com/huggingface/pytorch-pretrained-BERT/blob/3fc63f126ddf883ba9659f13ec046c3639db7b7e/examples/run_squad.py#L853-L857)). In this case `loss = loss.mean()` is not triggered. There is no averaging of the loss over the GPU. What happens here is that it's the gradients which are averaged (read PyTorch [DDP doc](https://pytorch.org/docs/stable/nn.html#torch.nn.parallel.DistributedDataParallel) for instance). The batch size should be the same for each script so no problem here also.\r\n\r\nSo I don't think the problem you are describing will show up in practice.\r\n", "@thomwolf I'm using the first multi-gpu setting `DataParallel`. And I tried to find some code in `run_squad.py` to ensure that the `batch_size` be a multiple of the number of gpu, but I didn't find.\r\n\r\nAnd I just run this command [here](https://github.com/huggingface/pytorch-pretrained-BERT#squad) with `--train_batch_size 13` using 2 GPUs, and print the size of `start_logits` and its device, the first dimension (`batch_size`) on each GPUs are actually different (7 vs 6).\r\n\r\nhttps://github.com/huggingface/pytorch-pretrained-BERT/blob/3fc63f126ddf883ba9659f13ec046c3639db7b7e/pytorch_pretrained_bert/modeling.py#L1204-L1206", "Oh you are right, this error is not there anymore.\r\nWe should probably just check for this again, I'll fix the examples.\r\n\r\n`DataParallel` is in the process of being replaced by `DistributedDataParallel` which is pretty much always faster due to releasing the GiL so maybe you should try DDP.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

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

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

[]

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

[ { "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" ]

[]

[]

[ "Looks good.", "Ok merging thanks, sorry for the delay!" ]

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

[]

[ "i have same question ", "How best to fine-tune or pool BERT is an open question in bertology :p \r\n\r\n[\"How to Fine-Tune BERT for Text Classification?\"](https://arxiv.org/pdf/1905.05583.pdf) has a comprehensive overview. Look at table 3 specifically which found that taking the max of the last 4 layers achieves the best performance (in their testing configuration with some fine tuning), this would be my guess. \r\n\r\nYou can also look at the approaches taken by [bert-as-a-service](https://github.com/hanxiao/bert-as-service#q-what-are-the-available-pooling-strategies) which are explained pretty well, in this repo they use MEAN as the default pooling strategy.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

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

[ "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", "@AdamDanielKing did you manage to fix the divergence somehow? ", "Was this finally an issue? It seems important and it was closed due to inactivity. @AdamDanielKing did you work it around?", "@Damiox While sampling with mixed precision gives different results, they seem to still be of high quality. I've been using mixed precision on [talktotransformer.com](https://talktotransformer.com) for at least 6-7 months now and the quality has been excellent.", "I am not using the \"past\" output at the moment, just doing topk=30. I ran into this issue as I am trying to improve my inferences times.\r\nI have apex installed, and using the gpt2-medium model I get in average per sentence an inference time of 11ms for ~50 tokens within a batch size of 30 sentences on Tesla T4. Bigger batches aren't increasing the throughput. I just tried turning on fp16 into my model via \".half()\" and it seems to be 3x faster. Is it possible? I am wondering whether that is fine or I need to do anything else (e.g. initializing apex, re-training my model with fp16). I feel I may be losing something. What do you think?", "Talktotransformer.com just uses Apex to cast the model to fp16--no\nretraining. I use the opt level that casts the weights themselves as well\n(allows larger batch sizes). It seems to work well.\n\nIf you're getting 3x improvement from calling .half() instead of\ninitializing with Apex, that is strange and I can't imagine why. I've found\nthat this methods diverges more quickly from fp32 than using Apex so I\nhaven't tested much with it.\n\nOn a separate note, I'd consider trying cheaper GPUs. It may be\ncounter-intuitive, but I've found for example on Google that cheap P4 GPUs\ngive greater throughput _for their cost_ than a T4 even though they don't\nhave tensor cores. I think this is for the following reason: in generating\neach token after the first, only the values for a single position are being\ncomputed at each iteration, which is very little computation for the amount\nof memory being used. I think this results in a lot of time being spent on\nGPU-CPU round trips rather than actual work. Batch size becomes more\nimportant than GPU speed.\n\nOn Mon, Mar 9, 2020, 5:56 PM Damian Nardelli, <[email protected]>\nwrote:\n\n> I am not using the \"past\" output at the moment, just doing topk=30. I ran\n> into this issue as I am trying to improve my inferences times.\n> I have apex installed, and using the gpt2-medium model I get in average\n> per sentence an inference time of 11ms for ~50 tokens within a batch size\n> of 30 sentences on Tesla T4. Bigger batches aren't increasing the\n> throughput. I just tried adding turning fp16 into my model via \".half()\"\n> and it seems to be 3x faster. Is it possible? I am wondering whether that\n> is fine or I need to do anything else (e.g. initializing apex, re-training\n> my model with fp16). I feel I may be losing something. What do you think?\n>\n> —\n> You are receiving this because you were mentioned.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/602?email_source=notifications&email_token=ABKUZHOUB74EQKNVRORWKB3RGVQX7A5CNFSM4HMG65OKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEOJG2FA#issuecomment-596798740>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/ABKUZHOS54LHVYM4OSGCR7LRGVQX7ANCNFSM4HMG65OA>\n> .the\n>\n", "I haven't tried initializing apex explicitly in my code. I believe I just thought that was being done by `GPT2LMHeadModel` behind the scenes, but it doesn't look like... I had installed it because I saw the following warning when running my inference app: `Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex`. Does this mean that the gpt2-medium is compiled with apex support? I'm not certain about that.\r\n\r\nShould I try doing the amp.initialize() with the opt level in my inference app? Are you using O1? From https://nvidia.github.io/apex/amp.html#opt-levels-and-properties looks like I shouldn't be explicitly calling `half()`, so probably I should try initializing apex. What do you think?\r\n\r\nThanks in advance for all your help!", "I had a question related to this: would the outputs in generation with GPT-2 change if the batch size changes?", "Currently generation only allows `batch_size=1`" ]

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

[]

[ "You need to retrain it with your embeddings replacing `BertModel.embeddings.word_embeddings` and the model size being your embeddings size.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

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

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

[]

[ "Hi Adrian, BERT already has a few unused tokens that can be used similarly to the `special_tokens` of GPT/GPT-2.\r\nFor more details see https://github.com/google-research/bert/issues/9#issuecomment-434796704 and issue #405 for instance.", "In case we use an unused special token from the vocabulary, is it enough to finetune a classification task or do we need to train an embedding from scratch? Did anyone already do this?\r\n\r\nTwo different and somehow related questions I had when looking into the implementation:\r\n\r\n1) The Bert paper mentions a (learned) positional embedding. How is this implemented here? examples/extract_features/convert_examples_to_features() defines tokens (representation), input_type_ids (the difference between the first and second sequence) and an input_mask (distinguishing padding/real tokens) but no positional embedding. Is this done internally?\r\n\r\n2) Can I use a special token as input_type_ids for Bert? In the classification example, only values of [0,1] are possible and I'm wondering what would happen if I would choose a special token instead? Is this possible with a pretrained embedding or do i need to retrain the whole embedding as a consequence?\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" ]

[]

[]

[ "Oh great thanks Burc!" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597?src=pr&el=h1) Report\n> :exclamation: No coverage uploaded for pull request base (`master@f9cde97`). [Click here to learn what that means](https://docs.codecov.io/docs/error-reference#section-missing-base-commit).\n> The diff coverage is `81%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #597 +/- ##\n=========================================\n Coverage ? 67.04% \n=========================================\n Files ? 18 \n Lines ? 3835 \n Branches ? 0 \n=========================================\n Hits ? 2571 \n Misses ? 1264 \n Partials ? 0\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/tokenization\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX29wZW5haS5weQ==) | `81.34% <0%> (ø)` | |\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `82.44% <20%> (ø)` | |\n| [pytorch\\_pretrained\\_bert/modeling\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfb3BlbmFpLnB5) | `78.3% <68.51%> (ø)` | |\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `79.04% <73.11%> (ø)` | |\n| [pytorch\\_pretrained\\_bert/modeling.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmcucHk=) | `88.57% <98.09%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597?src=pr&el=footer). Last update [f9cde97...35e6baa](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/597?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n", "Hi, is the script run_openai.py updating weights of all layers or just the last few? Thanks in advance!" ]

[]

[]

[ "Hi @shoegazerstella well XLM is already pretty much as powerful as BERT and focused on cross-lingual sentence representations so I would go directly for it instead of BERT.", "Thanks @thomwolf, \r\nAre you considering integrating something for cross-lingual representations in the `pytorch-pretrained-BERT` library in the near future?", "Not in the short-term" ]

[]

[]

[ "Yes, this error message hides several potential sources, I'll see if I can disentangle the error messages :) ", "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 is still an issue. I suggest to improve the message and raise an exception if unable to load any of the models, instead of silently returning `None`.", "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", "Still an issue, as far as I know.", "@czyzby Hello，How do you solve the cache problem? I have the same problem but can't fix it", "@twothousand To be honest, I don't remember, but I think the directory did not exist or lacked write permission. Are you sure the cache is causing the problem? If you changed the cache directory, make sure the folder exists and has appropriate permissions - otherwise I'd debug model loading and see which exception is being ignored.\r\n\r\n@thomwolf It seems that it's still an issue, will you look into proper error handling?", "Yes, it should have been improved on the latest release 2.1.1 (with the merge of #1480).\r\n\r\nMaybe open a new issue with clear details of the current issue?" ]

[ { "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", "```Traceback (most recent call last):\r\n File \"question-generation/interact.py\", line 238, in <module>\r\n run()\r\n File \"question-generation/interact.py\", line 144, in run\r\n model = GPT2LMHeadModel.from_pretrained(args.model_checkpoint)\r\n File \"/usr/local/lib/python3.6/dist-packages/pytorch_pretrained_bert/modeling_gpt2.py\", line 475, in from_pretrained\r\n \"Error(s) in loading state_dict for {}:\\n\\t{}\".format(model.__class__.__name__, \"\\n\\t\".join(error_msgs))\r\nRuntimeError: Error(s) in loading state_dict for GPT2LMHeadModel:\r\n\tsize mismatch for transformer.wte.weight: copying a param with shape torch.Size([50265, 768]) from checkpoint, the shape in current model is torch.Size([50257, 768]).\r\n\tsize mismatch for lm_head.decoder.weight: copying a param with shape torch.Size([50265, 768]) from checkpoint, the shape in current model is torch.Size([50257, 768]).\r\n```\r\nI am having a similar issue? Any idea how to solve this? @Wingie ", "+1", "+1", "+1\r\n", "> ```\r\n> File \"question-generation/interact.py\", line 238, in <module>\r\n> run()\r\n> File \"question-generation/interact.py\", line 144, in run\r\n> model = GPT2LMHeadModel.from_pretrained(args.model_checkpoint)\r\n> File \"/usr/local/lib/python3.6/dist-packages/pytorch_pretrained_bert/modeling_gpt2.py\", line 475, in from_pretrained\r\n> \"Error(s) in loading state_dict for {}:\\n\\t{}\".format(model.__class__.__name__, \"\\n\\t\".join(error_msgs))\r\n> RuntimeError: Error(s) in loading state_dict for GPT2LMHeadModel:\r\n> \tsize mismatch for transformer.wte.weight: copying a param with shape torch.Size([50265, 768]) from checkpoint, the shape in current model is torch.Size([50257, 768]).\r\n> \tsize mismatch for lm_head.decoder.weight: copying a param with shape torch.Size([50265, 768]) from checkpoint, the shape in current model is torch.Size([50257, 768]).\r\n> ```\r\n> \r\n> I am having a similar issue? Any idea how to solve this? @Wingie\r\n\r\nWere you able to fix this? I am having this issue, with the same sizes as well.", "same here!\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 } ]

[]

[ "@Dhanachandra Same issue with another pre-trained BERT model.\r\nHave you managed to solve that?", "> @Dhanachandra Same issue with another pre-trained BERT model.\r\n> Have you managed to solve that?\r\n\r\n@Dhanachandra I've just found the solution: this exception occurs at \"modeling.py\" module. It's because of tensorflow->pytorch transformation. You need to find rows where 'shape' is used in modeling.py (you'll see its path in error logs) and delete it (it's somewhere in try... assert ... except..., just delete it).", "You can use the following code: \r\n\r\ntf_path = 'pubmed_pmc_470k/biobert_model.ckpt'\r\nconfig_path = 'pubmed_pmc_470k/bert_config.json'\r\npytorch_dump_path = 'pytorch_model/pytorch_model.bin'\r\n# Save pytorch-model\r\n\r\nimport os\r\nimport re\r\nimport argparse\r\nimport tensorflow as tf\r\nimport torch\r\nimport numpy as np\r\n\r\nfrom pytorch_pretrained_bert import BertConfig, BertForPreTraining\r\n\r\ndef convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytorch_dump_path):\r\n config_path = os.path.abspath(bert_config_file)\r\n tf_path = os.path.abspath(tf_checkpoint_path)\r\n print(\"Converting TensorFlow checkpoint from {} with config at {}\".format(tf_path, config_path))\r\n # Load weights from TF model\r\n init_vars = tf.train.list_variables(tf_path)\r\n excluded = ['BERTAdam','_power','global_step']\r\n init_vars = list(filter(lambda x:all([True if e not in x[0] else False for e in excluded]),init_vars))\r\n names = []\r\n arrays = []\r\n for name, shape in init_vars:\r\n print(\"Loading TF weight {} with shape {}\".format(name, shape))\r\n array = tf.train.load_variable(tf_path, name)\r\n names.append(name)\r\n arrays.append(array)\r\n\r\n # Initialise PyTorch model\r\n config = BertConfig.from_json_file(bert_config_file)\r\n print(\"Building PyTorch model from configuration: {}\".format(str(config)))\r\n model = BertForPreTraining(config)\r\n\r\n for name, array in zip(names, arrays):\r\n name = name.split('/')\r\n # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v\r\n # which are not required for using pretrained model\r\n if any(n in [\"adam_v\", \"adam_m\", \"global_step\"] for n in name):\r\n print(\"Skipping {}\".format(\"/\".join(name)))\r\n continue\r\n pointer = model\r\n for m_name in name:\r\n if re.fullmatch(r'[A-Za-z]+_\\d+', m_name):\r\n l = re.split(r'_(\\d+)', m_name)\r\n else:\r\n l = [m_name]\r\n if l[0] == 'kernel' or l[0] == 'gamma':\r\n pointer = getattr(pointer, 'weight')\r\n elif l[0] == 'output_bias' or l[0] == 'beta':\r\n pointer = getattr(pointer, 'bias')\r\n elif l[0] == 'output_weights':\r\n pointer = getattr(pointer, 'weight')\r\n else:\r\n pointer = getattr(pointer, l[0])\r\n if len(l) >= 2:\r\n num = int(l[1])\r\n pointer = pointer[num]\r\n if m_name[-11:] == '_embeddings':\r\n pointer = getattr(pointer, 'weight')\r\n elif m_name == 'kernel':\r\n array = np.transpose(array)\r\n try:\r\n assert pointer.shape == array.shape\r\n except AssertionError as e:\r\n e.args += (pointer.shape, array.shape)\r\n raise\r\n print(\"Initialize PyTorch weight {}\".format(name))\r\n pointer.data = torch.from_numpy(array)\r\n\r\n # Save pytorch-model\r\n print(\"Save PyTorch model to {}\".format(pytorch_dump_path))\r\n torch.save(model.state_dict(), pytorch_dump_path)\r\n\r\nconvert_tf_checkpoint_to_pytorch(tf_path, config_path, pytorch_dump_path) `", "@DmLitov4 ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "I think so. Ultimately you should have s1 and s2 in input, your [CLS] + s1 + s2 + [SEP] will be equivalent to `[CLS] + s1 + [SEP]` in `[CLS] + s1 + [SEP] + s2 + [SEP]` where `s1` now is the concatenation of `s1` and `s2`. I don't think that's what you want to do." ]

[]

[]

[ "@RomanShen What is your observation on your question" ]

[]

[]

[ "Closed in favour of #604 " ]

[ { "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" ]

[]

[]

[ "Which commands are you running?", "I am so sorry that I took your time, I accidentally posted this here. " ]

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

[]

[ "When BERT is fine-tuned, all layers are trained - this is quite different from fine-tuning in a lot of other ML models, but it matches what was described in the paper and works quite well (as long as you only fine-tune for a few epochs - it's very easy to overfit if you fine-tune the whole model for a long time on a small amount of data!)", "Thank you. May I ask what the difference of this approach is from pre-training? ", "The original model weights are used for initialization, whereas for a model trained from scratch, the weights are initialized randomly.", "Thank you, I got it now.", "Came upon this when searching for an answer to a related question. \r\n\r\nWhen adding a dense layer on top for a classification task, do the model weights for BERT get updated or only the dense layer(are the BERT layers frozen or unfrozen during training)? I ask b/c when training a classifier on the stack overflow tags dataset which contains 40.000 posts with tags in 20 classes I got some unusual results. I trained base-uncased and base-cased and what is weird is that after the first epoch, the test set prediction remain the same. By that I mean exactly the same. In other words, with a 80/20 split (32.000 posts in train set / 8.000 posts in test set) it doesn't matter if you are doing 1, 2 or 3 epochs, the test set prediction don't change. It stays at 83.875% for uncased and 83.224% for cased. The weird thing is that the training loss goes down. \r\n\r\nI have put the actual predictions in a pandas dataframe and the predictions in epoch 1, 2 and 3 are exactly the same.\r\n\r\n", "When a classifier is trained, all the model weights get updated, not just the weights in the classifier layer, so I would expect some overfitting if you train on a small labelled dataset for a lot of epochs. \r\nThe behaviour you've described is unusual - have you tried varying the learning rate, or making a much smaller training set, training on it for 1 epoch only and seeing what the results look like? It might indicate a problem with the data.", "That's what I thought. I tested training the uncased version with 20% of the dataset (training set 6400 and testing set 1600) which gave me an eval accuracy of 0.76875 after epoch 1 and 2. The eval loss is even the excact same value ( 0.7407131800800562 )\r\n\r\nI ran eval before starting the training which gave an accuracy of 0.05 which makes sense with 20 classes and random weights. Then after epoch 1 it jumps up to aforementioned values and stays the same in epoch 2 and 3.\r\n\r\nAny pointers on how to debug this? Might it help checking the gradients?", "Yeah, that's where I'd look. If forced to guess, I'd say the network isn't really getting any input, and is therefore just learning the bias in the last layer. So you could try inspecting the data batches in the training loop right before they enter the network, to make sure there's actually data there and that the data matches the labels, and also checking that most network parameters are getting some gradient after each batch. If your code is in a repo somewhere, feel free to link it and I'll take a look.", "I went through the training data and it appears that its formatted the right way. I also checked the gradients and they are adjusted after each back() call. I think this might be related to the warm_up part of the adjustable learning rate.\r\n\r\nIt happens after epoch 3:\r\n\r\n<img width=\"413\" alt=\"Screen Shot 2019-05-17 at 20 13 33\" src=\"https://user-images.githubusercontent.com/3185711/57953675-49f04000-78e0-11e9-8976-a5367bc5b0f3.png\">\r\n\r\nThen I also get a warning: \r\n\r\n05/17/2019 20:15:02 - WARNING - pytorch_pretrained_bert.optimization - Training beyond specified 't_total'. Learning rate multiplier set to 0.0. Please set 't_total' of WarmupLinearSchedule correctly.\r\n\r\nI am using a default value of 0.1. I plotted the learning rate over 4 epochs and in epoch 4 the learning rate becomes negative: \r\n\r\n<img width=\"436\" alt=\"Screen Shot 2019-05-17 at 21 03 27\" src=\"https://user-images.githubusercontent.com/3185711/57956376-482a7a80-78e8-11e9-987b-657198f19ef5.png\">\r\n\r\n\r\n3 epochs is more then enough for this dataset as it starts to overfit quickly. I just want to understand why this happens, it doesn't make sense to me. The loss and accuracy in the evaluation phase is exactly the same(and the training loss drops in epoch no 4 when the LR is negative). I put the code on Kaggle if you want to take a look ( no pressure :-) )\r\n\r\nhttps://www.kaggle.com/stoddur/bert-so-classification-test/edit\r\n\r\nIm going to play a bit with the warm_up function and see which learning rates are set with different values. Will let you know if I find out anything else.", "In the BERT paper, and in this repo, the learning rate is 'warmed up' from 0 to the maximum over the first 10% of training, and then linearly decays back to 0 for the remaining 90% of training. In order for that to work, the learning rate scheduler needs to know how many steps there will be in training in total (i.e. steps_per_epoch * num_epochs). It seems like that value is being passed incorrectly, causing the LR to decay to zero too quickly and therefore freezing all the weights.\r\n\r\nAlso, I can't see the code at your link - is it possibly private?", "Yeah, I noticed that now reading through the paper :)\r\n\r\nMade the kernel public, the code is a bit rough, hope it makes sense to you. ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "I get error below while running the program.. Did I do any mistake?\r\n\r\n warmup_linear = WarmupLinearSchedule( warmup=args.warmup_proportion,\r\n t_total=num_train_optimization_steps)\r\n\r\nlr_this_step = args.learning_rate * warmup_linear.get_lr(num_train_optimization_steps, \r\n args.warmup_proportion)\r\n\r\n \r\n\r\n**WARNING - pytorch_pretrained_bert.optimization - Training beyond specified 't_total'. Learning rate multiplier set to 0.0. Please set 't_total' of WarmupLinearSchedule correctly.**\r\n\r\n\r\n", "@kbulutozler \r\n\r\n\r\n@steindor .. did you solve WarmupLinearSchedule issue ? I am getting same error .. I tried your kaggle code but getting error that \" the link does not exists\"\r\n\r\nI get error below while running the program.. Did I do any mistake?\r\n\r\nwarmup_linear = WarmupLinearSchedule( warmup=args.warmup_proportion,\r\nt_total=num_train_optimization_steps)\r\n\r\nlr_this_step = args.learning_rate * warmup_linear.get_lr(num_train_optimization_steps,\r\nargs.warmup_proportion)\r\n\r\nWARNING - pytorch_pretrained_bert.optimization - Training beyond specified 't_total'. Learning rate multiplier set to 0.0. Please set 't_total' of WarmupLinearSchedule correctly." ]

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

[]

[ "For these causal models that consider the left-context only, it's ok not to worry too much about padding since the attention modules only look to the previous tokens. Just be careful when you compute the loss to ignore the out-of-sentence-tokens (using loss functions `ignore_index` for instance).", "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 have a related doubt. In the example code [run_language_modeling.py](https://github.com/huggingface/transformers/blob/master/examples/run_language_modeling.py), GPT-2 and transformer-xl throw error due to lacking padding token.\r\n\r\nPlease check recent comments in this [Issue](https://github.com/huggingface/transformers/issues/3021)" ]

[]

[]

[ "Ok, good to go, thanks @huntzhan!" ]

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

[]

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

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

[]

[ "Do u mean this one? [link](https://news.developer.nvidia.com/nvidia-achieves-4x-speedup-on-bert-neural-network/)", "No, I mean this repo\nhttps://github.com/pytorch/xla/tree/master\n\nLooks like Facebook and Google want to make pytorch on TPU\n\n\nOn May 6, 2019 9:14:25 AM GMT+03:00, chunbo dai <[email protected]> wrote:\n>Do u mean this one?\n>[link](https://news.developer.nvidia.com/nvidia-achieves-4x-speedup-on-bert-neural-network/)\n>\n>-- \n>You are receiving this because you authored the thread.\n>Reply to this email directly or view it on GitHub:\n>https://github.com/huggingface/pytorch-pretrained-BERT/issues/583#issuecomment-489510379\n\n-- \nSent from my Android device with K-9 Mail. Please excuse my brevity.", "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 } ]

[]

[ "For convenience to others, here's the config file for 345M:\r\n\r\n```\r\n{\r\n \"initializer_range\": 0.02,\r\n \"layer_norm_epsilon\": 1e-05,\r\n \"n_ctx\": 1024,\r\n \"n_embd\": 1024,\r\n \"n_head\": 16,\r\n \"n_layer\": 24,\r\n \"n_positions\": 1024,\r\n \"vocab_size\": 50257\r\n}\r\n```", "Here are the concrete steps if you'd like to run the 345M.\r\n\r\nGrab OpenAI's download script from here https://github.com/openai/gpt-2/blob/master/download_model.py. and then run `python download_model.py 345M` to get the model checkpoint.\r\n\r\nThen use the conversion script here https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/convert_gpt2_checkpoint_to_pytorch.py using `python convert_gpt2_checkpoint_to_pytorch.py --gpt2_checkpoint_path gpt2_checkpoint_folder --gpt2_config_file config_file --pytorch_dump_folder_path output_dir`\r\n\r\nwhere config_file is the json posted by @daemon above. \r\n\r\nThen inside https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling_gpt2.py modify the PRETRAINED_MODEL_ARCHIVE_MAP and PRETRAINED_CONFIG_ARCHIVE_MAP to point to the converted pytorch file\r\n\r\n", "Thanks!\r\n\r\n> Then inside https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling_gpt2.py modify the PRETRAINED_MODEL_ARCHIVE_MAP and PRETRAINED_CONFIG_ARCHIVE_MAP to point to the converted pytorch file\r\n\r\nOr GPT2LMHeadModel.from_pretrained(pytorch_dump_folder_path) without changing modeling_gpt2.py?", "Why not add this in the module?\r\n\r\nThanks for the instruction, I will likely try if its not integrated soon.", "When running \"convert_gpt2_checkpoint_to_pytorch.py --gpt2_checkpoint_path gpt2_checkpoint_folder --gpt2_config_file config_file --pytorch_dump_folder_path output_dir\" I get the following error:\r\n\r\n_runfile('C:/Users/nietop1/Desktop/anaconda/trying to generate text/convert_checkpoint_gtp2.py', wdir='C:/Users/nietop1/Desktop/anaconda/trying to generate text')\r\nConverting TensorFlow checkpoint from C:\\Users\\nietop1\\Desktop\\anaconda\\models\\345M\r\nTraceback (most recent call last):\r\n\r\n File \"<ipython-input-32-bd0ca7f018f3>\", line 1, in <module>\r\n runfile('C:/Users/nietop1/Desktop/anaconda/trying to generate text/convert_checkpoint_gtp2.py', wdir='C:/Users/nietop1/Desktop/anaconda/trying to generate text')\r\n\r\n File \"C:\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\spyder\\utils\\site\\sitecustomize.py\", line 705, in runfile\r\n execfile(filename, namespace)\r\n\r\n File \"C:\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\spyder\\utils\\site\\sitecustomize.py\", line 102, in execfile\r\n exec(compile(f.read(), filename, 'exec'), namespace)\r\n\r\n File \"C:/Users/nietop1/Desktop/anaconda/trying to generate text/convert_checkpoint_gtp2.py\", line 81, in <module>\r\n 'C:/Users/nietop1/Desktop/anaconda/models/345M')\r\n\r\n File \"C:/Users/nietop1/Desktop/anaconda/trying to generate text/convert_checkpoint_gtp2.py\", line 47, in convert_gpt2_checkpoint_to_pytorch\r\n load_tf_weights_in_gpt2(model, gpt2_checkpoint_path)\r\n\r\n File \"C:\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\pytorch_pretrained_bert\\modeling_gpt2.py\", line 60, in load_tf_weights_in_gpt2\r\n init_vars = tf.train.list_variables(tf_path)\r\n\r\nAttributeError: module 'tensorflow.python.training.training' has no attribute 'list_variables'_\r\n\r\nHow can this be solved?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "Closing this because this is merged." ]

[ { "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" ]

[]

[]

[ "The `convert_tf_checkpoint_to_pytorch` script is made to convert the Google pre-trained weights in `BertForPretraining` model, you have to modify it to convert another type model.\r\n\r\nIn your case, you want to load the passage re-ranking model in a `BertForSequenceClassification` model which has the same structure (BERT + a classifier on top of the pooled output) as the NYU model.\r\n\r\nhere is a quick way to do that:\r\n- install pytorch-pretrained-bert from source so you can modify it\r\n- change https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py#L34 to initialize a `BertForSequenceClassification` model instead of the `BertForPreTraining` model in the conversion script.\r\n- the structure is not exactly identical so you need to ADD a line that say `pointer = getattr(pointer, 'cls')` in the TWO if-conditions related to `output_weights` and `output_bias` (between L89 and L90 and between L91 and L92 in modeling.py here: https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling.py#L90 and https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling.py#L92).\r\n- this should let you convert the tensorflow model in a pytorch one using the scripts.", "Thanks so much! Your comment saved me a lot of time. However there was a small issue I got around by just changing the tf variable names.\r\n\r\nFor anyone else out there the solution was\r\n* https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py#L34 CHANGE `model = BertForSequenceClassification(config, 2)`\r\n\r\n* https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling.py#L70 ADD\r\n``` \r\n if name in ['output_weights' , 'output_bias']:\r\n name = 'classifier/' + name\r\n ```\r\n\r\n", "Hello @oisin-dolphin and @thomwolf \r\nI followed above suggestions but getting following error.\r\ntensorflow.python.framework.errors_impl.NotFoundError: Key classifier/output_bias not found in checkpoint\r\n\r\nAlso what is significance of following line of code\r\npointer = getattr(pointer, 'cls') \r\n\r\nPlease suggest.\r\n\r\nThanks\r\nMahesh", "> The `convert_tf_checkpoint_to_pytorch` script is made to convert the Google pre-trained weights in `BertForPretraining` model, you have to modify it to convert another type model.\r\n> \r\n> In your case, you want to load the passage re-ranking model in a `BertForSequenceClassification` model which has the same structure (BERT + a classifier on top of the pooled output) as the NYU model.\r\n> \r\n> here is a quick way to do that:\r\n> \r\n> * install pytorch-pretrained-bert from source so you can modify it\r\n> * change https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py#L34 to initialize a `BertForSequenceClassification` model instead of the `BertForPreTraining` model in the conversion script.\r\n> * the structure is not exactly identical so you need to ADD a line that say `pointer = getattr(pointer, 'cls')` in the TWO if-conditions related to `output_weights` and `output_bias` (between L89 and L90 and between L91 and L92 in modeling.py here: https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling.py#L90 and https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/pytorch_pretrained_bert/modeling.py#L92).\r\n> * this should let you convert the tensorflow model in a pytorch one using the scripts.\r\n\r\nI followed these instructions for the SequenceClassification model but I still end up getting the same error for 'BertForSequenceClassification' object has no attribute 'bias'.", "Update for latest transformers, add modeling_bert.py:78: \r\n```python\r\n for name, array in zip(names, arrays):\r\n if name in ['output_weights', 'output_bias']:\r\n name = 'classifier/' + name\r\n```\r\nand convert_bert_original_tf_checkpoint_to_pytorch.py\r\n```python\r\nconfig.num_labels = 2\r\n print(\"Building PyTorch model from configuration: {}\".format(str(config)))\r\n model = BertForSequenceClassification(config)\r\n\r\n```", "you are my lifesaver @pertschuk Thank you for the instructions", "glad they helped @Soonhwan-Kwon. \r\n\r\nI used a similar reranking model as part of a project I just released which hooks in to Elasticsearch and reranks search results out of the box, [check it out]( https://medium.com/koursaros-ai/boost-search-api-performance-e-g-410868e82b22) if this sounds like it would be useful! repo: https://github.com/koursaros-ai/nboost ", "You can create a subclass of `BertForSequenceClassification` and add `self.weight` and `self.bias` to the` __init__` method. Then instantiate your new class and it is ready to use it:\r\n\r\n```\r\nclass BertForPassageRanking(BertForSequenceClassification):\r\n def __init__(self, config):\r\n super().__init__(config)\r\n self.weight = torch.autograd.Variable(torch.ones(2, config.hidden_size),\r\n requires_grad=True)\r\n self.bias = torch.autograd.Variable(torch.ones(2), requires_grad=True)\r\n\r\n\r\nbert_ranking = BertForPassageRanking.from_pretrained(BERT_PASSAGE_RANKING_PATH,\r\n from_tf=True)\r\n```\r\n\r\n`BERT_PASSAGE_RANKING_PATH` is the path where your tf checkpoints files and config json file are stored. You will need to rename the files as follows:\r\n\r\n```\r\nconfig.json\r\nmodel.ckpt.index\r\nmodel.ckpt.meta\r\n```\r\n\r\nAnother option if you do not want to change the file names is to load the json config file with `BertConfig.from_json_file()` and then pass to `BertForPassageRanking.from_pretained()` the path + ckpt file name and the configuration that you have already loaded with `BertConfig.from_json_file()` .\r\n", "I added passage pytorch msmarco reranking models to the huggingface / transformers bucket, no need for subclassing / modifications. \r\n\r\nhttps://huggingface.co/nboost", "> I added passage pytorch msmarco reranking models to the huggingface / transformers bucket, no need for subclassing / modifications.\r\n> \r\n> https://huggingface.co/nboost\r\n\r\nHi, I have a question regarding the output of your models. In transformers library, the bert_base model (`transformers.BertModel` class) has as output a tuple, where the first element is the last hidden state and the 2nd element is the pooler output. The last hidden state is a tensor of size `(batch_size, sequence_length, hidden_dim)`. For example for a batch size of 64 and 512 tokens we obtain for BERT an output of size `(64x512x768)`. The pooler output has size `(batch_size, hidden_size)`. This output is obtained training a linear layer with tanh activation function which had as input the `CLS` token hidden state (last layer hidden-state of the first oken of the sequence). Those weights have been trained from the next sentence prediction.\r\n\r\nYour model follows similar structure, at least `nboost/pt-biobert-base-msmarco`. However, a passage re-ranking model is a sequence classification model. Basically, the passage re-ranking model proposed by https://github.com/nyu-dl/dl4marco-bert is the BERT model fine-tuned with a dense layer on top to learn to classify a sequence as relevant or not relevant. Their first element of the tuple output is a tensor of size `(batch_size, num_classes)`, where num_classes is two (whether the sequence to classify is a relevant document).\r\n\r\nHow should we use your model for passage re-ranking?\r\nThanks a lot", "> > I added passage pytorch msmarco reranking models to the huggingface / transformers bucket, no need for subclassing / modifications.\r\n> > https://huggingface.co/nboost\r\n> \r\n> Hi, I have a question regarding the output of your models. In transformers library, the bert_base model (`transformers.BertModel` class) has as output a tuple, where the first element is the last hidden state and the 2nd element is the pooler output. The last hidden state is a tensor of size `(batch_size, sequence_length, hidden_dim)`. For example for a batch size of 64 and 512 tokens we obtain for BERT an output of size `(64x512x768)`. The pooler output has size `(batch_size, hidden_size)`. This output is obtained training a linear layer with tanh activation function which had as input the `CLS` token hidden state (last layer hidden-state of the first oken of the sequence). Those weights have been trained from the next sentence prediction.\r\n> \r\n> Your model follows similar structure, at least `nboost/pt-biobert-base-msmarco`. However, a passage re-ranking model is a sequence classification model. Basically, the passage re-ranking model proposed by https://github.com/nyu-dl/dl4marco-bert is the BERT model fine-tuned with a dense layer on top to learn to classify a sequence as relevant or not relevant. Their first element of the tuple output is a tensor of size `(batch_size, num_classes)`, where num_classes is two (whether the sequence to classify is a relevant document).\r\n> \r\n> How should we use your model for passage re-ranking?\r\n> Thanks a lot\r\n\r\nI found where was the problem. As pointed in the model's page (https://huggingface.co/nboost/pt-biobert-base-msmarco#) to load the model you have to do the following:\r\n\r\n`model = AutoModel.from_pretrained(\"nboost/pt-biobert-base-msmarco\")`\r\nThis creates as output a tuple where the first element is a tensor of size `(64x512x768)`.\r\n\r\nHowever, we should do the following, since our problem is a sequence classification:\r\n\r\n`model = AutoModelForSequenceClassification.from_pretrained(\"nboost/pt-biobert-base-msmarco\")`\r\nThis creates the correct output, a tuple where the first element is a tensor of size `(batch_size, num_classes)`\r\n\r\nI suggest to the authors to change the model info and model card in https://huggingface.co/nboost/pt-biobert-base-msmarco#, since it is little bit confusing", "> You can create a subclass of `BertForSequenceClassification` and add `self.weight` and `self.bias` to the` __init__` method. Then instantiate your new class and it is ready to use it:\r\n> \r\n> ```\r\n> class BertForPassageRanking(BertForSequenceClassification):\r\n> def __init__(self, config):\r\n> super().__init__(config)\r\n> self.weight = torch.autograd.Variable(torch.ones(2, config.hidden_size),\r\n> requires_grad=True)\r\n> self.bias = torch.autograd.Variable(torch.ones(2), requires_grad=True)\r\n> \r\n> \r\n> bert_ranking = BertForPassageRanking.from_pretrained(BERT_PASSAGE_RANKING_PATH,\r\n> from_tf=True)\r\n> ```\r\n> \r\n> `BERT_PASSAGE_RANKING_PATH` is the path where your tf checkpoints files and config json file are stored. You will need to rename the files as follows:\r\n> \r\n> ```\r\n> config.json\r\n> model.ckpt.index\r\n> model.ckpt.meta\r\n> ```\r\n> \r\n> Another option if you do not want to change the file names is to load the json config file with `BertConfig.from_json_file()` and then pass to `BertForPassageRanking.from_pretained()` the path + ckpt file name and the configuration that you have already loaded with `BertConfig.from_json_file()` .\r\n\r\n\r\nThanks a lot. I was having the same question about 'nboost' and was trying this method. However, the output seems to change when I run the same code multiple times, even though i am in the eval mode. Do you have any hint about what I am doing wrong here?\r\n\r\n```\r\nbert_ranking = BertForPassageRanking.from_pretrained(BERT_PASSAGE_RANKING_PATH,\r\n from_tf=True)\r\n\r\ndummy_query = [\r\n 'Rutgers is a good university. I like my experience there.', \r\n \"Hello, my dog is cute. My cute dog is amazing.\",\r\n 'Florida is a nice place but tiger king may be better',\r\n]\r\n\r\ndummy_passage = [\r\n 'My cat is really cute but my dog is even better.',\r\n 'My cat is really cute but my dog is even better.',\r\n 'My cat is really cute but my dog is even better.',\r\n]\r\nbert_ranking.eval()\r\nwith torch.no_grad():\r\n for idx in range(len(dummy_query)):\r\n input_ids = torch.tensor(tokenizer.encode(text=dummy_query[idx], \\\r\n text_pair=dummy_passage[idx], add_special_tokens=True)).unsqueeze(0)\r\n outputs = bert_ranking(input_ids)\r\n print(outputs)\r\n```\r\n", "> Thanks a lot. I was having the same question about 'nboost' and was trying this method. However, the output seems to change when I run the same code multiple times, even though i am in the eval mode. Do you have any hint about what I am doing wrong here?\r\n> \r\n> ```\r\n> bert_ranking = BertForPassageRanking.from_pretrained(BERT_PASSAGE_RANKING_PATH,\r\n> from_tf=True)\r\n> \r\n> dummy_query = [\r\n> 'Rutgers is a good university. I like my experience there.', \r\n> \"Hello, my dog is cute. My cute dog is amazing.\",\r\n> 'Florida is a nice place but tiger king may be better',\r\n> ]\r\n> \r\n> dummy_passage = [\r\n> 'My cat is really cute but my dog is even better.',\r\n> 'My cat is really cute but my dog is even better.',\r\n> 'My cat is really cute but my dog is even better.',\r\n> ]\r\n> bert_ranking.eval()\r\n> with torch.no_grad():\r\n> for idx in range(len(dummy_query)):\r\n> input_ids = torch.tensor(tokenizer.encode(text=dummy_query[idx], \\\r\n> text_pair=dummy_passage[idx], add_special_tokens=True)).unsqueeze(0)\r\n> outputs = bert_ranking(input_ids)\r\n> print(outputs)\r\n> ```\r\n\r\nSorry, I have no idea. Finally I am not using this approximation. I did not achieve good results for my purpose. Intead, I am using the model provided by nboost (https://huggingface.co/nboost/pt-tinybert-msmarco) and it works fine for me. Remember to load the model as follows:\r\n\r\n`model = AutoModelForSequenceClassification.from_pretrained(\"nboost/pt-tinybert-msmarco\")`\r\n\r\nI am using tinybert-msmarco, however you can use one of the following models:\r\n\r\n```\r\nnboost/pt-bert-base-uncased-msmarco\r\nnboost/pt-bert-large-msmarco\r\nnboost/pt-biobert-base-msmarco\r\nnboost/pt-tinybert-msmarco\r\n```", "Hi, I have fine tuned a multilingual model, taken from hugging face, on the passage reranking task. Now I am facing difficulties with converting the tensorflow checkpoint to a pytorch model, so that I can use the model using `BertForSequenceClassification`.\r\nI am using the following conversion function, but I get this error \r\n\r\n ```\r\nFile \"<ipython-input-50-1e24e5635ec9>\", line 1, in <module>\r\n convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytorch_dump_path)\r\n\r\n File \"<ipython-input-49-22827240b095>\", line 63, in convert_tf_checkpoint_to_pytorch\r\n assert pointer.shape == array.shape\r\n\r\n File \"/home/igli/anaconda3/envs/search-boost/lib/python3.8/site-packages/torch/nn/modules/module.py\", line 593, in __getattr__\r\n raise AttributeError(\"'{}' object has no attribute '{}'\".format(\r\n\r\nAttributeError: 'LayerNorm' object has no attribute 'shape'\r\n```\r\n\r\nThe conversion method:\r\n```\r\ndef convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytorch_dump_path):\r\n config_path = os.path.abspath(bert_config_file)\r\n tf_path = os.path.abspath(tf_checkpoint_path)\r\n print(\"Converting TensorFlow checkpoint from {} with config at {}\".format(tf_path, config_path))\r\n # Load weights from TF model\r\n init_vars = tf.train.list_variables(tf_path)\r\n names = []\r\n arrays = []\r\n for name, shape in init_vars:\r\n print(\"Loading TF weight {} with shape {}\".format(name, shape))\r\n array = tf.train.load_variable(tf_path, name)\r\n names.append(name)\r\n arrays.append(array)\r\n \r\n # Initialise PyTorch model\r\n config = BertConfig.from_json_file(bert_config_file)\r\n config.num_labels = 2\r\n\r\n print(\"Building PyTorch model from configuration: {}\".format(str(config)))\r\n model = BertForSequenceClassification()(config=config)\r\n\r\n \r\n for name, array in zip(names, arrays):\r\n if name in ['output_weights' , 'output_bias']:\r\n name = 'classifier/' + name\r\n name = name.split('/')\r\n # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v\r\n # which are not required for using pretrained model\r\n if name[-1] in [\"adam_v\", \"adam_m\"]:\r\n print(\"Skipping {}\".format(\"/\".join(name)))\r\n continue\r\n pointer = model\r\n \r\n for m_name in name: \r\n\r\n if re.fullmatch(r'[A-Za-z]+_\\d+', m_name):\r\n l = re.split(r'_(\\d+)', m_name)\r\n else:\r\n l = [m_name]\r\n if l[0] == 'kernel':\r\n pointer = getattr(pointer, 'weight')\r\n elif l[0] == 'output_bias':\r\n pointer = getattr(pointer, 'bias')\r\n pointer = getattr(pointer, 'cls')\r\n elif l[0] == 'output_weights':\r\n pointer = getattr(pointer, 'weight')\r\n pointer = getattr(pointer, 'cls') \r\n else:\r\n try:\r\n pointer = getattr(pointer, l[0])\r\n except:\r\n pass\r\n\r\n if len(l) >= 2:\r\n num = int(l[1])\r\n pointer = pointer[num]\r\n if m_name[-11:] == '_embeddings':\r\n pointer = getattr(pointer, 'weight')\r\n elif m_name == 'kernel':\r\n array = np.transpose(array)\r\n try:\r\n assert pointer.shape == array.shape\r\n except AssertionError as e:\r\n e.args += (pointer.shape, array.shape)\r\n raise\r\n #pass\r\n \r\n print(\"Initialize PyTorch weight {}\".format(name))\r\n array = np.array(array)\r\n print(array)\r\n print(type(array))\r\n pointer.data = torch.from_numpy(array)\r\n \r\n # Save pytorch-model\r\n print(\"Save PyTorch model to {}\".format(pytorch_dump_path))\r\n torch.save(model.state_dict(), pytorch_dump_path)\r\n```\r\nI have currently no clue, where the problem might be. Thanks in advanvce!", "> Update for latest transformers, add modeling_bert.py:78:\r\n> \r\n> ```python\r\n> for name, array in zip(names, arrays):\r\n> if name in ['output_weights', 'output_bias']:\r\n> name = 'classifier/' + name\r\n> ```\r\n> \r\n> and convert_bert_original_tf_checkpoint_to_pytorch.py\r\n> \r\n> ```python\r\n> config.num_labels = 2\r\n> print(\"Building PyTorch model from configuration: {}\".format(str(config)))\r\n> model = BertForSequenceClassification(config)\r\n> ```\r\n\r\nAs of 26/Mar/2021, \r\n`modeling_bert.py:78` is now around `modeling_bert.py:118`\r\n`convert_bert_original_tf_checkpoint_to_pytorch.py` is now around `convert_bert_original_tf_checkpoint_to_pytorch.py:33`. BTW, don't forget `from transformers import BertForSequenceClassification`" ]

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

[]

[ "Hmm maybe @Rocketknight1 have an insight on this?", "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": 1260952223, "node_id": "MDU6TGFiZWwxMjYwOTUyMjIz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Discussion", "name": "Discussion", "color": "22870e", "default": false, "description": "Discussion on a topic (keep it focused or open a new issue though)" }, { "id": 1314768611, "node_id": "MDU6TGFiZWwxMzE0NzY4NjEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/wontfix", "name": "wontfix", "color": "ffffff", "default": true, "description": null } ]

[]

[ "I join this issue.\r\n\r\nAlso I have a question related to the p.3\r\n> Optionally, modify run_classifier.py to allow loading of fine-tuned BERT language models from the lm_finetuning/ scripts\r\n\r\n`finetune_on_pregenerated.py` script uses `BertForPreTraining` with 2 heads and this is like vanilla training from the original paper. But `run_classifier.py` uses `BertForSequenceClassification` which is being learned only for predicting labels, not masked tokens and isNextSeq. Am I right?\r\nIf so, how can I merge these two approaches? I want to fine-tune the pretrained bert for my dataset and also train a classifier on the top of it.\r\n\r\nThank you.", "It's a good question and a good discussion @Rocketknight1.\r\n\r\nI think your suggestion of \"splitting\" the present repo in two by extracting the examples in a separate repo and refactoring them to have a better workflow is a good idea.\r\n\r\nAt the present stage, it's important to keep the core library stable as it is used in downstream libraries but the examples are an area where many people would like to contribute to the community and it would be nice to have a more flexible structure that allows such contributions more easily than the present monolithic setup.\r\n\r\nSo if you want to start a new repo splitting the \"usage-part\" of this repo it could be a very interesting idea I think. I'm also happy to help and host it close to the present repo if it's what you had in mind (maybe shoot me an email in that case).", "Understood! I'm tight on time right now, but if I find time I'll try to build an interface like that and contact you to ensure we sync things up between the two repos.", "Keeping the discussion open on this", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Maybe you get the error because of the position_ids that are most likely wrong. \r\n\r\nI believe positional ids are not needed - you can use this:\r\npredictions, past = model(tokens_tensor,position_ids=None token_type_ids=None, lm_labels=None, past=None)\r\n\r\nand the use the parameters you want at the place you want or leave None if you do not want to use anything.\r\n\r\n", "With zeros, it is not padded, zeros are \"!\" not \"[PAD]\"", "tokenizer.convert_tokens_to_ids('[PAD]') this results in 0. I am confused why it happens", "Really? When I use decode tokenizer.decode(0) results in !.\r\n\r\ntokenizer.encode(x) = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(x)) [or should be], but when i try to use tokenizer.tokenize or tokenizer.convert_tokens_to_ids I am not sure which one, I get that its nto defined even tokenizer.encode works properly.", "Does the 0 index actually stand for unknown tokens as well?\r\n`tokenizer.convert_tokens_to_ids([\"qrz\"]) ` where \"qrz\" is supposed to be an unknown word. This will give [0]\r\nBut `tokenizer.convert_ids_to_tokens([0])` gives [\"!\"].\r\n", "That strange - we together get that ! == [PAD]", "There was some issue fine-tuning GPT-2 with the master branch. This should now be fixed with the merge of #560 (which also add access to GPT-2 medium model).", "But to actually answer the discussion on this issue (sorry I was reading to quickly), there is no padding token in GPT-2 vocabulary.\r\nSo either you manage to not pad (which is how the model is pretrained) or you need to add a new token to the vocabulary using the special_token functions. This method explained for instance in our blog post on fine-tuning GPT/GPT-2 here: https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313", "Thanks. So we need to add the special token and then fine tune it?\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" ]

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

[]

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

[]

[]

[]

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

[]

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

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

[]

[ "@thomwolf I was wondering what are your thought on this issue?", "The bug in the library causing the index out of range error comes from masking (-1) the LM labels.\r\n\r\n> lm_labels: optional language modeling labels: torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss is only computed for the labels set in [0, ..., vocab_size].\r\n\r\nWhen I pad the lm_labels to a fixed size using -1's, the library throws an error, on the other hand, using any other positive value works but acts like a vocabulary piece and thus makes the input incorrect.\r\n@thomwolf Could someone please fix this?", "Yes there is a PR (#560) fixing a number of issues for fine-tuning GPT-2.\r\nShould be merged soon (this week hopefully).", "Any update on #560? ", "I guess that you should try add pading at the begining - it predicts next word not first so the padding should be added in front.\r\n[50258, 1212, 318, 617, 43086, 50258, 0] should be\r\n[0,50258, 1212, 318, 617, 43086, 50258]", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "I've checked that outputs from right-padded input tensors and from no-padding input tensors are different. Personally, the latter makes little more sense. A module for input masking and its corresponding position embeddings need to be implemented. ", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

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

[]

[]

[ "Thanks!" ]

[]

[]

[ "Great, thanks @MottoX!" ]

[]

[]

[ "Just found it's under Apache v2 in the Google bert repo. Closing." ]

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

[]

[ "I was facing the same issue when finetuning using `finetune_on_pregenerated.py`. The problem was in the fact that I have some empty sentences in my dataset. Also there are some special characters, like `\\t` (tabulation) which can make a mess and should be cleared. \r\nI preprocess the text like this:\r\n```\r\nfor_train = for_train.dropna(subset=['text'])\r\n\r\nimport re\r\nwith open('for_pretraining_full.txt', \"w\", encoding='utf-8') as writer:\r\n for doc in for_train['text'].tolist():\r\n doc = doc.replace(u'\\xa0', u' ').replace(u'\\u200b', u' ').replace(u'\\u206f', u' ').replace(u'\\u206e', u' ').replace(u'\\u206b', u' ').replace(u'\\u206c', u' ').replace(u'\\u2063', u' ').replace(u'\\u200d', u' ').strip() # replace some special unicode chars\r\n doc = re.sub('\\t+', '', doc) # replace tabs\r\n doc = doc.replace('. ', '\\n')\r\n doc = re.sub('\\n+( )*(\\n+)*', '\\n', doc) # replace several consecutive new lines by a single one\r\n doc = doc.strip()\r\n if (doc != ''):\r\n writer.write(doc)\r\n writer.write('\\n\\n')\r\n```\r\n\r\nAlso you can try to find problem sentances, debuging your `simple_lm_finetuning.py`.", "I ran into an issue on this where some sneaky \\n were still sneaking through even with the above code. I remedied this by just doing \r\n\r\n`if \"\\n\" in doc:\r\n\r\n doc = re.sub('\\n', ' ', doc)`", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "Hi,\r\n\r\nThe repo is compatible with PyTorch 1.1.0.\r\n\r\nBut, we probably won't switch to PyTorch Multi-headed-Attention module since this would mean refactoring all the models and adding complexity to the tensorflow conversion codes for unclear gains.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

[]

[]

[ "I'm asking the same question", "I had been dealing with the issue of low and decreasing accuracy when I use fp16, as shown below,\r\n\r\n```\r\nEpoch 1 - Batch 1600/287417 - Training Acc. 0.106250 - Training Loss 2.295977\r\nEpoch 1 - Batch 3200/287417 - Training Acc. 0.098125 - Training Loss 2.299707\r\nEpoch 1 - Batch 4800/287417 - Training Acc. 0.094792 - Training Loss 2.304948\r\nEpoch 1 - Batch 6400/287417 - Training Acc. 0.093125 - Training Loss 2.307725\r\nEpoch 1 - Batch 8000/287417 - Training Acc. 0.092625 - Training Loss 2.305703\r\nEpoch 1 - Batch 9600/287417 - Training Acc. 0.091667 - Training Loss 2.306758\r\nEpoch 1 - Batch 11200/287417 - Training Acc. 0.092589 - Training Loss 2.306116\r\nEpoch 1 - Batch 12800/287417 - Training Acc. 0.092969 - Training Loss 2.307227\r\nEpoch 1 - Batch 14400/287417 - Training Acc. 0.091458 - Training Loss 2.310017\r\nEpoch 1 - Batch 16000/287417 - Training Acc. 0.091000 - Training Loss 2.308750\r\nEpoch 1 - Batch 17600/287417 - Training Acc. 0.090795 - Training Loss 2.309631\r\nEpoch 1 - Batch 19200/287417 - Training Acc. 0.090625 - Training Loss 2.310771\r\nEpoch 1 - Batch 20800/287417 - Training Acc. 0.090433 - Training Loss 2.310832\r\nEpoch 1 - Batch 22400/287417 - Training Acc. 0.090625 - Training Loss 2.311030\r\nEpoch 1 - Batch 24000/287417 - Training Acc. 0.090083 - Training Loss 2.311357\r\nEpoch 1 - Batch 25600/287417 - Training Acc. 0.089883 - Training Loss 2.311748\r\nEpoch 1 - Batch 27200/287417 - Training Acc. 0.089449 - Training Loss 2.312302\r\nEpoch 1 - Batch 28800/287417 - Training Acc. 0.088993 - Training Loss 2.312582\r\nEpoch 1 - Batch 30400/287417 - Training Acc. 0.088651 - Training Loss 2.313187\r\nEpoch 1 - Batch 32000/287417 - Training Acc. 0.088656 - Training Loss 2.313006\r\nEpoch 1 - Batch 33600/287417 - Training Acc. 0.088750 - Training Loss 2.313333\r\nEpoch 1 - Batch 35200/287417 - Training Acc. 0.088665 - Training Loss 2.314015\r\nEpoch 1 - Batch 36800/287417 - Training Acc. 0.088641 - Training Loss 2.313631\r\nEpoch 1 - Batch 38400/287417 - Training Acc. 0.088854 - Training Loss 2.313276\r\nEpoch 1 - Batch 40000/287417 - Training Acc. 0.089325 - Training Loss 2.312648\r\nEpoch 1 - Batch 41600/287417 - Training Acc. 0.089183 - Training Loss 2.312943\r\nEpoch 1 - Batch 43200/287417 - Training Acc. 0.089051 - Training Loss 2.312587\r\nEpoch 1 - Batch 44800/287417 - Training Acc. 0.088929 - Training Loss 2.313172\r\nEpoch 1 - Batch 46400/287417 - Training Acc. 0.088793 - Training Loss 2.312671\r\nEpoch 1 - Batch 48000/287417 - Training Acc. 0.088479 - Training Loss 2.313255\r\nEpoch 1 - Batch 49600/287417 - Training Acc. 0.088972 - Training Loss 2.312710\r\nEpoch 1 - Batch 51200/287417 - Training Acc. 0.088906 - Training Loss 2.312372\r\n```\r\n\r\nHowever, after I made the change in `lr_this_step` that you indicated, I've started to get normal results, as follows,\r\n\r\n```\r\nEpoch 1 - Batch 1600/287417 - Training Acc. 0.156250 - Training Loss 2.224727\r\nEpoch 1 - Batch 3200/287417 - Training Acc. 0.200937 - Training Loss 2.166289\r\nEpoch 1 - Batch 4800/287417 - Training Acc. 0.245833 - Training Loss 2.098184\r\nEpoch 1 - Batch 6400/287417 - Training Acc. 0.299063 - Training Loss 2.018706\r\nEpoch 1 - Batch 8000/287417 - Training Acc. 0.351625 - Training Loss 1.937730\r\nEpoch 1 - Batch 9600/287417 - Training Acc. 0.400833 - Training Loss 1.855378\r\nEpoch 1 - Batch 11200/287417 - Training Acc. 0.444018 - Training Loss 1.768468\r\nEpoch 1 - Batch 12800/287417 - Training Acc. 0.481875 - Training Loss 1.685869\r\nEpoch 1 - Batch 14400/287417 - Training Acc. 0.513889 - Training Loss 1.606483\r\nEpoch 1 - Batch 16000/287417 - Training Acc. 0.536937 - Training Loss 1.537816\r\nEpoch 1 - Batch 17600/287417 - Training Acc. 0.556364 - Training Loss 1.477735\r\nEpoch 1 - Batch 19200/287417 - Training Acc. 0.576146 - Training Loss 1.418323\r\nEpoch 1 - Batch 20800/287417 - Training Acc. 0.592019 - Training Loss 1.367327\r\nEpoch 1 - Batch 22400/287417 - Training Acc. 0.606429 - Training Loss 1.321059\r\nEpoch 1 - Batch 24000/287417 - Training Acc. 0.617542 - Training Loss 1.281488\r\nEpoch 1 - Batch 25600/287417 - Training Acc. 0.627109 - Training Loss 1.246746\r\nEpoch 1 - Batch 27200/287417 - Training Acc. 0.637500 - Training Loss 1.211883\r\nEpoch 1 - Batch 28800/287417 - Training Acc. 0.645938 - Training Loss 1.182604\r\nEpoch 1 - Batch 30400/287417 - Training Acc. 0.652204 - Training Loss 1.158571\r\nEpoch 1 - Batch 32000/287417 - Training Acc. 0.658875 - Training Loss 1.134463\r\nEpoch 1 - Batch 33600/287417 - Training Acc. 0.665179 - Training Loss 1.111719\r\nEpoch 1 - Batch 35200/287417 - Training Acc. 0.671023 - Training Loss 1.089363\r\nEpoch 1 - Batch 36800/287417 - Training Acc. 0.676848 - Training Loss 1.068860\r\nEpoch 1 - Batch 38400/287417 - Training Acc. 0.681536 - Training Loss 1.050721\r\nEpoch 1 - Batch 40000/287417 - Training Acc. 0.685775 - Training Loss 1.034663\r\nEpoch 1 - Batch 41600/287417 - Training Acc. 0.690361 - Training Loss 1.017672\r\nEpoch 1 - Batch 43200/287417 - Training Acc. 0.693866 - Training Loss 1.004058\r\nEpoch 1 - Batch 44800/287417 - Training Acc. 0.698013 - Training Loss 0.990084\r\nEpoch 1 - Batch 46400/287417 - Training Acc. 0.701552 - Training Loss 0.977086\r\nEpoch 1 - Batch 48000/287417 - Training Acc. 0.704854 - Training Loss 0.965735\r\nEpoch 1 - Batch 49600/287417 - Training Acc. 0.708266 - Training Loss 0.953387\r\nEpoch 1 - Batch 51200/287417 - Training Acc. 0.712012 - Training Loss 0.940919\r\nEpoch 1 - Batch 52800/287417 - Training Acc. 0.714697 - Training Loss 0.929941\r\n```\r\n\r\nThanks!", "@burcturkoglu \r\nHow about performance comparison with fp32?", "@yeontaek \r\nI trained BERT for classification with my own data by _run_classifier_ script.\r\n\r\nHere are the benchmarks for fp32 vs fp16 in both single Tesla V100 and in 4 Tesla V100 with _DataParallel_,\r\n\r\n*_fp32:_*\r\n\r\n- Single Tesla V100 - Training Duration 17,739 seconds\r\n\r\n- 4 Tesla V100 - Training Duration 9,342 seconds\r\n\r\n*_fp16:_*\r\n\r\n- Single Tesla V100 - Training Duration 12,297 seconds\r\n\r\n- 4 Tesla V100 - Training Duration 6,330 seconds\r\n\r\nIn both types of instances, it gives approximately 30% increase in speed without a change in accuracy. ", "@burcturkoglu \r\nThank you so much. It was a big help. " ]

[ { "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" ]

[]

[]

[ "Thanks a lot for that @8enmann!" ]

[ { "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" ]

[]

[]

[ "Awesome, thanks!" ]

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

[]

[ "This looks good to me", "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", "@anshuman1992 could you share a code snippet/gist used for training TransformerXL model?\r\n", "@anshuman1992 this will be great for me too" ]

[]

[]

[ "# [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560?src=pr&el=h1) Report\n> :exclamation: No coverage uploaded for pull request base (`master@b832d5b`). [Click here to learn what that means](https://docs.codecov.io/docs/error-reference#section-missing-base-commit).\n> The diff coverage is `70.37%`.\n\n[](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560?src=pr&el=tree)\n\n```diff\n@@ Coverage Diff @@\n## master #560 +/- ##\n=========================================\n Coverage ? 66.04% \n=========================================\n Files ? 18 \n Lines ? 3673 \n Branches ? 0 \n=========================================\n Hits ? 2426 \n Misses ? 1247 \n Partials ? 0\n```\n\n\n| [Impacted Files](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560?src=pr&el=tree) | Coverage Δ | |\n|---|---|---|\n| [pytorch\\_pretrained\\_bert/modeling\\_openai.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfb3BlbmFpLnB5) | `79.68% <100%> (ø)` | |\n| [pytorch\\_pretrained\\_bert/tokenization\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvdG9rZW5pemF0aW9uX2dwdDIucHk=) | `84.86% <100%> (ø)` | |\n| [pytorch\\_pretrained\\_bert/modeling\\_gpt2.py](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560/diff?src=pr&el=tree#diff-cHl0b3JjaF9wcmV0cmFpbmVkX2JlcnQvbW9kZWxpbmdfZ3B0Mi5weQ==) | `80.16% <61.9%> (ø)` | |\n\n------\n\n[Continue to review full report at Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560?src=pr&el=continue).\n> **Legend** - [Click here to learn more](https://docs.codecov.io/docs/codecov-delta)\n> `Δ = absolute <relative> (impact)`, `ø = not affected`, `? = missing data`\n> Powered by [Codecov](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560?src=pr&el=footer). Last update [b832d5b...db98a4a](https://codecov.io/gh/huggingface/pytorch-pretrained-BERT/pull/560?src=pr&el=lastupdated). Read the [comment docs](https://docs.codecov.io/docs/pull-request-comments).\n" ]

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

[]

[ "Can you give the exact log of (and before) the error message?", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n" ]

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "2 years late, but can anyone figure it out?" ]

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

[]

[ "@thomwolf Could you or someone from your team point me in the right direction to get the gtp2 model running with a small number of newly defined special tokens?\r\nAny help very appreciated as I really need to move on with my research project.", "Hi @adigoryl, I'm adding this feature with PR #560\r\n\r\nYou can have a look.\r\n\r\nIt should be merged soon I guess.", "Hi @thomwolf, first of all, I would like to thank you for the quick response and solution. I have had a look at the added lines and have replaced the 'modelling_gpt2.py' file in my pytorch_pretrained_bert lib. Running the code: `model = GPT2LMHeadModel.from_pretrained(args.model_name, num_special_tokens=len(special_tokens))` gives me:\r\n\r\n> model = cls(config, *inputs, **kwargs)\r\nTypeError: __init__() got an unexpected keyword argument 'num_special_tokens'\r\n\r\nI am not sure whether this happens because of the way I have updated my lib or there still is something missing.\r\nWhat is the best way to update my lib with the freshly made changes?\r\n\r\n--------------------------UPDATE-------------------------------\r\nCopy and paste seemed to work. The problem was that I needed to add a new line after the updated code paste since python is space sensitive. Having to fix it the num_special_tokens works in an anticipated way as I can see in the debugger that it sets the n_special field and updates total_tokens_embeddings. However, having this all fixed I still end up with the same issue I started with:\r\n\r\n> Traceback (most recent call last):\r\n File \"/Users/aw678/PycharmProjects/BERT/gtp2_train_lyrics_LM_copy.py\", line 202, in <module>\r\n main()\r\n File \"/Users/aw678/PycharmProjects/BERT/gtp2_train_lyrics_LM_copy.py\", line 178, in main\r\n losses, past = model(input_ids, lm_labels, past=past)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/torch/nn/modules/module.py\", line 489, in __call__\r\n result = self.forward(*input, **kwargs)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/pytorch_pretrained_bert/modeling_gpt2.py\", line 661, in forward\r\n hidden_states, presents = self.transformer(input_ids, position_ids, token_type_ids, past)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/torch/nn/modules/module.py\", line 489, in __call__\r\n result = self.forward(*input, **kwargs)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/pytorch_pretrained_bert/modeling_gpt2.py\", line 587, in forward\r\n position_embeds = self.wpe(position_ids)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/torch/nn/modules/module.py\", line 489, in __call__\r\n result = self.forward(*input, **kwargs)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/torch/nn/modules/sparse.py\", line 118, in forward\r\n self.norm_type, self.scale_grad_by_freq, self.sparse)\r\n File \"/Users/aw678/anaconda3/envs/pytorch_conda_env/lib/python3.6/site-packages/torch/nn/functional.py\", line 1454, in embedding\r\n return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)\r\nRuntimeError: index out of range at /Users/soumith/mc3build/conda-bld/pytorch_1549593514549/work/aten/src/TH/generic/THTensorEvenMoreMath.cpp:191\r\n\r\nNot sure why it complains about \"position_ids\" since they are not compulsory. I believe this may not be an issue with my code (just in case someone wants to have a look):\r\n[gtp2_train_lyrics_LM_copy.pdf](https://github.com/huggingface/pytorch-pretrained-BERT/files/3131933/gtp2_train_lyrics_LM_copy.pdf)\r\n\r\nIf you could provide a simplified working example of running GPT2 with new tokens then this should resolve my issue.\r\n", "To replicate the error use the simplified version:\r\n[gpt2_simplified.py.zip](https://github.com/huggingface/pytorch-pretrained-BERT/files/3132183/gpt2_simplified.py.zip)\r\n", "You have to install the repo from source from the PR branch (see the instructions in the readme to install from source and after cloning git checkout to the PR branch before installing).\r\n\r\nIf it looks too complicated maybe the best is to wait for the PR to be merged.", "I have managed to update the library on my machine but it seems that there is an incompatibility in the lib code. If you could provide a working toy example on how to fine-tune GPT2 with special symbols then I am sure the community would appreciate it and my issue would be resolved. I have attached such toy example above in the zip file, however, it has an issue which I believe is caused by the lib.\r\n\r\nI am sorry to bother you so much but I just want to get on with my work. \r\nRegards, Adrian.", "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 } ]

[]

[ "Actually, shouldn't the `int()` be a `ceiling()`? Because let's say `args.gradient_accumulation_steps` is 1, then it is `ceiling(len(train_examples) / args.train_batch_size)` that is the number of batches in an epoch.", "I am having the same problem with my finetuned model for gpt2", "I am having the same issue in partially changed run_squad.py code.", "I have the same issue.", "Humm yes, we should probably change `int` to `ceiling` in this example indeed.", "Is this a significant issue? If it's only the last few batches in the last epoch that are not being trained on, it shouldn't be a huge problem, right?\r\n\r\nAlso I find it strange that suddenly a lot of people are running into this bug in this past week (according to the replies to this issue) even though the `int` code was written 3 months ago. Is this also related to some other more recent changes?", "Yes there was a huge refactoring of the `BertAdam` optimizer by @lukovnikov (#389, #445, #531)", "Hi, this warning is printed to avoid wasted computations with warmup-linear or other surprises with other schedules due to a t_total set too low.\r\nAnd I think that line should be `int( math.ceil(len(train_examples) / args.train_batch_size) / args.gradient_accumulation_steps) * args.num_train_epochs` (@thomwolf) ?", "Hi. I figured out the source of the problem: t_total, aka num_train_optimization_steps \r\n\r\nhttps://github.com/huggingface/pytorch-pretrained-BERT/blob/3fc63f126ddf883ba9659f13ec046c3639db7b7e/examples/run_squad.py#L903\r\n\r\n is computed over the length of the train examples, while the true number of steps is determined by whatever convert_examples_to_features returns\r\n\r\nhttps://github.com/huggingface/pytorch-pretrained-BERT/blob/3fc63f126ddf883ba9659f13ec046c3639db7b7e/examples/run_squad.py#L970\r\n\r\nA print(len(train_examples), len(train_features)) in line 980 returns:\r\n\r\n87599 191597 \r\n\r\n", "You could also add the option `drop_last=True` to the `DataLoader`, then the number of samples will be calculated correctly.", "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "I get error below while running the program.. Did I do any mistake?\r\n\r\nwarmup_linear = WarmupLinearSchedule( warmup=args.warmup_proportion,\r\nt_total=num_train_optimization_steps)\r\n\r\nlr_this_step = args.learning_rate * warmup_linear.get_lr(num_train_optimization_steps,\r\nargs.warmup_proportion)\r\n\r\nWARNING - pytorch_pretrained_bert.optimization - Training beyond specified 't_total'. Learning rate multiplier set to 0.0. Please set 't_total' of WarmupLinearSchedule correctly." ]

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

[]

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

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

[]

[ "Please give more information: the command used (arguments passed), traceback (the command line output), and version (can use `pip show pytorch_pretrained_bert`)\r\n\r\nI faced a similar problem with `read_squad_examples` when passing `input_file=dev.json` and `is_training=True`. ", "I have this problem when training on SQUADv2 without `--version_2_with_negative` option. Basically for squad 2.0, it is possible there is no answer for questions. Adding this option in training command fixed the problem for me.", "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,\r\n\r\nI see the same error. Some of the questions have multiple answer spans. \r\n\r\nThe error suggests that the current preprocessing codes cannot handle multiple answer spans for a given question.\r\n\r\n Has anyone fixed this ??\r\n\r\nThanks", "I also receive this error when using the `--version_2_with_negative` flag paired with training data from SQuAD 2.0. It looks like it may be caused by some logic in `utils_squad.py`, lines 150-152; it seems that answerable questions are expected to only have one answer. I'm not familiar enough with the task and data set to know if this is a correct assumption, but since I'm using data from the SQuAD 2.0 web site I would think it should train fine.", "For training, the assumption is true.\n\nOn Mon, Nov 18, 2019 at 09:56 Allen Kim <[email protected]> wrote:\n\n> I also receive this error when using the --version_2_with_negative flag\n> paired with training data from SQuAD 2.0. It looks like it may be caused by\n> some logic in utils_squad.py, lines 150-152; it seems that answerable\n> questions are expected to only have one answer. I'm not familiar enough\n> with the task and data set to know if this is a correct assumption, but\n> since I'm using data from the SQuAD 2.0 web site I would think it should\n> train fine.\n>\n> —\n> You are receiving this because you are subscribed to this thread.\n> Reply to this email directly, view it on GitHub\n> <https://github.com/huggingface/transformers/issues/554?email_source=notifications&email_token=AIEAE4GLKEIMFR3BUJQK5VTQUHY4BA5CNFSM4HI7NMRKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEEI5VNI#issuecomment-554818229>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AIEAE4H7OARRY4INFJTGU6DQUHY4BANCNFSM4HI7NMRA>\n> .\n>\n", "Thanks for clarifying!" ]

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

[]

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

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

[]

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

[ { "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" ]

[]

[]

[ "Thanks, this is closed now with #564" ]

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

[]

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

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

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.\n", "@shawnkx Hi! Have you found a solution？", "@all, any updates on this?" ]

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

[]

[ "I'm interested in an answer, too. A score/probability would help to select the best word for a masked token.", "You are looking for the softmax function: https://pytorch.org/docs/stable/nn.html?highlight=softmax#torch.nn.functional.softmax", "Thanks Thomas, I'll give it a try.", "Thanks,\r\nSo you say that for score x1 (where all the scores are x1,x2,.. xn) :\r\nprobability_x1 = (exp(^x1)/(exp(^x1) + exp(^x2) + .. exp(^xn))\r\n\r\n", "@Oxi84 could you share how you obtained the masked word probabilities? I have been trying to do that on my custom data. That is, I want to pretrain my own model and then do masked word prediction on new data.", "@rvoak The quickstart guide [here](https://github.com/huggingface/pytorch-transformers/blob/master/docs/source/quickstart.md#bert-example) shows a nice example of how to do masked word prediction. \r\nReplace \r\n```\r\n# confirm we were able to predict 'henson'\r\npredicted_index = torch.argmax(predictions[0, masked_index]).item()\r\npredicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]\r\n```\r\nwith something like this (e.g. if you want the top k predicted tokens):\r\n\r\n```\r\ntop_k = 10\r\nprobs = torch.nn.functional.softmax(predictions[0, mask_idx], dim=-1)\r\ntop_k_weights, top_k_indices = torch.topk(probs, top_k, sorted=True)\r\n\r\nfor i, pred_idx in enumerate(top_k_indicies):\r\n predicted_token = tokenizer.convert_ids_to_tokens([pred_idx])[0]\r\n token_weight = top_k_weights[i]\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", "> @rvoak The quickstart guide [here](https://github.com/huggingface/pytorch-transformers/blob/master/docs/source/quickstart.md#bert-example) shows a nice example of how to do masked word prediction.\r\n> Replace\r\n> \r\n> ```\r\n> # confirm we were able to predict 'henson'\r\n> predicted_index = torch.argmax(predictions[0, masked_index]).item()\r\n> predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]\r\n> ```\r\n> \r\n> with something like this (e.g. if you want the top k predicted tokens):\r\n> \r\n> ```\r\n> top_k = 10\r\n> probs = torch.nn.functional.softmax(predictions[0, mask_idx], dim=-1)\r\n> top_k_weights, top_k_indices = torch.topk(probs, top_k, sorted=True)\r\n> \r\n> for i, pred_idx in enumerate(top_k_indicies):\r\n> predicted_token = tokenizer.convert_ids_to_tokens([pred_idx])[0]\r\n> token_weight = top_k_weights[i]\r\n> ```\r\n\r\nvery good example thanks! is there a version for RoBERTa and other models?", "Hi @yuchenlin, you can use the recently added `fill-mask` pipeline to do so:\r\n\r\n```py\r\n>>> from transformers import pipeline\r\n>>> nlp = pipeline(\"fill-mask\", model=\"roberta-base\")\r\n>>> nlp(f\"This is the best thing I've {nlp.tokenizer.mask_token} in my life.\")\r\n[\r\n {'sequence': \"<s> This is the best thing I've done in my life.</s>\", 'score': 0.8024354577064514, 'token': 626}, \r\n {'sequence': \"<s> This is the best thing I've heard in my life.</s>\", 'score': 0.031355079263448715, 'token': 1317}, \r\n {'sequence': \"<s> This is the best thing I've learned in my life.</s>\", 'score': 0.027319395914673805, 'token': 2435}, \r\n {'sequence': \"<s> This is the best thing I've seen in my life.</s>\", 'score': 0.026892054826021194, 'token': 450}, \r\n {'sequence': \"<s> This is the best thing I've experienced in my life.</s>\", 'score': 0.02160099521279335, 'token': 2984}\r\n]\r\n```\r\n\r\nWe're in the process of adding example usage for common tasks (question answering, sequence classification, mask filling etc), you can follow the progress in https://github.com/huggingface/transformers/pull/2850. There already is an example for mask filling.", "Hey @LysandreJik, does the fill-mask also support whole word mask prediction, or does it only work on subword level?", "> Hi @yuchenlin, you can use the recently added `fill-mask` pipeline to do so:\r\n> \r\n> ```python\r\n> >>> from transformers import pipeline\r\n> >>> nlp = pipeline(\"fill-mask\", model=\"roberta-base\")\r\n> >>> nlp(f\"This is the best thing I've {nlp.tokenizer.mask_token} in my life.\")\r\n> [\r\n> {'sequence': \"<s> This is the best thing I've done in my life.</s>\", 'score': 0.8024354577064514, 'token': 626}, \r\n> {'sequence': \"<s> This is the best thing I've heard in my life.</s>\", 'score': 0.031355079263448715, 'token': 1317}, \r\n> {'sequence': \"<s> This is the best thing I've learned in my life.</s>\", 'score': 0.027319395914673805, 'token': 2435}, \r\n> {'sequence': \"<s> This is the best thing I've seen in my life.</s>\", 'score': 0.026892054826021194, 'token': 450}, \r\n> {'sequence': \"<s> This is the best thing I've experienced in my life.</s>\", 'score': 0.02160099521279335, 'token': 2984}\r\n> ]\r\n> ```\r\n> \r\n> We're in the process of adding example usage for common tasks (question answering, sequence classification, mask filling etc), you can follow the progress in #2850. There already is an example for mask filling.\r\n\r\nIs it possible to give this an input word for the mask and get probabilities back for that specific word?", "Also is it possible to request the top N sentences rather than the default returned?\r\n\r\nEdit: Never mind on this specific question! I found out by setting:\r\n\r\n`nlp.topk = 20`\r\n\r\nbefore doing:\r\n\r\n` nlp(f\"This is the best thing I've {nlp.tokenizer.mask_token} in my life.\")`\r\n\r\nIt now returns 20.", "Sure, you can do that using the recently added `targets` (in `v3.1.0`):\r\n\r\n```py\r\n>>> from transformers import pipeline\r\n>>> nlp = pipeline(\"fill-mask\", model=\"roberta-base\")\r\n>>> nlp(f\"This is the best thing I've {nlp.tokenizer.mask_token} in my life.\", targets=[' experienced'])\r\n[\r\n {\r\n 'sequence': \"<s>This is the best thing I've experienced in my life.</s>\", \r\n 'score': 0.022622672840952873, \r\n 'token': 2984, \r\n 'token_str': 'Ġexperienced'\r\n }\r\n]\r\n```\r\n\r\nPlease note the space before the word, because we're using the [RoBERTa tokenizer](https://huggingface.co/transformers/model_doc/roberta.html#robertatokenizer) which is a Byte-level BPE tokenizer that has a different behaviour according to the spaces before tokens.", "@LysandreJik So very helpful! Thank you so much!", "@LysandreJik If a word is at the start of a sentence, should it also have a space in front of it?:\r\n\r\n```\r\nnlp(f\"{nlp.tokenizer.mask_token} talk about the rules of the game first.\", targets=[' We\\'ll'])\r\n```\r\nWhich gives me:\r\n```\r\nThe specified target token ` We'll` does not exist in the model vocabulary. Replacing with `ĠWe`.\r\n[{'sequence': '<s> We talk about the rules of the game first</s>', 'score': 8.493712812196463e-06, 'token': 166, 'token_str': 'ĠWe'}]\r\n```\r\n\r\nOr\r\n\r\n```\r\nnlp(f\"{nlp.tokenizer.mask_token} talk about the rules of the game first.\", targets=['We\\'ll'])\r\n```\r\nWhich gives me:\r\n```\r\nThe specified target token `We'll` does not exist in the model vocabulary. Replacing with `We`.\r\n[{'sequence': '<s>We talk about the rules of the game first</s>', 'score': 0.12082401663064957, 'token': 170, 'token_str': 'We'}]\r\n```", "How to predict a word that is seperated into several tokens. For example, DOTA2(name for a popular game)?" ]

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

[]

[ "This should be fixed with the new release (0.6.2).", "Unfortunately, I still get this error with the new release. Could that be because I had installed the package before some time ago (and removed it afterwards)? \r\n\r\nNever mind, got it running by cleaning up the environments/paths.", "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" ]

[]

[]

[ "Ok, looks good, thanks @ailzhang!" ]

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

[]

[ "I also get this problem when predicting，Did you solved the problem？", "Here is the problem during initialization of the optimizer:\r\n` t_total=num_train_optimization_steps)`\r\n\r\nThis var is initialized with `None` for the first time `num_train_optimization_steps = None`\r\nand it's initialized correctly only when `--do_train` flag is passed to the script\r\n```\r\n if args.do_train:\r\n train_examples = processor.get_train_examples(args.data_dir)\r\n num_train_optimization_steps = int(\r\n len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs\r\n if args.local_rank != -1:\r\n num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()\r\n```\r\nbut in case of `--do_eval` this var is `None` and you got an error from description.\r\n\r\nIt's a bug, I think, and should be fixed. For you local needs, just initialize the optimizer \"somehow\" - it's not used while evaluating.", "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" ]