Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[ "Duplicate of https://github.com/huggingface/transformers/issues/10285" ]

[]

[]

[]

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

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

[ "Incidentally a bug fix was just merged as part of: https://github.com/huggingface/transformers/pull/10310\r\n- the scheduler step was getting run twice.\r\n\r\nCould you please re-test with `transformers` master?\r\n\r\nThank you!\r\n", "Thank you for your response.\r\n\r\nI have tested the latest version, but the following error occurred. There is something wrong in the lr initialization.\r\n```\r\n..\r\n File \"run_clm.py\", line 376, in main\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1054, in train\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n.. \r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/deepspeed/runtime/lr_schedules.py\", line 728, in get_last_lr\r\n assert getattr(self, '_last_lr', None) is not None, \"need to call step() first\"\r\nAssertionError: need to call step() first\r\n``` ", "Thank you for testing with the master version.\r\n\r\nPlease always post the full backtrace and the full command line you used - otherwise it's impossible to reproduce the problem and know how to fix it.\r\n\r\nThank you.", "Sorry. I just ran the same command shown above, and the full error is as follows:\r\n\r\n```\r\n File \"run_clm.py\", line 417, in <module>\r\n main()\r\n File \"run_clm.py\", line 376, in main\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1054, in train\r\nTraceback (most recent call last):\r\n File \"run_clm.py\", line 417, in <module>\r\nTraceback (most recent call last):\r\n File \"run_clm.py\", line 417, in <module>\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1135, in _maybe_log_save_evaluate\r\nTraceback (most recent call last):\r\n File \"run_clm.py\", line 417, in <module>\r\n main()\r\n File \"run_clm.py\", line 376, in main\r\n if version.parse(torch.__version__) >= version.parse(\"1.4\")\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/deepspeed/runtime/lr_schedules.py\", line 728, in get_last_lr\r\n main()\r\n File \"run_clm.py\", line 376, in main\r\n assert getattr(self, '_last_lr', None) is not None, \"need to call step() first\"\r\nAssertionErrortrain_result = trainer.train(resume_from_checkpoint=checkpoint): need to call step() first\r\n\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1054, in train\r\n main()\r\n File \"run_clm.py\", line 376, in main\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1054, in train\r\n train_result = trainer.train(resume_from_checkpoint=checkpoint)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1054, in train\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1135, in _maybe_log_save_evaluate\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1135, in _maybe_log_save_evaluate\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/transformers/trainer.py\", line 1135, in _maybe_log_save_evaluate\r\n if version.parse(torch.__version__) >= version.parse(\"1.4\")\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/deepspeed/runtime/lr_schedules.py\", line 728, in get_last_lr\r\n if version.parse(torch.__version__) >= version.parse(\"1.4\")\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/deepspeed/runtime/lr_schedules.py\", line 728, in get_last_lr\r\n if version.parse(torch.__version__) >= version.parse(\"1.4\")\r\n File \"/home/.../anaconda3/envs/transformers-4.3.2/lib/python3.7/site-packages/deepspeed/runtime/lr_schedules.py\", line 728, in get_last_lr\r\n assert getattr(self, '_last_lr', None) is not None, \"need to call step() first\"\r\nAssertionError: need to call step() first\r\n assert getattr(self, '_last_lr', None) is not None, \"need to call step() first\"\r\nAssertionError: need to call step() first\r\n assert getattr(self, '_last_lr', None) is not None, \"need to call step() first\"\r\nAssertionError: need to call step() first\r\n 1%|▋ | 10/1095 [00:08<16:13, 1.11it/s]\r\n```\r\n\r\nThanks.", "Great, thank you, I'm able to reproduce this problem. Let me investigate and I will get back to you with a solution. ", "I understand the problem.\r\n\r\nThe optimizer doesn't kick in until a much later step, so `lr_scheduler` doesn't get its first `step()` yet and `._maybe_log_save_evaluate` fails to retrieve `get_last_lr` since there wasn't any yet.\r\n\r\nA quick workaround is to add `\"initial_scale_power\": 1,`, which will force the optimizer to churn from step one.\r\n```\r\n \"fp16\": {\r\n \"enabled\": true,\r\n \"loss_scale\": 0,\r\n \"loss_scale_window\": 1000,\r\n \"initial_scale_power\": 1,\r\n \"hysteresis\": 2,\r\n \"min_loss_scale\": 1\r\n },\r\n```\r\nbut it might not be an optimal solution. https://www.deepspeed.ai/docs/config-json/#fp16-training-options\r\n\r\nI will think of how to resolve this correctly, but meanwhile please let me know if that resolves the scheduler issue.\r\n\r\nto explain - when you use deepspeed's fp16 it skips the optimizer/scheduler calls until the OVERFLOW is no more. And you'd see the following in the log:\r\n\r\n```\r\nOVERFLOW! Rank 0 Skipping step. Attempted loss scale: 4294967296, reducing to 4294967296\r\nOVERFLOW! Rank 0 Skipping step. Attempted loss scale: 4294967296, reducing to 2147483648.0\r\nOVERFLOW! Rank 0 Skipping step. Attempted loss scale: 2147483648.0, reducing to 1073741824.0\r\n```\r\n\r\nThis is also probably why you see an odd behavior as you reported originally (besides the double step bug I fixed)", "OK, this PR should work too if you would like to try it instead: https://github.com/huggingface/transformers/pull/10362\r\n", "Thank you for your work.\r\n\r\nI set `\"initial_scale_power\": 1`, re-ran the command, and the training finished without errors.\r\n \r\n\r\n\r\nAfter the PR #10362 is merged into master, I will try it. Thanks.", "FYI, it has been merged.", "Thanks.\r\n\r\nI have tested the latest version (without setting `\"initial_scale_power\": 1`), and the learning rate behavior is as expected! \r\n\r\n\r\n\r\nThanks for your work. It is very useful to use deepspeed in transformers.", "Thank you for your feedback and supporting this problem fixing process, @tomohideshibata " ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Today, I have encountered a bug in our codebase due to this issue. The model loading wasn't ok for some reason and the code just logged the warnings. I would like to get an error instead of a warning. I can't monitor thousands of models that are running in production for this particular warning message. Could you help me to work on this issue?" ]

[]

[]

[]

[]

[ { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "Hi @codingnoobneedshelp , thank for reporting this issue. Right now MBart50Tokenizer does not work with `AutoTokenizer`.\r\nThere will be a new script for translation in the next ~2 weeks that will handle this issue. For now, you could just modify the script to use `MBart50Tokenizer`, instead of `AutoTokenizer`.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Hi @codingnoobneedshelp \r\n\r\nThis is now resolved, the `run_translation.py` script now supports fine-tuning mBART-50." ]

[]

[]

[]

[]

[]

[ "Hello!\r\n\r\nAs first, I can see several issues on the way you want to train the model:\r\n1. The way you build your dataset is not correct. More precisely, in the `tokenize` function, the first element of the tuple (`a.ids`) is taken as the input, and the second (`a.attention_mask`) is taken as the label. Hence the error you get.\r\n2. When you instantiate your `tf.keras.models.Model` you define the `inputs` and the `outputs` to be the same, this is not correct either, you have to run the model once and then give this output.", "@jplu I realized my mistake and I changed the code to this\r\n\r\n```\r\ndef tokenize(sentence):\r\n sentence = sentence.numpy().decode('utf-8')\r\n a = tokenizer.encode(sentence)\r\n return tf.constant(a.ids,tf.int32), tf.constant(a.attention_mask, tf.int32)\r\n\r\ndef get_tokenized(sentence):\r\n return tf.py_function(tokenize, inp=[sentence], Tout=[tf.int32,tf.int32])\r\n\r\ndef get_tokenized_final(a,b):\r\n return (a,b), None\r\n\r\ndataset = tf.data.Dataset.from_tensor_slices(lines)\r\ndataset = dataset.map(get_tokenized, num_parallel_calls=tf.data.AUTOTUNE).map(get_tokenized_final, num_parallel_calls=tf.data.AUTOTUNE)\r\n\r\nimport tensorflow as tf\r\n\r\nconfig = DistilBertConfig(vocab_size=30000)\r\nmodel = TFDistilBertForMaskedLM(config)\r\ninp1 = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name=\"input_ids\")\r\ninp2 = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name=\"attention_mask\")\r\nop = model([inp1,inp2])\r\nmodel = tf.keras.models.Model(inputs=[inp1, inp2], outputs=model.output)\r\n```\r\nNow the model throws two warnings \r\n```\r\nWARNING:tensorflow:The parameters `output_attentions`, `output_hidden_states` and `use_cache` cannot be updated when calling a model.They have to be set to True/False in the config object (i.e.: `config=XConfig.from_pretrained('name', output_attentions=True)`).\r\nWARNING:tensorflow:The parameter `return_dict` cannot be set in graph mode and will always be set to `True`.\r\n```\r\nand then throws the final error\r\n```\r\nValueError: in user code:\r\n\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:805 train_function *\r\n return step_function(self, iterator)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:795 step_function **\r\n outputs = model.distribute_strategy.run(run_step, args=(data,))\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:1259 run\r\n return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2730 call_for_each_replica\r\n return self._call_for_each_replica(fn, args, kwargs)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:3417 _call_for_each_replica\r\n return fn(*args, **kwargs)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:788 run_step **\r\n outputs = model.train_step(data)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:757 train_step\r\n self.optimizer.minimize(loss, self.trainable_variables, tape=tape)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:498 minimize\r\n return self.apply_gradients(grads_and_vars, name=name)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:598 apply_gradients\r\n grads_and_vars = optimizer_utils.filter_empty_gradients(grads_and_vars)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/utils.py:79 filter_empty_gradients\r\n ([v.name for _, v in grads_and_vars],))\r\n\r\n ValueError: No gradients provided for any variable: ['tf_distil_bert_for_masked_lm_1/distilbert/embeddings/word_embeddings/weight:0', 'tf_distil_bert_for_masked_lm_1/distilbert/embeddings/position_embeddings/embeddings:0', 'tf_distil_bert_for_masked_lm_1/distilbert/embeddings/LayerNorm/gamma:0', 'tf_distil_bert_for_masked_lm_1/distilbert/embeddings/LayerNorm/beta:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/q_lin/kernel:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/q_lin/bias:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/k_lin/kernel:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/k_lin/bias:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/v_lin/kernel:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/v_lin/bias:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/out_lin/kernel:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/attention/out_lin/bias:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/sa_layer_norm/gamma:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/sa_layer_norm/beta:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/ffn/lin1/kernel:0', 'tf_distil_bert_for_masked_lm_1/distilbert/transformer/layer_._0/ffn/lin1/bias:0', 'tf_distil_bert_for_masked_lm_1/distilbert/tra...\r\n```\r\nAny idea what I am doing wrong?", "You cannot do `model.output`, as said in my previous message you have to run the model once to get how the output looks like :)", "@jplu Could you tell me exactly what you mean by \"run\" the model? If I pass a sample array with all ones, it gives me a Broadcasting error as follows \r\n```\r\nconfig = DistilBertConfig(vocab_size=30000)\r\nmodel = TFDistilBertForMaskedLM(config)\r\ninp1 = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name=\"input_ids\")\r\ninp2 = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name=\"attention_mask\")\r\n_ = model([inp1,inp2])\r\n\r\n# Error is thrown for this call\r\na = tf.ones((128,),dtype=tf.int32)\r\nmodel((a,a))\r\n```\r\nError is as attached\r\n```\r\nInvalidArgumentError: Incompatible shapes: [512,768] vs. [128,768] [Op:BroadcastTo]\r\n```\r\nMore specifically the error is raised in `modeling_tf_distilbert.py `\r\n```\r\n 183 if position_ids is None:\r\n--> 184 position_embeds = self.position_embeddings(position_ids=inputs_embeds)\r\n 185 else:\r\n 186 position_embeds = self.position_embeddings(position_ids=position_ids)\r\n```\r\n-----------------------------------------------------------------------------------------\r\nIf by \"run\" you mean calling fit on the model then it raises the same gradient error\r\n", "Here a dummy example:\r\n```python\r\nimport tensorflow as tf\r\nfrom transformers import TFDistilBertForMaskedLM, DistilBertTokenizer, DistilBertConfig\r\n\r\nconfig = DistilBertConfig(vocab_size=30000)\r\nmodel = TFDistilBertForMaskedLM(config)\r\ninp1 = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name=\"input_ids\")\r\ninp2 = tf.keras.layers.Input(shape=(128,), dtype=tf.int32, name=\"attention_mask\")\r\noutput = model([inp1,inp2])\r\nmodel = tf.keras.models.Model(inputs=[inp1,inp2], outputs=[output])\r\ntokenizer = DistilBertTokenizer.from_pretrained(\"distilbert-base-uncased\")\r\ndata = tokenizer([\"Hello1\", \"Hello2\", \"Hello3\"], truncation=True, max_length=128, padding=\"max_length\", return_tensors=\"tf\")\r\nlabels = tf.ones((3, 128), dtype=tf.int32)\r\nX = tf.data.Dataset.from_tensor_slices((dict(data), labels)).batch(1)\r\nloss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\r\nmodel.compile(loss=loss, optimizer=\"adam\")\r\nmodel.fit(X, epochs=1)\r\n```", "@jplu Thanks for this but this tokenizes the data and then loads it as a tf.data.Dataset. I was looking for an implementation where the tokenization can be integrated in the pipeline itself and can be done on the fly. I found [this](https://github.com/tensorflow/tensorflow/issues/38762) issue on tensorflow but there are no fixes for it yet. Do you have any idea how to do this because my dataset is big enough to fit in colab memory but cannot be fully tokenized in memory?", "Sorry you cannot do this.", "Okay. Thanks for all the help!" ]

[]

[]

[ "> This approach looks great and doesn't seem limiting at all. Implementing it for Wav2Vec2/SpeechToTextTransformer and refactoring/upstreaming methods down the road seems like a good implementation roadmap.\r\n> \r\n> Regarding the implementation of `FeatureProcessors`, what do you have in mind regarding understandability/explicitness? Do you expect something like models, where we aim for maximum accessibility, with copy/pastes and single-file containers, or do you expect something like tokenizers, where some tokenizers inherit from others while modifying certain aspects, and some level of abstraction, making them harder to decypher?\r\n> \r\n> I'm asking because I think it's relevant to the different preprocessing that can be handled by the feature processors. For example, normalizing or converting to MFCCs seems like it would be something quite widespread among speech-based feature processors, do we want to have that in each implementation (abstraction-free) or will the goal be to upstream these methods in the parent class once we identify similarities among feature processors?\r\n\r\n\r\nYeah good question! To be honest, I'm not really sure yet. I would like to enforce the rule that feature extractors can only inherit from `PreTrainedFeatureExtractor` and no other feature extractor. IMO, the best approach to begin with is to limit (as you've suggested) the user-facing API for `FeatureProcessor` to `__call__`, `from_pretrained()`, `save_pretrained()` and maybe something like `from_file()` and then in the beginning. \r\nI think a method like `pad()` is general enough to have this method in the beginning be implemented in `PreTrainedFeatureExtractor` because every extractor will need to do padding no matter what. \r\n\r\nFor pretty much all other methods (actually including `normalization()`), I would copy-paste them into each feature processor and make sure that they are private methods `_normalize()` so that we can later still do some refactoring here if needed.\r\n\r\nSo in general my strategy would be to have as little abstraction as possible - *e.g.* copy-paste classes such as those: https://github.com/huggingface/transformers/blob/19c14579f0c7f5f15c5a5115b2fd18582e61ac3b/src/transformers/models/speech_to_text_transformer/tokenization_speech_to_text_transformer.py#L239 to each feature extractor - and then when having more models maybe move some things upstream into the `PretrainedFeatureExtractor` file ", "Thanks for explaining, sounds like a good approach to me! Thanks for drafting the proposal." ]

[ { "id": 2760822153, "node_id": "MDU6TGFiZWwyNzYwODIyMTUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Tensor%20Parallel", "name": "Tensor Parallel", "color": "1AD0A8", "default": false, "description": "" }, { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

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

[ "@stas00 thanks for starting this thread!\r\n\r\nI guess, in order for everyone to be on the same page, a brief explanation of horizontal parallelism is needed. This would be a good place for future reference and introduce other contributors to the core concepts.\r\n\r\n**NOTE for everyone reading:** If you find any of the explanations below confusing, you can read about Megatron-LM in much more detail in its original paper: https://arxiv.org/pdf/1909.08053.pdf\r\n\r\n\r\n## The core idea\r\n\r\nThe main thing that separates Megatron-style (horizontal) parallelism from vertical parallelism is the way that it splits the model layers between GPUs without the need for idle time during training/inference (i.e. waiting while the previous GPUs complete their work on the previous layers of the model). This makes the whole process much more asynchronous, just like in MapReduce. Here's my rough sketch of how it looks:\r\n\r\n\r\nNow the question is, how do we split the computation of those layers so that the parallelized model weights would be equivalent to the CPU ones?\r\n\r\n## Parallelized layers\r\nLet's start with a simple building block of any transformer: a fully connected layer (nn.Linear) followed by a nonlinear activation (GeLU). Following the Megatron's paper notation, we can write the dot-product part of it as `Y = GeLU(XA)`, where `X` and `Y` are the input and output vectors, and `A` is the weight matrix.\r\n\r\nIf we look at the computation in matrix form, it's easy to see how the matrix multiplication can be split between multiple GPUs:\r\n\r\nBasically, if we split the weight matrix `A` column-wise across `N` GPUs and perform matrix multiplications `XA_1` through `XA_n` in parallel, then we will end up with `N` output vectors `Y_1, Y_2, ..., Y_n` which can be fed into GeLU independently:\r\n\r\n\r\nUsing this principle, we can update an MLP of arbitrary depth, without the need for any synchronization between GPUs until the very end, where we need to reconstruct the output vector from shards. The authors provide a helpful illustration for that:\r\n\r\n\r\n### Quick note on self-attention\r\nParallelizing the multiheaded attention layers is even simpler, since they are already inherently parallel, due to having multiple independent heads! \r\n\r\n\r\n## Practical implementation\r\nIf you want to just dive right in, here are the basic building blocks implemented in Megatron-LM:\r\n\r\n- [ColumnParallelLinear](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/mpu/layers.py#L195)\r\n- [RowParallelLinear](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/mpu/layers.py#L290)\r\n- [ParallelMLP](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/transformer.py#L58)\r\n- [ParallelSelfAttention](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/transformer.py#L112)\r\n\r\nAll of these rely on basic `Scatter`, `Gather` and `Reduce` ops to split and aggregate the weight matrices. Thanks to [PyTorch Distributed](https://pytorch.org/tutorials/intermediate/dist_tuto.html), we can use `torch.distributed.all_reduce` and `all_gather` for that, without having to worry about GPU synchronization. The scatter and gather layers just have to define appropriate forward and backward passes like so:\r\n```python\r\ndef _split(input_):\r\n world_size = get_tensor_model_parallel_world_size()\r\n input_list = split_tensor_along_last_dim(input_, world_size)\r\n rank = get_tensor_model_parallel_rank()\r\n output = input_list[rank].contiguous()\r\n return output\r\n\r\ndef _gather(input_):\r\n world_size = get_tensor_model_parallel_world_size()\r\n last_dim = input_.dim() - 1\r\n rank = get_tensor_model_parallel_rank()\r\n tensor_list = [torch.empty_like(input_) for _ in range(world_size)]\r\n tensor_list[rank] = input_\r\n torch.distributed.all_gather(tensor_list, input_, group=get_tensor_model_parallel_group())\r\n output = torch.cat(tensor_list, dim=last_dim).contiguous()\r\n return output\r\n\r\nclass ScatterToModelParallelRegion(torch.autograd.Function):\r\n def forward(ctx, input_):\r\n return _split(input_)\r\n\r\n def backward(ctx, grad_output):\r\n return _gather(grad_output)\r\n\r\nclass GatherFromModelParallelRegion(torch.autograd.Function):\r\n def forward(ctx, input_):\r\n return _gather(input_)\r\n\r\n def backward(ctx, grad_output):\r\n return _split(grad_output)\r\n```\r\n\r\nIn a single transformer layer, there are 4 communication operations in total, for the forward and backward passes:\r\n\r\n\r\n\r\n## Other things to consider\r\n\r\n#### Parallelized embeddings and output logits\r\nSince the weights of input and output embeddings of BERT/GPT2 are tied, they require a coordinated modification. In the original implementation, the input embedding matrix is parallelized along the vocabulary dimension (column-wise), and the output embeddings' matrix multiplications is parallelized _together with the cross-entropy loss_ to reduce the communication size (see end of section 3 in the paper):\r\n\r\n- [VocabParallelEmbedding](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/mpu/layers.py#L123)\r\n- [parallel_lm_logits](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/language_model.py#L28)\r\n#### Model parallelism-aware Dropout\r\nTransformers have dropout layers outside the model parallel regions before residual connections and within model parallel regions in the self attention block. Because some dropout layers are in a model parallel region, while others are not, we need to treat random number generation carefully to ensure dropout works correctly. See appendix B.2 in the paper for reference.\r\nThe necessary RNG state tracking is implemented in [random.py](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/mpu/random.py)\r\n\r\n#### Hybrid model and data parallelism\r\nCombining horizontal parallelism with data parallelism requires grouping the GPUs in a specific way, as described in appendix B.1:\r\n\r\n\r\n", "Phew! That felt like a start of a whole blog post :smile: \r\n\r\nAs for porting all of this, I would follow [fairseq's example](https://github.com/pytorch/fairseq/blob/master/fairseq/model_parallel/models/transformer.py) and copy Megatron-LM's parallel layers verbatim into an existing (but separate) implementation of `BertModel` or `GPT2Model` as a proof-of-concept and then work from there. \r\n\r\nAfter the first semi-working prototype we could figure out how to implement the switching mechanism between a homogeneous model and a parallelized one, but it's too early to think about that, IMO. What do you think, @stas00 ?", "Amazing! Thank you for this awesome presentation, @anton-l! This could totally be a great blog post - I agree!\r\n\r\nLet me study the information you shared and I will follow up then!\r\n\r\nUntil then I have a quick suggestion: Do you have an easy access to 2 gpus? That would be enough to make a\r\nPoC work and then we can find a larger cluster with more gpus to experiment on and eventually port the 8 splits from fairseq. \r\n\r\nI suppose it'd be easier to implement this for Megatron-LM, but the main use would be t5 and gpt2 where we have most huge models at the moment. So we could start there as well. If it works for you. Which also can be worked on independently of your Megatron-LM PR.", "Regarding the setup: I can borrow a second gpu for the time being, that shouldn't be a problem :)\n\nAs for the models, I think GPT2 is a good candidate for our experiments, since the transformers' implementation is already stable and has multiple smaller checkpoints for quick demos.\n\nAlso, I don't think we should even be too concerned about porting the 8 original splits of fairseq's megatron, since I've already concatenated them for the model's PR. If everything was done correctly, this potentially allows us to create an arbitrary split across 2^n devices, not just 8.", "Sounds good on all accounts. GPT2 would be perfect, @anton-l! \r\n\r\nI had the same thought about just splitting your merged model if needed.\r\n\r\nPlease let us know how we can support you in this endeavor.\r\n\r\njust for you to be aware, I mentioned in the other thread the DeepSpeed version of their Megatron-LM port - perhaps theirs is newer - I haven't had a chance to study it yet. https://github.com/jeffra/DSE/tree/master/megatron-lm . You can diff the different versions against the baseline - that is I assume it has been changed - perhaps it hasn't. If you want to have a look, if not, it is good too. It will be good to start anywhere.\r\n\r\n", "@anton-l Thanks for the great work on this, its really nice to be able to load the pretrained model so thanks for that too! Did you have any progress on fine-tuning across multiple GPUs? Would love to see if the results get any better with some fine-tuning...", "@anton-l, let's do it if you have resources and interest? Let me know how I can be of help.\r\n\r\nNow having used Megatron-LM in [big science experiments](https://github.com/bigscience-workshop/bigscience/blob/master/experiments/gpt2.md) it's time to port it to transformers.", "@stas00 @anton-l Just curious, is Megatron-LM now ported to transformers? Or the proof of concept mentioned in: \r\n\r\n> As for porting all of this, I would follow [fairseq's example](https://github.com/pytorch/fairseq/blob/master/fairseq/model_parallel/models/transformer.py) and copy Megatron-LM's parallel layers verbatim into an existing (but separate) implementation of `BertModel` or `GPT2Model` as a proof-of-concept and then work from there.\r\n\r\nI would love to work on this issue, if there is anything I could do!\r\n", "Thanks for nice the overview.\r\n\r\nHaving read [the paper](https://arxiv.org/pdf/1909.08053.pdf), I disagree with the following statement (emphasis mine)\r\n\r\n> Using this principle, we can update an MLP of *arbitrary depth*, without the need for any synchronization between GPUs until the very end \r\n\r\nIf you split one layer inputs across rows, then the outputs are split across columns, so you need to split the second layer weights across rows, then you need to gather outputs before applying a non-linearity. This is explained in Section 3 of the paper.\r\n" ]

[]

[]

[ "Hello, thanks for opening an issue! We try to keep the github issues for bugs/feature requests.\r\nCould you ask your question on the [forum](https://discusss.huggingface.co) instead?\r\n\r\nAlso, [here's the doc](https://huggingface.co/transformers/model_doc/wav2vec2.html#transformers.Wav2Vec2ForCTC) for `Wav2Vec2ForCTC` which seems to be the model you're interested in.\r\n\r\nThanks!" ]

[]

[]

[ "Hello! You should use the [resize_token_embeddings](https://huggingface.co/transformers/main_classes/model.html?highlight=resize_token_embeddings#transformers.PreTrainedModel.resize_token_embeddings) method for that. Will add that to the documentation." ]

[]

[]

[ "also it is my questions, thanks ", "denoising function is a part of T5 pretraining as well, is there a denoising function implementation in Huggingface repo ? Any advice is appreciated. thanks ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "It would be great to have a `BartForTokenClassification`. Does it use the same head as `BertForTokenClassification`, etc.? \r\n\r\nFeel free to open a PR :) ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Wonderful! Thank you for this fix, @tagucci! \r\n\r\n(I tweaked your PR to run `make style` to appease to auto-formatters to have CI pass)" ]

[]

[]

[ "Hey @dorost1234, \r\n\r\ncould you post your question on the [forum](https://discuss.huggingface.co/) and see whether you can get help from the community there? We try to keep GitHub issues for bug reports mostly. \r\n\r\nIt would also be very important that you attach a notebook or something that allows people to understand what you have done and what you might have missed...", "Hi Patrick\nthanks, sure, I posted here because I mainly wanted to ask if you have in\nthe past compared the performance of original mt5 with the HuggingFace\nmodel side by side on one setting?\nthis is unfortunately a lot of codes for me to share, since I build on top\nof my codebase, and this is not very easy to make a small example showing\nthese differences.\nOverall, knowing if in the past such comparison is done is great.\nthanks.\n\nOn Mon, Feb 22, 2021 at 3:02 PM Patrick von Platen <[email protected]>\nwrote:\n\n> Hey @dorost1234 <https://github.com/dorost1234>,\n>\n> could you post your question on the forum\n> <https://discuss.huggingface.co/> and see whether you can get help from\n> the community there? We try to keep GitHub issues for bug reports mostly.\n>\n> It would also be very important that you attach a notebook or something\n> that allows people to understand what you have done and what you might have\n> missed...\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/10315#issuecomment-783396814>,\n> or unsubscribe\n> <https://github.com/notifications/unsubscribe-auth/AS37NMXVTAFKQLNWCBFUG43TAJPXBANCNFSM4X65FR7Q>\n> .\n>\n", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "I'll push the fixed weights and remove from_pt in test.", "> Ok with the change!\r\n> \r\n> Should we do a patch release for this?\r\n\r\nThink it's a good idea", "Thanks @patrickvonplaten ", "Ok will do a patch this afternoon", "cc @stefan-it and @mrm8488 that have been playing with the model. We'll release v4.3.3 very soon which will contain that patch.", "Should I re-convert the TF model then :thinking: ", "I think you should, once v4.3.3 is released (in a few minutes)", "v4.3.3 has been released!" ]

[]

[]

[ "Please post the full error stacktrace. Which version of the script are you using?", "Hi thank you for your reply. I did post the trace back for the error.\n\nLooking forward to your reply\n\nHadeel\n\n\n\n> On 21 Feb 2021, at 10:01 am, cronoik <[email protected]> wrote:\n> \n> \n> Please post the full error stacktrace. Which version of the script are you using?\n> \n> —\n> You are receiving this because you authored the thread.\n> Reply to this email directly, view it on GitHub, or unsubscribe.\n", "Please use the [forums](https://discuss.huggingface.co/) to debug custom code with help from the community and only when you have isolated a part that is a bug in the library use a GitHub issue. \r\nIn this particular case, you should include the full stack trace, as pointed out before (it's not in your message, just the last frame) and the code that created your model: the error seems to point out that it only returns one value when you need two (since you write `loss, logits = model(...)`). You should debug that by looking at the model return on one batch.", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Sure, I can guide you if you want.\r\n\r\nAs LayoutLM is only a slight adaptation from BERT, I guess you can define `modeling_tf_layoutlm.py` based on `modeling_tf_bert.py`. Note that all layers should be renamed, e.g. `TFBertEmbeddings` -> `TFLayoutLMEmbeddings`. LayoutLM adds position embeddings for the tokens based on the bounding boxes, so I guess this is the only thing that needs to be added in the [embedding layer](https://github.com/huggingface/transformers/blob/88605f37a6fe7bde336f52700229d619b5ffa0f6/src/transformers/models/bert/modeling_tf_bert.py#L131). \r\n\r\nIn PyTorch, we have:\r\nhttps://github.com/huggingface/transformers/blob/88605f37a6fe7bde336f52700229d619b5ffa0f6/src/transformers/models/layoutlm/modeling_layoutlm.py#L65-L68\r\n\r\nSo this will need to be added to the `TFLayoutLMEmbeddings` class. Regarding a conversion script to convert the PyTorch weights into the TF version, there's a general script to convert PyTorch weights to TF 2 models: https://github.com/huggingface/transformers/blob/master/src/transformers/convert_pytorch_checkpoint_to_tf2.py\r\n\r\nThis works if all weights names are identical between the PT and TF implementations.", "@NielsRogge Thanks for the guidance, I think I know where to start. \r\nI'll comment here if I had more questions. I am hoping to have a PR by the end of this week ", "Should I upload TF weights under `https://huggingface.co/microsoft/` in the same place PT weights are stored?", "Yes, pinging @julien-c here to give you access", "pinging @LysandreJik and @sgugger ", "We don't have granular write access to model repos so (unless you're affiliated with Microsoft in some way) I would suggest you upload the files to a model repo under your HF user account and then we (or them) can copy the relevant files to the main repos!\r\n\r\nLet me know if this is a suitable workflow.", "@julien-c Thanks, the suggested workflow sounds good.", "closing this issue as the code is already merged." ]

[]

[]

[ "Hey @xe442,\r\n\r\nActually Reformer cannot make use of `tie_word_embeddings=True` because the output word embedding layer is twice as big as the input layer (because of Reformer's architecture, see section 3) in this blog: https://huggingface.co/blog/reformer", "But, we should in this case give a better error message! Feel free to open a PR to add such an error message :-)", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[ { "id": 2659267025, "node_id": "MDU6TGFiZWwyNjU5MjY3MDI1", "url": "https://api.github.com/repos/huggingface/transformers/labels/DeepSpeed", "name": "DeepSpeed", "color": "4D34F7", "default": false, "description": "" } ]

[]

[ "> Cool that they added it! This all looks pretty good to me!\r\n\r\nWell, it has been there all this time, this PR just bolts it on correctly.\r\n\r\n> Absolutely no problems with moving the regression trainer somewhere accessible, it was just in `test_trainer` because only used there.\r\n\r\nAh, that makes sense. I will rework it then next time I touch on this code.\r\n\r\nThank you for the feedback, @sgugger \r\n" ]

[]

[]

[ "Can reproduce locally, here is a short reproducer from the root of the repo:\r\n```\r\npython examples/token-classification/run_ner.py \\\r\n --model_name_or_path bert-base-uncased \\\r\n --train_file tests/fixtures/tests_samples/conll/sample.json \\\r\n --validation_file tests/fixtures/tests_samples/conll/sample.json \\\r\n --output_dir /tmp/test-ner \\\r\n --overwrite_output_dir \\\r\n --do_train \\\r\n --do_eval \\\r\n --label_smoothing_factor 0.1\r\n```\r\n\r\nWill look into it tomorrow." ]

[]

[]

[]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Which code caused this error?", "This issue comes from the hosted API https://huggingface.co/bert-base-german-cased?text=Ich+bin+%5BMASK%5D", "@tholor The url above currently loads for me, but to be future-proof should we cp the files currently loaded from that S3 bucket to the corresponding model repo (here, https://huggingface.co/bert-base-german-cased)?\r\n\r\ncc'ing @LysandreJik ", "@julien-c Sure, let's copy them from our S3 to the model repo. ", "copied to the model repo in \r\nhttps://huggingface.co/bert-base-german-cased/commit/876457621368b8c955478cfe1cdee634f47ea34c\r\n\r\nChanged hardcoded url in https://github.com/huggingface/transformers/pull/10353", "@Narsil could you please check if the inference widget works for this model when you get a chance to upgrade the transformers dependency in the API? Thanks!", "⚠️ Can't load tokenizer using from_pretrained, please update its configuration: 400 Client Error: Bad Request for url: https://int-deepset-models-bert.s3.eu-central-1.amazonaws.com/pytorch/bert-base-german-cased-vocab.txt\r\nIt's still not working!", "Hi @George-Ogden, the change was merged two days ago and is therefore available on the `master` branch, but not yet in a release.\r\n\r\nDo you still get the same error when installing from source?", "This is on the inference API on the website I haven't tried it from source." ]

[]

[]

[ "This has been fixed a few days ago, I believe. Look at the [master doc tokenizer page](https://huggingface.co/transformers/master/main_classes/tokenizer.html) (the stable documentation is only updated at each release).", "Yes, you are right." ]

[]

[]

[ "Thanks for fixing!" ]

[]

[]

[]

[]

[]

[ "What is the output of:\r\n```\r\nimport tensorflow\r\nprint(tensorflow.__version__)\r\n```\r\n?", "> What is the output of:\r\n> \r\n> ```\r\n> import tensorflow\r\n> print(tensorflow.__version__)\r\n> ```\r\n> \r\n> ?\r\n\r\n'2.4.0-rc0'", "Hello!\r\n\r\nWe currently don't support other implementations than the Google's pypi versions. The reason is because we don't tests on other versions and then we cannot guarantee it will works on those \"extra\" versions.\r\n\r\nTo make `transformers` works on Mac I suggest you to use the official version of TensorFlow as proposed in their documentation https://www.tensorflow.org/install/pip", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[ { "id": 2796628563, "node_id": "MDU6TGFiZWwyNzk2NjI4NTYz", "url": "https://api.github.com/repos/huggingface/transformers/labels/WIP", "name": "WIP", "color": "234C99", "default": false, "description": "Label your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress" } ]

[]

[ "That's very neat, @anton-l! thank you for the port\r\n\r\nYou demonstrated a very good creativity by finding a way to recompose the model shards!\r\n\r\n> This one will probably be fun to test with DeepSpeed, as @stas00 mentioned it's referenced a lot in its docs \r\n\r\nAs you correctly noticed studying Megatron-LM's horizontal model parallel sharding is on my TODO list. \r\n\r\nI suppose since `transformers` currently doesn't provide this feature you didn't port that part of the model, correct? i.e. you unsharded it. I had a brief read through the PR and didn't see anything of a sort - unless I somehow missed it? And without this feature, this is like any other `transformers` model - It's its horizontal model parallel feature that is needed to complete 3D parallelism with Deepspeed. Your PR is an excellent start.\r\n\r\nI think the part that deals with sharding is here in the original:\r\nhttps://github.com/jeffra/DSE/blob/79888e162425e8d64043a9597ee14751bd4b53d1/megatron/data/realm_index.py\r\nThough this is the NVIDIA version. \r\n\r\nSo if the horizontal MP is eventually re-ported (I hope it will be so) the model will need to know when to load the flattened version and when the sharded one. But `transformers` doesn't even have a framework for loading multiple-part models at the moment, so I guess we will cross that bridge when we get to it.\r\n\r\nI'm just just thinking aloud here, considering different options, not making any requests ;)\r\n\r\n-------\r\n\r\nThe fp32 weights are ~41GB https://huggingface.co/anton-l/megatron-11b/tree/main - i.e. it's quite similar to t5-11b, so it should be possible to load it on a 40GB gpu w/ DeepSpeed ZeRO-Offload if there are some 256GB of RAM available.\r\n\r\n-----\r\n\r\nAlso, FYI, Deepspeed are making a new port of Megatron-LM to work with DeepSpeed. https://github.com/jeffra/DSE/tree/master/megatron-lm\r\n", "@stas00 you're correct, I didn't port the model-parallel implementation. Fairseq uses an older Megatron-LM version as a submodule [here](https://github.com/pytorch/fairseq/tree/master/fairseq/model_parallel) for its MP map-reduce fuctions. This makes it quite cumbersome to reproduce, since it requires compiling an older `apex` library among other dependencies with broken versioning. It would also require a patched version of faiseq's state loader, since right now it requires exactly 8 GPUs available to load the sharded checkpoint correctly.\r\n\r\nHowever, on the surface it seems like adding support for model parallelism comes down to porting `VocabParallelEmbedding`, `ColumnParallelLinear` and `RowParallelLinear` layers as implemented [here](https://github.com/ngoyal2707/Megatron-LM/blob/adb23324c222aad0aad89308e70302d996a5eaeb/mpu/layers.py). This seems doable, but I don't have multiple GPUs to test it out :(\r\n\r\nI guess a proper MP implementation should also take care of splitting the checkpointed layers regardless of how many GPUs are available (i.e. 2, 4 or 8). That would remove the requirement to have a full DGX setup if the user is willing to use gradient checkpointing/accumulation instead.\r\n\r\n", "@anhon-l, in order not to make your and reviewers' lives unnecessarily difficult, let's take the discussion of the Horizontal MP to a dedicated issue, since it could take some time to figure and none of is required for you to complete this PR and I trust @patil-suraj and @patrickvonplaten will support you at completing this awesome effort. \r\n\r\nSo if you could re-post your last comment here: https://github.com/huggingface/transformers/issues/10321 and I will follow up there. Thank you!", "> ```\r\n> ['Before boarding your rocket to Mars, remember to pack these items: 1. A parachute.',\r\n> 'Before boarding your rocket to Mars, remember to pack these items: 1. A parachute $100 bill2. A copy of your passport3. A copy of your passport444',\r\n> 'Before boarding your rocket to Mars, remember to pack these items: 1. A parachute $1 million dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars dollars']\r\n> ```\r\n> \r\n> To be honest, I'm not too impressed with its text-generation power. 😄 I guess it's either that the model was too large to train it for enough steps, or I missed something during the conversion. The original implementation does not have a text-generation script (or any non-wikitext results, for that matter), so I'm kinda in the dark here.\r\n\r\nThis is amazing work, big kudos! The seemingly low text-generation quality surprises me though, because of the crazy good output you get from https://inferkit.com/ which is also just Megatron11b, according to their docs (https://inferkit.com/docs/generation). Their output seems to be much better than GPT2.", "@anton-l, would you like to complete this PR? For it to be reviewed it needs to be a normal PR and not a draft.\r\n\r\nI marked it as WIP so that the stale bot won't try to close it.\r\n\r\nThank you.", "pinging @anton-l - let's revisit this? Please let us know what you need.\r\n\r\nI know meanwhile someone else did the porting of the original GPT2-345M checkpoint https://huggingface.co/nvidia/megatron-gpt2-345m and I see from the docs they use straight GPT2 transformers model to operate it. \r\nhttps://huggingface.co/nvidia/megatron-gpt2-345m#text-generation\r\n\r\nAll they have is a conversion script:\r\nhttps://github.com/huggingface/transformers/tree/master/src/transformers/models/megatron_gpt2\r\nCan the same be done with the fairseq version - i.e. reuse some of the existing models for that? or is it unique enough to warrant its own?\r\n\r\nPlease bear with me, I'm just starting to figure out Megatron-LM and its variants (there is also a Deepspeed variant), so I'm just slightly above clueless at the moment - should have a better understanding in a few days once I had a chance working with it.", "@stas00 sorry for the late reply!\r\n\r\nIt's great that someone figured out a way to post the original megatron models. When I was looking into that, it wasn't exactly straightforward due to the differences between the attention block implementations in HF GPT2 and Megatron, which was probably patched/parameterized in the meantime.\r\n\r\nI chose to implement a separate model for the fairseq megatron because the model uses the same code as the existing MBART & FSMT, but there's only an encoder model, without the decoder. However, we could take a different route and convert the fairseq weights to fit GPT2, since it's clearly possible now. I'll try that tomorrow, and if it works out, we can discard this PR and just add a simple conversion script :+1: ", "This PR seems very promising and I know the model would be really useful to many.\r\n\r\nAs it was earlier pointed out, the converted model doesn't seem to have the same quality of generation as the model elsewhere. Perhaps the conversion script could have caused it somehow? Just curious if there was any success with converting the fairseq weights to fit GPT2. " ]

[]

[]

[ "hi @IamAdiSri \r\n\r\nWhat is your Transformers version ? mBART-50 currently only works on master.", "@patil-suraj I'm on version 4.3.2, but I tried it with the modules in master branch. I searched through the repository but as far as I can tell, none of the relevant mbart modules take `forced_bos_token_id` as a parameter in their generate function.\r\n\r\nI'm looking the example on [this](https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt) page btw.", "`forced_bos_token_id` is included on master, https://github.com/huggingface/transformers/blob/master/src/transformers/generation_utils.py#L549\r\n\r\nyou should install from source to use mBART-50", "Oh okay, thank you." ]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "I too ran into this problem and its caused by turning on evaluation strategy which then adds metrics in the log_history of the models state, which is using numpy data types and causes the JSON encoder issue. That was the case with 4.3.3. There appear to be a bunch of changes in the trainer in the works, whether this has been fixed as a result of those i've not checked.", "As a temporary work around you can modify trainer.py at line 1260 \"output = {**logs, **{\"step\": self.state.global_step}}\" and add the following three lines after. If the metrics are being calculated the same in the latest code as in 4.3.3 then something like this may also be needed going forward, or things calling the log method will need to ensure they safely cast data points beforehand if its going to be added to the trainer state still.\r\n\r\n```\r\n for k,v in output.items():\r\n if isinstance(v, np.generic):\r\n output[k]=v.item()\r\n```", "I confirm I can reproduce in master. Will investigate more tomorrow.", "My only comment on the fix submitted is that it targets the metrics output, but will not stop others putting things into the log history in the model state which later on cause the same problem if serializing the state to json. " ]

[ { "id": 2357479466, "node_id": "MDU6TGFiZWwyMzU3NDc5NDY2", "url": "https://api.github.com/repos/huggingface/transformers/labels/fsmt", "name": "fsmt", "color": "d0e884", "default": false, "description": "" } ]

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

[ "Pinging @stas00 here", "Thank you for the ping, @NielsRogge \r\n\r\n@tagucci, when you file an issue you will find a list of who to tag for what topic, so please use it to tag the right people. Otherwise it's hard for everybody to try to follow all issues.\r\n\r\nalso when you link to a line of code in github, always hit `y` first to get the exact sha (it rewrites the url to embed the current git sha). Otherwise your links quickly become invalid, e.g. I have no idea where you were trying to link to in your link to [transformer_wmt_en_de](https://github.com/pytorch/fairseq/blob/master/fairseq/models/transformer.py#L1046) as the code was modified today.\r\n\r\n--------------------------------\r\n\r\nOK, could you first clarify where do you get \"decoder.embed_out weight is missing\" - the command line and the backtrace please. Also a dump of the model (i.e. `print(model)`.\r\n\r\nNow to the guess work.\r\n\r\nDoes your model miss `output_projection` weight key?\r\n\r\nThe context is here:\r\nhttps://github.com/pytorch/fairseq/blob/ab560669cd9baaa4009e1fd01c970f8ffccd1ee0/fairseq/models/transformer.py#L950-L960\r\n\r\nfairseq has different versions of their code, and some have keys renamed or added, that's why they have all that logic.\r\n\r\nYou can see that it's a simple alias - i.e. in fsmt decoder embed and output are always shared.\r\n\r\nhttps://github.com/huggingface/transformers/blob/461e8cacf94d1f76367cc9ba2cfd5b9bd3641c81/src/transformers/models/fsmt/modeling_fsmt.py#L651\r\n\r\nSo if it's missing you can assign it in the conversion script:\r\n```\r\n model_state_dict[\"model.decoder.output_projection.weight\"] = model_state_dict[\"model.decoder.embed_tokens.weight\"]\r\n```\r\nadd this to this line:\r\nhttps://github.com/huggingface/transformers/blob/461e8cacf94d1f76367cc9ba2cfd5b9bd3641c81/src/transformers/models/fsmt/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py#L247\r\n\r\nbut again I could have guessed wrong and will need to see the model dump to tell you more.\r\n\r\nYou can see the dump of original model I converted from here: https://github.com/stas00/porting/blob/master/transformers/fairseq-wmt19/nbs/config.ipynb\r\n", "@NielsRogge\r\nThanks for pinging @stas00!\r\n@stas00\r\nSorry for the inconvenience of linking the code.\r\nFollowing your advice, my model args and model dump are as below.\r\n\r\n> in fsmt decoder embed and output are always shared.\r\n\r\nAs you said, fsmt does not have decoder embed and output seperately, my fairseq `transformer_wmt_en_de` without `share_decoder_input_output_embed` cannot fit fsmt in transformers. In this case, do I need to retrain fairseq model with `share_decoder_input_output_embed` or modify [FSMTDecoer](https://github.com/huggingface/transformers/blob/461e8cacf94d1f76367cc9ba2cfd5b9bd3641c81/src/transformers/models/fsmt/modeling_fsmt.py#L622)? \r\n\r\n\r\n```python\r\nimport torch\r\nfrom pprint import pprint\r\nchkpt = torch.load(\"model/checkpoint_best.pt\")\r\nmodel = chkpt[\"model\"]\r\npprint(vars(chkpt[\"args\"]))\r\nprint(\"\\n\".join(model.keys()))\r\n```\r\n\r\n```\r\n# args\r\n{'activation_dropout': 0.0,\r\n 'activation_fn': 'relu',\r\n 'adam_betas': '(0.9, 0.98)',\r\n 'adam_eps': 1e-08,\r\n 'adaptive_input': False,\r\n 'adaptive_softmax_cutoff': None,\r\n 'adaptive_softmax_dropout': 0,\r\n 'arch': 'transformer_wmt_en_de',\r\n 'attention_dropout': 0.0,\r\n 'best_checkpoint_metric': 'loss',\r\n 'bpe': None,\r\n 'bucket_cap_mb': 25,\r\n 'clip_norm': 0.0,\r\n 'cpu': False,\r\n 'criterion': 'label_smoothed_cross_entropy',\r\n 'cross_self_attention': False,\r\n 'curriculum': 0,\r\n 'data': './data/data.src_trg',\r\n 'dataset_impl': None,\r\n 'ddp_backend': 'no_c10d',\r\n 'decoder_attention_heads': 8,\r\n 'decoder_embed_dim': 512,\r\n 'decoder_embed_path': None,\r\n 'decoder_ffn_embed_dim': 2048,\r\n 'decoder_input_dim': 512,\r\n 'decoder_layerdrop': 0,\r\n 'decoder_layers': 6,\r\n 'decoder_layers_to_keep': None,\r\n 'decoder_learned_pos': False,\r\n 'decoder_normalize_before': False,\r\n 'decoder_output_dim': 512,\r\n 'device_id': 0,\r\n 'disable_validation': False,\r\n 'distributed_backend': 'nccl',\r\n 'distributed_init_method': 'tcp://localhost:16441',\r\n 'distributed_no_spawn': False,\r\n 'distributed_port': -1,\r\n 'distributed_rank': 0,\r\n 'distributed_world_size': 4,\r\n 'dropout': 0.1,\r\n 'empty_cache_freq': 0,\r\n 'encoder_attention_heads': 8,\r\n 'encoder_embed_dim': 512,\r\n 'encoder_embed_path': None,\r\n 'encoder_ffn_embed_dim': 2048,\r\n 'encoder_layerdrop': 0,\r\n 'encoder_layers': 6,\r\n 'encoder_layers_to_keep': None,\r\n 'encoder_learned_pos': False,\r\n 'encoder_normalize_before': False,\r\n 'fast_stat_sync': False,\r\n 'find_unused_parameters': False,\r\n 'fix_batches_to_gpus': False,\r\n 'fixed_validation_seed': None,\r\n 'fp16': False,\r\n 'fp16_init_scale': 128,\r\n 'fp16_scale_tolerance': 0.0,\r\n 'fp16_scale_window': None,\r\n 'keep_interval_updates': 20,\r\n 'keep_last_epochs': -1,\r\n 'label_smoothing': 0.1,\r\n 'layer_wise_attention': False,\r\n 'layernorm_embedding': False,\r\n 'lazy_load': False,\r\n 'left_pad_source': True,\r\n 'left_pad_target': False,\r\n 'load_alignments': False,\r\n 'log_format': 'json',\r\n 'log_interval': 50,\r\n 'lr': [0.0007],\r\n 'lr_scheduler': 'inverse_sqrt',\r\n 'max_epoch': 100,\r\n 'max_sentences': None,\r\n 'max_sentences_valid': None,\r\n 'max_source_positions': 1024,\r\n 'max_target_positions': 1024,\r\n 'max_tokens': 4096,\r\n 'max_tokens_valid': 4096,\r\n 'max_update': 0,\r\n 'maximize_best_checkpoint_metric': False,\r\n 'memory_efficient_fp16': False,\r\n 'min_loss_scale': 0.0001,\r\n 'min_lr': 1e-09,\r\n 'no_cross_attention': False,\r\n 'no_epoch_checkpoints': True,\r\n 'no_last_checkpoints': False,\r\n 'no_progress_bar': True,\r\n 'no_save': False,\r\n 'no_save_optimizer_state': False,\r\n 'no_scale_embedding': False,\r\n 'no_token_positional_embeddings': False,\r\n 'num_workers': 1,\r\n 'optimizer': 'adam',\r\n 'optimizer_overrides': '{}',\r\n 'raw_text': False,\r\n 'required_batch_size_multiple': 8,\r\n 'reset_dataloader': False,\r\n 'reset_lr_scheduler': False,\r\n 'reset_meters': False,\r\n 'reset_optimizer': False,\r\n 'restore_file': 'checkpoint_last.pt',\r\n 'save_dir': './data/models',\r\n 'save_interval': 1,\r\n 'save_interval_updates': 1000,\r\n 'seed': 1,\r\n 'sentence_avg': False,\r\n 'share_all_embeddings': False,\r\n 'share_decoder_input_output_embed': False,\r\n 'skip_invalid_size_inputs_valid_test': True,\r\n 'source_lang': 'src',\r\n 'target_lang': 'trg',\r\n 'task': 'translation',\r\n 'tensorboard_logdir': '',\r\n 'threshold_loss_scale': None,\r\n 'tokenizer': None,\r\n 'train_subset': 'train',\r\n 'truncate_source': False,\r\n 'update_freq': [16],\r\n 'upsample_primary': 1,\r\n 'use_bmuf': False,\r\n 'user_dir': None,\r\n 'valid_subset': 'valid',\r\n 'validate_interval': 1,\r\n 'warmup_init_lr': 1e-07,\r\n 'warmup_updates': 4000,\r\n 'weight_decay': 0.0}\r\n```\r\n\r\n```\r\n# model dump\r\nencoder.version\r\nencoder.embed_tokens.weight\r\nencoder.embed_positions._float_tensor\r\nencoder.layers.0.self_attn.k_proj.weight\r\nencoder.layers.0.self_attn.k_proj.bias\r\nencoder.layers.0.self_attn.v_proj.weight\r\nencoder.layers.0.self_attn.v_proj.bias\r\nencoder.layers.0.self_attn.q_proj.weight\r\nencoder.layers.0.self_attn.q_proj.bias\r\nencoder.layers.0.self_attn.out_proj.weight\r\nencoder.layers.0.self_attn.out_proj.bias\r\nencoder.layers.0.self_attn_layer_norm.weight\r\nencoder.layers.0.self_attn_layer_norm.bias\r\nencoder.layers.0.fc1.weight\r\nencoder.layers.0.fc1.bias\r\nencoder.layers.0.fc2.weight\r\nencoder.layers.0.fc2.bias\r\nencoder.layers.0.final_layer_norm.weight\r\nencoder.layers.0.final_layer_norm.bias\r\nencoder.layers.1.self_attn.k_proj.weight\r\nencoder.layers.1.self_attn.k_proj.bias\r\nencoder.layers.1.self_attn.v_proj.weight\r\nencoder.layers.1.self_attn.v_proj.bias\r\nencoder.layers.1.self_attn.q_proj.weight\r\nencoder.layers.1.self_attn.q_proj.bias\r\nencoder.layers.1.self_attn.out_proj.weight\r\nencoder.layers.1.self_attn.out_proj.bias\r\nencoder.layers.1.self_attn_layer_norm.weight\r\nencoder.layers.1.self_attn_layer_norm.bias\r\nencoder.layers.1.fc1.weight\r\nencoder.layers.1.fc1.bias\r\nencoder.layers.1.fc2.weight\r\nencoder.layers.1.fc2.bias\r\nencoder.layers.1.final_layer_norm.weight\r\nencoder.layers.1.final_layer_norm.bias\r\nencoder.layers.2.self_attn.k_proj.weight\r\nencoder.layers.2.self_attn.k_proj.bias\r\nencoder.layers.2.self_attn.v_proj.weight\r\nencoder.layers.2.self_attn.v_proj.bias\r\nencoder.layers.2.self_attn.q_proj.weight\r\nencoder.layers.2.self_attn.q_proj.bias\r\nencoder.layers.2.self_attn.out_proj.weight\r\nencoder.layers.2.self_attn.out_proj.bias\r\nencoder.layers.2.self_attn_layer_norm.weight\r\nencoder.layers.2.self_attn_layer_norm.bias\r\nencoder.layers.2.fc1.weight\r\nencoder.layers.2.fc1.bias\r\nencoder.layers.2.fc2.weight\r\nencoder.layers.2.fc2.bias\r\nencoder.layers.2.final_layer_norm.weight\r\nencoder.layers.2.final_layer_norm.bias\r\nencoder.layers.3.self_attn.k_proj.weight\r\nencoder.layers.3.self_attn.k_proj.bias\r\nencoder.layers.3.self_attn.v_proj.weight\r\nencoder.layers.3.self_attn.v_proj.bias\r\nencoder.layers.3.self_attn.q_proj.weight\r\nencoder.layers.3.self_attn.q_proj.bias\r\nencoder.layers.3.self_attn.out_proj.weight\r\nencoder.layers.3.self_attn.out_proj.bias\r\nencoder.layers.3.self_attn_layer_norm.weight\r\nencoder.layers.3.self_attn_layer_norm.bias\r\nencoder.layers.3.fc1.weight\r\nencoder.layers.3.fc1.bias\r\nencoder.layers.3.fc2.weight\r\nencoder.layers.3.fc2.bias\r\nencoder.layers.3.final_layer_norm.weight\r\nencoder.layers.3.final_layer_norm.bias\r\nencoder.layers.4.self_attn.k_proj.weight\r\nencoder.layers.4.self_attn.k_proj.bias\r\nencoder.layers.4.self_attn.v_proj.weight\r\nencoder.layers.4.self_attn.v_proj.bias\r\nencoder.layers.4.self_attn.q_proj.weight\r\nencoder.layers.4.self_attn.q_proj.bias\r\nencoder.layers.4.self_attn.out_proj.weight\r\nencoder.layers.4.self_attn.out_proj.bias\r\nencoder.layers.4.self_attn_layer_norm.weight\r\nencoder.layers.4.self_attn_layer_norm.bias\r\nencoder.layers.4.fc1.weight\r\nencoder.layers.4.fc1.bias\r\nencoder.layers.4.fc2.weight\r\nencoder.layers.4.fc2.bias\r\nencoder.layers.4.final_layer_norm.weight\r\nencoder.layers.4.final_layer_norm.bias\r\nencoder.layers.5.self_attn.k_proj.weight\r\nencoder.layers.5.self_attn.k_proj.bias\r\nencoder.layers.5.self_attn.v_proj.weight\r\nencoder.layers.5.self_attn.v_proj.bias\r\nencoder.layers.5.self_attn.q_proj.weight\r\nencoder.layers.5.self_attn.q_proj.bias\r\nencoder.layers.5.self_attn.out_proj.weight\r\nencoder.layers.5.self_attn.out_proj.bias\r\nencoder.layers.5.self_attn_layer_norm.weight\r\nencoder.layers.5.self_attn_layer_norm.bias\r\nencoder.layers.5.fc1.weight\r\nencoder.layers.5.fc1.bias\r\nencoder.layers.5.fc2.weight\r\nencoder.layers.5.fc2.bias\r\nencoder.layers.5.final_layer_norm.weight\r\nencoder.layers.5.final_layer_norm.bias\r\ndecoder.embed_out\r\ndecoder.version\r\ndecoder.embed_tokens.weight\r\ndecoder.embed_positions._float_tensor\r\ndecoder.layers.0.self_attn.k_proj.weight\r\ndecoder.layers.0.self_attn.k_proj.bias\r\ndecoder.layers.0.self_attn.v_proj.weight\r\ndecoder.layers.0.self_attn.v_proj.bias\r\ndecoder.layers.0.self_attn.q_proj.weight\r\ndecoder.layers.0.self_attn.q_proj.bias\r\ndecoder.layers.0.self_attn.out_proj.weight\r\ndecoder.layers.0.self_attn.out_proj.bias\r\ndecoder.layers.0.self_attn_layer_norm.weight\r\ndecoder.layers.0.self_attn_layer_norm.bias\r\ndecoder.layers.0.encoder_attn.k_proj.weight\r\ndecoder.layers.0.encoder_attn.k_proj.bias\r\ndecoder.layers.0.encoder_attn.v_proj.weight\r\ndecoder.layers.0.encoder_attn.v_proj.bias\r\ndecoder.layers.0.encoder_attn.q_proj.weight\r\ndecoder.layers.0.encoder_attn.q_proj.bias\r\ndecoder.layers.0.encoder_attn.out_proj.weight\r\ndecoder.layers.0.encoder_attn.out_proj.bias\r\ndecoder.layers.0.encoder_attn_layer_norm.weight\r\ndecoder.layers.0.encoder_attn_layer_norm.bias\r\ndecoder.layers.0.fc1.weight\r\ndecoder.layers.0.fc1.bias\r\ndecoder.layers.0.fc2.weight\r\ndecoder.layers.0.fc2.bias\r\ndecoder.layers.0.final_layer_norm.weight\r\ndecoder.layers.0.final_layer_norm.bias\r\ndecoder.layers.1.self_attn.k_proj.weight\r\ndecoder.layers.1.self_attn.k_proj.bias\r\ndecoder.layers.1.self_attn.v_proj.weight\r\ndecoder.layers.1.self_attn.v_proj.bias\r\ndecoder.layers.1.self_attn.q_proj.weight\r\ndecoder.layers.1.self_attn.q_proj.bias\r\ndecoder.layers.1.self_attn.out_proj.weight\r\ndecoder.layers.1.self_attn.out_proj.bias\r\ndecoder.layers.1.self_attn_layer_norm.weight\r\ndecoder.layers.1.self_attn_layer_norm.bias\r\ndecoder.layers.1.encoder_attn.k_proj.weight\r\ndecoder.layers.1.encoder_attn.k_proj.bias\r\ndecoder.layers.1.encoder_attn.v_proj.weight\r\ndecoder.layers.1.encoder_attn.v_proj.bias\r\ndecoder.layers.1.encoder_attn.q_proj.weight\r\ndecoder.layers.1.encoder_attn.q_proj.bias\r\ndecoder.layers.1.encoder_attn.out_proj.weight\r\ndecoder.layers.1.encoder_attn.out_proj.bias\r\ndecoder.layers.1.encoder_attn_layer_norm.weight\r\ndecoder.layers.1.encoder_attn_layer_norm.bias\r\ndecoder.layers.1.fc1.weight\r\ndecoder.layers.1.fc1.bias\r\ndecoder.layers.1.fc2.weight\r\ndecoder.layers.1.fc2.bias\r\ndecoder.layers.1.final_layer_norm.weight\r\ndecoder.layers.1.final_layer_norm.bias\r\ndecoder.layers.2.self_attn.k_proj.weight\r\ndecoder.layers.2.self_attn.k_proj.bias\r\ndecoder.layers.2.self_attn.v_proj.weight\r\ndecoder.layers.2.self_attn.v_proj.bias\r\ndecoder.layers.2.self_attn.q_proj.weight\r\ndecoder.layers.2.self_attn.q_proj.bias\r\ndecoder.layers.2.self_attn.out_proj.weight\r\ndecoder.layers.2.self_attn.out_proj.bias\r\ndecoder.layers.2.self_attn_layer_norm.weight\r\ndecoder.layers.2.self_attn_layer_norm.bias\r\ndecoder.layers.2.encoder_attn.k_proj.weight\r\ndecoder.layers.2.encoder_attn.k_proj.bias\r\ndecoder.layers.2.encoder_attn.v_proj.weight\r\ndecoder.layers.2.encoder_attn.v_proj.bias\r\ndecoder.layers.2.encoder_attn.q_proj.weight\r\ndecoder.layers.2.encoder_attn.q_proj.bias\r\ndecoder.layers.2.encoder_attn.out_proj.weight\r\ndecoder.layers.2.encoder_attn.out_proj.bias\r\ndecoder.layers.2.encoder_attn_layer_norm.weight\r\ndecoder.layers.2.encoder_attn_layer_norm.bias\r\ndecoder.layers.2.fc1.weight\r\ndecoder.layers.2.fc1.bias\r\ndecoder.layers.2.fc2.weight\r\ndecoder.layers.2.fc2.bias\r\ndecoder.layers.2.final_layer_norm.weight\r\ndecoder.layers.2.final_layer_norm.bias\r\ndecoder.layers.3.self_attn.k_proj.weight\r\ndecoder.layers.3.self_attn.k_proj.bias\r\ndecoder.layers.3.self_attn.v_proj.weight\r\ndecoder.layers.3.self_attn.v_proj.bias\r\ndecoder.layers.3.self_attn.q_proj.weight\r\ndecoder.layers.3.self_attn.q_proj.bias\r\ndecoder.layers.3.self_attn.out_proj.weight\r\ndecoder.layers.3.self_attn.out_proj.bias\r\ndecoder.layers.3.self_attn_layer_norm.weight\r\ndecoder.layers.3.self_attn_layer_norm.bias\r\ndecoder.layers.3.encoder_attn.k_proj.weight\r\ndecoder.layers.3.encoder_attn.k_proj.bias\r\ndecoder.layers.3.encoder_attn.v_proj.weight\r\ndecoder.layers.3.encoder_attn.v_proj.bias\r\ndecoder.layers.3.encoder_attn.q_proj.weight\r\ndecoder.layers.3.encoder_attn.q_proj.bias\r\ndecoder.layers.3.encoder_attn.out_proj.weight\r\ndecoder.layers.3.encoder_attn.out_proj.bias\r\ndecoder.layers.3.encoder_attn_layer_norm.weight\r\ndecoder.layers.3.encoder_attn_layer_norm.bias\r\ndecoder.layers.3.fc1.weight\r\ndecoder.layers.3.fc1.bias\r\ndecoder.layers.3.fc2.weight\r\ndecoder.layers.3.fc2.bias\r\ndecoder.layers.3.final_layer_norm.weight\r\ndecoder.layers.3.final_layer_norm.bias\r\ndecoder.layers.4.self_attn.k_proj.weight\r\ndecoder.layers.4.self_attn.k_proj.bias\r\ndecoder.layers.4.self_attn.v_proj.weight\r\ndecoder.layers.4.self_attn.v_proj.bias\r\ndecoder.layers.4.self_attn.q_proj.weight\r\ndecoder.layers.4.self_attn.q_proj.bias\r\ndecoder.layers.4.self_attn.out_proj.weight\r\ndecoder.layers.4.self_attn.out_proj.bias\r\ndecoder.layers.4.self_attn_layer_norm.weight\r\ndecoder.layers.4.self_attn_layer_norm.bias\r\ndecoder.layers.4.encoder_attn.k_proj.weight\r\ndecoder.layers.4.encoder_attn.k_proj.bias\r\ndecoder.layers.4.encoder_attn.v_proj.weight\r\ndecoder.layers.4.encoder_attn.v_proj.bias\r\ndecoder.layers.4.encoder_attn.q_proj.weight\r\ndecoder.layers.4.encoder_attn.q_proj.bias\r\ndecoder.layers.4.encoder_attn.out_proj.weight\r\ndecoder.layers.4.encoder_attn.out_proj.bias\r\ndecoder.layers.4.encoder_attn_layer_norm.weight\r\ndecoder.layers.4.encoder_attn_layer_norm.bias\r\ndecoder.layers.4.fc1.weight\r\ndecoder.layers.4.fc1.bias\r\ndecoder.layers.4.fc2.weight\r\ndecoder.layers.4.fc2.bias\r\ndecoder.layers.4.final_layer_norm.weight\r\ndecoder.layers.4.final_layer_norm.bias\r\ndecoder.layers.5.self_attn.k_proj.weight\r\ndecoder.layers.5.self_attn.k_proj.bias\r\ndecoder.layers.5.self_attn.v_proj.weight\r\ndecoder.layers.5.self_attn.v_proj.bias\r\ndecoder.layers.5.self_attn.q_proj.weight\r\ndecoder.layers.5.self_attn.q_proj.bias\r\ndecoder.layers.5.self_attn.out_proj.weight\r\ndecoder.layers.5.self_attn.out_proj.bias\r\ndecoder.layers.5.self_attn_layer_norm.weight\r\ndecoder.layers.5.self_attn_layer_norm.bias\r\ndecoder.layers.5.encoder_attn.k_proj.weight\r\ndecoder.layers.5.encoder_attn.k_proj.bias\r\ndecoder.layers.5.encoder_attn.v_proj.weight\r\ndecoder.layers.5.encoder_attn.v_proj.bias\r\ndecoder.layers.5.encoder_attn.q_proj.weight\r\ndecoder.layers.5.encoder_attn.q_proj.bias\r\ndecoder.layers.5.encoder_attn.out_proj.weight\r\ndecoder.layers.5.encoder_attn.out_proj.bias\r\ndecoder.layers.5.encoder_attn_layer_norm.weight\r\ndecoder.layers.5.encoder_attn_layer_norm.bias\r\ndecoder.layers.5.fc1.weight\r\ndecoder.layers.5.fc1.bias\r\ndecoder.layers.5.fc2.weight\r\ndecoder.layers.5.fc2.bias\r\ndecoder.layers.5.final_layer_norm.weight\r\ndecoder.layers.5.final_layer_norm.bias\r\n```", "Thank you for the model dump, so my guess was correct - it's missing `output_projection` and I gave you the solution at the end of my previous comment.\r\n\r\nI still don't know what the error you get, when and the backtrace, but perhaps my guessed solution is all you need.\r\n\r\nBut no, you don't need to re-train.\r\n\r\nif it works could you adapt the script to check if the checkpoint that is being loaded doesn't have this key and if so to copy it as I suggested?", "@stas00 \r\nRunning [convert_fsmt_original_pytorch_checkpoint_to_pytorch.py]( https://github.com/huggingface/transformers/blob/461e8cacf94d1f76367cc9ba2cfd5b9bd3641c81/src/transformers/models/fsmt/convert_fsmt_original_pytorch_checkpoint_to_pytorch.py) is successful, but there is something wrong.\r\nIn comparing fairseq model provided by `torch.hub` and converted HF model, the translation result is matched.\r\n\r\n```python\r\nfrom transformers import FSMTForConditionalGeneration, FSMTTokenizer, TranslationPipeline\r\nimport torch\r\n\r\ninput_text = \"Machine learning is great!\"\r\n# fairseq\r\nen2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de', checkpoint_file='model1.pt:model2.pt:model3.pt:model4.pt',\r\n tokenizer='moses', bpe='fastbpe')\r\nfairseq_res = en2de.translate(input_text)\r\n# tranformers\r\nfsmt_path = \"./fairseq2hf/data/wmt19-en-de/\"\r\ntokenizer = FSMTTokenizer.from_pretrained(fsmt_path)\r\nmodel = FSMTForConditionalGeneration.from_pretrained(fsmt_path)\r\nnlp = TranslationPipeline(model=model, tokenizer=tokenizer)\r\nfsmt_res = nlp(input_text)[0][\"translation_text\"]\r\n\r\nprint(\"fairseq: {}\".format(fairseq_res))\r\nprint(\"transformer: {}\".format(fsmt_res))\r\nprint(\"match: {}\".format(fairseq_res == fsmt_res))\r\n\"\"\"\r\nfairseq: Maschinelles Lernen ist großartig!\r\ntransformer: Maschinelles Lernen ist großartig!\r\nmatch: True\r\n\"\"\"\r\n```\r\n\r\nHowever, my fairseq model and converted HF model have wrong result with same parameter (beam_size=5). Do you have any idea to debug why tranlation results are different?\r\n\r\n### fairseq result\r\n```\r\n# encoded token by hypo_token by fairseq-interactive\r\ntensor([[5269, 2069, 5, 1154, 9, 4, 1823, 3382, 5, 3128, 116, 167,\r\n 1582, 7, 2192, 914, 63, 6, 1823, 2807, 124, 1219, 1106, 8,\r\n 53, 2175, 2007, 483, 4, 660, 708, 5229, 33, 44, 4, 6049,\r\n 1430, 5, 1806, 2050, 2282, 1908, 4, 334, 3229, 4808, 6102, 5,\r\n 5031, 11, 5, 291, 4214, 6485, 10, 5784, 1908, 23, 1765, 4916,\r\n 6, 2]])\r\n\r\n# hypo_token by fairseq-interactive\r\ntensor([ 924, 4938, 6, 3056, 59, 503, 1497, 4, 5835, 847, 6, 592,\r\n 2], dtype=torch.int32)\r\n```\r\n\r\n### transformers result\r\n```python\r\nencoded_token = torch.tensor([[5269, 2069, 5, 1154, 9, 4, 1823, 3382, 5, 3128, 116, 167, 1582, 7, 2192, 914, 63, 6, 1823, 2807, 124, 1219, 1106, 8, 53, 2175, 2007, 483, 4, 660, 708, 5229, 33, 44, 4, 6049, 1430, 5, 1806, 2050, 2282, 1908, 4, 334, 3229, 4808, 6102, 5, 5031, 11, 5, 291, 4214, 6485, 10, 5784, 1908, 23, 1765, 4916, 6, 2]])\r\n\r\nfsmt = FSMTForConditionalGeneration.from_pretrained(\"./fairseq2HF/\")\r\nhypo = fsmt.generate(encoded_token, num_beams=5)\r\nprint(hypo)\r\n# tensor([[ 2, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 21, 2]])\r\n```\r\n", "I'm a bit lost - we were discussing a missing state dict key, now we are discussing invalid translation.\r\n\r\nDid my suggestion help to resolve the problem of the missing key and now you're presenting the next issue?\r\n\r\nWrt to your transformers result with your model, do you get any better behavior if you encode the tokens via transformers and then feed it to generate? perhaps the dict has somehow changed? though a repeated 21 is suspiciously bad.\r\n\r\n", "@stas00\r\n> Did my suggestion help to resolve the problem of the missing key and now you're presenting the next issue?\r\n\r\nYes, thanks for the helpful comments. \r\nSorry, I should post it as another issue.\r\n\r\n\r\n> do you get any better behavior if you encode the tokens via transformers and then feed it to generate?\r\n\r\nI do not use transformers tokenizer because my fairseq model has a different vocab size, and it's impossible to encode/decode by a single tokenizer model. Converting token to id is used by fairseq's `Dictionary`.\r\nI'll post another issue if necessary after scrutinizing my code.\r\n\r\n\r\nThanks for the big help!", "Thank you for clarifying that your original issue has been resolved. Please feel free to close this issue when you feel it's working for you.\r\n\r\nBased on your comments, I'm concerned about 2 things:\r\n1. your different dictionaries - a model has to come with the exact dict it was trained on, after conversion too. So it sounds that something isn't right there. If you're not sure what's happening perhaps try to clarify how it came to be that your fairseq model has a different vocab size.\r\n2. perhaps that `output_projection` layer is getting in the way of your model if it was trained without it. You could try to hunt down the few lines where it's used in the code and and bypass it and test whether your translation works then. If you're comfortable editing the source code that is." ]

[]

[]

[ "Hello! Thank you for opening an issue with a reproducible example, it helps a lot.\r\n\r\nThe issue here is that you're using the `encode` method to encode a batch, which it can't do. Encode only encodes single sequences, and can accept a \"batch\" of two because it processes them as two independent sequences that should be joined together, for example for text-classification where you would want to classify the relationship between two sequences (tasks like Next Sentence Prediction from BERT or Sentence Ordering Prediction ALBERT).\r\n\r\nThe method you're looking for is the `__call__` method of the tokenizer, which handles exactly all the use-cases you've mentioned, and is the recommended API for tokenizers:\r\n\r\n```py\r\nfrom transformers import AutoTokenizer\r\n\r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\r\ntokenizer([\"hello\", \"world\"]) # <--- This works\r\ntokenizer([\"hello\"]) # <--- This works too :)\r\ntokenizer([\"dette\", \"er\", \"en\", \"sø\"]) # <--- This works as well!\r\n```\r\n\r\n[Here is the documentation](https://huggingface.co/transformers/main_classes/tokenizer.html#transformers.PreTrainedTokenizer.__call__) for that method, hope that helps!", "Thank you very much for this good explanation which clearly resolves my problem.\r\n\r\nDo you by any chance know whether this behaviour changed in the last years time?\r\nThe transformers-based repos [NERDA](https://github.com/ebanalyse/NERDA) and [danlp](https://github.com/alexandrainst/danlp) seem to rely on `tokenizer.encode` to be working as you show the call method does, and as such fail on the current version, but work on 3.5.1 (https://github.com/alexandrainst/danlp/issues/113)", "I believe the `encode` method never accepted batches as inputs. We introduced `encode_plus` and `batch_encode_plus` down the road, the latter being the first to handle batching.\r\n\r\nWhile these two methods are deprecated, they're still tested and working, and they're used under the hood when calling `__call__`.\r\n\r\nWhat is happening here is that v3.5.1 is treating your input as individual words (but by all means it shouldn't as the `is_split_into_words` argument is `False` by default), rather than as different batches, I was mistaken in my first analysis. Something did change between version v3.5.1 and v4.0.0, all the breaking changes are documented in the [migration guide](https://huggingface.co/transformers/migration.html).\r\n\r\nIf you want to get back to the previous behavior, you have two ways of handling it:\r\n\r\n- Specify that you don't want a fast tokenizer. The main change affecting you here is that the `AutoTokenizer` returns a fast tokenizer by default (in Rust) rather than the python-based tokenizer. You can change that behavior with the following:\r\n```py\r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\", use_fast=False)\r\n```\r\n- The behavior you're relying on here is the `is_split_into_words` parameter: you're passing it a list of words, rather than a sequence of words. That it worked in previous versions seems like a bug to me, here's how you would handle it now (works with a fast tokenizer):\r\n```py\r\ntokenizer([\"hello\", \"world\"], is_split_into_words=True)\r\ntokenizer([\"hello\"], is_split_into_words=True)\r\ntokenizer([\"dette\", \"er\", \"en\", \"sø\"], is_split_into_words=True)\r\n```", "Thank you, the `is_split_into_words` clears up the confusion between batches and tokens clearly for me!", "Hi @LysandreJik \r\n\r\nI am excpericing the same error:\r\n`'TypeError: TextEncodeInput must be Union[TextInputSequence,Tuple[InputSequence, InputSequence]]'`\r\n\r\nwhile running below code:\r\n```python\r\nself.tokenizer.encode_plus(example[0],\r\n add_special_tokens=True,\r\n padding='max_length',\r\n max_length=max_length,\r\n return_attention_mask=True,\r\n return_tensors='pt')\r\n```\r\n\r\n`example[0]` is list of Int which I encoded:\r\n[49518, 111, 22560, 20, 1112, 128, 29, 568, 7, 7244, 10, 10905, 111, 12396, 3781, 111, 4878, 1087, 396, 10, 812, 111, 3077, 629, 847, 202, 3607, 490, 5, 3302, 9, 17890, 154, 10, 3077, 629, 4878, 42, 76, 479, 10130, 273, 363, 2156, 5, 1112, 2763, 8176, 111, 262, 7, 7244, 41, 2319, 68, 508, 4, 245, 325, 2450, 14, 56, 57, 12850, 9, 2213, 9, 7668, 14, 74, 33, 23398, 2156, 1195, 87, 20546, 2156, 5, 752, 1229, 3781, 479, 2589, 6040, 17811, 28455, 5, 1087, 7, 18720, 3633, 14, 24, 21, 33602, 19, 780, 111, 773, 629...\r\n\r\nNow i want to pad it and get the attention back.\r\nIn the docs it mentioned that i can send List[Int]\r\nwhat I am missing ?\r\n", "Hi @shon-otmazgin could you open a new issue with a reproducible code example + full stack trace so that we can take a look? Thanks!", "Taking a look at it, I believe the documentation is wrong here and the fast tokenizers handle strings as inputs. Have you tried using `prepare_for_model` for your use-case?", "I will take a look on `prepare_for_model ` this is new to me.\r\n`prepare_for_model ` accept list of input_ids, can pad and return attention mask? ", "So we dived into it with @n1t0 and actually the problem here is slightly complex. The slow & fast tokenizers have roughly the same API with a few excptions, and this is one of them: the fast tokenizers are great at handling strings and at being extremely efficient with a bunch of features (offsets is one example of a really powerful feature), but they're not made to handle lists of ints.\r\n\r\nIn this particular case, while I think it is theoretically possible with fast tokenizers methods by using some private methods, it seems you would be way better off to use a slow tokenizer to achieve what you're looking for.\r\n\r\nBut this begs the question: is there a way you could share your use-case so that we could study it and understand why you need to pass already processed lists of ints to the tokenizer, instead of tokenizing the text and relying on the information within the encoding?\r\n\r\nHere the fast tokenizers would probably be way more efficient at handling this use-case in a one step process, rather than the two step process we're trying to achieve here.", "Also, I'm seeing the following in the docs for `encode_plus`:\r\n\r\n\r\nand for `batch_encode_plus`:\r\n\r\n\r\nIs there a docstring we've forgotten somewhere that tells this is also supported for fast tokenizers?\r\n\r\n", "I tell you what happened:\r\nI worked on version 3.3.1 which by default `use_fast=False` for `AutoTokenizer`.\r\nI upgraded to version 4.4.2 and that break. -> `use_fast `changed to `True `for `AutoTokenizer`.\r\n\r\n@LysandreJik thank you very much for your help. appreciate that :)" ]

[]

[]

[ "`metrics_format` was recently introduced on master, you should update the transformers version to master.", "Thanks! I'll try it again!" ]

[]

[]

[]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false }, { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "Hey @Mehrad0711,\r\n\r\nThanks a lot for the very clean & easy to understand issue!\r\nI can reproduce the error and would be super happy about a PR to fix it! Your fix to let the context manager handle the `spm_target` sounds like the correct solution to me!", "Hi @patrickvonplaten!\r\nThank you for your feedback. I just submitted a PR fixing this issue.\r\nThanks ahead for reviewing." ]

[ { "id": 1990918270, "node_id": "MDU6TGFiZWwxOTkwOTE4Mjcw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20First%20Issue", "name": "Good First Issue", "color": "bbf794", "default": false, "description": "" } ]

[]

[ "Ah indeed, that's a good request ! There's no reason, we could definitely raise a warning when loading the weights by checking the model type in the configuration against the arch's model type. Do you want to open a PR?", "Why a warning and not an assert? If the code throws a totally unrelated long backtrace how would a user know to search for an earlier warnings?\r\n\r\nDo you see a use-case where someone may need to load mismatching arch for the given model?", "After thinking about it, you're right that an error would be better. I can't think of use-cases where that would affect someone's workflow negatively.", "> Do you want to open a PR?\r\n\r\nI could, but realistically it might not happen soon. But since it's not a complicated task perhaps asking the community to help? I guess it'd be as simple as:\r\n\r\n1. read the config of the downloaded model as soon as the config got downloaded\r\n2. compare `config.arch` with model's arch\r\n3. assert if mismatch", "Hi @LysandreJik Does someone work on that ? I'd like to make my first contribution to the project", "Hi @ankh6, feel free to work on it! The issue is not reserved until a PR is opened with some progress made towards solving the issue.", "And when you solve it, one test can be:\r\n```\r\npython -c 'from transformers import PegasusForConditionalGeneration; PegasusForConditionalGeneration.from_pretrained(\"patrickvonplaten/t5-tiny-random\")'\r\n```\r\nbut this one doesn't crash, just spits a lot of warnings. \r\n\r\nThis one does crash:\r\n```\r\npython -c 'from transformers import BartForConditionalGeneration; BartForConditionalGeneration.from_pretrained(\"prajjwal1/bert-tiny\")'\r\n```\r\n\r\nSo it'd be a better candidate to go into the test suite. \r\n\r\nWe want a tiny model so that it runs the test fast.", "@LysandreJik If I understand correctly we should check that the input is in the PRETRAINED_VOCAB_FILES_MAP object (for this issue). Should the assertion occur when we call is_torch_available method, i.e. in src/transformers/models/gpt2/__init__.py, ? ", "As soon as you retrieved the config file and you know which model's class is used, so that you have the 2 things to compare.\r\n\r\nIt definitely shouldn't happen in the specific model files, but inside the common library.\r\n\r\nMost likely there should be one check inside one of the super-classes for model's `from_pretrained` and the same for the tokenizer. Since either may have this conflict.", "Hi, \r\nI've made some progress on this issue. Think I've fixed it for initiating models. \r\nTo show if my approach is fine shall I submit a PR? \r\n\r\nI've essentially added an assert statement in the `from_pretrained` method in the `PretrainedConfig` class. ", "That sounds about right, and yes PR please - thank you!", "Added a pull request #10586 " ]

[]

[]

[ "#10287 contains the fix.", "Confirmed that it works, albeit the cl args changed so tested with:\r\n```\r\nPYTHONPATH=src python examples/seq2seq/run_translation.py --model_name_or_path facebook/mbart-large-en-ro --do_train --do_eval --dataset_name wmt16 --dataset_config_name ro-en --output_dir /tmp/tst-translation --per_device_train_batch_size=2 --per_device_eval_batch_size=2 --overwrite_output_dir --predict_with_generate --source_lang en_XX --target_lang ro_RO --max_val_samples 10 --max_train_samples 10\r\n```" ]

[]

[]

[]

[]

[]

[ "Indeed, I can see the problem. I'm not sure there is an easy fix however and I don't have time right now to build a proper callback checkpointing system. Will have to wait a little bit to be fixed!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@sgugger Could you please reopen this issue? The issue persists despite being automatically closed. " ]

[]

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

[ "Hey @xxbidiao, \r\n\r\nFor batched generation GPT2 has to be used in quite a special way... -> could you check out [this](https://discuss.huggingface.co/t/batch-generation-with-gpt2/1517/2) forum post to see whether this makes sense for you?", "```\r\nimport torch,transformers\r\ngpt2_model = transformers.GPT2LMHeadModel.from_pretrained(\"gpt2\")\r\nprint(gpt2_model.generate(torch.tensor([[100,200,300]]),do_sample=False))\r\nprint(gpt2_model.generate(torch.tensor([[100,200,300,50256]]),attention_mask=torch.tensor([[1,1,1,0]]),do_sample=False))\r\nprint(gpt2_model.generate(torch.tensor([[50256,100,200,300]]),attention_mask=torch.tensor([[0,1,1,1]]),do_sample=False))\r\n```\r\n\r\n```\r\nSetting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\r\nSetting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\r\ntensor([[100, 200, 300, 84, 12, 75, 84, 12, 75, 84, 12, 75, 84, 12,\r\n 75, 84, 12, 75, 84, 12]])\r\nSetting `pad_token_id` to `eos_token_id`:50256 for open-end generation.\r\ntensor([[ 100, 200, 300, 50256, 198, 198, 7, 16, 8, 383,\r\n 3381, 366, 75, 1, 1724, 262, 4129, 286, 262, 4731]])\r\ntensor([[50256, 100, 200, 300, 84, 12, 75, 84, 12, 75,\r\n 84, 12, 75, 84, 12, 75, 84, 12, 75, 84]])\r\n```\r\n\r\nLooks like it works! Will double check. Thank you!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Thanks for flagging! Those have not been updated in a while so I made a pass over that file.", "I still see the bad links. Is the change getting pushed/merged later?", "It will only be seen in the [master documentation](https://huggingface.co/transformers/master/) for now. At the next release, it will become visible in the stable documentation.", "Will the new run_ner.py work with PyTorch and TF? PyTorch-Lightening too?\r\n" ]

[]

[]

[ "Hi all! Sorry, but this seems to be cleaner: (Some feature request: #14255)\r\n```python\r\nencoded_train_dataset = train_dataset.map(\r\n lambda batch: tokenizer.prepare_seq2seq_batch(\r\n batch['text'], batch['summary'], padding='max_length', truncation=True, max_length=256, max_target_length=64\r\n ),\r\n batched=True,\r\n remove_columns=train_dataset.column_names,\r\n)\r\n```" ]

[]

[]

[ "Hi @sgugger, \r\nGot some time to raise the changes we talked about in [my previous PR](https://github.com/huggingface/transformers/pull/10267). \r\nDo let me know if I missed something. \r\nThanks!", "Hi @LysandreJik / @sgugger,\r\n\r\nThanks for your inputs! I can think of some better names, _but before that_: \r\nIs deprecating usage of `EvaluationStrategy` and keeping its definition along with the `TimeStrategy` (or whatever the name would be) in `trainer_utils.py` for the time being a good option? Can also throw a FutureWarning when `EvaluationStrategy` is used.", "Yes that would be the preferred option: not use it anymore but still keep it until v5, and each time someone uses it a `FutureWarning` indicating it's deprecated and will be removed in version 5 is thrown.\r\nLet me know if you have other questions!", "Hi @sgugger,\r\nSeems like the latest changes are failing a certain `make modified_only_fixup` test.\r\nI'm not entirely sure where this test is failing.\r\nGiven that this test is passing on my local machine, this could be due to erratic hard-updation of some test/doc?\r\n\r\n> 2021-02-26T19:45:22.6161202Z Checking/fixing src/transformers/__init__.py src/transformers/integrations.py src/transformers/models/__init__.py src/transformers/models/auto/configuration_auto.py src/transformers/models/auto/modeling_auto.py src/transformers/models/auto/modeling_tf_auto.py src/transformers/trainer_callback.py src/transformers/trainer_tf.py src/transformers/trainer_utils.py src/transformers/training_args.py src/transformers/training_args_tf.py src/transformers/utils/dummy_pt_objects.py src/transformers/utils/dummy_tf_objects.py src/transformers/utils/dummy_tokenizers_objects.py src/transformers/utils/notebook.py tests/test_trainer.py tests/test_trainer_callback.py utils/check_repo.py\r\n> 2021-02-26T19:45:24.8027639Z All done! ✨ 🍰 ✨\r\n> 2021-02-26T19:45:24.8028839Z 18 files left unchanged.\r\n> 2021-02-26T19:45:28.4905837Z tests/test_trainer.py:1060:37: F821 undefined name 'EvaluationStrategy'\r\n> 2021-02-26T19:45:28.5127548Z make: *** [modified_only_fixup] Error 1\r\n> 2021-02-26T19:45:28.5129201Z Makefile:7: recipe for target 'modified_only_fixup' failed", "Oh and for the failing test, you missed an `EvaluationStrategy` toward the end of `tests/test_trainer.py`, that's why you have the error.", "Thanks! Fixed it now.\r\nAlthough I see some approaches mentioned on other forums, I'm not entirely sure what would be the best approach to print a warning on usage of enum `EvaluationStrategy`.\r\nIf not too complex, you can point me towards how to do it.\r\nOtherwise, you can merge this PR :-).", "I'm not finding anything easy to do that, so I think we can merge for now and I'll keep looking." ]

[]

[]

[ "Nevermind, didn't see the not for replaced indices so it's 50% of the remaining 20% after masking" ]

[]

[]

[]

[]

[]

[]

[ { "id": 1834088753, "node_id": "MDU6TGFiZWwxODM0MDg4NzUz", "url": "https://api.github.com/repos/huggingface/transformers/labels/Tests", "name": "Tests", "color": "a6fcca", "default": false, "description": "Related to tests" } ]

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

[ "One other solution - since this is a single node we could use a unique file rather than port for setting up the distributed process group.\r\n\r\nThat is `init_process_group()` with `init_method=\"file:///tmp/unique_file\"` - but the trainer currently hardcodes the `env://` method so we may need to make it more flexible around that.\r\n\r\nReference: https://pytorch.org/docs/master/distributed.html#torch.distributed.init_process_group", "since we are switching to docker runs, this becomes moot, as there will be no processes from previous runs." ]

[]

[]

[]

[]

[ { "login": "sgugger", "id": 35901082, "node_id": "MDQ6VXNlcjM1OTAxMDgy", "avatar_url": "https://avatars.githubusercontent.com/u/35901082?v=4", "gravatar_id": "", "url": "https://api.github.com/users/sgugger", "html_url": "https://github.com/sgugger", "followers_url": "https://api.github.com/users/sgugger/followers", "following_url": "https://api.github.com/users/sgugger/following{/other_user}", "gists_url": "https://api.github.com/users/sgugger/gists{/gist_id}", "starred_url": "https://api.github.com/users/sgugger/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/sgugger/subscriptions", "organizations_url": "https://api.github.com/users/sgugger/orgs", "repos_url": "https://api.github.com/users/sgugger/repos", "events_url": "https://api.github.com/users/sgugger/events{/privacy}", "received_events_url": "https://api.github.com/users/sgugger/received_events", "type": "User", "site_admin": false } ]

[ "Instead of adding a new argument, I would use the existing `resume_from_checkpoint` and change its type to bool or str/PathLike. If it's a bool and if it's `True`, we then use `get_last_checkpoint` to get the last checkpoint in `args.output_dir`. Does that sound good to you?", "Yes, SGTM. I have raised [a PR](https://github.com/huggingface/transformers/pull/10334) doing the same. Do let me know if there is any other change required as well!\r\n\r\n**PS**: Can you also review my [other PR](https://github.com/huggingface/transformers/pull/10286) introducing `save_strategy` in `TrainingArguments`? This PR is the last one to round-up the `save_strategy`, `evaluation_strategy` and `logging_strategy` enhancements. \r\n\r\nThanks!", "> # 🚀 Feature request\r\n> `Trainer.train` accepts `resume_from_checkpoint` argument, which requires the user to explicitly provide the checkpoint location to continue training from. `resume_from_last_checkpoint` can be useful to resume training by picking the latest checkpoint from `output_dir` of the `TrainingArguments` passed.\r\n> \r\n> ## Motivation\r\n> 1. The checkpoint directory is created by the library, so user needs to navigate to the directory to find the value to provide for `resume_from_checkpoint`\r\n> 2. User may just want to resume from the last valid checkpoint since their training got disrupted previously (a common scenario for someone to want to resume training). All they know is the `output_dir` they provided initially\r\n> \r\n> This motivates to provide a `resume_from_last_checkpoint=True` to the `Trainer.train(...)` call, which will pick the latest checkpoint from `args.output_dir`. FYI `get_last_checkpoint` function from `trainer_utils` can be used to do exactly the same.\r\n> \r\n> ## Your contribution\r\n> I can raise a PR if it is a useful feature to have!\r\nIs it possible to train while adding a new category to the dataset using this resume_from_checkpoint argument?\r\n" ]

[]

[ { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "Unstale", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "+1 I found the same problem. The bottleneck seems to be huggingface/datasets. Hence I switched back to use the old customized dataset, which was way more faster.", "Hi ! There's currently an issue in huggingface/datasets that makes iterating through the dataset slow if your dataset is big.\r\nWe're working on a fix and we'll do a new release soon to address this :)\r\n\r\nI'll ping you when the fix is ready if you want to try it out !\r\nedit: https://github.com/huggingface/datasets/pull/2122 fixed it", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored.", "@gyin-ai could you share your custom solution? I'm running into the same problem" ]

[]

[]

[ "I'm on MacOS btw.", "Possibly duplicate of https://github.com/huggingface/transformers/issues/9939", "> Possibly duplicate of https://github.com/huggingface/transformers/issues/9939\n\nI have installed TF 2.0 right at the start. Is there a version to update to resolve this error?", "So after install TF 2.0 with conda, I performed pip install --upgrade tensorflow to v2.4.1 and it works now.", "The issue happens again with latest version of tensorflow and transformers.\r\n\r\n`>>> import transformers`\r\n`>>> from transformers import pipeline`\r\n`Traceback (most recent call last):`\r\n ` File \"<stdin>\", line 1, in <module>`\r\n`ImportError: cannot import name 'pipeline' from 'transformers' (unknown location)`\r\n`>>> tensorflow.__version__`\r\n`'2.5.0'`\r\n`>>> transformers.__version__`\r\n`'4.7.0'`", "I had the same problem but used the false transformer:\r\n**Initial**\r\n`conda update -c conda-forge transformers`\r\n\r\n**Before**\r\n> tensorflow.__version__ # '2.3.0'\r\n> transformers.__version__ # '2.1.1'\r\n\r\n**Solution**\r\n`conda install -c huggingface transformers `\r\n\r\n**After**\r\n> tensorflow.__version__ # '2.3.0'\r\n> transformers.__version__ # '4.11.3'\r\n> torch.__version__ # '1.10.0'\r\n\r\n", "It is related to the solution. I have to solve it by explicit stating the version. Otherwise, it keep installing the conda-forge version.\r\n\r\n`conda install -c huggingface transformers =4.11.3`", "What is your python files name? Mine was _tokenizers_ and when I changed it to _use_tokenizers_ it works. Maybe using names like \"tokenizers\", \"pipeline\" for files not the best idea.", "> What is your python files name? Mine was _tokenizers_ and when I changed it to _use_tokenizers_ it works. Maybe using names like \"tokenizers\", \"pipeline\" for files not the best idea.\r\n\r\nI made the mistake of calling mine tokenize.py - once changed the code worked. " ]

[]

[]

[ "Nice! What means \"same way of training\", same way than what?", "Same way as the current `run_ner` script." ]

[]

[]

[ "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "Yes, `bfloat16` is only for TPU. Hence, we cannot really test it elsewhere than inside a TPU context. I have added the `bfloat16` condition only if XLNet is run on TPU because we were handling a specific case when the model is run under AMP." ]

[]

[]

[ "Hello! We would welcome a PR that offers this. Maybe instead of renaming the current QA model to `Simple` (which would break backwards-compatibility), we could add a new model called `ElectraForQuestionAnsweringBeamSearch`? What do you think?", "Agreed. We should use a new name for the model for backward-compatibility. I will submit a PR soon.", "This sounds great, thanks @bkiat1123!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[ { "login": "patrickvonplaten", "id": 23423619, "node_id": "MDQ6VXNlcjIzNDIzNjE5", "avatar_url": "https://avatars.githubusercontent.com/u/23423619?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patrickvonplaten", "html_url": "https://github.com/patrickvonplaten", "followers_url": "https://api.github.com/users/patrickvonplaten/followers", "following_url": "https://api.github.com/users/patrickvonplaten/following{/other_user}", "gists_url": "https://api.github.com/users/patrickvonplaten/gists{/gist_id}", "starred_url": "https://api.github.com/users/patrickvonplaten/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patrickvonplaten/subscriptions", "organizations_url": "https://api.github.com/users/patrickvonplaten/orgs", "repos_url": "https://api.github.com/users/patrickvonplaten/repos", "events_url": "https://api.github.com/users/patrickvonplaten/events{/privacy}", "received_events_url": "https://api.github.com/users/patrickvonplaten/received_events", "type": "User", "site_admin": false }, { "login": "patil-suraj", "id": 27137566, "node_id": "MDQ6VXNlcjI3MTM3NTY2", "avatar_url": "https://avatars.githubusercontent.com/u/27137566?v=4", "gravatar_id": "", "url": "https://api.github.com/users/patil-suraj", "html_url": "https://github.com/patil-suraj", "followers_url": "https://api.github.com/users/patil-suraj/followers", "following_url": "https://api.github.com/users/patil-suraj/following{/other_user}", "gists_url": "https://api.github.com/users/patil-suraj/gists{/gist_id}", "starred_url": "https://api.github.com/users/patil-suraj/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/patil-suraj/subscriptions", "organizations_url": "https://api.github.com/users/patil-suraj/orgs", "repos_url": "https://api.github.com/users/patil-suraj/repos", "events_url": "https://api.github.com/users/patil-suraj/events{/privacy}", "received_events_url": "https://api.github.com/users/patil-suraj/received_events", "type": "User", "site_admin": false } ]

[ "Hey @db1981,\r\n\r\ncould you please post a fully reproducible code snippet in the following format:\r\n\r\n```python\r\nfrom transformers import T5ForConditionalGeneration\r\n\r\nmodel = T5ForConditionalGeneration.from_pretrained(\"...\")\r\n\r\n...\r\n```\r\n\r\nso that we can help you better?", "Hi @patrickvonplaten thanks for you reply! Below the full code!\r\n\r\nfrom transformers import T5ForConditionalGeneration, T5Tokenizer\r\n\r\n```python\r\nmodel_str = \"t5-base\"\r\n\r\nmodel = T5ForConditionalGeneration.from_pretrained(model_str)\r\ntokenizer = T5Tokenizer.from_pretrained(model_str)\r\n\r\nfull_path = \"./brief.txt\"\r\n\r\nwith open(full_path) as file: # Use file to refer to the file object\r\n text = file.read()\r\n\r\n proc_text = text.strip().replace(\"\\n\",\"\")\r\n len_text = len(proc_text.split())\r\n\r\n inputs = tokenizer.encode(\"summarize: \" + proc_text, return_tensors=\"pt\", max_length= (len_text + 1), truncation=True)\r\n outputs = model.generate(\r\n inputs, \r\n max_length=round(len_text/3), \r\n min_length=round(len_text/5), \r\n no_repeat_ngram_size=2,\r\n length_penalty=2.0, \r\n num_beams=4, \r\n early_stopping=False)\r\n\r\n final = tokenizer.decode(outputs[0], skip_special_tokens=True)\r\n\r\n print(\"%d words: %s\" % (len(final.split()), final))\r\n", "Yeah actually your input text is too long here I think -> T5 was only trained to handle up to 512 tokens (which corresponds to less than 512 words), so T5 will definitely not perform well for > 1500 words", "@patrickvonplaten correct, T5 was originally trained to handle up to 512 tokens. But recently the Transformers library was updated to handle longer texts, using the Longformer apparoach. I thought that behind the scenes a longer text would have triggered automagically Longformer...I'll now give. it a shot with that class directly.\r\n\r\nThanks! ", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hi Colin. I ran into this same issue when I switched over to using the datasets library to load my poetry corpus, where line breaks are super important. \r\n\r\nI ended up making a slightly modified version of the built-in [text](https://github.com/huggingface/datasets/blob/master/src/datasets/packaged_modules/text/text.py) loader called text_with_linebreaks, changing line 62 to `batch = batch.splitlines(True)` to keep the newlines. \r\n", "@jncasey Thanks for the rapid reply! I figured the culprit here might be the switch over to huggingface/datasets. How did you end up incorporating this into your workflow? Did you modify other scripts to reference text_with_linebreaks?", "Yes, my training script is a sloppily modified version of the run_clm.py example. I added a new training arg for whether to keep the line breaks, and check for that arg in the section where the script determines which loader to use based on the file extension of the data files. ", "Cc @lhoestq to see how we could surface that functionality more easily.", "Maybe let's add a `keep_linebreaks` parameter to the text loader ? What do you think ?\r\nThis is already a feature request: https://github.com/huggingface/datasets/issues/870", "Thanks for the rapid replies, and relevant updates. would there be interest then in surfacing this new functionality an extra level to the run_[c]lm.py script? or should we just modify the relevant load_dataset call in that script?", "We will do that as soon as there is a new release of datasets to pin in the requirements! For now changing the `load_dataset` in the script if you have a source install is the best way.", "That seems a fine enough solution to me. Thanks again for the assistance. I'll close the issue for now." ]

[]

[]

[]

[]

[]

[ "Currently WIP. \r\nThanks!", "Thanks, yes that worked out! \r\nIMO, defaulting `eval_steps` to `logging_steps` is not a good decision any longer. \r\nWith `logging_strategy` introduced, it seems more intuitive to decouple both. In case user chooses `logging_strategy=\"epoch\"`, `logging_steps` is no longer a valid quantity. \r\nWhat is your take on this?", "> In case user chooses logging_strategy=\"epoch\", logging_steps is no longer a valid quantity.\r\n\r\nIt will just be ignored in that case, so there is no weird behavior for the user.", "Sure! That makes sense. \r\nYou can review the changes and let me know if any changes required. \r\nThanks", "Yes, something like `TimeInterval` (or `TimeStrategy`) will be a good generic enum. \r\nI can work on this generic enum and `saving_strategy` earlier next week most probably. Will raise a PR soon enough. \r\nFor now I've made the amends to introduce `LoggingStrategy.NO`.", "Great! We can merge this in the meantime.\r\nLooking forward to your next PR!" ]

[]

[]

[ "@sgugger, are you ok if we merge this and I will ask Second Good Issue to help with this - I'm not sure I will have time to do this and test all the scripts in the coming days, and since you guys discuss changing this script again, we should probably merge this first.", "I'm fine with that!", "Started an issue here: https://github.com/huggingface/transformers/issues/10337 - This is an easy task so I think First Good Issue might work. Let me know if I should bump it to Second.\r\n" ]

[]

[]

[ "Hi,\r\n\r\nCan you provide the table on which you tried this?\r\n\r\nTo add numeric value information to the table, each cell in the table is replaced by a `Cell` object. A `Cell` object has 2 attributes: `text` (the original string corresponding to the cell value) and an optional `numeric_value` (which can be a `float_value` or a `date`). \r\n\r\nDid you apply `.astype(str)` on your Pandas dataframe before providing it to `TapasTokenizer`? Since this is required before encoding the table.", "Hi Niels,\r\nIndeed, the code changes the DataFrame content to Cell format... but the `iterrows() `returns sometimes a correct row format which is transformed into Cell format... but sometimes a Cell object which is transformed into a Cell (of a Cell) object with text attribute initialized to the original Cell object !! :( \r\n\r\nI have applied the `.astype(str)` yeap, before and after the sample() call... just to be sure :)\r\n\r\nHere is the table with ; as separator : \r\n\r\nWater injected volume (% P.V.);Oil recovery (% I.O.I.P.);Watercut (%)\r\n214.11;61;98\r\n215.23;61;98\r\n216.36;61.1;99\r\n217.49;61.1;99\r\n218.62;61.2;98\r\n219.75;61.2;99\r\n220.88;61.2;98\r\n222.02;61.3;98\r\n223.15;61.3;99\r\n224.28;61.4;98\r\n225.41;61.4;99\r\n226.55;61.4;98\r\n227.67;61.4;99\r\n228.8;61.5;99\r\n229.94;61.5;99\r\n231.07;61.6;98\r\n232.2;61.6;99\r\n233.33;61.6;98\r\n234.47;61.7;99\r\n235.6;61.7;98\r\n236.73;61.8;99\r\n237.86;61.8;99\r\n239;61.8;99\r\n240.11;61.9;99\r\n241.24;61.9;99\r\n242.39;61.9;99\r\n243.51;62;99\r\n244.64;62;99\r\n245.77;62;99\r\n246.9;62.1;99\r\n248.03;62.1;99\r\n\r\nHow to repeat : \r\n\r\ntable = pd.read_csv(os.path.join(DATA_PATH, \"Table_01.csv\"), sep=\";\").astype(str)\r\ntable = table.sample(frac=1.0, random_state=42, replace=False).astype(str)\r\n\r\ninputs = tokenizer(table=table, queries=questions, padding='max_length', return_tensors=\"pt\")\r\n\r\nGood luck :)", "I was able to reproduce it, however when I reset the indices after sampling, it works:\r\n\r\n`table = table.sample(frac=1.0, random_state=42, replace=False).reset_index(drop=True).astype(str)`\r\n\r\nWill look into why it can't handle without resetting the row indices.\r\n", "Hi Niels,\r\nI have also tried the reset index... and in my side it was crashing the same. But it was withtout the drop=True.\r\nThis behavior is very strange. I have check the Pandas documentation and normally sample() should returned a DataFrame object... nothing fancy here. \r\nAnd it does because the iterrows() outside the Tapas Tokenizer works fine :)\r\nSo Tapas Tokenizer is doing something on the DataFrame modified by the sample() function :)", "Here's a notebook illustrating the issue, and fixing it:\r\n\r\nhttps://colab.research.google.com/drive/10MbZiMKyEWUGk2Y1fvIj0Y0_lB42NO38?usp=sharing\r\n\r\nThe reason why you're getting the error is because in the part where each cell of the table is replaced by a Cell object:\r\n\r\nhttps://github.com/huggingface/transformers/blob/97e688bc220514cd5ea072f06b186401c9cfbbd0/src/transformers/models/tapas/tokenization_tapas.py#L2742-L2745, the row indices are used. \r\n\r\nThis can be fixed by replacing `table` with `table.reset_index(drop=True)` in the first line (or resetting the index of the table before providing it to the tokenizer). Another solution is to replace the final line by `table.iloc[row_index, col_index] = Cell(text=table.iloc[row_index, col_index])`. Will make a small PR to add this.\r\n\r\nThank you for spotting the error!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]