Datasets:

---

DanielPFlorian

/

Transformers-Github-Issues

like 0

[]

[]

[]

[]

[]

[ "As per current implementation, any non-zero value given for `warmup_steps` will override any effects of `warmup_ratio`. It made sense for me to give higher precedence to `warmup_steps` as it seems to be the more inconvenient argument of the 2 to provide from user perspective. Please let me know if this default behaviour is to be changed.\r\n\r\nPS: this is my first PR, so feel free to correct me. I will be happy to accommodate 😄 ", "Thanks for the review! I've incorporated the comments. \r\nDo let me know if anything else needs to be addressed.", "Thanks for your comments! \r\nAgreed, the code looks much more readable now. \r\nDo let me know if there can be any more improvement. \r\n\r\nThanks.", "It would indeed be better in `TrainingArguments.__post_init__`: the rational for that is that when instantiating an object with wrong values, we want the error to be raised as soon as possible and as close as possible to the source for easy debugging.\r\n\r\nIn this case, the problem should appear at the line that parses the `TrainingArguments` or when they are created.", "Thanks! Taken care of it.", "Thanks for adding this functionality!" ]

[]

[]

[ "Hi,\r\n\r\nthis file exists, it can be found here: https://github.com/huggingface/transformers/blob/master/src/transformers/models/t5/convert_t5_original_tf_checkpoint_to_pytorch.py\r\n\r\n", "Thank you! I tried the script and it misses a `config.json` file. Where can I find this?", "The config.json should be part of the original T5 files, which can be found [here](https://github.com/google-research/text-to-text-transfer-transformer#released-model-checkpoints).\r\n\r\nHowever, I wonder why you want to convert the original checkpoints yourself, because this has already been done by the authors of HuggingFace. You can find all T5 checkpoints on the [hub](https://huggingface.co/models?search=google/t5). ", "Because I finetuned them on TPU which is not possible in Transformers yet (at least not in TF) and I want to use Transformers for prediction.", "...I think you linked this issue as location for original T5 files", "Apologies, updated the URL. The `config.json` file should look something like [this](https://huggingface.co/google/t5-large-ssm-nq/blob/main/config.json), containg all the hyperparameter values. When you fine-tuned T5 on TPUs, do you have a configuration available?", "Thanks! (you are a lifesaver by the way with these response times :)). I finetuned using the original repo which uses Mesh Tensorflow and it exports checkpoints in the same format as the original published checkpoints. And there is no `config.json` file, not even in the original published checkpoints you linked. For future reference: you can look at the files by going to this url: https://console.cloud.google.com/storage/browser/t5-data/pretrained_models/small if you have a google cloud account.", "I see that they store configurations in .gin files, like this one: https://console.cloud.google.com/storage/browser/_details/t5-data/pretrained_models/small/operative_config.gin\r\n\r\nWhen opening this on my laptop in Notepad, this looks like this:\r\n\r\n```\r\nimport t5.models.mesh_transformer\r\nimport t5.data.sentencepiece_vocabulary\r\nimport mesh_tensorflow.optimize\r\nimport mesh_tensorflow.transformer.dataset\r\nimport mesh_tensorflow.transformer.learning_rate_schedules\r\nimport mesh_tensorflow.transformer.t2t_vocabulary\r\nimport mesh_tensorflow.transformer.transformer_layers\r\nimport mesh_tensorflow.transformer.utils\r\n\r\n# Macros:\r\n# ==============================================================================\r\nd_ff = 2048\r\nd_kv = 64\r\nd_model = 512\r\ndropout_rate = 0.1\r\ninputs_length = 512\r\nmean_noise_span_length = 3.0\r\nMIXTURE_NAME = 'all_mix'\r\nnoise_density = 0.15\r\nnum_heads = 8\r\nnum_layers = 6\r\ntargets_length = 512\r\ninit_checkpoint = \"gs://t5-data/pretrained_models/small/model.ckpt-1000000\"\r\ntokens_per_batch = 1048576\r\n\r\n# Parameters for AdafactorOptimizer:\r\n# ==============================================================================\r\nAdafactorOptimizer.beta1 = 0.0\r\nAdafactorOptimizer.clipping_threshold = 1.0\r\nAdafactorOptimizer.decay_rate = None\r\nAdafactorOptimizer.epsilon1 = 1e-30\r\nAdafactorOptimizer.epsilon2 = 0.001\r\nAdafactorOptimizer.factored = True\r\nAdafactorOptimizer.min_dim_size_to_factor = 128\r\nAdafactorOptimizer.multiply_by_parameter_scale = True\r\n\r\n# Parameters for Bitransformer:\r\n# ==============================================================================\r\nBitransformer.shared_embedding = True\r\n\r\n# Parameters for denoise:\r\n# ==============================================================================\r\ndenoise.inputs_fn = @preprocessors.noise_span_to_unique_sentinel\r\ndenoise.noise_density = %noise_density\r\ndenoise.noise_mask_fn = @preprocessors.random_spans_noise_mask\r\ndenoise.targets_fn = @preprocessors.nonnoise_span_to_unique_sentinel\r\n\r\n# Parameters for decoder/DenseReluDense:\r\n# ==============================================================================\r\ndecoder/DenseReluDense.dropout_rate = %dropout_rate\r\ndecoder/DenseReluDense.hidden_size = %d_ff\r\n\r\n# Parameters for encoder/DenseReluDense:\r\n# ==============================================================================\r\nencoder/DenseReluDense.dropout_rate = %dropout_rate\r\nencoder/DenseReluDense.hidden_size = %d_ff\r\n\r\n# Parameters for decoder/EncDecAttention:\r\n# ==============================================================================\r\n# None.\r\n\r\n# Parameters for get_sentencepiece_model_path:\r\n# ==============================================================================\r\nget_sentencepiece_model_path.mixture_or_task_name = %MIXTURE_NAME\r\n\r\n# Parameters for get_variable_dtype:\r\n# ==============================================================================\r\nget_variable_dtype.activation_dtype = 'bfloat16'\r\n\r\n# Parameters for decoder/LayerStack:\r\n# ==============================================================================\r\ndecoder/LayerStack.dropout_rate = %dropout_rate\r\ndecoder/LayerStack.norm_epsilon = 1e-06\r\n\r\n# Parameters for encoder/LayerStack:\r\n# ==============================================================================\r\nencoder/LayerStack.dropout_rate = %dropout_rate\r\nencoder/LayerStack.norm_epsilon = 1e-06\r\n\r\n# Parameters for learning_rate_schedule_noam:\r\n# ==============================================================================\r\nlearning_rate_schedule_noam.linear_decay_fraction = 0.1\r\nlearning_rate_schedule_noam.multiplier = 1.0\r\nlearning_rate_schedule_noam.offset = 0\r\nlearning_rate_schedule_noam.warmup_steps = 10000\r\n\r\n# Parameters for make_bitransformer:\r\n# ==============================================================================\r\nmake_bitransformer.decoder_name = 'decoder'\r\nmake_bitransformer.encoder_name = 'encoder'\r\n\r\n# Parameters for decoder/make_layer_stack:\r\n# ==============================================================================\r\ndecoder/make_layer_stack.block_scope = True\r\ndecoder/make_layer_stack.layers = \\\r\n [@mesh_tensorflow.transformer.transformer_layers.SelfAttention,\r\n @mesh_tensorflow.transformer.transformer_layers.EncDecAttention,\r\n @mesh_tensorflow.transformer.transformer_layers.DenseReluDense]\r\ndecoder/make_layer_stack.num_layers = %num_layers\r\n\r\n# Parameters for encoder/make_layer_stack:\r\n# ==============================================================================\r\nencoder/make_layer_stack.block_scope = True\r\nencoder/make_layer_stack.layers = \\\r\n [@mesh_tensorflow.transformer.transformer_layers.SelfAttention,\r\n @mesh_tensorflow.transformer.transformer_layers.DenseReluDense]\r\nencoder/make_layer_stack.num_layers = %num_layers\r\n\r\n# Parameters for mesh_train_dataset_fn:\r\n# ==============================================================================\r\nmesh_train_dataset_fn.mixture_or_task_name = %MIXTURE_NAME\r\n\r\n\r\n# Parameters for noise_span_to_unique_sentinel:\r\n# ==============================================================================\r\n# None.\r\n\r\n# Parameters for nonnoise_span_to_unique_sentinel:\r\n# ==============================================================================\r\n# None.\r\n\r\n# Parameters for pack_dataset:\r\n# ==============================================================================\r\n\r\n\r\n# Parameters for pack_or_pad:\r\n# ==============================================================================\r\n# None.\r\n\r\n# Parameters for random_spans_helper:\r\n# ==============================================================================\r\nrandom_spans_helper.extra_tokens_per_span_inputs = 1\r\nrandom_spans_helper.extra_tokens_per_span_targets = 1\r\nrandom_spans_helper.inputs_length = %inputs_length\r\nrandom_spans_helper.mean_noise_span_length = %mean_noise_span_length\r\nrandom_spans_helper.noise_density = %noise_density\r\n\r\n# Parameters for targets_length/random_spans_helper:\r\n# ==============================================================================\r\ntargets_length/random_spans_helper.extra_tokens_per_span_inputs = 1\r\ntargets_length/random_spans_helper.extra_tokens_per_span_targets = 1\r\ntargets_length/random_spans_helper.inputs_length = %inputs_length\r\ntargets_length/random_spans_helper.mean_noise_span_length = %mean_noise_span_length\r\ntargets_length/random_spans_helper.noise_density = %noise_density\r\n\r\n# Parameters for random_spans_noise_mask:\r\n# ==============================================================================\r\nrandom_spans_noise_mask.mean_noise_span_length = %mean_noise_span_length\r\n\r\n# Parameters for targets_length/random_spans_targets_length:\r\n# ==============================================================================\r\n# None.\r\n\r\n# Parameters for random_spans_tokens_length:\r\n# ==============================================================================\r\n# None.\r\n\r\n# Parameters for rate_num_examples:\r\n# ==============================================================================\r\nrate_num_examples.maximum = 1000000.0\r\nrate_num_examples.scale = 1.0\r\nrate_num_examples.temperature = 1.0\r\n\r\n# Parameters for rate_unsupervised:\r\n# ==============================================================================\r\nrate_unsupervised.value = 710000.0\r\n\r\n# Parameters for reduce_concat_tokens:\r\n# ==============================================================================\r\nreduce_concat_tokens.batch_size = 128\r\nreduce_concat_tokens.feature_key = 'targets'\r\n\r\n# Parameters for run:\r\n# ==============================================================================\r\nrun.autostack = True\r\nrun.batch_size = ('tokens_per_batch', %tokens_per_batch)\r\nrun.dataset_split = 'train'\r\nrun.ensemble_inputs = None\r\nrun.eval_checkpoint_step = None\r\nrun.eval_dataset_fn = None\r\nrun.eval_summary_dir = None\r\nrun.export_path = ''\r\nrun.iterations_per_loop = 100\r\nrun.keep_checkpoint_max = None\r\nrun.layout_rules = \\\r\n 'ensemble:ensemble,batch:batch,d_ff:model,heads:model,vocab:model,experts:batch'\r\nrun.learning_rate_schedule = @learning_rate_schedules.learning_rate_schedule_noam\r\nrun.mesh_shape = @mesh_tensorflow.transformer.utils.tpu_mesh_shape()\r\nrun.mode = 'train' \r\nrun.init_checkpoint = %init_checkpoint\r\nrun.model_type = 'bitransformer'\r\nrun.optimizer = @optimize.AdafactorOptimizer\r\nrun.perplexity_eval_steps = 10\r\nrun.predict_fn = None\r\nrun.save_checkpoints_steps = 2400\r\nrun.sequence_length = {'inputs': %inputs_length, 'targets': %targets_length}\r\nrun.train_dataset_fn = \\\r\n @t5.models.mesh_transformer.mesh_train_dataset_fn\r\nrun.train_steps = 1000000000\r\nrun.variable_filter = None\r\nrun.vocabulary = \\\r\n @t5.data.sentencepiece_vocabulary.SentencePieceVocabulary()\r\n\r\n# Parameters for select_random_chunk:\r\n# ==============================================================================\r\nselect_random_chunk.feature_key = 'targets'\r\nselect_random_chunk.max_length = 65536\r\n\r\n# Parameters for decoder/SelfAttention:\r\n# ==============================================================================\r\ndecoder/SelfAttention.attention_kwargs = None\r\ndecoder/SelfAttention.dropout_rate = %dropout_rate\r\ndecoder/SelfAttention.key_value_size = %d_kv\r\ndecoder/SelfAttention.num_heads = %num_heads\r\ndecoder/SelfAttention.num_memory_heads = 0\r\ndecoder/SelfAttention.relative_attention_num_buckets = 32\r\ndecoder/SelfAttention.relative_attention_type = 'bias_shared'\r\ndecoder/SelfAttention.shared_kv = False\r\n\r\n# Parameters for encoder/SelfAttention:\r\n# ==============================================================================\r\nencoder/SelfAttention.attention_kwargs = None\r\nencoder/SelfAttention.dropout_rate = %dropout_rate\r\nencoder/SelfAttention.key_value_size = %d_kv\r\nencoder/SelfAttention.num_heads = %num_heads\r\nencoder/SelfAttention.num_memory_heads = 0\r\nencoder/SelfAttention.relative_attention_num_buckets = 32\r\nencoder/SelfAttention.relative_attention_type = 'bias_shared'\r\nencoder/SelfAttention.shared_kv = False\r\n\r\n# Parameters for SentencePieceVocabulary:\r\n# ==============================================================================\r\nSentencePieceVocabulary.extra_ids = 100\r\nSentencePieceVocabulary.sentencepiece_model_file = \\\r\n @t5.models.mesh_transformer.get_sentencepiece_model_path()\r\n\r\n# Parameters for serialize_num_microbatches:\r\n# ==============================================================================\r\nserialize_num_microbatches.tokens_per_microbatch_per_replica = 8192\r\n\r\n# Parameters for split_tokens:\r\n# ==============================================================================\r\nsplit_tokens.feature_key = 'targets'\r\nsplit_tokens.max_tokens_per_segment = @preprocessors.random_spans_tokens_length()\r\nsplit_tokens.min_tokens_per_segment = None\r\n\r\n# Parameters for tpu_estimator_model_fn:\r\n# ==============================================================================\r\ntpu_estimator_model_fn.init_checkpoint = %init_checkpoint\r\ntpu_estimator_model_fn.outer_batch_size = 1\r\ntpu_estimator_model_fn.tpu_summaries = False\r\n\r\n# Parameters for tpu_mesh_shape:\r\n# ==============================================================================\r\ntpu_mesh_shape.ensemble_parallelism = None\r\ntpu_mesh_shape.model_parallelism = 1\r\ntpu_mesh_shape.tpu_topology = '8x8'\r\n\r\n# Parameters for decoder/Unitransformer:\r\n# ==============================================================================\r\ndecoder/Unitransformer.d_model = %d_model\r\ndecoder/Unitransformer.ensemble = None\r\ndecoder/Unitransformer.input_full_attention = False\r\ndecoder/Unitransformer.label_smoothing = 0.0\r\ndecoder/Unitransformer.loss_denominator = None\r\ndecoder/Unitransformer.loss_fn = None\r\ndecoder/Unitransformer.loss_on_targets_only = False\r\ndecoder/Unitransformer.max_length = 512\r\ndecoder/Unitransformer.positional_embedding = False\r\ndecoder/Unitransformer.shared_embedding_and_softmax_weights = True\r\ndecoder/Unitransformer.vocab_divisor = 128\r\ndecoder/Unitransformer.z_loss = 0.0001\r\ndecoder/Unitransformer.loss_denominator = 233472\r\n\r\n# Parameters for encoder/Unitransformer:\r\n# ==============================================================================\r\nencoder/Unitransformer.d_model = %d_model\r\nencoder/Unitransformer.ensemble = None\r\nencoder/Unitransformer.input_full_attention = False\r\nencoder/Unitransformer.label_smoothing = 0.0\r\nencoder/Unitransformer.loss_denominator = None\r\nencoder/Unitransformer.loss_fn = None\r\nencoder/Unitransformer.loss_on_targets_only = False\r\nencoder/Unitransformer.max_length = 512\r\nencoder/Unitransformer.positional_embedding = False\r\nencoder/Unitransformer.shared_embedding_and_softmax_weights = True\r\nencoder/Unitransformer.vocab_divisor = 128\r\nencoder/Unitransformer.z_loss = 0.0001\r\n\r\n# Parameters for unsupervised:\r\n# ==============================================================================\r\nunsupervised.preprocessors = \\\r\n [@preprocessors.select_random_chunk,\r\n @preprocessors.reduce_concat_tokens,\r\n @preprocessors.split_tokens,\r\n @preprocessors.denoise]\r\n```\r\n\r\n=> the relevant part here seems to be only the model hyperparameters:\r\n```\r\nd_ff = 2048\r\nd_kv = 64\r\nd_model = 512\r\ndropout_rate = 0.1\r\ninputs_length = 512\r\nmean_noise_span_length = 3.0\r\nMIXTURE_NAME = 'all_mix'\r\nnoise_density = 0.15\r\nnum_heads = 8\r\nnum_layers = 6\r\ntargets_length = 512\r\ninit_checkpoint = \"gs://t5-data/pretrained_models/small/model.ckpt-1000000\"\r\ntokens_per_batch = 1048576\r\n```\r\nSo maybe you can create a config.json based on those?\r\n\r\n\r\n\r\n> Thanks! (you are a lifesaver by the way with these response times :)).\r\n\r\nAnd happy to hear this :) you're welcome", "...actually the link you sent for the example config file proved to be extremely useful! Starting from there I've found all related files. Here is everything (including the config file) for T5 Small: https://huggingface.co/t5-small. Also an example workflow for future reference:\r\n```\r\nmkdir t5\r\ngsutil -m cp -r gs://t5-data/pretrained_models/small t5\r\npython ~/transformers/src/transformers/models/t5/convert_t5_original_tf_checkpoint_to_pytorch.py --tf_checkpoint_path t5/small/model.ckpt-1000000 --pytorch_dump_path t5-small-pt --config_file t5/small_config.json\r\n```" ]

[]

[]

[ "Hey @monmanuela,\r\n\r\nThanks for checking out the post! We try to keep the repository for github issues and kindly ask you to post these kinds of questions on the [forum](https://discuss.huggingface.co/). Feel free to tag me there (@patrickvonplaten) :-)", "@patrickvonplaten thanks for your quick reply! Posted a topic on the forum, closing this issue." ]

[]

[]

[ "Hi,\r\n\r\nFor training-related issues, it might be better to ask your question on the [forum](https://discuss.huggingface.co/).\r\n\r\nThe authors of HuggingFace (and community members) are happy to help you there!\r\n" ]

[]

[]

[ "> Thanks for adding this functionality! One general comment I have is on the type of the `stage` argument. Since it has only four possible values from what I can see, it would be better to create an enum for those (to avoid typos and have auto-complete in an IDE).\r\n\r\nOh, let me make it absolutely automatic with `inspect` so it won't need a `stage` argument at all. \r\n\r\nAnd I will collapse the two calls into one in all but `__init__`, so it'll be less noisy.\r\n\r\n", "So, the API has been simplified to remove the need for naming the stages in the caller, tests added. \r\n\r\nI'm sure we will think of further improvements down the road, please let me know if this is good for the first iteration.\r\n\r\nI'm not sure if anybody else wants to review before we merge this. " ]

[]

[]

[ "I think you did not clone the repository properly or are not running the command from the folder `examples/legacy/token-classification`, since that folder does have a task.py file.", "Ah, I was searching for \"tasks.py\" instead. Just user error, thanks for the fast reply." ]

[]

[]

[ "The answers to those two points is not necessarily yes.\r\n- DDP training is only faster if you have NVLinks between your GPUs, otherwise the slow communication between them can slow down training.\r\n- DDP training takes more space on GPU then a single-process training since there is some gradients caching.\r\n\r\nBoth issues come from PyTorch and not us, the only thing we can check on our side is if there is something in our script that would introduce a CPU-bottleneck, but I doubt this is the reason here (all tokenization happens before the training so there is nothing I could think of there).\r\n\r\nYou should also try regular `DataParalell` which does not have the memory problem IIRC, but I don't remember the comparison in terms of speed. I think @stas00 may have more insight there.", "Thanks. I’ll give DP a try.", "What @sgugger said, \r\n\r\nplease see this excellent benchmark: https://github.com/huggingface/transformers/issues/9371#issuecomment-768656711 You can see that a single GPU beats DDP over 2 gpus there if it's not NVLink-connected. I haven't tried it on 3 gpus though. Surely it should be somewhat faster at least. \r\n\r\nDid you check you're feeding the gpus fast enough? - i.e. check their utilization %, it they are under 90% then you probably have an issue with loading - add more dataloader workers.\r\n\r\nAlso please consider using DeepSpeed ZeRO-DP, which should be even faster. https://huggingface.co/blog/zero-deepspeed-fairscale", "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" } ]

[]

[ "@Narsil @LysandreJik How do you suggest we go about with the targets param? At the moment, targets can either be a list of strings or a string. In case of multiple masks, there are 2 ways to go about with it.\r\n\r\n1. Provide a way for the user to define targets for each mask. \r\n2. One single target list that can be uniformly applied across all the positions. \r\n\r\nThe first method would be best implemented by expecting a dict as argument in the keyword param. Something like\r\n{ \"0\" : \"str or list of strings\" , \"2\" : \"str or list of strings\" ... } \r\n\r\nThis way the user can decide to skip explicitly defining candidate keywords in some of the mask positions if needed ( skipped mask 1 in the example above). \r\n\r\n", "Tough question indeed regarding targets! Switching to a dict sounds a bit non intuitive to me, but I don't see any other choice. I guess eventually the API would be the following:\r\n\r\nGiven a single input string, with a single mask:\r\n- A candidate as a string returns the candidate score for the mask\r\n- A candidate list of strings returns the candidate scores for the mask\r\n\r\nGiven a single input string, with multiple masks:\r\n- A candidate as a string returns the candidate scores for all masks\r\n- A candidate list of strings returns the candidate scores for all masks, on all candidates\r\n- A candidate dict of strings returns the candidate scores for the masks which are concerned by the dictionary keys. Their candidates is the dictionary value linked to that dictionary key.\r\n- A candidate dict of list of strings returns the candidate scores for the masks which are concerned by the dictionary keys. Their candidates are the dictionary values linked to that dictionary key.\r\n\r\nThen there are also lists of input strings, with single masks, and lists of input strings, with multiple masks. This results in a very large amount of possibilities, with different returns, which sounds overwhelming. I'm not too sure that's the best way to handle the issue, I'll give it a bit more thought.", "@LysandreJik I had a question. From what I can understand, one can only define a single set of targets at the moment irrespective of how many input texts are given right? For both the case of a single input text and multiple input texts for even the base case of a single mask, we can only define a single target or a list of targets that applies across them all right? Essentially, it is a many to one relation for the input texts to the target. If that is the case, targets functionality is currently not designed in a useful manner right?", "Hi, sorry for getting back to you so late on this. I agree with you that we can improve the `targets`. I'm pinging @joeddav as he's the author of the PR that added them.\r\n\r\n@joeddav your input on this PR would be more than welcome! Thank you.", "Personally, I think the simplest solution would be best: only support `targets` in the single-mask case. If `targets` is passed and there are multiple mask tokens, raise a `ValueError`. It's a pretty narrow use case to need to pass a string with multiple masked tokens while also needing to evaluate possible target tokens for each. In my opinion, that's a complicated and rare use case and we don't need to muddle pipelines code by attempting to support it. It can always be accomplished by using the core modules instead of a pipeline.", "@joeddav That does make sense to me! The objective of a pipeline should only be to accommodate for some quick use test cases. Making it cumbersome misses the point altogether. @LysandreJik What do you think? ", "Yes, I agree with @joeddav as well!", "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.", "> [...] If you think this still needs to be addressed please comment on this thread.\r\n\r\nThis feature would have many applications and would enable comparison of MLMs in gloze tests beyond the restricted setting of targeting words in the intersection of the vocabularies of the models to be compared. There are some open questions how `top_k` predictions should be made, see issue #3609, so I think it would be good to wait a few more weeks to give everybody time to read the linked paper and discuss ideas.", "@jowagner Just to clarify it for others who might be following, the paper you are referring to is this one https://arxiv.org/abs/2002.03079 right?", "> @jowagner Just to clarify it for others who might be following, the paper you are referring to is this one https://arxiv.org/abs/2002.03079 right?\r\n\r\nYes. I hope to read it soon and get a more clear picture what is needed here. I tend to think that producing `top_k` predictions for multiple masked tokens is outside the scope of the BERT model and really needs an extra model on top of it, e.g. a model that predicts a ranked list of best crystallisation points and can then be used to perform a beam search, fixing on subword unit at a time and producing a k-best list of best crystallisation processes.\r\n\r\n", "@jowagner I have a doubt in that case coming back to the basics of BERT. when some of the words are masked and a prediction is to be made on multiple masks during pre-training step in BERT, does BERT not face the same issue? Or are the masks predicted one mask at a time in each training sentence fed to BERT?", "Looking at Devlin et al 2018 again, I don't see the pre-training objective stated but certainly they try to push as much probability mass as possible to the one completion attested in the training data. BERT is trained to get the top prediction right. Good secondary predictions for individual tokens are only a by-product. Nothing pushes the model to make the k-th predictions consistent across multiple masked subword units for k > 1.\r\n\r\nYes, making predictions can be expected to be harder when there are multiple masked subword units but that also happens in pre-training and BERT therefore learns to do this. Maybe BERT does this in steps, crystallising only a few decisions in each layer. A way to find out would be to fix the BERT layers, add MLM heads to each layer, tune these heads and then see how the predictions (and probabilities) change from layer to layer. (This would make a nice paper, or maybe somebody has done this already.)", "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.", "Do we have a final verdict yet on the approach to be followed? @mitramir55 had suggested a code proposal I believe in #3609 ", "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.", "@LysandreJik Shall i replace this with the implementation suggested earlier in #3609 and raise a PR? Though I dont quite think we have discussed on what scoring would be ideal for the beam search used to sort the predictions.\r\n\r\n", "@Narsil had good insights about your previous implementation - @Narsil could you let us know what you think of the solution proposed here https://github.com/huggingface/transformers/issues/3609#issuecomment-854005760?", "The design in https://github.com/huggingface/transformers/issues/3609#issuecomment-854005760 seems very interesting !\r\n\r\nMain comments: \r\n- I would be curious to see (and probably it would need to become a test) to prove that doing `n` inference instead of 1 will produce better results (because it should be close to the real joint probabilities) that's the main interest of this proposed approach.\r\n\r\n- I think it should output the same tokens as fill-mask pipeline in the degenerate case (when there's only 1 mask).\r\n I don't think it's correct right now (see below what I tried)\r\n\r\n- Because we iteratively do `topk` for each `mask` it's a bit of an exponential if I understand correctly. I would probably add some kind of cleanup to limit the number of \"beams\" to topk (I may have overlooked but it seems to be currently missing)\r\n\r\n- the proposed code could probably be refactored a bit for clarity and avoid integer indexing and deep nesting.\r\n\r\n\r\n```python\r\nimport torch \r\nfrom transformers import AutoModelForMaskedLM, AutoTokenizer, pipeline \r\nimport random \r\n \r\n \r\ndef predict_seqs_dict(sequence, model, tokenizer, top_k=5, order=\"right-to-left\"): \r\n \r\n ids_main = tokenizer.encode(sequence, return_tensors=\"pt\", add_special_tokens=False) \r\n \r\n ids_ = ids_main.detach().clone() \r\n position = torch.where(ids_main == tokenizer.mask_token_id) \r\n \r\n positions_list = position[1].numpy().tolist() \r\n \r\n if order == \"left-to-right\": \r\n positions_list.reverse() \r\n \r\n elif order == \"random\": \r\n random.shuffle(positions_list) \r\n \r\n # print(positions_list) \r\n predictions_ids = {} \r\n predictions_detokenized_sents = {} \r\n \r\n for i in range(len(positions_list)): \r\n predictions_ids[i] = [] \r\n predictions_detokenized_sents[i] = [] \r\n \r\n # if it was the first prediction, \r\n # just go on and predict the first predictions \r\n \r\n if i == 0: \r\n model_logits = model(ids_main)[\"logits\"][0][positions_list[0]] \r\n top_k_tokens = torch.topk(model_logits, top_k, dim=0).indices.tolist() \r\n \r\n for j in range(len(top_k_tokens)): \r\n # print(j) \r\n ids_t_ = ids_.detach().clone() \r\n ids_t_[0][positions_list[0]] = top_k_tokens[j] \r\n predictions_ids[i].append(ids_t_) \r\n \r\n pred = tokenizer.decode(ids_t_[0]) \r\n predictions_detokenized_sents[i].append(pred) \r\n \r\n # append the sentences and ids of this masked token \r\n \r\n # if we already have some predictions, go on and fill the rest of the masks \r\n # by continuing the previous predictions \r\n if i != 0: \r\n for pred_ids in predictions_ids[i - 1]: \r\n \r\n # get the logits \r\n model_logits = model(pred_ids)[\"logits\"][0][positions_list[i]] \r\n # get the top 5 of this prediction and masked token \r\n top_k_tokens = torch.topk(model_logits, top_k, dim=0).indices.tolist() \r\n \r\n for top_id in top_k_tokens: \r\n \r\n ids_t_i = pred_ids.detach().clone() \r\n ids_t_i[0][positions_list[i]] = top_id \r\n \r\n pred = tokenizer.decode(ids_t_i[0]) \r\n \r\n # append the sentences and ids of this masked token \r\n \r\n predictions_ids[i].append(ids_t_i) \r\n predictions_detokenized_sents[i].append(pred) \r\n \r\n return predictions_detokenized_sents \r\n \r\n \r\nsequence = \"This is some super neat [MASK] !\" \r\ntokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\") \r\nmodel = AutoModelForMaskedLM.from_pretrained(\"bert-base-uncased\") \r\n \r\npipe = pipeline(task=\"fill-mask\", tokenizer=tokenizer, model=model) \r\nprint(predict_seqs_dict(sequence, model, tokenizer)) \r\nprint(pipe(sequence)) \r\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.\r\n\r\nYes, we need more time or help from somebody with time to review the discussion and make recommendations.\r\n\r\nMy thoughts re-reading a few comments, including some of my own:\r\n\r\n- Producing k-best predictions for multiple masked tokens requires choices, i.e. a model, separate from the underlying transformer model. This is where the PR is stalled. A quick way forward would be to support only `k=1` when there are multiple masked tokens for the time being. For `k=1`, it is undisputed that the prediction should be the transformer's top prediction for each token.\r\n- This PR/feature does not directly allow comparison of cloze test predictions of models with different vocabularies. Users would have to probe with continuous sequences of masked tokens of varying length and somehow decide between the candidate predictions.", "After reading this thread and skimming through #14716, I must confess I still a little unsure how the scores for multi-masked prompts are computed. Based on my understanding, for a prompt with k-masks, it seems like you want to do a beam search over over the Cartesian product `mask_1_targets x mask_2_targets x ... x mask_k_targets` and return the top-n most likely tuples maximizing `P(mask_1=token_i_k, mask_2=token_i_2, ... m_k=token_i_k)`, i.e.:\r\n\r\n```\r\n{\r\n T_1=[(token_1_1, ..., token_1_k), score_t_1],\r\n T_2=[(token_2_1, ..., token_2_k), score_t_2],\r\n ...\r\n T_n=[(token_n_1, ..., token_n_k), score_t_n]\r\n}\r\n```\r\n\r\nIs this accurate? Perhaps you could try to clarify the design intent and limitations of the current API in the documentation somewhere. If you intend to eventually support computing the joint probability, I think would be beneficial to provide a way for consumers to supply a set of per-mask targets and configure the beam search parameters, e.g. beam width. Thanks!", "> After reading this thread and skimming through #14716, I must confess I still a little unsure how the scores for multi-masked prompts are computed. Based on my understanding, for a prompt with k-masks, it seems like you want to do a beam search over over the Cartesian product `mask_1_targets x mask_2_targets x ... x mask_k_targets` and return the top-n most likely tuples maximizing `P(mask_1=token_i_k, mask_2=token_i_2, ... m_k=token_i_k)`, i.e.:\r\n> \r\n> ```\r\n> {\r\n> T_1=[(token_1_1, ..., token_1_k), score_t_1],\r\n> T_2=[(token_2_1, ..., token_2_k), score_t_2],\r\n> ...\r\n> T_n=[(token_n_1, ..., token_n_k), score_t_n]\r\n> }\r\n> ```\r\n> \r\n> Is this accurate? \r\n\r\nActually no, this was the intent of *this* PR which never got merged. Instead of trying to make educated guess about mask combinations, https://github.com/huggingface/transformers/pull/14716 added what seems the most appropriate, which is what the models really answers, which is various tokens at mask locations, without ANY information about correlations.\r\n\r\nThis is how the model is built, and as such, we return it raw.\r\n\r\n```python\r\nfrom transformers import pipeline\r\n\r\n\r\npipe = pipeline(model=\"bert-base-uncased\")\r\n\r\nprint(pipe(\"This is a [MASK] and a [MASK]\", top_k=3))\r\n```\r\n```\r\n[[{'score': 0.5048776268959045,\r\n 'sequence': '[CLS] this is a. and a [MASK] [SEP]',\r\n 'token': 1012,\r\n 'token_str': '.'},\r\n {'score': 0.07435218244791031,\r\n 'sequence': '[CLS] this is a ; and a [MASK] [SEP]',\r\n 'token': 1025,\r\n 'token_str': ';'},\r\n {'score': 0.05109349265694618,\r\n 'sequence': '[CLS] this is a, and a [MASK] [SEP]',\r\n 'token': 1010,\r\n 'token_str': ','}],\r\n [{'score': 0.8665121793746948,\r\n 'sequence': '[CLS] this is a [MASK] and a. [SEP]',\r\n 'token': 1012,\r\n 'token_str': '.'},\r\n {'score': 0.05160374939441681,\r\n 'sequence': '[CLS] this is a [MASK] and a | [SEP]',\r\n 'token': 1064,\r\n 'token_str': '|'},\r\n {'score': 0.046446096152067184,\r\n 'sequence': '[CLS] this is a [MASK] and a ; [SEP]',\r\n 'token': 1025,\r\n 'token_str': ';'}]]\r\n```\r\n\r\nYou are then free to do all the complex attempts to make the suggestions combined. But we don't attempt to hide it since, the model really doesn't model that.\r\n\r\n", "I appreciate this implementation for the support of multiple [MASK] tokens in the input.\r\nHowever, I cannot figure out why the pipeline output is kept nested only in those cases. It forces me to do some additional coding to make it unnested.\r\nIs there any specific reason for this?\r\n\r\nhttps://github.com/huggingface/transformers/blob/4975002df50c472cbb6f8ac3580e475f570606ab/src/transformers/pipelines/fill_mask.py#L142-L144", "> Is there any specific reason for this?\r\n\r\nBackward compatibility, the first pipeline wasn't built with that option in mind making it harder to support multi mask seamlessly like you would expect. The removal of such quirks might happen in 5.0 though. We know it's not convenient as it is, but breaking user code is even less convenient.\r\n" ]

[]

[]

[ "The implementation of T5 is based on the original implementation which can be found [here](https://github.com/google-research/text-to-text-transfer-transformer).\r\n\r\nThe implementation you are referring to is a general Transformer implementation from Tensorflow Mesh. It seems like this repo does not implement T5, but other models like the Funnel Transformer and the Evolved Transformer.\r\n\r\n", "But the T5 repo actually calls the Tensorflow Mesh implementation internally. AFAIK there is no standalone t5 implementation (apart from the HF one).", "Ok I see. In that case I'll leave it on to Patrick to help you. ", "Hey @maurice-g, we compute it ones and then forward it to all the follow-up layers since the result will be the same for all layers. 2 month ago, I made sure that our implementation is exactly the same as the original implementation - see those tests: https://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/tests/test_modeling_t5.py#L734 and https://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/tests/test_modeling_tf_t5.py#L492 \r\nSo I'm confident that the model behaves as expected, but in case you have a reproducible code snippet showcasing differences between the 2 implementations, I'm more than happy to take a look :-)", "You're right, missed that they were returned and forwarded to the other layers. Sorry for that.", "No worries! Thanks for checking in-detail. I think it's always a very good practice to check things in-detail or more often than not you will find subtle bugs in Transformers that will help us improve the code :-)" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "Hi, \r\n\r\nT5 is an encoder-decoder Transformer. The `run_mlm.py` script can only be used for encoder-only models, such as BERT, RoBERTa, DeBERTa, etc. \r\n\r\nBesides this, T5 does not use the regular [MASK] token as BERT. Rather than masked language modeling, T5 is pre-trained on \"unsupervised denoising training\". This is explained [here](https://huggingface.co/transformers/model_doc/t5.html#training).", "Hi @NielsRogge thanks, but I have checked the paper of T5 and this seems to be a unique token:\r\n\r\n\"We consider two strategies to achieve this: First, instead of replacing each corrupted token with a mask token, we replace\r\nthe entirety of each consecutive span of corrupted tokens with a unique mask token.\"\r\n\r\nAs for using run_mlm.py script, I do not think T5 model can be an issue as if we could add T5ForConditionalGeneration it could work to me in run_mlm.py out of the box.\r\n\r\nIs there any place I could look see how to create datasets the way you mentioned to do T5 pretraining with huggingface codes? thanks\r\n", "Hi \r\n@patrickvonplaten, @patil-suraj could you give me some advice how to do T5 pretraining with denoising objective? thanks ", "Here is an old issue on this subject: https://github.com/huggingface/transformers/issues/5079\r\n\r\nAlso @NielsRogge is correct - T5 replaces each span of tokens with a unique mask token -> the so-called sentinel tokens.\r\nCurrently, there is sadly no script showcasing pertaining for T5. Maybe you have some luck when asking this question on the [forum](https://discuss.huggingface.co/)", "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 @NielsRogge thanks, but I have checked the paper of T5 and this seems to be a unique token:\r\n> \r\n> \"We consider two strategies to achieve this: First, instead of replacing each corrupted token with a mask token, we replace\r\n> the entirety of each consecutive span of corrupted tokens with a unique mask token.\"\r\n> \r\n> As for using run_mlm.py script, I do not think T5 model can be an issue as if we could add T5ForConditionalGeneration it could work to me in run_mlm.py out of the box.\r\n> \r\n> Is there any place I could look see how to create datasets the way you mentioned to do T5 pretraining with huggingface codes? thanks\r\n\r\nHI, Did you successfully run the Huggingface T5 pretraining? Can you give me some advice?\r\n" ]

[]

[]

[]

[]

[]

[]

[]

[]

[ "@sgugger @patrickvonplaten verified this fixed the issue in #10214" ]

[]

[]

[ "Indeed! I can reproduce, will fix.", "Can you tell me if https://github.com/huggingface/transformers/pull/10215 fixes it? You can try by installing the following:\r\n\r\n```\r\npip install git+https://github.com/huggingface/transformers@parameter-device-dtype\r\n```", "Thanks a lot for the quick fix. I'm running it right now. I will post whether the training is done in ca. 7 hours.", "> Can you tell me if #10215 fixes it? You can try by installing the following:\r\n> \r\n> ```\r\n> pip install git+https://github.com/huggingface/transformers@parameter-device-dtype\r\n> ```\r\n\r\nHi, the issue is fixed. Thanks a lot. " ]

[]

[]

[]

[]

[]

[ "This comes from the `_total_flos` being stored as long and overflowing in a big training. Will fix this by storing them as floats (hoping for a PR by the end of today)." ]

[]

[]

[]

[]

[]

[ "Pinging @joeddav on this one, since he wrote this tutorial :-)", "Thank you @sgugger for the reply.\r\nOk I can wait for the answer from @joeddav.\r\n\r\nHave a nice day. ", "Figured it out. `answer_end` is the character position immediately _after_ the answer, so end_position should be derived from `answer_end - 1`. I'm not sure why I was able to run it without this error previously (perhaps a resolved tokenizer bug?), but this should be correct.\r\n\r\n```python\r\ndef add_token_positions(encodings, answers):\r\n start_positions = []\r\n end_positions = []\r\n for i in range(len(answers)):\r\n start_positions.append(encodings.char_to_token(i, answers[i]['answer_start']))\r\n end_positions.append(encodings.char_to_token(i, answers[i]['answer_end'] - 1))\r\n \r\n # if start position is None, the answer passage has been truncated\r\n if start_positions[-1] is None:\r\n start_positions[-1] = tokenizer.model_max_length\r\n end_positions[-1] = tokenizer.model_max_length\r\n\r\n encodings.update({'start_positions': start_positions, 'end_positions': end_positions})\r\n```", "Closed by #10217 ", "Thank you @joeddav the posted code works perfectly. \r\n", "Sorry for bothering you @joeddav again, I have a question related to the code posted by you here. \r\nI am still getting None with the dataset built by myself using this code. My dataset works perfectly with the run_squad original script.\r\nIn this snipped posted by you I encounter None in the vector of end_positions and I don't know how fix it. I saw the condition in which there's a None the start_positions but what I have to do in the case the None is only in the end_positions vector?\r\n\r\nKind regards,\r\nAndrea" ]

[]

[]

[]

[]

[]

[ "Hello!\r\n\r\nThis is because in the 4.3 version the implementation of the embedding have changed and `get_input_embeddings()` returns only the word embeddings layer, hence only `input_ids` can be passed.\r\n\r\nThis was an unexpected behavior and will be fixed for the next release (the fix is already in master if you want). Sorry for the inconvenience.", "thank you for the quick reply!\r\n\r\nOk, i was unaware it was already fixed in master\r\n\r\nThank you!" ]

[]

[]

[]

[]

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

[ "Hi! Could you provide a reproducible example? I don't understand what's your `TOKENIZER` here. Thanks", "Hi @LysandreJik , \r\n\r\nThanks for your response , I have done deep-dive and modified code a bit to reproduce the same issue. I am extending the vocabulary of tokenizer and using some automated logic to add the new words. I don't not have complete control on the words which are adding (but, I am making sure those are not the noise.)\r\n\r\n```python\r\nfrom transformers import AutoTokenizer, AutoModelForMaskedLM\r\nmodel_checkpoint_tok = \"microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext\"\r\ntokenizer = AutoTokenizer.from_pretrained(model_checkpoint_tok)\r\nenc_ids = tokenizer.encode('vancomycin')\r\ntokenizer.decode(enc_ids)\r\n```\r\n```out\r\n[CLS] vancomycin [SEP]\r\n```\r\n\r\nAdding the new vocabulary through some automated logic, I added thousands of words, To reproduce the same issue, I am manually adding one of the word in next line\r\n\r\n```python\r\ntokenizer.add_tokens(['vanco'])\r\nenc_ids = tokenizer.encode('vancomycin')\r\ntokenizer.decode(enc_ids)\r\n```\r\n```out\r\n`Output : '[CLS] vanco mycin [SEP]'\r\n```\r\n\r\nThe word 'vancomycin' is present in vocabulary, The word 'vanco' got added in vocabulary (by some automated logic), Now when I am again tokenizing \"vancomycin\" , It is splitting it into token id's of 'vanco' and 'mycin' (Mycin is not exactly present, So it got splitted into subwords). \r\n\r\nMy question is - \r\n\r\n- Is it the expected functionality of tokenizer ( I know that its a Subword tokenization technique) , I am not sure but what I guess is there should be some word boundary detection technique and if the exact word in word boundary is not present, then only sub word tokenization should happen? \r\n\r\n\r\n \r\nPlease suggest how I can handle these scenarios ?\r\n\r\nPlease Find Below Screenshot for the same - \r\n\r\n\r\n", "Hi @LysandreJik , Are there any updates on this ? ", "Hi! That isn't a bug of itself, but expected behavior (which could be better documented). The tokens added to the tokenizers don't get added to the \"vocabulary\" that resulted from the tokenizer training.\r\n\r\nInstead, they get added to the \"added tokens\", and these tokens take priority over the vocabulary tokens. It is, unfortunately, complex or near impossible to add tokens to the vocabulary itself in a way that all tokenizers could benefit. This is particularly complex in the case of subwords tokenizers, which is the case of the BERT tokenizer here.", "You can check this thread https://github.com/huggingface/tokenizers/issues/370 for a similar issue.", "Hi @LysandreJik , So, if we add the tokens and then continually pretrain the Bert Model (By Fine Tuning on Masked Language Modeling) on specific corpus , Then , Does the model learn embeddings for these added tokens ?", "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 a lot for fixing this!", "Other models/tasks have this issue as well https://github.com/huggingface/transformers/issues/10292\r\n\r\nThese features require tests. Without tests this is an endless work-work-work - otherwise we keep on breaking what was working before.\r\n" ]

[]

[]

[ "cc @lhoestq ", "Related to https://github.com/huggingface/datasets/issues/861\r\nMaybe on-the-fly tokenization can help.\r\nOr if we stick to having the tokenization in the preprocessing, at least reduce the precision of the integers stored on disk and maybe do the padding on the fly.", "@lhoestq Are there any minor changes that could fix this temporarily? Will changing the map function to set transform as mentioned [here](https://github.com/huggingface/datasets/issues/1825) help?", "Currently the Trainer doesn't handle `set_transform` but this will be supported soon.\r\n\r\nAnother think you could try is specify `features=` in the parameters of the map function to specify the precision of the integers that are written on disk. For example\r\n```python\r\nfrom datasets import Features, Sequence, Value\r\n\r\nfeatures = Features({\r\n \"input_ids\": Sequence(Value(\"int32\")),\r\n \"token_type_ids\": Sequence(Value(\"bool\")),\r\n \"attention_mask\": Sequence(Value(\"bool\")),\r\n \"special_tokens_mask\": Sequence(Value(\"bool\")),\r\n})\r\ntokenized_datasets = datasets.map(\r\n tokenize_function,\r\n batched=True,\r\n num_proc=data_args.preprocessing_num_workers,\r\n remove_columns=[text_column_name],\r\n load_from_cache_file=not data_args.overwrite_cache,\r\n features=features,\r\n)\r\n```\r\nThe tokenization will still be done during the preprocessing and store the tokenized texts on disk, but this time it will take much less space since you'll store int32 and booleans instead of int64 by default.", "@lhoestq Nope. I get a casting error as attached below\r\n```\r\nException in device=TPU:4: Could not convert 1 with type int: tried to convert to boolean\r\nTraceback (most recent call last):\r\n File \"/opt/conda/lib/python3.7/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 330, in _mp_start_fn\r\n _start_fn(index, pf_cfg, fn, args)\r\n File \"/opt/conda/lib/python3.7/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 324, in _start_fn\r\n fn(gindex, *args)\r\n File \"/kaggle/working/run_mlm_custom.py\", line 461, in _mp_fn\r\n main()\r\n File \"/kaggle/working/run_mlm_custom.py\", line 355, in main\r\n features=features\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/dataset_dict.py\", line 386, in map\r\n for k, dataset in self.items()\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/dataset_dict.py\", line 386, in <dictcomp>\r\n for k, dataset in self.items()\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1140, in map\r\n update_data=update_data,\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 167, in wrapper\r\n out: Union[\"Dataset\", \"DatasetDict\"] = func(self, *args, **kwargs)\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/fingerprint.py\", line 312, in wrapper\r\n out = func(self, *args, **kwargs)\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/arrow_dataset.py\", line 1411, in _map_single\r\n writer.write_batch(batch)\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/arrow_writer.py\", line 279, in write_batch\r\n pa_table = pa.Table.from_pydict(typed_sequence_examples)\r\n File \"pyarrow/table.pxi\", line 1474, in pyarrow.lib.Table.from_pydict\r\n File \"pyarrow/array.pxi\", line 322, in pyarrow.lib.asarray\r\n File \"pyarrow/array.pxi\", line 222, in pyarrow.lib.array\r\n File \"pyarrow/array.pxi\", line 110, in pyarrow.lib._handle_arrow_array_protocol\r\n File \"/opt/conda/lib/python3.7/site-packages/datasets/arrow_writer.py\", line 100, in __arrow_array__\r\n out = pa.array(self.data, type=type)\r\npyarrow.lib.ArrowInvalid: Could not convert 1 with type int: tried to convert to boolean\r\nTraceback (most recent call last):\r\n File \"transformers/examples/xla_spawn.py\", line 85, in <module>\r\n main()\r\n File \"transformers/examples/xla_spawn.py\", line 81, in main\r\n xmp.spawn(mod._mp_fn, args=(), nprocs=args.num_cores)\r\n File \"/opt/conda/lib/python3.7/site-packages/torch_xla/distributed/xla_multiprocessing.py\", line 395, in spawn\r\n start_method=start_method)\r\n File \"/opt/conda/lib/python3.7/site-packages/torch/multiprocessing/spawn.py\", line 157, in start_processes\r\n while not context.join():\r\n File \"/opt/conda/lib/python3.7/site-packages/torch/multiprocessing/spawn.py\", line 112, in join\r\n (error_index, exitcode)\r\nException: process 5 terminated with exit code 17\r\n```\r\n\r\nFixed this using \r\nfeatures = Features({\r\n \"input_ids\": Sequence(Value(\"int32\")),\r\n \"attention_mask\": Sequence(Value(\"int32\")),\r\n \"special_tokens_mask\": Sequence(Value(\"int32\")),\r\n })\r\nbut it still ends up taking a lot of space. Runs out of space again on kaggle notebook with 19GB free", "Looks like PyArrow doesn't know how to convert 1 to True ^^'\r\nI created an issue on Apache Arrow's JIRA [here](https://issues.apache.org/jira/browse/ARROW-11646) to track the issue.\r\nAlso I think we can add support for uint16 in the pytorch integration of `datasets` to reduce the size even more. Feel free to open an issue on the `datasets` repo about this if you want.\r\n\r\nIn the end I think the best solution is probably to not do padding during preprocessing and do it on the fly in the Trainer. This can use `dataset.set_transform` or a data collator. As long as padding is done before sending the data to the TPU it should be good.", "@lhoestq \r\nYou said in the comment before that ```set-transform``` support hasn't been added to the trainer yet. How exactly do I do it on the fly then?", "Yes you're right, we first need to make the Trainer support `set_transform`, sorry if it was confusing.", "`Trainer` in master completely supports `set_transform`. If there are some columns removed that should not be, you just have to set the training arguments `remove_unused_columns` to `False` for the time being.", "Nice ! so one could try to replace the .map with .set_transform(tokenize_function)", "> `Trainer` in master completely supports `set_transform`. If there are some columns removed that should not be, you just have to set the training arguments `remove_unused_columns` to `False` for the time being.\r\n\r\nI tried changing remove_unused_columns to False but it gives me an error during the Trainer call. The set_transform function returns a NoneType object and so the Trainer complains of getting a None instead of training data. The run_mlm_custom.py file below is the same run_mlm file, just with set_transform instead of map (you can have a look at it [here](https://drive.google.com/file/d/1--ijV3UK-Rq9TnzkcWFAjkk2A-0j_XWK/view?usp=sharing))\r\n\r\n```\r\nTraceback (most recent call last):\r\n File \"run_mlm_custom.py\", line 452, in <module>\r\n main()\r\n File \"run_mlm_custom.py\", line 397, in main\r\n train_dataset=tokenized_datasets[\"train\"] if training_args.do_train else None, # print(tokenized_datasets) gives None\r\nTypeError: 'NoneType' object is not subscriptable\r\n```\r\n\r\nDo you have any template code for passing the data to the trainer?", "Indeed `set_transform` is in-place. For example you can do\r\n```python\r\ndataset.set_transform(tokenize) # return None, but sets the transform of the current dataset\r\n```\r\n\r\nIf you want to use a non in-place function like what was doing map, you can do\r\n```python\r\ndataset = dataset.with_transform(tokenize) # return a new dataset object with the specified transform\r\n```\r\n\r\nAlso I'm not a big fan of having two functions that does the same thing (except one is in-place) so we might deprecate one or the other in the future. I guess the second one is more convenient and is more aligned with the other Dataset functions. Let me know what you think", "Returning the dataset is more intuitive I feel. Anyway, this is some really good news. I will try to modify the script and make it work. If it does then maybe, if you want, I can clean the code and create a pull request for the same.", "@DarshanDeshpande did it work for you after you made the changes? \r\nI have the same issue, trying to train a roberta mlm on 1.3G of text data on cloud TPU and got the no space on device error (the code works with 300M of data though). I changed the run_mlm.py code based on your PR to do tokenization on the fly, but now I get this error:\r\n\r\n```\r\n[INFO|trainer.py:946] 2021-03-16 17:51:34,455 >> ***** Running training *****\r\n[INFO|trainer.py:947] 2021-03-16 17:51:34,456 >> Num examples = 14602056\r\n[INFO|trainer.py:948] 2021-03-16 17:51:34,456 >> Num Epochs = 2\r\n[INFO|trainer.py:949] 2021-03-16 17:51:34,456 >> Instantaneous batch size per device = 8\r\n[INFO|trainer.py:950] 2021-03-16 17:51:34,456 >> Total train batch size (w. parallel, distributed & accumulation) = 256\r\n[INFO|trainer.py:951] 2021-03-16 17:51:34,456 >> Gradient Accumulation steps = 4\r\n[INFO|trainer.py:952] 2021-03-16 17:51:34,456 >> Total optimization steps = 114078\r\n 0%| | 0/114078 [00:00<?, ?it/s]Exception in thread Thread-2:\r\nTraceback (most recent call last):\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/threading.py\", line 916, in _bootstrap_inner\r\n self.run()\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/threading.py\", line 864, in run\r\n self._target(*self._args, **self._kwargs)\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/torch_xla/distributed/parallel_loader.py\", line 141, in _loader_worker\r\n _, data = next(data_iter)\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/torch/utils/data/dataloader.py\", line 435, in __next__\r\n data = self._next_data()\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/torch/utils/data/dataloader.py\", line 475, in _next_data\r\n data = self._dataset_fetcher.fetch(index) # may raise StopIteration\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py\", line 44, in fetch\r\n data = [self.dataset[idx] for idx in possibly_batched_index]\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py\", line 44, in <listcomp>\r\n data = [self.dataset[idx] for idx in possibly_batched_index]\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 1130, in __getitem__\r\n format_kwargs=self._format_kwargs,\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/datasets/arrow_dataset.py\", line 1117, in _getitem\r\n pa_subtable, key, formatter=formatter, format_columns=format_columns, output_all_columns=output_all_columns\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/datasets/formatting/formatting.py\", line 375, in format_table\r\n return formatter(pa_table, query_type=query_type)\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/datasets/formatting/formatting.py\", line 173, in __call__\r\n return self.format_row(pa_table)\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/datasets/formatting/formatting.py\", line 239, in format_row\r\n formatted_batch = self.format_batch(pa_table)\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/datasets/formatting/formatting.py\", line 268, in format_batch\r\n return self.transform(batch)\r\n File \"/home/aida_delfan/pretrain/run_mlm.py\", line 363, in tokenize_function\r\n return_special_tokens_mask=True,\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/transformers/tokenization_utils_base.py\", line 2266, in __call__\r\n **kwargs,\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/transformers/tokenization_utils_base.py\", line 2451, in batch_encode_plus\r\n **kwargs,\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/transformers/models/gpt2/tokenization_gpt2_fast.py\", line 163, in _batch_encode_plus\r\n return super()._batch_encode_plus(*args, **kwargs)\r\n File \"/anaconda3/envs/torch-xla-1.7/lib/python3.6/site-packages/transformers/tokenization_utils_fast.py\", line 411, in _batch_encode_plus\r\n for key in tokens_and_encodings[0][0].keys():\r\nIndexError: list index out of range\r\n\r\n```\r\nwondering if you saw a similar error and if there is a fix for it.\r\n\r\nI have transformers 4.4.0 and datasets 1.4.0.\r\n\r\nThanks!", "@aidad This looks like a tokenizer issue. The script in the PR works for me. Try checking your tokenizer files or altering the script to use the Roberta tokenizer specifically instead of the included AutoTokenizer" ]

[]

[]

[ "The CI failed but it seems irrelevant. Could you please give it a check? @LysandreJik ", "The CI issue is with `test_hf_api`, which is fixed on master now, rebasing should make the CI green!", "@sgugger Sorry but I can't really agree with you here. It is a problem introduced by mistake and we shouldn't just try to ignore or downplay that. It just adds a few lines of if-else and I don't think it'll add too much burden for users. I can add some comments if it helps explain the code.\n\nI am never a fan of trainer-style code since IMO in the context of ML, capsulation is not simplicity, instead, transparency is simplicity (I believe our user survey supports my claim here). Thus, I don't think our example here is meant for beginners at all. Besides, I do think back-compatibility matters especially when our success depends heavily on the prosperity of the model hub. This PR is meant to fix a bug though it's harsh. I want the users won't feel confused when they load a community model and wonder why it doesn't work. (Because I myself spent one whole day debugging this)\n\n\nHappy to have more discussion here.", "So we discussed it a bit more internally. This particular example script will stay as is with the backward-compatibility problem (which isn't one IMO since the problem is in the model config not having the right labels). As I said before it's an example aimed at data scientists that shouldn't necessarily have all the functionality.\r\n\r\nThere will be another script for GLUE (probably by the end of the month) very soon that doesn't use `Trainer` and has the training loop exposed, where we can integrate your fix.", "Looking forward to the new script you mentioned!" ]

[]

[]

[ "Indeed, I can reproduce! Do you know what might be causing this @n1t0?", "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 think this is happening because when you load it from the vocab and merge files, it doesn't know `<|endoftext|>` is a special token. For the `skip_special_tokens` to work, I believe it would be necessary to add them to the tokenizer:\r\n```python\r\ntokenizer_fast.add_special_tokens({\r\n \"additional_special_tokens\": \"<|endoftext|>\"\r\n})\r\n```\r\n\r\nThe `tokenizer.json` file on the hub, available for `gpt2` does have this special token registered, that's why it works in this case.", "That workaround is sufficient for my needs and appears to have done the trick. Thanks!" ]

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

[]

[ "Would be lovely to see it, seems promising!", "Any progress? @LysandreJik " ]

[]

[]

[ "Thanks for the PR @mingruimingrui !\r\nBut I'm not sure how filling the embeddings with 0 will avoid them being learnable, and `nn.Embedding` actually handles this itself, if pad index is specified then the output of the embedding layer at that index will be all zeros.\r\n\r\nPlus BART's offest is very specific to the padding index, so I'm not sure if it's a good idea to change the padding index of BART.", "@patil-suraj\r\n\r\nZeroing of weights at positions 0 and 1 is merely code sugar, they do not carry any implication to model behavior.\r\nMy point is that the current implementation causes some weights to be untrainable.\r\n\r\nThe issue this PR is attempting to solve is `padding_idx`.\r\n\r\nHere's a demonstration of the effects of `padding_idx` on the training of `torch.nn.Embedding`.\r\n```python\r\nimport torch\r\n\r\nfor padding_idx in [None, 0, 1]:\r\n print(f'padding_idx: {padding_idx}')\r\n\r\n module = torch.nn.Embedding(2, 1, padding_idx=padding_idx)\r\n print(f'Starting weight: {module.weight.data.tolist()}')\r\n\r\n x = torch.LongTensor([0, 1])\r\n y = torch.FloatTensor([1, 1]).reshape(2, 1)\r\n\r\n optimizer = torch.optim.Adam(module.parameters(), lr=1.0)\r\n for _ in range(10):\r\n optimizer.zero_grad()\r\n pred = module(x)\r\n loss = torch.sum((pred - y) ** 2)\r\n loss.backward()\r\n optimizer.step()\r\n\r\n print(f'Ending weight: {module.weight.data.tolist()}', end='\\n\\n')\r\n```\r\n\r\nYou can expect output similar to the following.\r\n```txt\r\npadding_idx: None\r\nStarting weight: [[-0.696786642074585], [1.2698755264282227]]\r\nEnding weight: [[0.39832496643066406], [0.7758995294570923]]\r\n\r\npadding_idx: 0\r\nStarting weight: [[0.0], [0.7031512260437012]]\r\nEnding weight: [[0.0], [1.2048938274383545]]\r\n\r\npadding_idx: 1\r\nStarting weight: [[-0.8316971659660339], [0.0]]\r\nEnding weight: [[0.48533281683921814], [0.0]]\r\n```\r\n\r\nNotice the weight stays 0 at the respective `padding_idx` position.\r\n\r\n`padding_idx` can easily be 2 or greater as this is dependent on how the user train/config their tokenizer.", "Hey @mingruimingrui, \r\n\r\nThe positional embedding for the padding token is not really relevant since padding tokens are only used in training for tokens that are discarded anyways during training/evaluation (all padding tokens by design cannot influence the attention mechanism). Could you give us an example/use case which would require your fix here?", "@patrickvonplaten \r\n\r\nI am aware that the masks used for attention and loss computation allow model output to not get affected by the positions containing the padding token.\r\n\r\nBut positional embedding is not gathered using token id but rather sequence position.\r\n\r\nGiven padding_idx = 2, the positional embedding of the first token of every sequence will be untrainable regardless of what it is.", "@patrickvonplaten try this out. This is tested on transformers 4.3.2 (latest release as time of writing)\r\n\r\n```python\r\nimport torch\r\nimport transformers\r\nfrom transformers.models.bart.modeling_bart import \\\r\n BartLearnedPositionalEmbedding\r\n\r\nprint(f'Running script on transformers=={transformers.__version__}')\r\n\r\n# Init positional embedding with padding_idx = 2\r\npe = BartLearnedPositionalEmbedding(128, 1, padding_idx=2)\r\n\r\n# Fix input embedding to a tensor of seq_len = 5\r\ninput_ids = torch.randint(0, 32000, (4, 5))\r\n\r\n# Print out pos_embs of input_ids\r\npos_embs = pe.forward(input_ids.shape)\r\nprint('Initial pos_embs:', pos_embs.tolist())\r\n\r\n# Backprop to make positional embeddings = 1.0\r\noptimizer = torch.optim.Adam(pe.parameters(), lr=1.0)\r\nfor _ in range(100):\r\n optimizer.zero_grad()\r\n\r\n pos_embs = pe.forward(input_ids.shape)\r\n target = torch.ones_like(pos_embs)\r\n loss = torch.sum((pos_embs - target) ** 2)\r\n\r\n loss.backward()\r\n optimizer.step()\r\n\r\n# Print out pos_embs of input_ids after optimization\r\n# Expectation: A tensor of arppox. 1.0\r\n# Result: A tensor arppox. 1.0 but first embedding has a value of 0.0\r\npos_embs = pe.forward(input_ids.shape)\r\nprint('Initial pos_embs:', pos_embs.tolist())\r\n```\r\n\r\nstdout\r\n```txt\r\nRunning script on transformers==4.3.2\r\nInitial pos_embs: [[0.0], [-0.2676548659801483], [-0.31950631737709045], [-0.9886524081230164], [0.6115532517433167]]\r\nInitial pos_embs: [[0.0], [1.0013480186462402], [0.9954316020011902], [1.0088629722595215], [0.9995125532150269]]\r\n```", "A major problem with the current behavior is when a user uploads a model with padding_idx >= 2, the position embedding at the padding_idx will be zeroed out.\r\n\r\nMulti-head-attention-based encoder and decoders can be hugely affected (since the concept of the sequence is represented using this embedding).\r\n\r\nUsing an extreme but realistic example, an English to German translation model translates \"hello world\" to \"halo welt\" correctly.\r\nHowever, when exported to huggingface BART, the model can produce \"welt halo\" instead (due to the mix-up of position).\r\n\r\nWhen this happens, the root cause (this issue) can be extremely difficult to discover/debug.", "Hey @mingruimingrui,\r\n\r\nThanks for clarifying! And yes, I agree with you now - sorry I missed your point the first time! This bug is then actually in multiple spots in the library...It would be awesome if you could fix the bug also for the following models:\r\n\r\n- `modeling_mbart.py`: https://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/src/transformers/models/mbart/modeling_mbart.py#L117\r\n- `modeling_blenderbot.py`: https://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/src/transformers/models/blenderbot/modeling_blenderbot.py#L115\r\n- `modeling_blenderbot_small.py`:\r\nhttps://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/src/transformers/models/blenderbot_small/modeling_blenderbot_small.py#L115\r\n- `modleing_led.py`:\r\nhttps://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/src/transformers/models/led/modeling_led.py#L115\r\n- The cookie-cutter:\r\nhttps://github.com/huggingface/transformers/blob/e94d63f6cbf5efe288e41d9840a96a5857090617/templates/adding_a_new_model/cookiecutter-template-%7B%7Bcookiecutter.modelname%7D%7D/modeling_%7B%7Bcookiecutter.lowercase_modelname%7D%7D.py#L1619\r\n\r\nGreat catch :-)", "Thanks @patrickvonplaten\r\n\r\nBut before I make the changes, I thought it might be good to do something about backward compatibility when users use `XXLearnedPositionalEmbedding` directly.\r\nI suggest that we raise a warning when `padding_idx` is not `None` so that the function interface can be kept as is.\r\n\r\n```python\r\ndef __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None):\r\n assert padding_idx is not None, \"`padding_idx` should not be None, but of type int\"\r\n if padding_idx is not None:\r\n warnings.warn(\r\n f'padding_idx should not be provided for {self.__class__.__name__}. '\r\n 'An exception will be raised in future versions of transformers.'\r\n )\r\n\r\n ...\r\n```\r\n\r\nLet me know if this is alright.", "@patrickvonplaten Btw can I also ask how to quote code chunks from repos in github comments like above?\r\n> It is very cool and useful", "> Thanks @patrickvonplaten\r\n> \r\n> But before I make the changes, I thought it might be good to do something about backward compatibility when users use `XXLearnedPositionalEmbedding` directly.\r\n> I suggest that we raise a warning when `padding_idx` is not `None` so that the function interface can be kept as is.\r\n> \r\n> ```python\r\n> def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None):\r\n> assert padding_idx is not None, \"`padding_idx` should not be None, but of type int\"\r\n> if padding_idx is not None:\r\n> warnings.warn(\r\n> f'padding_idx should not be provided for {self.__class__.__name__}. '\r\n> 'An exception will be raised in future versions of transformers.'\r\n> )\r\n> \r\n> ...\r\n> ```\r\n> \r\n> Let me know if this is alright.\r\n\r\n`XXLearnedPositionalEmbedding` is not available in init, *e.g.* one cannot do:\r\n\r\n```python\r\nfrom transformers import BartLearnedPositionalEmbedding\r\n```\r\n\r\ndoesn't work, so in this case we don't have to care about backwards compatibility. As you pointed out, it is simply wrong to pass a padding_idx to the position embeddings.\r\n", "@patrickvonplaten That should be all the changes that are required.", "Hey @mingruimingrui,\r\n\r\nThanks a lot for applying the changes also to the other models - it looks very nice already :-) Could you also completely remove the `padding_idx` from the call to the `PositionalEmbeddings`, *e.g.* `BartLearnedPositionalEmbedding`", "> Hey @mingruimingrui,\r\n> \r\n> Thanks a lot for applying the changes also to the other models - it looks very nice already :-) Could you also completely remove the `padding_idx` from the call to the `PositionalEmbeddings`, _e.g._ `BartLearnedPositionalEmbedding`\r\n\r\nNoted, what do you say about leaving `TFLearnedPositionalEmbeddings` as is?\r\nThis way, padding_idx wouldn't sneak its way into `kwargs`.", "Anything else I've missed out on?", "Looks great to me! I'll run the slow tests of the respective models just to be sure and then I think we are good to go! ", "Slow tests are all passing. \r\n\r\nPinging @LysandreJik @sgugger @patil-suraj for review. To be concise, this PR removes the dependency of the `LearnedPositionalEmbedding` on the `pad_token_id`. After discussion with @mingruimingrui , I think positional encodings should **not** be dependent on the padding idx and the padding idx should not get a special positional embedding. The reason is the following:\r\n- The padding idx can be at every position so it doesn't make sense to have one positional embedding be reserved for the padding idx\r\n- For a padded input - let's say `<pad><pad> Hello <pad> <pad>` my name, the position ids should be `0 0 0 1 1 1 2` IMO and not `pad_pos pad_pos 0 pad_pos pad_pos 1 2`\r\n- Forcing one token to be a padding token, let's say 3 means that the positional embedding for 3 doesn't exist anymore which can then be very awkward when fine-tuning a model or training it from scratch\r\n\r\n=> This PR fixes this behavior without any breaking changes." ]

[]

[]

[ "Hi! It seems that you're on a very old transformers library version. I would recommend you upgrade to a more recent versions, as this particular has been patched several months ago.", "@LysandreJik Hi! thank you for answering. But I'm using version 4.3.2", "Are you sure? The error happens on the following line:\r\n\r\n```py\r\n extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility\r\n```\r\n\r\nThis does not exist in recent Transformers versions.", "It was changed in may: https://github.com/huggingface/transformers/pull/4300", "@LysandreJik Oh.. sorry. i was confused. I installed and uninstalled because it interupted run the code, saying it can't import pytorch_transformers.\r\n\r\nThe code can run with transformers from \"https://github.com/huggingface/transformers/tree/067923d3267325f525f4e46f357360c191ba562e.\"\r\n\r\nIs there any way that i can run?\r\nthis is the full traceback when i install transformers\r\n```\r\nTraceback (most recent call last):\r\n File \"oscar/run_captioning.py\", line 20, in <module>\r\n from oscar.utils.cbs import ConstraintFilter, ConstraintBoxesReader\r\n File \"/home/u2/바탕화면/Oscar/oscar/utils/cbs.py\", line 13, in <module>\r\n from oscar.modeling.modeling_utils import BeamHypotheses\r\n File \"/home/u2/바탕화면/Oscar/oscar/modeling/modeling_utils.py\", line 8, in <module>\r\n from transformers.pytorch_transformers.modeling_bert import (BertConfig,\r\nModuleNotFoundError: No module named 'transformers.pytorch_transformers'\r\n```\r\n\r\n", "Oh that's a very old version indeed! Unfrotunately, without knowing what's in your `oscar` folder it's a bit complicated to help you.\r\n\r\nFor example, the following line is erroneous:\r\n\r\n```\r\nfrom transformers.pytorch_transformers.modeling_bert import (BertConfig,\r\n```\r\n\r\nIt should simply be \r\n\r\n```\r\nfrom transformers import BertConfig\r\n```", "@LysandreJik It imports many .py files from pytorch_transformers/modeling_bert.\r\n\r\nDo you think if i chage below code, I can fix it?\r\n` extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility`\r\n\r\n\r\n\r\n\r\n\r\n", "This is the try/except we implemented to prevent that error from happening. You can add it to your file:\r\n\r\n```py\r\n try:\r\n dtype = next(self.parameters()).dtype\r\n except StopIteration:\r\n # For nn.DataParallel compatibility in PyTorch 1.5\r\n\r\n def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:\r\n tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]\r\n return tuples\r\n\r\n gen = self._named_members(get_members_fn=find_tensor_attributes)\r\n first_tuple = next(gen)\r\n dtype = first_tuple[1].dtype\r\n\r\n extended_attention_mask = extended_attention_mask.to(dtype)\r\n```", "@LysandreJik I added it on my modeling_bert.py. \r\n\r\nI got this traceback.\r\nI guess it's hard to fix it just editing some code from original file.\r\n\r\nI really appreciate your help!\r\n\r\nif you have any idea to solve it, please share with me. \r\nThank you again.\r\n```\r\nTraceback (most recent call last):\r\n File \"oscar/run_captioning.py\", line 886, in <module>\r\n main()\r\n File \"oscar/run_captioning.py\", line 865, in main\r\n global_step, avg_loss = train(args, train_dataset, val_dataset, model, tokenizer)\r\n File \"oscar/run_captioning.py\", line 436, in train\r\n outputs = model(**inputs)\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py\", line 161, in forward\r\n outputs = self.parallel_apply(replicas, inputs, kwargs)\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/parallel/data_parallel.py\", line 171, in parallel_apply\r\n return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)])\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py\", line 86, in parallel_apply\r\n output.reraise()\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/_utils.py\", line 428, in reraise\r\n raise self.exc_type(msg)\r\nNameError: Caught NameError in replica 0 on device 0.\r\nOriginal Traceback (most recent call last):\r\n File \"/home/u2/바탕화면/Oscar/oscar/modeling/modeling_bert.py\", line 224, in forward\r\n dtype = next(self.parameters()).dtype\r\nStopIteration\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/parallel/parallel_apply.py\", line 61, in _worker\r\n output = module(*input, **kwargs)\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/u2/바탕화면/Oscar/oscar/modeling/modeling_bert.py\", line 453, in forward\r\n return self.encode_forward(*args, **kwargs)\r\n File \"/home/u2/바탕화면/Oscar/oscar/modeling/modeling_bert.py\", line 461, in encode_forward\r\n encoder_history_states=encoder_history_states)\r\n File \"/home/u2/anaconda3/envs/oscar/lib/python3.7/site-packages/torch/nn/modules/module.py\", line 727, in _call_impl\r\n result = self.forward(*input, **kwargs)\r\n File \"/home/u2/바탕화면/Oscar/oscar/modeling/modeling_bert.py\", line 227, in forward\r\n def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:\r\nNameError: name 'List' is not defined\r\n\r\n```\r\n", "These are the type hints. Just remove them if you don't need them or don't want to import them:\r\n\r\n```py\r\n try:\r\n dtype = next(self.parameters()).dtype\r\n except StopIteration:\r\n # For nn.DataParallel compatibility in PyTorch 1.5\r\n\r\n def find_tensor_attributes(module):\r\n tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]\r\n return tuples\r\n\r\n gen = self._named_members(get_members_fn=find_tensor_attributes)\r\n first_tuple = next(gen)\r\n dtype = first_tuple[1].dtype\r\n\r\n extended_attention_mask = extended_attention_mask.to(dtype)\r\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." ]

[]

[]

[ "Sorry I think I misinterpreted the operators table https://github.com/onnx/onnx/blob/master/docs/Operators.md and `Shape` should have been available since opset 1, getting an update in opset 13. \r\n\r\nThis seems to be an issue with `onnxjs` not implementing the full set of operators in opset 11." ]

[]

[]

[ "Maybe @patrickvonplaten or @patil-suraj can chime in here!", "Hey @MorenoLaQuatra \r\n\r\nThe `run_seq2seq.py` supports translation with custom dataset. And you can use T5, mT5, MarianMT, mBART, mBART-50 for fine-tuning.\r\nhttps://github.com/huggingface/transformers/tree/master/examples/seq2seq\r\n\r\nAnd the best place to ask this question is our forum https://discuss.huggingface.co/\r\n\r\nHope this helps :) ", "Hi @patil-suraj, thank you for the feedback.\r\n\r\nActually we are trying to train it in a python script (the overall architecture is more complex than the single network). We were asking about the translation loss. If we do as reported in the above snippet, do we optimize the translation loss?\r\nThe finetuning for MarianMT use the same head and loss of the training phase?\r\n\r\nThank you for your time.", "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": 1843244711, "node_id": "MDU6TGFiZWwxODQzMjQ0NzEx", "url": "https://api.github.com/repos/huggingface/transformers/labels/New%20model", "name": "New model", "color": "fbca04", "default": false, "description": "" } ]

[]

[ "Hi @zolekode ,\r\n\r\nthe folder structure is not quite correct:\r\n\r\nhttps://huggingface.co/flexudy/t5-small-wav2vec2-grammar-fixer/tree/main\r\n\r\nYou just need to move everything from the `t5-small-wav2vec2-grammar-fixer` folder to the root folder. Then it should work :hugs: ", "Ah I see. awesome. Thanks alot @stefan-it " ]

[ { "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": "" }, { "id": 2392046359, "node_id": "MDU6TGFiZWwyMzkyMDQ2MzU5", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20Second%20Issue", "name": "Good Second Issue", "color": "dd935a", "default": false, "description": "Issues that are more difficult to do than \"Good First\" issues - give it a try if you want!" } ]

[]

[ "Hi @sgugger I'm up for the mission. After looking though the code I think I have a basic understanding of what you mean. However, would you be able to provide an example on one model just for clarification? I'm a bit confused on the second step\r\n\r\n\r\n>modeling_bert.py\r\n\r\n```\r\n_CONFIG_FOR_DOC = \"BertConfig\"\r\n_TOKENIZER_FOR_DOC = \"BertTokenizer\"\r\n_CHECKPOINT_FOR_DOC = \"bert-base-uncased\"\r\n```\r\n...\r\n\r\n```\r\nclass BertForSequenceClassification(BertPreTrainedModel):\r\n...\r\n...\r\n @add_code_sample_docstrings(\r\n tokenizer_class=_TOKENIZER_FOR_DOC,\r\n checkpoint=_CHECKPOINT_FOR_DOC,\r\n output_type=SequenceClassifierOutput,\r\n config_class=_CONFIG_FOR_DOC,\r\n )\r\n\r\n\r\n```\r\n\r\nThank you", "I've given an example of the first step in the issue above. Sadly the second step can't fully work with our utils just yet, I need to make some adjustments to our internal tooling. If you want to begin on step 1 though, don't hesitate!" ]

[]

[]

[ "`--pad_to_max_length False` is the reason you have a very slow training: this creates batches of different sequence lengths but TPUs need fixed shapes to be efficient.\r\n\r\nThere was a bug in our argument parser before that ignored bool setting like this, so it may be the reason you are seeing that slow down now instead of before (but it was applying `pad_to_max_length=True` before because of that bug, even if you said the opposite). If you remove that option, you should see a faster training.", "Perfect! Thank you so much! Closing this issue" ]

[]

[]

[ "I succeed to fix Marian and Pegasus, and my first guess was the good one. I basically reworked a bit how the embedding was created, and now it works in XLA_GPU. Of course, all the corresponding slow tests are passing, and the weights are properly loaded." ]

[]

[]

[ "I tried decorating a function that contains the tranier command like this:-\r\n```\r\n\r\[email protected](num_cpus=3, num_gpus=1, accelerator_type=ray.accelerators.NVIDIA_TESLA_V100)\r\ndef search():\r\n trainer.hyperparameter_search(n_trials=100, compute_objective='accuracy', direction=\"maximize\", backend='ray',\r\n scheduler=pbt)\r\n\r\nsearch.remote()\r\n```\r\nbut I am constantly getting:\r\n`AttributeError: module 'ray' has no attribute 'accelerators'`\r\nwhich I think is there because I may have written it the wrong way. can anyone shed any light on this?\r\n", "Hi @neel04 can you add this arg to `trainer.hyperparameter_search`:\r\n\r\n```\r\nresources_per_trial={\"cpu\": 1, \"gpu\": 1}\r\n```\r\n\r\nThis will let Tune know to reserve 1 CPU and 1 GPU for each trial.\r\n\r\nAlso, after instantiating your training_args, but before passing it into the `Trainer` can you also add this: `training_args._n_gpu = 1`.\r\n\r\nHere is a more up to date example if you want to try it out and see if it works for you: https://github.com/amogkam/ray/blob/hf-pbt/python/ray/tune/examples/pbt_transformers/pbt_transformers.py\r\n", "@amogkam Bless you sir! I am now getting mostly 100% GPU usage (but only around 6-7GB VRAM usage out of available 16Gb).\r\nHowever, each of my trial is failing with this error:-\r\n```\r\n\r\nTraceback (most recent call last):\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trial_runner.py\", line 586, in _process_trial\r\n results = self.trial_executor.fetch_result(trial)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/ray_trial_executor.py\", line 609, in fetch_result\r\n result = ray.get(trial_future[0], timeout=DEFAULT_GET_TIMEOUT)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/_private/client_mode_hook.py\", line 47, in wrapper\r\n return func(*args, **kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/worker.py\", line 1456, in get\r\n raise value.as_instanceof_cause()\r\nray.exceptions.RayTaskError(TuneError): ray::ImplicitFunc.train_buffered() (pid=415, ip=172.28.0.2)\r\n File \"python/ray/_raylet.pyx\", line 480, in ray._raylet.execute_task\r\n File \"python/ray/_raylet.pyx\", line 432, in ray._raylet.execute_task.function_executor\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trainable.py\", line 167, in train_buffered\r\n result = self.train()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trainable.py\", line 226, in train\r\n result = self.step()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 366, in step\r\n self._report_thread_runner_error(block=True)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 513, in _report_thread_runner_error\r\n (\"Trial raised an exception. Traceback:\\n{}\".format(err_tb_str)\r\nray.tune.error.TuneError: Trial raised an exception. Traceback:\r\nray::ImplicitFunc.train_buffered() (pid=415, ip=172.28.0.2)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 248, in run\r\n self._entrypoint()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 316, in entrypoint\r\n self._status_reporter.get_checkpoint())\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 576, in _trainable_func\r\n output = fn()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 651, in _inner\r\n inner(config, checkpoint_dir=None)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 645, in inner\r\n fn(config, **fn_kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/integrations.py\", line 160, in _objective\r\n local_trainer.train(resume_from_checkpoint=checkpoint, trial=trial)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 983, in train\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 1059, in _maybe_log_save_evaluate\r\n self._report_to_hp_search(trial, epoch, metrics)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 640, in _report_to_hp_search\r\n self.objective = self.compute_objective(metrics.copy())\r\nTypeError: 'str' object is not callable\r\n```\r\nIs it something related to `ray/tune` or is this a wrong argument on my part?\r\nBTW would you also happen to know how to set a fixed batch size for all trials? for some reason, it is overriding the `batch_size` provided in the Trainer_arguments and trying out it's own random ones", "Ah this is because you are passing in `'accuracy'` as the `compute_objective` to `hyperparameter_search`. The `compute_objective` should actually be a function that computes the objective to minimize or maximize from the metrics returned by `evaluate`. You can also not pass one in, and it will default to `trainer_utils.default_compute_objective`. ", "@amogkam MIssed that :( thanx a lot for taking the time out to help me!! :+1: :100: :1st_place_medal: \r\nThanx to amogkam's comment, the issuse described here has been resolved, so I am closing this. but it is still giving the error:-\r\n```\r\n\r\n2021-02-15 18:58:56,343\tERROR worker.py:1053 -- Possible unhandled error from worker: ray::ImplicitFunc.train_buffered() (pid=2319, ip=172.28.0.2)\r\n File \"python/ray/_raylet.pyx\", line 480, in ray._raylet.execute_task\r\n File \"python/ray/_raylet.pyx\", line 432, in ray._raylet.execute_task.function_executor\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trainable.py\", line 167, in train_buffered\r\n result = self.train()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trainable.py\", line 226, in train\r\n result = self.step()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 366, in step\r\n self._report_thread_runner_error(block=True)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 513, in _report_thread_runner_error\r\n (\"Trial raised an exception. Traceback:\\n{}\".format(err_tb_str)\r\nray.tune.error.TuneError: Trial raised an exception. Traceback:\r\nray::ImplicitFunc.train_buffered() (pid=2319, ip=172.28.0.2)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 248, in run\r\n self._entrypoint()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 316, in entrypoint\r\n self._status_reporter.get_checkpoint())\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 576, in _trainable_func\r\n output = fn()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 651, in _inner\r\n inner(config, checkpoint_dir=None)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 645, in inner\r\n fn(config, **fn_kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/integrations.py\", line 160, in _objective\r\n local_trainer.train(resume_from_checkpoint=checkpoint, trial=trial)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 983, in train\r\n self._maybe_log_save_evaluate(tr_loss, model, trial, epoch)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 1059, in _maybe_log_save_evaluate\r\n self._report_to_hp_search(trial, epoch, metrics)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 640, in _report_to_hp_search\r\n self.objective = self.compute_objective(metrics.copy())\r\n File \"<ipython-input-20-4880cfa49052>\", line 5, in compute_metrics\r\nAttributeError: 'dict' object has no attribute 'label_ids'\r\n```\r\nIt's not clear to me why there is a whole string of errors on code copy/pasted from training models (which executes successfully). but there are 2 issues:-\r\n\r\n1. Firstly, the `train_bs` is randomly selected and almost always causes OOM's - arguments do not override this behavior\r\n2. The above \"dict\" error happens, if OOM does not get it.\r\n\r\ncan anyone also explain why the method in the official docs doesn't work even though it is the same task with the same libraries? Has the `ray` framework changed and the docs do not reflect it? Honestly, it is becoming difficult exactly where to post these errors - ray OR HuggingFace. They require a person intimately familiar with both.", "@neel04 are you running this example- https://github.com/amogkam/ray/blob/hf-pbt/python/ray/tune/examples/pbt_transformers/pbt_transformers.py? This is the most up to date one and should work with transformers v4.", "I think the example is a bit verbose and some of it goes over my head :) So I am having difficulty in identifying what steps I have configured wrong and what exactly needs to be corrected. Understanding things like passing the `tune_configs` are easy to get, but the errors are much more difficult to track", "OK got it. The stack trace you just posted is coming from the `compute_metrics` that is passed into your `Trainer`. What does that look like?", "`compute_metrics` is standard code from the official example:-\r\n```\r\nfrom sklearn.metrics import accuracy_score, precision_recall_fscore_support\r\n\r\ndef compute_metrics(pred):\r\n labels = pred.label_ids\r\n preds = pred.predictions.argmax(-1)\r\n precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='weighted') #none gives score for each class\r\n acc = accuracy_score(labels, preds)\r\n return {\r\n 'accuracy': acc,\r\n 'f1': f1,\r\n 'precision': precision,\r\n 'recall': recall\r\n }\r\n```\r\nIt worked perfectly when training the model without any tuning, so I doubt the true error originates from here.\r\nAlso, would you mind telling me how you tracked the problem to `compute_metrics` since I couldn't sniff any clues to there in the error?", "The stack trace says `line 5, in compute_metrics` so I thought it was coming from there. Do you mind sharing your full code to reproduce this?", "Sure. Sorry if the code is a bit verbose.\r\n```\r\n\r\n%%capture\r\n!pip install ray[tune]\r\n!pip install ray\r\n!pip install -q transformers\r\n\r\nfrom sklearn.metrics import accuracy_score, precision_recall_fscore_support\r\nfrom transformers import RobertaForSequenceClassification, Trainer, TrainingArguments\r\n\r\ndef compute_metrics(pred):\r\n labels = pred.label_ids\r\n preds = pred.predictions.argmax(-1)\r\n precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='weighted') #none gives score for each class\r\n acc = accuracy_score(labels, preds)\r\n return {\r\n 'accuracy': acc,\r\n 'f1': f1,\r\n 'precision': precision,\r\n 'recall': recall\r\n }\r\n\r\ntraining_args = TrainingArguments(\r\n output_dir='./results', # output directory\r\n overwrite_output_dir = True,\r\n num_train_epochs=20, # total number of training epochs\r\n per_device_train_batch_size=8, # batch size per device during training\r\n per_device_eval_batch_size=8, # batch size for evaluation,\r\n warmup_steps=500, # number of warmup steps for learning rate scheduler\r\n weight_decay=0.01, # strength of weight decay\r\n logging_dir='./logs', # directory for storing logs\r\n logging_steps=10,\r\n evaluation_strategy='steps',\r\n learning_rate=2e-5,\r\n fp16 = True,\r\n load_best_model_at_end = True,\r\n metric_for_best_model = 'accuracy',\r\n greater_is_better = True,\r\n seed = 101,\r\n do_eval = True,\r\n do_train = True,\r\n adam_beta1=0.9,\r\n adam_beta2=0.999,\r\n adam_epsilon=1e-8,\r\n max_grad_norm=1.0,\r\n adafactor = False\r\n\r\n)\r\n\r\ntraining_args._n_gpu = 1\r\n\r\ndef model_init():\r\n return RobertaForSequenceClassification.from_pretrained('/content/drive/MyDrive/checkpoint-2700', num_labels=20)\r\ntrainer = Trainer(\r\n model_init=model_init,\r\n args=training_args,\r\n train_dataset=train_dataset, # Indexing\r\n eval_dataset=val_dataset,\r\n tokenizer=tokenizer,\r\n compute_metrics=compute_metrics)\r\n\r\nfrom ray.tune.suggest.hyperopt import HyperOptSearch\r\nfrom ray.tune.schedulers import PopulationBasedTraining\r\nfrom ray.tune import CLIReporter\r\nfrom ray import tune\r\nimport random\r\n\r\npbt = PopulationBasedTraining(\r\n time_attr=\"training_iteration\",\r\n metric=\"accuracy\",\r\n mode=\"max\",\r\n perturbation_interval=10, # every 10 `time_attr` units\r\n # (training_iterations in this case)\r\n hyperparam_mutations={\r\n\r\n \"weight_decay\": tune.uniform(1, 0.0001),\r\n \"seed\": tune.uniform(1,20000),\r\n \"learning_rate\": tune.choice([1e-5, 2e-5, 3e-5, 4e-5, 5e-5, 6e-5, 2e-7, 1e-7, 3e-7, 2e-8]),\r\n \"adafactor\": tune.choice(['True','False']),\r\n \"adam_beta1\": tune.uniform(1.0, 0.0),\r\n \"adam_beta2\": tune.uniform(1.0, 0),\r\n \"adam_epsilon\": tune.choice([1e-8, 2e-8, 3e-8, 1e-9, 2e-9, 3e-10]),\r\n \"max_grad_norm\": tune.uniform(1.0, 0),\r\n\r\n })\r\n\r\n\r\nreporter = CLIReporter(\r\n parameter_columns={\r\n \"weight_decay\": \"w_decay\",\r\n \"learning_rate\": \"lr\",\r\n \"per_device_train_batch_size\": \"train_bs/gpu\",\r\n \"num_train_epochs\": \"num_epochs\"},\r\n\r\n metric_columns=[\"eval_acc\", \"eval_loss\", \"epoch\", \"training_iteration\"])\r\n\r\ntune_config = {\r\n \"per_device_train_batch_size\": 8,\r\n \"per_device_eval_batch_size\": 16,\r\n \"num_train_epochs\": tune.choice([15,20,25])\r\n }\r\n \r\nbest = trainer.hyperparameter_search(hp_space = lambda _: tune_config, \r\n n_trials=100, compute_objective=compute_metrics, direction=\"maximize\", backend='ray', #search_alg=HyperOptSearch(metric='accuracy', mode='max', use_early_stopped_trials=True)\r\n scheduler=pbt, resources_per_trial={\"cpu\": 2, \"gpu\": 1}, keep_checkpoints_num=1,\r\n name = \"tune_transformer_pbt\", progress_reporter=reporter)\r\n```", "Ah @neel04, the error message is happening because `compute_metrics` is being passed as the `compute_objective` arg in `trainer.hyperparameter_search`. If you remove this arg your code runs fine.\r\n\r\n`compute_objective` should be a function that takes in the output of `evaluate` (which is the dictionary returned `compute_metrics`) as an input and returns a single float value (see the docstring). It is not the same as `compute_metrics`. So here you should just be returning the \"accuracy\" value from the input dictionary. Something like this should work I believe:\r\n```\r\ndef compute_objective(metrics):\r\n return metrics[\"accuracy\"]\r\n```", "So I tried that above, but apparently `evaluate` does not return \"accuracy\", so as a workaround I switched to `eval_accuracy`.\r\nBut this creates a new problem; this error comes in the first trial **but** it doesn't go on to the next trial. Could be that it is training? GPU usage seems to be 0, so I doubt it is training but it is not terminating the process or moving on. Strange.\r\n```\r\n\r\n2021-02-17 10:57:12,244\tERROR worker.py:1053 -- Possible unhandled error from worker: ray::ImplicitFunc.train_buffered() (pid=1340, ip=172.28.0.2)\r\n File \"python/ray/_raylet.pyx\", line 480, in ray._raylet.execute_task\r\n File \"python/ray/_raylet.pyx\", line 432, in ray._raylet.execute_task.function_executor\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trainable.py\", line 167, in train_buffered\r\n result = self.train()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/trainable.py\", line 226, in train\r\n result = self.step()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 366, in step\r\n self._report_thread_runner_error(block=True)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 513, in _report_thread_runner_error\r\n (\"Trial raised an exception. Traceback:\\n{}\".format(err_tb_str)\r\nray.tune.error.TuneError: Trial raised an exception. Traceback:\r\nray::ImplicitFunc.train_buffered() (pid=1340, ip=172.28.0.2)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 248, in run\r\n self._entrypoint()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 316, in entrypoint\r\n self._status_reporter.get_checkpoint())\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 576, in _trainable_func\r\n output = fn()\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 651, in _inner\r\n inner(config, checkpoint_dir=None)\r\n File \"/usr/local/lib/python3.6/dist-packages/ray/tune/function_runner.py\", line 645, in inner\r\n fn(config, **fn_kwargs)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/integrations.py\", line 160, in _objective\r\n local_trainer.train(resume_from_checkpoint=checkpoint, trial=trial)\r\n File \"/usr/local/lib/python3.6/dist-packages/transformers/trainer.py\", line 925, in train\r\n for step, inputs in enumerate(epoch_iterator):\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py\", line 435, in __next__\r\n data = self._next_data()\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py\", line 475, in _next_data\r\n data = self._dataset_fetcher.fetch(index) # may raise StopIteration\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/fetch.py\", line 44, in fetch\r\n data = [self.dataset[idx] for idx in possibly_batched_index]\r\n File \"/usr/local/lib/python3.6/dist-packages/torch/utils/data/_utils/fetch.py\", line 44, in <listcomp>\r\n data = [self.dataset[idx] for idx in possibly_batched_index]\r\n File \"<ipython-input-12-cd510628f360>\", line 10, in __getitem__\r\nTypeError: new(): invalid data type 'str'\r\n```\r\n\r\nIt looks like it Is pointing to 'objective', which is the same function you wrote above:-\r\n\r\n```\r\ndef compute_objective(metrics):\r\n return metrics[\"eval_accuracy\"] #does not return accuracy\r\n```\r\nInterestingly, removing the args `compute_objective`and `direction` does not yield anything, so I figured the problem must be elsewhere. \r\n\r\nPutting `eval_accuracy` in the PBT parameters and making the `compute_objective` solves the issue.\r\n\r\nThanks a lot @amogkam for your support!! we need more people like you :+1: :rocket: :partying_face: \r\n" ]

[]

[]

[]

[]

[]

[ "I have tried on my machine and the test passes, so the bug is linked to the setup of the machine executing the multi-GPU tests. I have never seen that error before but I would guess there is something wrong with nccl/cuda?", "Retrieved the backtrace:\r\n\r\nThe error is: `RuntimeError: Address already in use`\r\n\r\nWhich means that more than one of these was running at the same time or one from a previous run is zombie and is holding this port. Probably need to try to catch that case and try a different port. I will have a look\r\n\r\nFull trace:\r\n```\r\ntests/test_trainer_distributed.py:72: \r\n_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ \r\n\r\ncmd = ['/home/github_actions/actions-runner/_work/transformers/transformers/.env/bin/python', '-m', 'torch.distributed.launc.../github_actions/actions-runner/_work/transformers/transformers/tests/test_trainer_distributed.py', '--output_dir', ...]\r\nenv = {'CI': 'true', 'GITHUB_ACTION': 'run7', 'GITHUB_ACTIONS': 'true', 'GITHUB_ACTION_REF': '', ...}\r\nstdin = None, timeout = 180, quiet = False, echo = True\r\n\r\n def execute_subprocess_async(cmd, env=None, stdin=None, timeout=180, quiet=False, echo=True) -> _RunOutput:\r\n \r\n loop = asyncio.get_event_loop()\r\n result = loop.run_until_complete(\r\n _stream_subprocess(cmd, env=env, stdin=stdin, timeout=timeout, quiet=quiet, echo=echo)\r\n )\r\n \r\n cmd_str = \" \".join(cmd)\r\n if result.returncode > 0:\r\n stderr = \"\\n\".join(result.stderr)\r\n raise RuntimeError(\r\n> f\"'{cmd_str}' failed with returncode {result.returncode}\\n\\n\"\r\n f\"The combined stderr from workers follows:\\n{stderr}\"\r\n )\r\nE RuntimeError: '/home/github_actions/actions-runner/_work/transformers/transformers/.env/bin/python -m torch.distributed.launch --nproc_per_node=2 /home/github_actions/actions-runner/_work/transformers/transformers/tests/test_trainer_distributed.py --output_dir /tmp/tmpsdmi_ca2' failed with returncode 1\r\nE \r\nE The combined stderr from workers follows:\r\nE Traceback (most recent call last):\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/tests/test_trainer_distributed.py\", line 82, in <module>\r\nE training_args = parser.parse_args_into_dataclasses()[0]\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/hf_argparser.py\", line 180, in parse_args_into_dataclasses\r\nE obj = dtype(**inputs)\r\nE File \"<string>\", line 59, in __init__\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/training_args.py\", line 479, in __post_init__\r\nE if is_torch_available() and self.device.type != \"cuda\" and self.fp16:\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/file_utils.py\", line 1346, in wrapper\r\nE return func(*args, **kwargs)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/training_args.py\", line 601, in device\r\nE return self._setup_devices\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/file_utils.py\", line 1336, in __get__\r\nE cached = self.fget(obj)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/file_utils.py\", line 1346, in wrapper\r\nE return func(*args, **kwargs)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/training_args.py\", line 586, in _setup_devices\r\nE torch.distributed.init_process_group(backend=\"nccl\")\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/.env/lib/python3.7/site-packages/torch/distributed/distributed_c10d.py\", line 436, in init_process_group\r\nE store, rank, world_size = next(rendezvous_iterator)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/.env/lib/python3.7/site-packages/torch/distributed/rendezvous.py\", line 179, in _env_rendezvous_handler\r\nE store = TCPStore(master_addr, master_port, world_size, start_daemon, timeout)\r\nE RuntimeError: Address already in use\r\nE Traceback (most recent call last):\r\nE File \"/usr/lib/python3.7/runpy.py\", line 193, in _run_module_as_main\r\nE \"__main__\", mod_spec)\r\nE File \"/usr/lib/python3.7/runpy.py\", line 85, in _run_code\r\nE exec(code, run_globals)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/.env/lib/python3.7/site-packages/torch/distributed/launch.py\", line 260, in <module>\r\nE main()\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/.env/lib/python3.7/site-packages/torch/distributed/launch.py\", line 256, in main\r\nE cmd=cmd)\r\nE subprocess.CalledProcessError: Command '['/home/github_actions/actions-runner/_work/transformers/transformers/.env/bin/python', '-u', '/home/github_actions/actions-runner/_work/transformers/transformers/tests/test_trainer_distributed.py', '--local_rank=1', '--output_dir', '/tmp/tmpsdmi_ca2']' returned non-zero exit status 1.\r\nE Traceback (most recent call last):\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/tests/test_trainer_distributed.py\", line 82, in <module>\r\nE training_args = parser.parse_args_into_dataclasses()[0]\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/hf_argparser.py\", line 180, in parse_args_into_dataclasses\r\nE obj = dtype(**inputs)\r\nE File \"<string>\", line 59, in __init__\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/training_args.py\", line 479, in __post_init__\r\nE if is_torch_available() and self.device.type != \"cuda\" and self.fp16:\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/file_utils.py\", line 1346, in wrapper\r\nE return func(*args, **kwargs)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/training_args.py\", line 601, in device\r\nE return self._setup_devices\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/file_utils.py\", line 1336, in __get__\r\nE cached = self.fget(obj)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/file_utils.py\", line 1346, in wrapper\r\nE return func(*args, **kwargs)\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/src/transformers/training_args.py\", line 586, in _setup_devices\r\nE torch.distributed.init_process_group(backend=\"nccl\")\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/.env/lib/python3.7/site-packages/torch/distributed/distributed_c10d.py\", line 455, in init_process_group\r\nE barrier()\r\nE File \"/home/github_actions/actions-runner/_work/transformers/transformers/.env/lib/python3.7/site-packages/torch/distributed/distributed_c10d.py\", line 1960, in barrier\r\nE work = _default_pg.barrier()\r\nE RuntimeError: NCCL error in: /pytorch/torch/lib/c10d/ProcessGroupNCCL.cpp:784, unhandled system error, NCCL version 2.7.8\r\n\r\nsrc/transformers/testing_utils.py:1062: RuntimeError\r\n\r\n```", "We can close this one, right, @LysandreJik?\r\n\r\nIt was happening because some zombies from previous jobs, which will now kill before the job starts. i.e. this test wasn't at fault.", "Yes, we can close! Thanks." ]

[]

[]

[]

[]

[]

[ "Hi, no workaround, we're working on the implementation now (https://github.com/huggingface/transformers/pull/10018). It should be available in a few days.", "Thanks @LysandreJik ", "@saichandrapandraju the PR was merged, so I think this issue can be closed now?", "ok @yaysummeriscoming ,\r\n\r\nMay I know when will be the stable release( I think 4.4.0) for these merges?", "Yes, this issue can be closed! v4.4.0 should be released in the next two weeks.", "Has this been fixed? I'm downloading the model direct from huggingface [here](https://huggingface.co/microsoft/deberta-v2-xlarge) and i still get this error thrown", "Are you using the `DebertaV2Model` or the `DebertaModel` ?", "@LysandreJik I am using the simpletransformers library, i'm not sure if you're familiar with it but i believe by default it uses the DebertaModel, not sure how and if i can change it to DebertaV2Model" ]

[]

[ { "login": "jplu", "id": 959590, "node_id": "MDQ6VXNlcjk1OTU5MA==", "avatar_url": "https://avatars.githubusercontent.com/u/959590?v=4", "gravatar_id": "", "url": "https://api.github.com/users/jplu", "html_url": "https://github.com/jplu", "followers_url": "https://api.github.com/users/jplu/followers", "following_url": "https://api.github.com/users/jplu/following{/other_user}", "gists_url": "https://api.github.com/users/jplu/gists{/gist_id}", "starred_url": "https://api.github.com/users/jplu/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/jplu/subscriptions", "organizations_url": "https://api.github.com/users/jplu/orgs", "repos_url": "https://api.github.com/users/jplu/repos", "events_url": "https://api.github.com/users/jplu/events{/privacy}", "received_events_url": "https://api.github.com/users/jplu/received_events", "type": "User", "site_admin": false } ]

[ "Hello @saboof!\r\n\r\nFirst of all, which version of Transformers and TF are you using? And can you share with us a Colab from which we can reproduce the issue. Thanks!!", "version of transformers: transformers==4.2.2\r\nversion of TF2: tensorflow==2.3.1\r\n\r\nafter editing, the code is attached to the main issue ", "Ok, Thanks for sharing! I now see better what you are trying to do. The results you get are different mostly because how the models are implemented don't support what you are trying to do and then some unexpected behavior might occur.\r\n\r\nUsually, a model should not belong to another model, and using a tokenizer inside a model is currently not recommended because not stable. Sorry, for your use case you will have to use `TFBertModel` and tokenize/pad your documents outside the model.\r\n\r\nIt is in our plans to give the possibility to integrate the tokenization process directly inside the model, and then be also part of a SavedModel, but we don't have an ETA yet. Sorry :(", "hi @jplu \r\n\r\nThanks for the quick reply :), I don't think your comment answers my question.\r\nEven when I take the tokenizer outside the bert class model, we still get the same error: \r\n\r\nAttributeError: Must set `config_class` to use @keras_serializable", "Sorry I should have been much clearrer. More specifically, you cannot use `TFBertMainLayer` publicly. You have to directly use `TFBertModel`.\r\n\r\nAlso, in case you don't know, TF doesn't recommend to use a model inside another model because some unstable behaviour might occur.", "Thanks,\r\n\r\nBut I don't use `TFBertMainLayer` at all, the opposite, I use `TFBertModel`. \r\n\r\nRegarding your second point I'll try to take the BERT model outside the general model and only pass it as an argument. \r\nGenerally - how would you go about creating additional linear layer on top of BERT CLS embedding? \r\n\r\n\r\n", "> Regarding your second point I'll try to take the BERT model outside the general model and only pass it as an argument.\r\n\r\nThis might bring unexpected behaviour as well.\r\n\r\n> Generally - how would you go about creating additional linear layer on top of BERT CLS embedding?\r\n\r\nI suggest you to take a look at how the `TFBertForSequenceClassification` is built :)", "Great, I'll take a look at this, thanks @jplu ", "Looking at the code of `TFBertForSequenceClassification` - \r\nhttps://github.com/huggingface/transformers/blob/d1b14c9b548de34b6606946482946008622967db/src/transformers/models/bert/modeling_bert.py#L1496 \r\nIt has a linear layer and the BERT model within the same model, (although written in Pytorch). This doesn't give me insights as to how to combine these two models/layers in the same Keras model so we would be able to save them both in the same command. \r\n\r\nThe only option that I see now is to completely adopt the `TFBertForSequenceClassification` as my full model. Unfortunately, This doesn't give the dynamic architecture generation which I would like to achieve. ", "What do you mean by dynamic architecture? Can you detail a bit more about what you would like to do?", "For example, having two types of outputs based on the same CLS embedding. \r\n\r\nThe first type would be based on a single linear layer and the second type would have two layers with some activation function between the layers.\r\n\r\nsketch of the model.\r\n\r\nInput -> BERT CLS embedded vector -> output based on a single layer -> another output based on a second layer\r\n\r\nMore generally, we would like to get the CLS embedded vector and to add on top of that whatever other layers. \r\nThese layers should be part of the Keras model so they all will be saved as part of the model.\r\n\r\nI hope this is somehow clearer now. \r\n", "Ok this is much clearer now, thanks!!\r\n\r\nThe best way to do that would be to build your model as we do for ours, it means something like:\r\n\r\n```python\r\nclass MyModel(TFBertPreTrainedModel):\r\n def __init__(self, config, *inputs, **kwargs):\r\n super().__init__(config, *inputs, **kwargs)\r\n\r\n self.bert = TFBertMainLayer(config, name=\"bert\")\r\n self.my_first_layer = .....\r\n self.my_second_layer = ....\r\n\r\n def call(inputs, training=None):\r\n bert_output = bert(inputs)\r\n # below all the processes that you have to implement for your models\r\n ....\r\n```\r\n\r\nOnce you have your model you can instantiate it like our models with ```model = MyModel.from_pretrained(\"model_path\")```", "Hi @jplu \r\n\r\nAgain, thanks for the help. \r\n\r\nSorry, I think we are back in square one. \r\nMy initial issue was how to save a Keras model which includes a HF model within it?\r\n\r\nAttached is an example of a code that generates a Keras model which incorporates a BERT model. \r\n[example.txt](https://github.com/huggingface/transformers/files/6015722/example.txt)\r\n\r\nHow would you go about saving this model in such a way that I'd be able to load the model with the additional layer on top of it? can I save the tokenizer in the same place? \r\n\r\n\r\n\r\n\r\n\r\n", "> Sorry, I think we are back in square one. My initial issue was how to save a Keras model which includes a HF model within it?\r\n\r\nOnce again, you should not integrate a model inside another model, What I proposed you to do is exactly what you would like to do because `TFBertMainLayer == TFBertModel`.\r\n\r\n> How would you go about saving this model in such a way that I'd be able to load the model with the additional layer on top of it? can I save the tokenizer in the same place?\r\n\r\nI would go to what I proposed :)", "Thanks @jplu \r\n\r\nFor the best of my understanding, your proposed code doesn't instantiate a Keres model. right?\r\nIf I get your suggestions correctly what you say is that I can't wrap HF model inside a Keras model since integrating a model inside another model, is discouraged by Keras. \r\n\r\n", "Yes! I proposed you to use such a way to do that is 100% equivalent to what you would like.", "Thanks! ", "Keras discouraging nested models is new information for me. \r\nAnd this is weird as in their Keras official examples they do exactly that.\r\nThey have a nested HF model inside a Keras model.\r\n\r\nhttps://keras.io/examples/nlp/text_extraction_with_bert/#preprocess-the-data\r\nPlease look under the function create_model(). \r\n\r\nDo you have any pointers in which Keras discouraging nested models is mentioned? ", "What you point out is a totally different way to what you are doing. In the link you proposed, the model is built in a functional manner which is totally different of building a model in a subclassing manner.\r\n\r\nPlease use the TFBertMainLayer, this is the exact same thing.", "Sorry, I can't find the difference between the Keras example and my previous example under the create_model function.\r\nAttached is the function I used in my example file.\r\n\r\n```\r\ndef create_model(bert_variant, max_len):\r\n ## BERT encoder\r\n encoder = TFBertModel.from_pretrained(bert_variant)\r\n\r\n ## findings classifer Model\r\n input_word_ids = tf.keras.layers.Input(shape=(max_len,), dtype=tf.int32,\r\n name=\"input_word_ids\")\r\n embedding = encoder([input_word_ids]).last_hidden_state\r\n cls_layer = embedding[:,0,:]\r\n\r\n logits = layers.Dense(4, name=\"prediction\", use_bias=True)(cls_layer)\r\n logits = layers.Reshape((1, 4))(logits)\r\n\r\n pred_probs = layers.Activation(tf.keras.activations.softmax)(logits)\r\n\r\n model = tf.keras.Model(\r\n inputs=[input_word_ids],\r\n outputs=[pred_probs],\r\n )\r\n\r\n loss = tf.keras.losses.CategoricalCrossentropy(from_logits=False)\r\n optimizer = tf.keras.optimizers.Adam(lr=5e-5)\r\n model.compile(optimizer=optimizer, loss=loss)\r\n return model\r\n\r\n```\r\n \r\n", "Sorry, this way of creating a model is fine, I was referencing to your first example script called `compile_model.txt`.\r\n\r\nNow to properly save the model you can do something like this:\r\n```python\r\ndef create_model(bert_variant, max_len):\r\n ## BERT encoder\r\n encoder = TFBertModel.from_pretrained(bert_variant).bert # ====> look at this change here :)\r\n\r\n ## findings classifer Model\r\n input_word_ids = tf.keras.layers.Input(shape=(max_len,), dtype=tf.int32,\r\n name=\"input_word_ids\")\r\n embedding = encoder([input_word_ids]).last_hidden_state\r\n cls_layer = embedding[:,0,:]\r\n\r\n logits = layers.Dense(4, name=\"prediction\", use_bias=True)(cls_layer)\r\n logits = layers.Reshape((1, 4))(logits)\r\n\r\n pred_probs = layers.Activation(tf.keras.activations.softmax)(logits)\r\n\r\n model = tf.keras.Model(\r\n inputs=[input_word_ids],\r\n outputs=[pred_probs],\r\n )\r\n\r\n loss = tf.keras.losses.CategoricalCrossentropy(from_logits=False)\r\n optimizer = tf.keras.optimizers.Adam(lr=5e-5)\r\n model.compile(optimizer=optimizer, loss=loss)\r\n return model\r\n\r\nmodel = create_model(\"bert-base-cased\", 128)\r\nmodel.save(\"model_path\")\r\n\r\nloaded_model = tf.keras.models.load_model(\"model_path\")\r\n```", "Or a much nicer version IMO:\r\n```python\r\ndef create_model(bert_variant, max_len):\r\n config = BertConfig.from_pretrained(bert_variant)\r\n input_word_ids = tf.keras.layers.Input(shape=(max_len,), dtype=tf.int32,\r\n name=\"input_word_ids\")\r\n embedding = TFBertMainLayer(config)(input_word_ids)\r\n cls_layer = embedding.last_hidden_state[:,0,:]\r\n logits = tf.keras.layers.Dense(4, name=\"prediction\", use_bias=True)(cls_layer)\r\n logits = tf.keras.layers.Reshape((1, 4))(logits)\r\n pred_probs = tf.keras.layers.Activation(tf.keras.activations.softmax)(logits)\r\n model = tf.keras.Model(inputs=[input_word_ids], outputs=[pred_probs])\r\n loss = tf.keras.losses.CategoricalCrossentropy(from_logits=False)\r\n optimizer = tf.keras.optimizers.Adam(lr=5e-5)\r\n model.compile(optimizer=optimizer, loss=loss)\r\n\r\n return model\r\n\r\nmodel = create_model(\"bert-base-cased\", 128)\r\nmodel.save(\"model_path\") \r\n```\r\n\r\nOnce saved you can load and use it like this:\r\n```python\r\nmodel = tf.keras.models.load_model(\"model_path\")\r\nl = [1]*128 \r\ninp = tf.constant([l])\r\nmodel(inp)\r\n```", "Thanks, @jplu ", "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." ]

[]

[]

[]

[]

[]

[ "Will BigBird-Pegasus be added, and then `BigBirdForConditionalGeneration` so that summarization will be possible?", "Yes, we will be adding that soon.\r\n\r\n> Will BigBird-Pegasus be added, and then `BigBirdForConditionalGeneration` so that summarization will be possible?\r\n\r\n", "Once pre-trained checkpoints are uploaded to `huggingface_hub`, model & tokenizer can be accessed this way:\r\n\r\n```python\r\nfrom transformers import BigBirdForMaskedLM, BigBirdForPreTraining, BigBirdTokenizer\r\n\r\ntokenizer = BigBirdTokenizer.from_pretrained(\"google/bigbird-roberta-base\")\r\n\r\n# model with LM head\r\nmodel_with_lm = BigBirdForMaskedLM.from_pretrained(\"google/bigbird-roberta-base\")\r\n\r\n# model with pertaining heads\r\nmodel_for_pretraining = BigBirdForPreTraining.from_pretrained(\"google/bigbird-roberta-base\")\r\n```", "```python\r\nfrom transformers import BigBirdConfig\r\n\r\n# config for bigbird base\r\nconfig = BigBirdConfig(hidden_size=768, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072)\r\n# or simply\r\nconfig = BigBirdConfig()\r\n\r\n# config for bigbird trivia ckpts (both ITC & ETC)\r\nconfig = BigBirdConfig(type_vocab_size=16)\r\n\r\n# config for bigbird large\r\nconfig = BigBirdConfig(hidden_size=1024, num_hidden_layers=24, num_attention_heads=16, intermediate_size=4096)\r\n```\r\n\r\nRunning this script will enable checkpoints conversion:\r\n\r\n```shell\r\npython src/transformers/models/big_bird/convert_bigbird_original_tf_checkpoint_to_pytorch.py --tf_checkpoint_path ./tf_checkpoint/ckpt/model.ckpt-0 --big_bird_config_file ./tf_checkpoint/config.json --pytorch_dump_path ./hf_ckpt\r\n```", "I will fix everything up & add tests for auto padding.", "Failing tests are unrelated to this PR.", "@sgugger, @LysandreJik I updated the code based on your suggestions. Please let me know if I have missed something.", "Thank you for taking care of the comments @vasudevgupta7 and for this PR altogether!", "@vasudevgupta7 great work, when are you planning to add the BigBirdForConditionalGeneration? And any plans on adding the pubmed pre-trained models?", "@sayakmisra I am currently working on it. You can track PR #10991.", "@vasudevgupta7 currently loading `vasudevgupta/bigbird-pegasus-large-bigpatent` into `BigBirdForConditionalGeneration` leads to some weights of the checkpoint not being used for initializing the model. Is there a workaround for this?\r\n\r\nCan we have separate pretrained checkpoints for BigBird and Pegasus without the finetuning, so that we can use the Pegasus decoder along with the BigBird encoder in our code?", "Hey @jigsaw2212, \r\n\r\nwe are still working on integrating `BigBirdPegasus` -> for now only the `google/bigbird-...` are fully supported. `BigBirdPegasus` will be merged in 1,2 weeks " ]

[]

[]

[ "Hi ! Thanks for reporting\r\n#10167 should fix this issue", "Convenient to see that the fix was already in the pipeline. Thanks!" ]

[]

[]

[ "Hello! If you look inside the `TFGPT2LMHeadModel`, you'll see it automatically shifts the labels for you.\r\n\r\nThe model generates tokens given a past. It then compares the generated token to the \"true\" token contained in the labels you passed to it. If you shift the tokens as it is done in the model, you should get identical results:\r\n\r\n```py\r\nimport tensorflow as tf\r\nfrom transformers import TFGPT2LMHeadModel\r\nfrom transformers.modeling_tf_utils import TFCausalLanguageModelingLoss\r\nimport numpy as np\r\n\r\nmodel = TFGPT2LMHeadModel.from_pretrained('gpt2')\r\n\r\n\r\none_line_dset = (np.random.rand(1, 1024)>.5)*1\r\ninput_ids = one_line_dset\r\ntarget_ids = one_line_dset\r\n\r\n# explicit loss calculation\r\nprediction = model.predict(input_ids).logits\r\nprediction = tf.convert_to_tensor(prediction)\r\n\r\n# internal loss calculation\r\noutputs = model(input_ids, labels=target_ids, training=False)\r\n\r\ntarget_ids = target_ids[:, 1:]\r\nprediction = prediction[:, :-1]\r\nl = TFCausalLanguageModelingLoss().compute_loss(target_ids, prediction)\r\n\r\nprint(tf.math.reduce_mean(l), tf.math.reduce_mean(outputs[0]))\r\n```\r\n\r\nYou'll see that the two results are very close to being equal. Note the very close, as they are not exactly equal. @jplu can chime in here but I believe this has to do with the switch to graph mode happening with the call to `.predict`.", "I entirely second what @LysandreJik said! Nice and clear explanation 👍 ", "@LysandreJik @jplu you are super fast guys! Thanks! a lot!", "This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.\n\nPlease note that issues that do not follow the [contributing guidelines](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md) are likely to be ignored." ]

[]

[]

[ "I figured out how to do it. Within python, you have to use `model.save_pretrained(\"path/to/output_dir\")` where `model` is the fine-tuned `model = DistilBertForTokenClassification.from_pretrained('distilbert-base-uncased', num_labels=...)`\r\n\r\nThen inside an empty directory, run\r\n`python -m transformers.convert_graph_to_onnx --model path/to/output_dir --framework pt --tokenizer distilbert-base-uncased out.onnx`\r\n" ]

[]

[]

[ "Hello, Maybe this comment can help you out https://github.com/huggingface/transformers/issues/1935#issuecomment-561305086!", "@LysandreJik Oh, I got it now, thank you!" ]

[]

[]

[]

[]

[]

[ "> The problem arises when using the model from local files. I have noticed that the accuracy of the local model for the exact same configuration (same data, number of epochs, lr, etc.) is around 20% lower than downloading the model for each experiment.\r\n\r\nThis is extremely vague and we can't help you solve your bug if you don't give us something more tangible than this. What code are you then running? What's the difference in accuracy?\r\n", "Thanks. It was my mistake. \r\nThe issue was that I was trying to dynamically change the model config file and reuse it for models with different number of output labels.\r\n" ]

[]

[]

[ "You can decode them back to a string using `T5Tokenizer`, like so:\r\n\r\n`tokenizer.decode(outputs.squeeze().tolist(), skip_special_tokens=True)`\r\n\r\nBtw, for a really good guide on the different generation strategies of models like T5, see this blog post: https://huggingface.co/blog/how-to-generate", "This post was really helpful, thanks!" ]

[]

[]

[ "@patrickvonplaten, @sgugger , @LysandreJik The PR is now finalized and ready for your review :) ", "I've added proper instructions to install the extra dependencies and addressed Patrick and Sylvain's comments regarding the docs and imports. All slow/non-slow tests are passing!\r\n\r\nMerging!", "Edit: please see this issue: #10631 ", "hi @xjdeng\r\n\r\nThanks for reporting this. This PR is now merged. So could you please open an issue with this error, we will discuss it there.\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\nThanks!" ]

[]

[]

[ "Ok, I got it. That means [CLS] token." ]

[]

[]

[ "Thanks for reporting that @lewtun!\r\nFeel free to open a PR when you have a working solution for higher versions of PyTorch!", "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." ]

[]

[]

[]

[]

[]

[ "Would like to help you here, I've created a [Colab notebook](https://colab.research.google.com/drive/1PtRxbK4oNUsm4lrsOvWoNYzA-BhOWwf2?usp=sharing) that illustrates how to fine-tune `TFT5ForConditionalGeneration` using Keras. However, I'm having the same issue as posted in #6817, namely:\r\n\r\n`ValueError: No gradients provided for any variable: ['shared/shared/weight:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._0/SelfAttention/q/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._0/SelfAttention/k/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._0/SelfAttention/v/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._0/SelfAttention/o/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._0/SelfAttention/relative_attention_bias/embeddings:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._0/layer_norm/weight:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._1/DenseReluDense/wi/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._1/DenseReluDense/wo/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._0/layer_._1/layer_norm/weight:0', 'tf_t5for_conditional_generation_2/encoder/block_._1/layer_._0/SelfAttention/q/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._1/layer_._0/SelfAttention/k/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._1/layer_._0/SelfAttention/v/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._1/layer_._0/SelfAttention/o/kernel:0', 'tf_t5for_conditional_generation_2/encoder/block_._1/layer_._0/layer_norm/weight:0', 'tf_t5for_conditional_generation_2/encoder/block_._1/layer_._1/DenseReluDense/wi/kernel:0', 'tf_t5for_conditional_generation_2/encoder/bloc...`\r\n\r\nUPDATE: this issue was resolved by providing the data in the correct format, namely a tuple of `(inputs, outputs)`. A forward pass on a random batch is now working. However, having the following error when calling `model.fit()`:\r\n\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:758 train_step\r\n self.compiled_metrics.update_state(y, y_pred, sample_weight)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/compile_utils.py:387 update_state\r\n self.build(y_pred, y_true)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/compile_utils.py:318 build\r\n self._metrics, y_true, y_pred)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/util/nest.py:1163 map_structure_up_to\r\n **kwargs)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/util/nest.py:1245 map_structure_with_tuple_paths_up_to\r\n expand_composites=expand_composites)\r\n /usr/local/lib/python3.6/dist-packages/tensorflow/python/util/nest.py:878 assert_shallow_structure\r\n input_length=len(input_tree), shallow_length=len(shallow_tree)))\r\n\r\n ValueError: The two structures don't have the same sequence length. Input structure has length 4, while shallow structure has length 3.\r\n```", "Hello!\r\n\r\nFor now T5 cannot be trained with usual `.compile()` and `.fit()` methods (such as multiple other models but we are currently working on this). You have to either use the TFTrainer or to update yourself the behavior of the internal training loop of Keras. An example of how to deal with T5 and properly training it, is showed in this nice [Colab](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb).", "Ok, I saw that Colab but now understand why the author defined his own `train_step` (this was failing for me). Will update my notebook to add this.\r\n\r\nThank you!", "@marton-avrios Here's an updated version of my Colab notebook, illustrating how to fine-tune `TFT5ForConditionalGeneration` on your data: https://colab.research.google.com/drive/1PtRxbK4oNUsm4lrsOvWoNYzA-BhOWwf2?usp=sharing", "@NielsRogge Your colab looks much better but will be buggy for a few cases. Your train step must look something like this:\r\n\r\n```python\r\ndef train_step(self, data):\r\n x, y = data\r\n\r\n with tf.GradientTape() as tape:\r\n y_pred = self(x, training=True)\r\n loss = self.compiled_loss(y, y_pred.logits, regularization_losses=self.losses)\r\n\r\n gradients = tape.gradient(loss, self.trainable_variables)\r\n\r\n self.optimizer.apply_gradients(zip(gradients, self.trainable_variables))\r\n\r\n self.compiled_metrics.update_state(y, y_pred.logits)\r\n\r\n return {m.name: m.result() for m in self.metrics}\r\n```\r\n\r\nNice adaptation in your Colab BTW :)", "Thank you guys, very useful resource! Will it work on TPU or with other distribution strategies? Keras handles that when I stick to using ```compile()``` and ```fit()``` even if I redefine ```train_step()```, right? I mean dividing loss with global batch size, etc.", "I doubt you will be able to train a T5 on TPU because T5 is not entirely XLA compliant so that you might encounter some unexpected issues. Sorry for that, it is also something on which we are currently working :)", "So PyTorch version won't work either on TPU? Any hints as to which parts? I might be able to look into it. I get this error:\r\n```\r\nInvalid argument: {{function_node __inference_distributed_training_steps_51234}} Compilation failure: Detected unsupported operations when trying to compile graph cluster_distributed_training_steps_2864851955408122598[] on XLA_TPU_JIT: StringFormat (No registered 'StringFormat' OpKernel for XLA_TPU_JIT devices compatible with node {{node tf_t5for_conditional_generation/encoder/StringFormat}}){{node tf_t5for_conditional_generation/encoder/StringFormat}}\r\n```", "There are still a lot of work to make TFT5 XLA compliant, so I don't suggest you to use it for this case. The Pytorch version is fully TPU compliant yes.", "@marton-avrios turns out that the Tensorflow implementation of T5 already creates the `decoder_input_ids` for you as seen [here](https://github.com/huggingface/transformers/blob/587197dcd2b50ad9e96aedbfa389bf4fcc294c3c/src/transformers/models/t5/modeling_tf_t5.py#L1376), you don't need to prepare them yourself (I thought this was only supported in the PyTorch version for now). So I've updated my notebook, it's simpler now \r\n\r\n", "Thank you @NielsRogge ! I think you need to provide either label input ids as ```labels``` or shifted label input ids as ```decoder_input_ids``` in input dictionary. At least it failed for me with only ```input_ids``` and ```attention_mask```.", "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\r\n\r\nSeems to me this script is the repetition of this other script: transformers/examples/language-modeling/run_mlm.py \r\nDo you mind adding T5 also to this script? thanks ", "Actually, we can not simply add T5 to this script, because `run_mlm.py` is for encoder-only models (such as BERT, RoBERTa, DeBERTa, etc.). T5 is an encoder-decoder (seq2seq) model, so this would require a new script. The [seq2seq scripts](https://github.com/huggingface/transformers/tree/master/examples/seq2seq) currently only support fine-tuning, not pre-training.\r\n\r\ncc @patil-suraj @sgugger ", "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." ]

[]

[]

[ "@Narsil, do you want to take a look at this?", "Took a look, it seems the issue was not padding, but argument handling.\r\n\r\n" ]

[]

[]

[]

[ { "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": "" }, { "id": 2392046359, "node_id": "MDU6TGFiZWwyMzkyMDQ2MzU5", "url": "https://api.github.com/repos/huggingface/transformers/labels/Good%20Second%20Issue", "name": "Good Second Issue", "color": "dd935a", "default": false, "description": "Issues that are more difficult to do than \"Good First\" issues - give it a try if you want!" } ]

[]

[ "Hi @stas00,\r\n\r\nI could not understand how to use `RUN_SLOW` in the windows command line, When I run it I was getting\r\n```\r\n'RUN_SLOW' is not recognized as an internal or external command,\r\noperable program or batch file.\r\n```\r\n\r\nIt was mentioned in the contribution guidelines but I don't know how to enable it", "Sorry, I don't know much about windows. \r\n\r\nAren't you using some unixy shell on windows to run this? In which case it should support it?\r\n\r\nOtherwise look up how to setup env vars in your windows shell.\r\n\r\nAnd of course the simplest hack is for the duration of your test to simply comment out the `@slow` decorator inside the test file.\r\n", "oh, but I actually just fixed it here https://github.com/huggingface/transformers/pull/10584 - thank you for unearthing this one. I can close it now." ]

[ { "id": 1936351150, "node_id": "MDU6TGFiZWwxOTM2MzUxMTUw", "url": "https://api.github.com/repos/huggingface/transformers/labels/Examples", "name": "Examples", "color": "d4c5f9", "default": false, "description": "Which is related to examples in general" }, { "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 } ]

[ "While what you say make sense, I'm unsure it warrants a new change of argument names on all example scripts as it seems more cosmetic to me.\r\n\r\nThe `TrainingArguments` have the proper mode already (`--do_train`, `--do_eval`, `--do_predict`) so it's only the examples. We're less attached to no breaking changes there for now but we will soon need them to be production ready, so if we want to do this, it should be done by end of next week if possible.\r\n\r\nAlso cc @LysandreJik ", "> we will soon need them to be production ready\r\n\r\nThis examples-are-not-production and examples-are-production back and forth depending on the context is incredibly difficult to sustain.\r\n\r\nI'm proposing to improve clarity and consistency so that the user can have an easier understanding, and if these are honest examples then the goal is to improve that - exemplification. If you make a slide presentation and people find typos in it, you don't say, but I already showed it to a group so the typos have to remain. Examples are that slide presentation that try to do the best exemplification of the core code. At least based on much feedback I received on my PRs.\r\n\r\nAnd if we want to have programs that are production quality see my rfc https://github.com/huggingface/transformers/issues/10155. I really hope this project will make a clear cut decision wrt this back and forth and stand by it. \r\n\r\nMy fantasy is that there will be:\r\n1. examples - change and improve those any time to make things easier - these are living and working tutorials. Nothing is fixed in stone here and which are a subject to evolve at any moment. No fixed API, but really easy to understand the code.\r\n2. apps - production quality programs that are part of the core, with well thought out API, clean refactored code, thorough testing and tests relying on these apps to do testing.\r\n", "Not really my area of expertise here, but I do agree with @stas00 -> I think we should keep the liberty of quickly adapting the examples", "> Are metrics reporting stats on a split or a mode?\r\n> A. split - rename all metrics keys to be `train|val|test`\r\n> B. mode - rename all metrics keys to be `train|eval|predict`\r\n\r\nSo what it should be? - either A or B - the current `train|eval|test` is a very odd amalgamation of A and B. Unless we just say that the `validation` set is really an `evaluation` set and let it be.\r\n\r\nThis impacts the results json files too in example scripts.\r\n", "This really is broken. I was just trying to write some code that was using the splits and again run into this some args are \"*val*\" others \"*eval*\" :( so can't even write automated attribute retrieval by split and have to write the code as:\r\n\r\n```\r\ndef get_actual_samples(self, split):\r\n\r\n if not split in [\"train\", \"eval\", \"test\"]:\r\n raise ValueError(f\"Unknown split {split}\")\r\n\r\n dataset = getattr(self, f\"{split}_dataset\")\r\n split_fixed = split if split != \"eval\" else \"val\"\r\n max_samples_arg = getattr(self.args, f\"max_{split_fixed}_samples\")\r\n max_samples = max_samples_arg if max_samples_arg is not None else len(dataset)\r\n return min(max, len(dataset))\r\n```\r\n\r\nas you can see this is so strange.\r\n\r\nPlease, please, please - make your vote and let's make the examples use either the splits or the modes and not the mix of both:\r\n\r\nA. split - rename all metrics keys to be `train|val|test`\r\nB. mode - rename all metrics keys to be `train|eval|predict`\r\n\r\nTo remind currently it's:\r\n- `train|val-eval|test`\r\n\r\nIf option B is chosen we rename all cl arg keys + metrics: \"val\" => \"eval\" and \"test\" => \"predict\"\r\nIf option A is chosen we rename all cl arg keys + metrics: \"eval\" => \"val\" \r\n\r\nMy vote is B: `train|eval|predict` because we are reporting on and configuring a specific Trainer mode and not the split.\r\n\r\n\r\n", "I vote for B, for consistency with `do_train`, `do_eval`, `do_predict`.\r\n\r\nFor examples: switching an arg name can be done without taking precautions for BC as long as the README is updated at the same time, but for `TrainingArguments `(if any is concerned), a proper deprecation cycle has to be made.", "@bhadreshpsavani, would this be something you'd like to work on by chance? If you haven't tired of examples yet.", "Hi @stas00,\r\nYa i will be happy to work more.\r\nActually I was looking for some issues to work on!", "Awesome! Thank you, @bhadreshpsavani!\r\n\r\nSo the changes we need are:\r\n\r\n1. use `eval` instead of `val`\r\n2. use `predict` instead of `test`\r\n\r\nin cl args and variable names in example scripts (only the active ones, please ignore legacy/research subdirs).\r\n\r\nI hope this will be a last rename in awhile.\r\n", "Hi @stas00, \r\nIn the dataset we have `validation` as a key for proper conversion shall we also need to change it to `evaluation`?\r\nFor `validation_file` we can either change it to `evalution_file` or `eval_file` or keep it as it is.", "No the key in the dataset dictionary is \"validation\", so it should be `validation_file`.", "While testing my changes I come to know that few example scripts are not working fine before my changes!\r\nHere is the List:\r\n```\r\n/language-modeling/run_clm.py \r\n/language-modeling/run_plm.py\r\n/question-answering/run_qa_beam_search.py\r\n```\r\nPlease check/run this [colab](https://github.com/bhadreshpsavani/UnderstandingNLP/blob/master/TestingAllHuggingfaceScripts_.ipynb) for instant testing.\r\nWhen I run the above script locally in ubuntu my system got freeze.", "I have no luck using colab today, it doesn't connect at all, so I can't test.\r\n\r\nI run the clm test as you posted it on my own machine and it worked just fine.\r\n\r\nIs it only colab that it's failing on? ", "Hi @stas00,\r\nIt actually not giving any error but after a few epochs of training, it gives something like this ^C, and it's not doing further stages like eval and predict. ", "Does it silently abort the run w/o any traceback? That often means the system run out of RAM and the kernel killed the process - often you get no response. colab is notorious for giving a tiny amount of RAM.\r\n\r\nI hope colab will start working for me again and I will see if I can reproduce this.\r\n\r\nfor hanging there is this trick, add this to the beginning of the program:\r\n\r\n```\r\nimport faulthandler\r\nfaulthandler.dump_traceback_later(20, repeat=True)\r\n```\r\nnow every 20 secs it will print out where each thread is (traceback). super handy!\r\n", "There is also the on demand version:\r\n```\r\n# register and then kill the process w/ stack trace\r\nimport faulthandler, signal\r\nfaulthandler.register(signal.SIGUSR1)\r\n# kill the stuck process\r\nkill -USR1 PID\r\n```\r\nbut it often doesn't work.\r\n\r\n`py-spy` is another handy one but it requires `sudo` so won't work on colab. Unless you start the program with it:\r\n```\r\n# trace a running python application - e.g. when it's hanging or very slow and you want to see the backtrace - one way is using a sighandler - but that requires killing it and already having it installed\r\npip install py-spy\r\nsudo py-spy top --pid PID\r\n# if one has no sudo, start the program via\r\npy-spy -- python myprogram.py\r\n# and then it will attached without sudo\r\n# https://github.com/benfred/py-spy#when-do-you-need-to-run-as-sudo\r\n```\r\n", "Ya, I think it silently aborted the run w/o any traceback. Might be because it is occupying the entire ram somehow.\r\nSimilar behavior I observed when I run a really big docker image locally.\r\n\r\nI will definitely try this command and dig more!\r\n\r\nThanks a lot for your input. This is really insightful! I will note down this as well :)", "Yes, this is almost always the case in colab. It's too bad they don't have a simple widget that shows real time memory usage.\r\n\r\nFor your personal machine, always have a huge swap file if you do ML dev. Like 100GB. It will save the day.\r\n\r\nAlso there is a way to protect your desktop from your RAM-hungry commands - you can look into `cgroups`. If you machine crashes a lot because of run away training or jupyter, this is another super useful addition. But there are other way to manage resources.", "ok, so I had to disable a `Privacy Badger` firefox extension and colab started working.\r\n\r\nFirst, make a habit to start colab with:\r\n```\r\n!free -h\r\n```\r\nsometimes I get 12GB RAM, other times 25GB, 12GB is typically too low for much.\r\n\r\nSo `run_clm` works just fine even on 12GB. I had to use a small bs so edited your cmd lines to limit bs:\r\n\r\n```\r\n!python examples/pytorch/language-modeling/run_clm.py \\\r\n--model_name_or_path gpt2 \\\r\n--dataset_name wikitext \\\r\n--max_train_samples 5 \\\r\n--max_val_samples 5 \\\r\n--dataset_config_name wikitext-2-raw-v1 \\\r\n--do_train \\\r\n--do_eval \\\r\n--output_dir /tmp/test-clm \\\r\n--per_device_eval_batch_size 2 \\\r\n--per_device_train_batch_size 2 \\\r\n--overwrite_output_dir\r\n```\r\nthis worked too:\r\n```\r\n!python examples/pytorch/language-modeling/run_plm.py \\\r\n--model_name_or_path xlnet-base-cased \\\r\n--dataset_name wikitext \\\r\n--max_train_samples 5 \\\r\n--max_val_samples 5 \\\r\n--dataset_config_name wikitext-2-raw-v1 \\\r\n--do_train \\\r\n--do_eval \\\r\n--output_dir /tmp/test-clm \\\r\n--per_device_eval_batch_size 2 \\\r\n--per_device_train_batch_size 2 \\\r\n--overwrite_output_dir\r\n```\r\nand so did:\r\n```\r\n!python examples/pytorch/question-answering/run_qa.py \\\r\n--model_name_or_path distilbert-base-uncased \\\r\n--train_file tests/fixtures/tests_samples/SQUAD/sample.json \\\r\n--validation_file tests/fixtures/tests_samples/SQUAD/sample.json \\\r\n--test_file tests/fixtures/tests_samples/SQUAD/sample.json \\\r\n--do_train \\\r\n--do_eval \\\r\n--do_predict \\\r\n--max_train_samples 5 \\\r\n--max_val_samples 5 \\\r\n--max_test_samples 5 \\\r\n--learning_rate 3e-5 \\\r\n--max_seq_length 384 \\\r\n--doc_stride 128 \\\r\n--version_2_with_negative \\\r\n--output_dir /tmp/debug_squad/ \\\r\n--per_device_eval_batch_size 2 \\\r\n--per_device_train_batch_size 2 \\\r\n--overwrite_output\r\n```\r\n" ]