Datasets:

0ppxnhximxr

/

CodeAI-valid

like 0

https://api.github.com/repos/huggingface/transformers/issues/11046

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11046/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11046/comments

https://api.github.com/repos/huggingface/transformers/issues/11046/events

https://github.com/huggingface/transformers/issues/11046

MDU6SXNzdWU4NDk1Njg0NTk=

Potential incorrect application of layer norm in BlenderbotSmallDecoder

open

NONE

In BlenderbotSmallDecoder, layer norm is applied only on the token embeddings, and not on the hidden_states, whereas in the BlenderbotSmallEncoder, layer norm is applied after adding the input_embeds and positional embeds BlenderbotSmallEncoder: `hidden_states = inputs_embeds + embed_pos` `hidden_states = self.layernorm_embedding(hidden_states)` BlenderbotSmallDecoder: `inputs_embeds = self.layernorm_embedding(inputs_embeds)` `hidden_states = inputs_embeds + positions`

https://api.github.com/repos/huggingface/transformers/issues/11045

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11045/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11045/comments

https://api.github.com/repos/huggingface/transformers/issues/11045/events

https://github.com/huggingface/transformers/issues/11045

MDU6SXNzdWU4NDk1NDQzNzQ=

Multi-GPU seq2seq example evaluation significantly slower than legacy example evaluation

open

NONE

### Who can help @patil-suraj @sgugger Models: T5 ## Information I've been doing multi-GPU evaluation for some weeks using a Transformers pull from Feb 12th, just using the example scripts for training/evaluating custom datasets (specifically `run_distributed_eval.py` , though that seq2seq example is now legacy: https://github.com/huggingface/transformers/tree/master/examples/legacy/seq2seq ) Today I grabbed a fresh pull and migrated the data over to the JSON lines format for the new seq2seq example `run_summarization.py` : https://github.com/huggingface/transformers/blob/master/examples/seq2seq/run_summarization.py run_summarization.py appears to use all visible GPUs to do the evaluation (great!), but it also appears significantly slower than the old run_distributed_eval.py . When examining GPU utilization using `nvtop`, it appears that it allocates GPU memory from all devices (much more from device 0), but only uses device 0 for processing:  ## Script In case it's my issue and I'm not invoking it correctly (I know the legacy one required being invoked with `torch.distributed.launch` for multi-GPU evaluation), the runscript I'm using is: ``` #/bin/bash python run_summarization.py \ --model_name_or_path mymodel-debug1000 \ --do_predict \ --train_file mydata/train.json \ --validation_file mydata/val.json \ --test_file mydata/val.json \ --max_source_length 256 \ --max_target_length 512 \ --num_beams 8 \ --source_prefix "" \ --output_dir tst-debug \ --overwrite_output_dir \ --per_device_train_batch_size=4 \ --per_device_eval_batch_size=4 \ --predict_with_generate ```

https://api.github.com/repos/huggingface/transformers/issues/11044

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11044/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11044/comments

https://api.github.com/repos/huggingface/transformers/issues/11044/events

https://github.com/huggingface/transformers/issues/11044

MDU6SXNzdWU4NDk1Mjk3NjE=

[DeepSpeed] ZeRO stage 3 integration: getting started and issues

open

COLLABORATOR

**[This is not yet alive, preparing for the release, so please ignore for now]** The DeepSpeed ZeRO-3 has been integrated into HF `transformers`. While I tried to write tests for a wide range of situations I'm sure I've missed some scenarios so if you run into any problems please file a separate issue. I'm going to use this issue to track progress on individual ZeRO3 issues. # Why would you want ZeRO-3 In a few words, while ZeRO-2 was very limited scability-wise - if `model.half()` couldn't fit onto a single gpu, adding more gpus won't have helped so if you had a 24GB GPU you couldn't train a model larger than about 5B params. Since with ZeRO-3 the model weights are partitioned across multiple GPUs plus offloaded to CPU, the upper limit on model size has increased by about 2 orders of magnitude. That is ZeRO-3 allows you to scale to huge models with Trillions of parameters assuming you have enough GPUs and general RAM to support this. ZeRO-3 can benefit a lot from general RAM if you have it. If not that's OK too. ZeRO-3 combines all your GPUs memory and general RAM into a vast pool of memory. If you don't have many GPUs but just a single one but have a lot of general RAM ZeRO-3 will allow you to fit larger models. Of course, if you run in an environment like the free google colab, while you can use run Deepspeed there, you get so little general RAM it's very hard to make something out of nothing. Some users (or some sessions) one gets 12GB of RAM which is impossible to work with - you want at least 24GB instances. Setting is up might be tricky too, please see this notebook for an example: https://github.com/stas00/porting/blob/master/transformers/deepspeed/DeepSpeed_on_colab_CLI.ipynb # Getting started Install the latest deepspeed version: ``` pip install deepspeed ``` You will want to be on a transformers master branch, if you want to run a quick test: ``` git clone https://github.com/huggingface/transformers cd transformers BS=4; PYTHONPATH=src USE_TF=0 deepspeed examples/seq2seq/run_translation.py \ --model_name_or_path t5-small --output_dir /tmp/zero3 --overwrite_output_dir --max_train_samples 64 \ --max_val_samples 64 --max_source_length 128 --max_target_length 128 --val_max_target_length 128 \ --do_train --num_train_epochs 1 --per_device_train_batch_size $BS --per_device_eval_batch_size $BS \ --learning_rate 3e-3 --warmup_steps 500 --predict_with_generate --logging_steps 0 --save_steps 0 \ --eval_steps 1 --group_by_length --adafactor --dataset_name wmt16 --dataset_config ro-en --source_lang en \ --target_lang ro --source_prefix "translate English to Romanian: " \ --deepspeed examples/tests/deepspeed/ds_config_zero3.json ``` You will find a very detailed configuration here: https://huggingface.co/transformers/master/main_classes/trainer.html#deepspeed Your new config file will look like this: ```json { "fp16": { "enabled": true, "loss_scale": 0, "loss_scale_window": 1000, "initial_scale_power": 16, "hysteresis": 2, "min_loss_scale": 1 }, "zero_optimization": { "stage": 3, "cpu_offload": true, "cpu_offload_params": true, "cpu_offload_use_pin_memory" : true, "overlap_comm": true, "contiguous_gradients": true, "stage3_max_live_parameters": 1e9, "stage3_max_reuse_distance": 1e9, "stage3_prefetch_bucket_size": 0.94e6, "stage3_param_persistence_threshold": 1e4, "reduce_bucket_size": 1e6, "prefetch_bucket_size": 3e6, "sub_group_size": 1e14, "stage3_gather_fp16_weights_on_model_save": true }, "optimizer": { "type": "AdamW", "params": { "lr": 3e-5, "betas": [0.8, 0.999], "eps": 1e-8, "weight_decay": 3e-7 } }, "scheduler": { "type": "WarmupLR", "params": { "warmup_min_lr": 0, "warmup_max_lr": 3e-5, "warmup_num_steps": 500 } }, "steps_per_print": 2000, "wall_clock_breakdown": false } ``` So if you were already using ZeRO-2 it's only the `zero_optimization` stage that has changed. One of the biggest nuances of ZeRO-3 is that the model weights aren't inside `model.state_dict`, as they are spread out through multiple gpus. The Trainer has been modified to support this but you will notice a slow model saving - as it has to consolidate weights from all the gpus. I'm planning to do more performance improvements in the future PRs, but for now let's focus on making things work. # Issues / Questions If you have any general questions or something is unclear/missing in the docs please don't hesitate to ask in this thread. But for any bugs or problems please open a new Issue and tag me there. You don't need to tag anybody else. Thank you!

https://api.github.com/repos/huggingface/transformers/issues/11043

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11043/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11043/comments

https://api.github.com/repos/huggingface/transformers/issues/11043/events

https://github.com/huggingface/transformers/issues/11043

MDU6SXNzdWU4NDk0OTk3MzQ=

Can't load model to estimater

open

NONE

I was trying to follow the Sagemaker instructions [here](https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model.html) to load the model I just trained and test an estimation. I get the error message: NotImplementedError: Creating model with HuggingFace training job is not supported. Can someone share some sample code to run to do this? Here is the basic thing I am trying to do: ``` from sagemaker.estimator import Estimator # job which is going to be attached to the estimator old_training_job_name='huggingface-sdk-extension-2021-04-02-19-10-00-242' # attach old training job huggingface_estimator_loaded = Estimator.attach(old_training_job_name) # get model output s3 from training job testModel = huggingface_estimator_loaded.model_data ner_classifier = huggingface_estimator.deploy(initial_instance_count=1, instance_type='ml.m4.xlarge') ``` I also tried some things with .deploy() and endpoints but didn't have any luck there either.

https://api.github.com/repos/huggingface/transformers/issues/11042

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11042/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11042/comments

https://api.github.com/repos/huggingface/transformers/issues/11042/events

https://github.com/huggingface/transformers/issues/11042

MDU6SXNzdWU4NDkyNzQzNjI=

[LXMERT] Unclear what img_tensorize does with color spaces

open

NONE

## Environment info - `transformers` version: Not using transformers directly, I'm loading a model "unc-nlp/frcnn-vg-finetuned" - Platform: MacOS - Python version: 3.8 - PyTorch version (GPU?): 1.6.0, no GPU - Tensorflow version (GPU?): don't have - Using GPU in script?: no - Using distributed or parallel set-up in script?: no ### Who can help @eltoto1219 probably Models: "LXMERT": "unc-nlp/frcnn-vg-finetuned" Library: https://github.com/huggingface/transformers/tree/master/examples/research_projects/lxmert ## Information Model I am using (Bert, XLNet ...): "LXMERT": "unc-nlp/frcnn-vg-finetuned" The problem arises when using: * [ ] the official example scripts: (give details below) * [x ] my own modified scripts: (give details below) The tasks I am working on is: * [ ] an official GLUE/SQUaD task: (give the name) * [x ] my own task or dataset: (give details below) ## Problem I don't get what img_tensorize in utils.py is doing with color spaces. I run the following code to load the model. ``` # load models and model components frcnn_cfg = Config.from_pretrained("unc-nlp/frcnn-vg-finetuned") frcnn = GeneralizedRCNN.from_pretrained("unc-nlp/frcnn-vg-finetuned", config=frcnn_cfg) image_preprocess = Preprocess(frcnn_cfg) ``` Turns out that frcnn_cfg.input.format is "BGR" so I wanted to know what exactly is going on? Here is where the image is loaded (utils.img_tensorize) ``` def img_tensorize(im, input_format="RGB"): assert isinstance(im, str) if os.path.isfile(im): img = cv2.imread(im) else: img = get_image_from_url(im) assert img is not None, f"could not connect to: {im}" img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) if input_format == "RGB": img = img[:, :, ::-1] return img ``` See, we seem to be converting the images to RGB, then if it's "RGB" format we flip the blue (?) channel? Is the image ever converted to "BGR"?

https://api.github.com/repos/huggingface/transformers/issues/11041

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11041/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11041/comments

https://api.github.com/repos/huggingface/transformers/issues/11041/events

https://github.com/huggingface/transformers/pull/11041

MDExOlB1bGxSZXF1ZXN0NjA4MDcxNjc1

wav2vec2 converter: create the proper vocab.json while converting fairseq wav2vec2 finetuned model

open

CONTRIBUTOR

# What does this PR do? While converting a finetuned wav2vec2 model we also need to convert the related dictionary `dict.ltr.txt` to hugging face `vocab.json` format. If a `dict_path` is specified: - Creates&saves the necessary vocab.json file - Modifies config file special token ids and vocab size accordingly - Creates a processor with the right special tokens and saves the processor `preprocessor_config.json` ## Before submitting - [x] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#start-contributing-pull-requests), Pull Request section? - [ ] Did you make sure to update the documentation with your changes? -> not sure if there are any docs related to this. - [ ] Did you write any new necessary tests? -> not sure if there are tests related to this. ## Who can review? Models: - wav2vec2: @patrickvonplaten @LysandreJik

https://api.github.com/repos/huggingface/transformers/issues/11040

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11040/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11040/comments

https://api.github.com/repos/huggingface/transformers/issues/11040/events

https://github.com/huggingface/transformers/issues/11040

MDU6SXNzdWU4NDkyNjU2MTU=

max_length in beam_search() and group_beam_search() does not consider beam_scorer.max_length

open

CONTRIBUTOR

## Environment info <!-- You can run the command `transformers-cli env` and copy-and-paste its output below. Don't forget to fill out the missing fields in that output! --> - `transformers` version: - Platform: 4.3.2 - Python version: 3.8.5 - PyTorch version (GPU?): 1.8.0 - Using GPU in script?: No - Using distributed or parallel set-up in script?: No ### Who can help <!-- Your issue will be replied to more quickly if you can figure out the right person to tag with @ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**. Please tag fewer than 3 people. Models: - albert, bert, xlm: @LysandreJik - blenderbot, bart, marian, pegasus, encoderdecoder, t5: @patrickvonplaten, @patil-suraj - longformer, reformer, transfoxl, xlnet: @patrickvonplaten - fsmt: @stas00 - funnel: @sgugger - gpt2: @patrickvonplaten, @LysandreJik - rag: @patrickvonplaten, @lhoestq - tensorflow: @LysandreJik Library: - benchmarks: @patrickvonplaten - deepspeed: @stas00 - ray/raytune: @richardliaw, @amogkam - text generation: @patrickvonplaten - tokenizers: @LysandreJik - trainer: @sgugger - pipelines: @LysandreJik Documentation: @sgugger Model hub: - for issues with a model report at https://discuss.huggingface.co/ and tag the model's creator. HF projects: - datasets: [different repo](https://github.com/huggingface/datasets) - rust tokenizers: [different repo](https://github.com/huggingface/tokenizers) Examples: - maintained examples (not research project or legacy): @sgugger, @patil-suraj - research_projects/bert-loses-patience: @JetRunner - research_projects/distillation: @VictorSanh --> ## Information I am using BART model in particular, but this problem exists for all the other models, using `beam_search()` and `group_beam_search()` for decoding the generated text. The `max_length` variable set using `BeamSearchScorer` is not used by `beam_search()` or `group_beam_search()` function in `generation_utils.py` script. Thus using a smaller `max_length` while initializing the object of class, for example: ``` beam_scorer = BeamSearchScorer( batch_size=1, max_length=5, num_beams=num_beams, device=model.device, ) ``` instead of ``` beam_scorer = BeamSearchScorer( batch_size=1, max_length=model.config.max_length, num_beams=num_beams, device=model.device, ) ``` in the example given [here](https://huggingface.co/transformers/main_classes/model.html#transformers.generation_utils.GenerationMixin.beam_search) results in an error: ``` File "temp.py", line 34, in <module> outputs = model.beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs) File "<conda_env_dir>/lib/python3.8/site-packages/transformers/generation_utils.py", line 1680, in beam_search sequence_outputs = beam_scorer.finalize( File "<conda_env_dir>/lib/python3.8/site-packages/transformers/generation_beam_search.py", line 328, in finalize decoded[i, : sent_lengths[i]] = hypo RuntimeError: The expanded size of the tensor (5) must match the existing size (6) at non-singleton dimension 0. Target sizes: [5]. Tensor sizes: [6] ``` The problem arises when using: * [x] the official example scripts: - [beam_scorer_example](https://huggingface.co/transformers/main_classes/model.html#transformers.generation_utils.GenerationMixin.beam_search) * [x] my own modified scripts: - Also, using `max_length` higher than `model.config.max_length` while initializing object (`beam_scorer`) of type `BeamSearchScorer` does not help in generating longer sequences, as `beam_scorer.max_length` is not used by `beam_search()` or `group_beam_search()` ## To reproduce Steps to reproduce the behavior: 1. The above mentioned modification in the [example](https://huggingface.co/transformers/main_classes/model.html#transformers.generation_utils.GenerationMixin.beam_search) <!-- If you have code snippets, error messages, stack traces please provide them here as well. Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting Do not use screenshots, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.--> ## Expected behavior 1. The program to run without any errors when lower `max_length` is set for object of type `BeamSearchScorer` 2. Generate longer length sequences (longer than `model.config.max_length`) when higher `max_length` is set for object of type `BeamSearchScorer` <!-- A clear and concise description of what you would expect to happen. -->

https://api.github.com/repos/huggingface/transformers/issues/11039

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11039/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11039/comments

https://api.github.com/repos/huggingface/transformers/issues/11039/events

https://github.com/huggingface/transformers/issues/11039

MDU6SXNzdWU4NDkyNDQ4MTk=

Trainer not logging into Tensorboard

open

NONE

## Environment info <!-- You can run the command `transformers-cli env` and copy-and-paste its output below. Don't forget to fill out the missing fields in that output! --> - `transformers` version: 4.5.0.dev0 - Platform: Ubuntu 18.04.5 LTS (x86_64) - Python version: 3.7.0 - PyTorch version (GPU?): 1.7.1+cu101 - Tensorflow version (GPU?): - Using GPU in script?: Yes - Using distributed or parallel set-up in script?: Yes ### Who can help @sgugger <!-- Your issue will be replied to more quickly if you can figure out the right person to tag with @ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**. Please tag fewer than 3 people. Models: - albert, bert, xlm: @LysandreJik - blenderbot, bart, marian, pegasus, encoderdecoder, t5: @patrickvonplaten, @patil-suraj - longformer, reformer, transfoxl, xlnet: @patrickvonplaten - fsmt: @stas00 - funnel: @sgugger - gpt2: @patrickvonplaten, @LysandreJik - rag: @patrickvonplaten, @lhoestq - tensorflow: @LysandreJik Library: - benchmarks: @patrickvonplaten - deepspeed: @stas00 - ray/raytune: @richardliaw, @amogkam - text generation: @patrickvonplaten - tokenizers: @LysandreJik - trainer: @sgugger - pipelines: @LysandreJik Documentation: @sgugger Model hub: - for issues with a model report at https://discuss.huggingface.co/ and tag the model's creator. HF projects: - datasets: [different repo](https://github.com/huggingface/datasets) - rust tokenizers: [different repo](https://github.com/huggingface/tokenizers) Examples: - maintained examples (not research project or legacy): @sgugger, @patil-suraj - research_projects/bert-loses-patience: @JetRunner - research_projects/distillation: @VictorSanh --> ## Information Model I am using (Bert, XLNet ...): EncoderDecoderModel The problem arises when using: * [ ] the official example scripts: (give details below) * [x] my own modified scripts: (give details below) The tasks I am working on is: * [ ] an official GLUE/SQUaD task: (give the name) * [x] my own task or dataset: (give details below) ## To reproduce Steps to reproduce the behavior: This is the tensorboard logs https://tensorboard.dev/experiment/caY7XIGGTbK2Zfr2DTeoJA/#scalars 1. go to the `Text` tab [here](https://tensorboard.dev/experiment/caY7XIGGTbK2Zfr2DTeoJA/#text), you can see that `"logging_first_step": true, "logging_steps": 2` 2. `epoch` graph is showing 75 total steps, but no scalars were logged except for the first_step <img width="413" alt="Screen Shot 2021-04-02 at 10 14 49 AM" src="https://user-images.githubusercontent.com/66082334/113423326-48c54080-939c-11eb-9c61-b8fde0d62d12.png"> ``` [INFO|trainer.py:402] 2021-04-02 10:05:50,085 >> Using amp fp16 backend [INFO|trainer.py:1013] 2021-04-02 10:05:50,181 >> ***** Running training ***** [INFO|trainer.py:1014] 2021-04-02 10:05:50,182 >> Num examples = 100 [INFO|trainer.py:1015] 2021-04-02 10:05:50,182 >> Num Epochs = 3 [INFO|trainer.py:1016] 2021-04-02 10:05:50,182 >> Instantaneous batch size per device = 1 [INFO|trainer.py:1017] 2021-04-02 10:05:50,182 >> Total train batch size (w. parallel, distributed & accumulation) = 4 [INFO|trainer.py:1018] 2021-04-02 10:05:50,182 >> Gradient Accumulation steps = 1 [INFO|trainer.py:1019] 2021-04-02 10:05:50,182 >> Total optimization steps = 75 {'loss': 13.7546, 'learning_rate': 2.5000000000000002e-08, 'epoch': 0.04} 100%|█████████████████████████████████████████████████████████████████████████████████| 75/75 [01:03<00:00, 1.28it/s][INFO|trainer.py:1196] 2021-04-02 10:06:53,931 >> Training completed. Do not forget to share your model on huggingface.co/models =) {'train_runtime': 63.7497, 'train_samples_per_second': 1.176, 'epoch': 3.0} 100%|█████████████████████████████████████████████████████████████████████████████████| 75/75 [01:03<00:00, 1.18it/s] [INFO|trainer.py:1648] 2021-04-02 10:06:54,265 >> Saving model checkpoint to ./pretrain_decoder/ ``` <!-- If you have code snippets, error messages, stack traces please provide them here as well. Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting Do not use screenshots, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.--> ## Expected behavior It should log training loss very other `logging_steps` right? or did I misunderstood? <!-- A clear and concise description of what you would expect to happen. -->

https://api.github.com/repos/huggingface/transformers/issues/11038

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11038/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11038/comments

https://api.github.com/repos/huggingface/transformers/issues/11038/events

https://github.com/huggingface/transformers/issues/11038

MDU6SXNzdWU4NDkxODAzODQ=

DeBERTa xlarge v2 throwing runtime error

open

NONE

- `transformers` version: 4.4.2 - Platform: Linux-3.10.0-1127.el7.x86_64-x86_64-with-redhat-7.8-Maipo - Python version: 3.6.13 - PyTorch version (GPU?): 1.7.1 (True) - Tensorflow version (GPU?): not installed (NA) - Using GPU in script: yes ``` RuntimeError: Error(s) in loading state_dict for DebertaForSequenceClassification: size mismatch for deberta.encoder.rel_embeddings.weight: copying a param with shape torch.Size([512, 1536]) from checkpoint, the shape in current model is torch.Size([1024, 1536]). ``` I've seen a previous post made about this error and i believe it is a known issue. On the thread i found it was mentioned that a fix was due to come out a month ago. Has the fix come out? [Downloaded from here](https://huggingface.co/microsoft/deberta-v2-xlarge)

https://api.github.com/repos/huggingface/transformers/issues/11036

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11036/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11036/comments

https://api.github.com/repos/huggingface/transformers/issues/11036/events

https://github.com/huggingface/transformers/issues/11036

MDU6SXNzdWU4NDg5OTYyNDA=

BertForTokenClassification class ignores long tokens when making predictions

open

NONE

# Goal I am trying to run the adapted version of `run_ner.py` hosted [here](https://github.com/huggingface/transformers/tree/master/examples/token-classification) (see MWE session for my code) on my custom dataset. The dataset I am using has some extra-long tokens (mainly URLs). When I obtained the predictions after running `run_ner.py`, I found that some tokens are missing. Concretely, in my experiment, 28970 - 28922 = 68 tokens are missing in the predictions - Here is the prediction statistics I obtained with `sklearn.metrics.classification_report`  - Here is the statistics of the dataset when I looked at the dataset  I further checked the threshold of token length (in characters) the `BertForTokenClassification` decided to ignore, it turns out that when the length of token is greater or equal to 28, it is ignored  I have searched the documentation to find if there is a parameter I could set to control this behavior but no luck. This made me suspect this might be a undocumented behavior worth noticing for the community. # Environment info - `transformers` version: 4.4.2 - Platform: Linux-5.4.0-70-generic-x86_64-with-debian-buster-sid - Python version: 3.7.0 - PyTorch version (GPU?): 1.7.1 (True) - Tensorflow version (GPU?): 2.2.0 (True) - Using GPU in script?: Yes - Using distributed or parallel set-up in script?: No # Who can help Model - albert, bert, xlm: @LysandreJik Library: - tokenizers: @LysandreJik - trainer: @sgugger # MWE The following is the main body of code and (anonymized) toy dataset used to show the behavior. After running the code, you will see the following, which shows that **model ignores many tokens when making predictions**.  To reproduce the result, simply copy and paste and dataset into `data.json` and put it in the same directory as the code. Then run the code ## Code ```python import json import itertools import numpy as np from datasets import load_dataset from transformers import AutoConfig, AutoTokenizer, AutoModelForTokenClassification from transformers import DataCollatorForTokenClassification, Trainer, TrainingArguments from collections import Counter from sklearn.metrics import classification_report model_name = "bert-base-uncased" label_list = ["O", "P", "Q"] label_to_id = {"O": 0, "P": 1, "Q": 2} tokenizer = AutoTokenizer.from_pretrained(model_name) def tokenize_and_align_labels(examples): padding = False text_column_name = "tokens" label_column_name = "tags" tokenized_inputs = tokenizer(examples[text_column_name], padding=padding, truncation=True, is_split_into_words=True) labels = list() for i, label in enumerate(examples[label_column_name]): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = list() for word_idx in word_ids: if word_idx is None: label_ids.append(-100) elif word_idx != previous_word_idx: label_ids.append(label_to_id[label[word_idx]]) else: label_ids.append(-100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs training_args = TrainingArguments(output_dir="output/ner", per_device_train_batch_size=2, per_device_eval_batch_size=2) config = AutoConfig.from_pretrained(model_name, num_labels=3) model = AutoModelForTokenClassification.from_pretrained(model_name, config=config) data_collator = DataCollatorForTokenClassification(tokenizer, pad_to_multiple_of=None) datasets = load_dataset("json", data_files={"test": "data.json"}) tokenized_datasets = datasets.map(tokenize_and_align_labels, batched=True, num_proc=None, load_from_cache_file=True) trainer = Trainer(model=model, args=training_args, tokenizer=tokenizer, train_dataset=None, eval_dataset=None, data_collator=data_collator) predictions, labels, metrics = trainer.predict(tokenized_datasets["test"]) predictions = np.argmax(predictions, axis=2) y_pred = [[label_list[p] for (p, l) in zip(prediction, label) if l != -100] for prediction, label in zip(predictions, labels)] y_true = [[label_list[l] for (p, l) in zip(prediction, label) if l != -100] for prediction, label in zip(predictions, labels)] y_pred_list = list(itertools.chain(*y_pred)) y_true_list = list(itertools.chain(*y_true)) print("predictions...") metric = classification_report(y_true=y_true_list, y_pred=y_pred_list, target_names=label_list, output_dict=False) print(metric) print("statistics of datasets...") tag_list = list() with open("data.json", "r") as fp: for line in fp.readlines(): tag_list.extend(json.loads(line)["tags"]) print(Counter(tag_list)) ``` ## Dataset ``` {"tokens": ["c", "HdI", "Op", "Ypgdm", "kssA", "gM", "azGFjtgAGDKJwahegGycdUsraeRvx", "Y", "z"], "tags": ["O", "O", "Q", "P", "P", "O", "P", "O", "P"]} {"tokens": ["HnRCVoBlyYjvWw", "JOQLTMrQSuPnB", "tj", "PjM", "dDMwaNYdwfgh", "kehjfOZa", "GG", "BGbWacckKOTSglSZpFsKssjnkqxuZzStYnFw", "Fu", "FPb", "yGvnkbGOAG", "WXxmmC", "KPD", "qgd", "wqGPK", "ulgmNz", "lDw", "P", "ee", "Rdrk", "mb", "rgZQnJGL", "YgOaUHjxik", "CzacME", "l", "RYFh", "C", "WscGhFK", "vcSldQFcbUdvg", "ijK", "MRD", "hnsPMqA", "tJn", "tkSuD", "sbJINmCL", "A", "XKtvHv", "NbrqNKuGA", "mF", "NDJf", "jcaodNHnUX", "bL", "bwIfI", "j", "mDPxyf", "Jp", "QvVBNmw", "W", "wBYzhr", "mzjxngTtvL", "y", "xZP", "ST", "KUcgzAUJswD", "vLir", "ZGUmN", "k", "kyoqdki", "YSGyV", "gfpy", "E"], "tags": ["O", "O", "O", "Q", "Q", "P", "P", "P", "O", "O", "O", "Q", "Q", "Q", "P", "P", "Q", "O", "O", "Q", "P", "Q", "Q", "P", "Q", "Q", "P", "Q", "O", "P", "Q", "Q", "Q", "P", "O", "P", "O", "O", "P", "P", "O", "Q", "Q", "Q", "Q", "Q", "Q", "P", "O", "Q", "O", "P", "P", "O", "Q", "O", "Q", "P", "Q", "Q", "Q"]} {"tokens": ["mHjjgX", "pv", "NlnLGhfJOjXQmdKBDoZbMJYbjMUPPpQVqLyj", "QwKjLiAVjZahrYjMsD", "ud", "fOYI", "wZCamIed", "V", "YNDFWpi", "n", "QLaToqF", "V", "P", "KG", "xk", "gHL", "to", "doYsYrgwC", "aP", "wijzV", "aR", "kZxvRsje", "eRSG", "moaijWxPGU", "IrFLx", "s"], "tags": ["Q", "O", "O", "P", "O", "O", "P", "O", "P", "P", "Q", "O", "O", "P", "O", "P", "P", "Q", "O", "Q", "Q", "O", "Q", "P", "Q", "O"]} {"tokens": ["Tq", "QnBu", "fOvqVK", "NlC", "JIBZwVk", "uL", "ceGY", "YQibS", "EI", "stIoTiWuwDCuLPBbZyVdxThcsjjTrXXbLZbPThg", "T", "plU", "yc", "pOOd", "bJKTECZM", "EcZHhimP", "rlxMVb", "wLj", "MAfob", "gT", "olvMEVNU", "JX", "uvhILBJSnxhrzBeEioHJuH", "j"], "tags": ["Q", "P", "O", "P", "O", "P", "O", "O", "P", "Q", "O", "P", "Q", "O", "P", "Q", "P", "O", "P", "O", "O", "Q", "O", "O"]} {"tokens": ["FqXSxwtS", "VzFPLNX", "NcTWHoHSv", "Rn", "uCj", "iodTKA", "cHLzTmFnR", "GK", "XqX", "T", "MIseQD", "hoY", "ws", "BFhME", "LDJJDlG", "nKkWW", "diEiWLHCSeAIIruHn", "MpYpfbTXQ", "QD", "ruHxEjF", "BTuuSVsCV", "IfsD", "GrM", "q", "f", "a", "F", "sMGEnatpNHMJBfinEzIzybvhPjKRnbd", "U", "v", "d", "n", "pBzXRQBdRcWphjmLVxmnBNtOJMceisw", "H", "GN", "S", "O", "n", "bICOosUadrGNlfAssbJOcpWJQLcCCKQq", "XliiPNU", "MKkF", "rN", "EBlhwak", "Lbato", "MjiPVtGMjR", "moD", "yTTxFb", "SW", "ossZZ", "gR", "sybT", "tq", "eKo", "mxQfeoi", "DZbe", "k", "uBvzS", "TFwxyIRx", "lXiv", "JrXcwr", "XdSfxLlDZR", "y"], "tags": ["O", "Q", "P", "Q", "P", "Q", "O", "O", "O", "P", "O", "Q", "O", "O", "O", "Q", "O", "Q", "P", "Q", "P", "Q", "O", "P", "O", "O", "O", "O", "Q", "O", "O", "P", "P", "O", "P", "Q", "Q", "Q", "O", "O", "Q", "O", "O", "O", "P", "P", "Q", "Q", "Q", "P", "P", "P", "O", "Q", "P", "O", "O", "O", "Q", "O", "Q", "Q"]} {"tokens": ["cMpSfp", "IOSq", "wizkn", "oEB", "Ux", "Gmord", "V", "RvkwzjJrkBOMsVEuoLvACZjFYDBrBUEnWkPuqnzY", "cgjplUK", "D", "Auj", "RGLDSSW", "uWRB", "y", "e", "JlTunC", "b", "GqYOtc", "CRXIL", "DOSndEb", "j", "C", "DVe", "Csp", "IjYeptL", "HWtK", "uDIKPJ", "E", "NIOshiUq", "KtD", "CGzcNNg", "Q", "e", "rzyYVX", "ncx", "yhGACxogyADau", "evP", "qbYK", "oigXl", "P"], "tags": ["O", "P", "Q", "P", "P", "P", "P", "O", "O", "O", "O", "Q", "O", "O", "P", "O", "Q", "O", "Q", "P", "O", "O", "P", "O", "O", "P", "Q", "O", "Q", "Q", "P", "O", "O", "Q", "Q", "Q", "O", "O", "Q", "P"]} {"tokens": ["hyAcXV", "RW", "iO", "e", "eTePUCB", "o", "grVNB", "M", "LX", "CWkrrMHaxHFNu", "JXDPpxS", "g", "iWciIRc", "RyZjwZU", "u", "plcdCecp", "LOWbDCO", "n", "CzAULxOi", "LMPyDWe", "N", "SfZysOoD", "gusPKRp", "R", "DldmjFoHIY", "wXOrbAq", "G", "zuDJFvSXWhgTNTgwoE", "OLFazlp", "B", "cWEZHEBOWss", "PMzyvWG", "q", "gHJEujHPPwHXneHYv", "XZalipR", "T", "JFBaJlhHuFDWLM", "IuutEum", "A", "TNVmvRlQUuD", "HmZsgzC", "X", "szBRrWzUoLAOGsfBohYPDIVLr", "EijeeQw", "R", "xUsvfNrjMNxBCYhg", "cEbAIsG", "F", "MhWZlEDNGBOXJaQuesNfXdydiwxaUST", "URImwm", "yFOKiYpM", "x", "eao", "O", "DsP", "g", "fs", "a", "Mo", "p", "LgA", "q", "r", "siDKIEPOZkyI", "L", "v", "EQTSNTOxrd", "epW", "K", "ziV", "p", "A", "eWra", "U", "hBGCLsYLllrR", "ZfcJg", "g", "c", "P", "oUJUef", "aqWS", "i", "AHs", "F", "St", "T", "TbDMAs", "fIN", "kSup", "c", "BgKBoKtxXFBf", "t", "b", "o", "hKRhT", "U", "N", "m", "pZfQLs", "fLn", "T", "Imx", "H", "nG", "H", "ocVNIs", "kgT", "nZedF", "J", "U", "tcmU", "aSYlJG", "vpA", "Xf", "V", "FdrvDP", "OHcc", "UntG", "YFyy", "NAHr", "m", "opzPb", "icgUoNo", "fWKI", "cI", "ybB", "XQbwSa", "pEDMAJZ", "Ajnqq", "yOLKPAE", "cEACM", "dT", "pau", "OjAT", "UCmZNFHQ", "Vb", "ECpULl", "KaJsF", "GEmg", "o", "kjEc", "gKkvZvj", "gSM", "HCaPbni", "p", "OhYz", "tdTKv", "Oeuresj", "Raqjcr", "Z", "sKCQiSjGWyqvJASuB", "HNiFpp", "i", "HHwkWlygESwwMBL", "GEpb", "R", "SjjRzuPXZEcgNqfgbJGUgAIiMi", "Iwwp", "P", "s", "FGqxybNIPJzVwhL", "G", "nfKb", "qEnHOJl", "kGT", "SLXDAvfwR", "n", "WEso", "puuYI", "nPdUJOJ", "NAVRZp", "e", "gMhFUtVnSZaZAPeBi", "NNQGMQ", "R", "mzIOmvrRVjzVBfS", "Zmtj", "B", "uvwYbcmunUbgoJabBLymVNXCyE", "wpmI", "Z", "fqOjYrLSDtKXbqcBCE", "uDXTCqJ", "L", "L", "IrSpLTyHmyZqJg", "P", "yaWdOfA", "xLMytL", "r", "uJtGGjajlrSAKJXnv", "INXREu", "t", "h", "ndrFkQBGoRkiIHlCp", "f", "NjJxWLK", "kirQOT"], "tags": ["P", "Q", "P", "P", "P", "P", "O", "Q", "P", "O", "O", "P", "P", "P", "P", "O", "Q", "O", "P", "O", "P", "Q", "Q", "O", "Q", "O", "Q", "P", "Q", "P", "Q", "O", "O", "O", "O", "O", "Q", "O", "P", "O", "O", "Q", "O", "O", "P", "Q", "P", "P", "Q", "Q", "P", "O", "O", "O", "Q", "Q", "O", "Q", "Q", "O", "O", "P", "Q", "O", "Q", "P", "Q", "Q", "P", "Q", "Q", "O", "O", "Q", "P", "Q", "O", "O", "Q", "Q", "O", "P", "O", "P", "O", "O", "O", "O", "P", "P", "P", "P", "O", "P", "Q", "O", "Q", "Q", "Q", "P", "P", "O", "O", "P", "O", "Q", "Q", "Q", "P", "O", "P", "O", "O", "Q", "O", "O", "O", "O", "Q", "O", "Q", "O", "O", "P", "Q", "P", "Q", "O", "O", "Q", "O", "P", "P", "Q", "Q", "P", "P", "Q", "Q", "O", "P", "P", "O", "Q", "Q", "Q", "Q", "Q", "Q", "Q", "Q", "Q", "O", "Q", "Q", "Q", "Q", "P", "Q", "P", "P", "O", "Q", "P", "P", "Q", "P", "P", "Q", "Q", "P", "P", "P", "O", "O", "O", "P", "P", "P", "P", "O", "P", "P", "O", "Q", "O", "O", "O", "O", "P", "P", "P", "Q", "O", "O", "Q", "P", "O"]} {"tokens": ["bJl", "cAsqzlymeBfFnO", "hMigNgVJ", "vD", "esTrrnMJBamvkOvjaLWARywfQiFwRwM", "Is", "Hhp", "lbn", "vnf", "wuWVkO", "Aw", "PnUPcoI", "AJE", "xnixjKF", "uRzEGyaDrRkjLd", "Qqel", "sLc", "ukyjdp", "Cyqn", "o", "fPBJSrC", "FqtzKpK", "Dw", "Vkl", "J", "HnnTKNsVP", "DPiK", "a", "ZUXAQ", "hRIbLv", "WWyWOq", "iyzEziDrS", "th", "VHyIL", "h", "eXFLCs", "xQ", "XandWFa", "W", "hncKOj", "KLdkXOrRyE", "b", "Oy", "mbbElWnm", "NHrDqsE", "nBtVcsPWY", "BdPX", "RX", "VHefpxJdxK", "a", "MRMAY", "iwQs", "hrZETMe", "lL", "E", "dQuLPWnub", "D"], "tags": ["O", "O", "Q", "O", "P", "O", "O", "Q", "P", "P", "Q", "O", "O", "O", "O", "P", "O", "O", "P", "Q", "Q", "P", "O", "O", "P", "O", "Q", "P", "P", "P", "P", "O", "Q", "P", "P", "P", "Q", "O", "O", "Q", "Q", "O", "O", "P", "Q", "O", "Q", "O", "O", "Q", "Q", "Q", "O", "Q", "P", "P", "Q"]} {"tokens": ["PjUgTUcniaaguyQczGZDlOAxudGEQUpxxTsr", "Rf", "zGC", "cdfOxLl", "vLwEb", "mBbiLKn", "EvhIfT", "KED", "tvc", "deUac", "lHaBcZ", "mWH", "W", "yw", "rjfX", "tH", "eDdqDHyJ", "MaUJzNVJyW", "AhQlsI", "i", "Eilm", "b", "jBZUzcA", "nJsSQzOcjtBoa", "fvR", "Vih", "HGnLEjG", "frT", "eJaEN", "sypBkIMw", "H", "HxwFFg", "QOjNHfklD", "DZ", "KYOWZ", "m", "MogbCy", "Wu", "QhzMXWx", "O", "sMIUQR", "YioflBLuit", "m", "Hv", "gKMaoRXH", "DwsM", "wOeAUwkSlbV", "zTOOv", "GlSZly", "TEz", "t", "fllFm", "VqEcjKZ", "d", "Qj", "jiZQwLFNNccV", "LL", "m", "PRxEnWeJvUhmtuFzZb", "GWYCJs", "RljauRaV", "J"], "tags": ["O", "O", "P", "O", "O", "Q", "P", "O", "P", "O", "Q", "Q", "Q", "P", "P", "Q", "P", "O", "O", "Q", "O", "O", "P", "Q", "P", "P", "P", "O", "P", "P", "O", "O", "Q", "O", "Q", "P", "O", "O", "P", "Q", "Q", "P", "Q", "P", "O", "Q", "O", "P", "O", "O", "O", "Q", "O", "P", "P", "P", "O", "Q", "Q", "P", "Q", "P"]} {"tokens": ["abX", "gDeEwdSlQFUXFWiWRIie", "emdFvnyR", "GZ", "jSm", "fagyuPSYnS", "qmosLR", "wL", "nLRINCl", "zUSyOZU", "AIO", "eLecAu", "ijzZSCWa", "r", "cmgXSrlFaoD", "ayF", "qfFQPi", "yxrvzDIe", "t", "dIK", "zweNjVRuf", "fmU", "vkIAaH", "hObrzx", "QlXqVW", "uZHCwn", "PeQWgisYg", "WV", "erxvr", "h", "oJivba", "gj", "ucRLQFQ", "z", "PCleg", "y", "Nkf", "oizdKKJ", "fzqexnO", "LSXd", "SdEfZcM", "uuWgbC", "rKKkDwPqWc", "b", "bw", "HNRCHCxUHbmV", "Wh", "qrsekbtnTJitvsENHpARJgKThtgEmXbv", "i"], "tags": ["Q", "P", "P", "P", "O", "P", "Q", "Q", "Q", "O", "Q", "Q", "O", "O", "O", "Q", "Q", "P", "O", "Q", "O", "P", "Q", "Q", "O", "O", "P", "O", "O", "O", "Q", "O", "O", "P", "Q", "O", "Q", "Q", "P", "P", "O", "O", "O", "P", "Q", "O", "O", "P", "P"]} ```

https://api.github.com/repos/huggingface/transformers/issues/11035

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11035/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11035/comments

https://api.github.com/repos/huggingface/transformers/issues/11035/events

https://github.com/huggingface/transformers/issues/11035

MDU6SXNzdWU4NDg5NzY0Njg=

404 Client Error: Not Found for url: https://huggingface.co/%5CHuggingface-Sentiment-Pipeline/resolve/main/config.json

open

NONE

I'm trying to use the hugging face sentimet-analysis pipeline. I've downloaded the pipeline using save.pretrained(model). And trying to load the pipeline with the help of below code ``` from transformers import pipeline model = '\Huggingface-Sentiment-Pipeline' classifier = pipeline(task='sentiment-analysis', model=model, tokenizer=model, from_pt=True) ``` The Huggingface-Sentiment-Pipeline contains 6 files. I'm mentioning below ``` -> Huggingface-Sentiment-Pipeline -> config.json -> modelcard.json -> pytorch_model.bin -> special_tokens_map.json -> tokenizer_config.json -> vocab.txt ``` The error I'm getting is given below ``` 404 Client Error: Not Found for url: https://huggingface.co/%5CHuggingface-Sentiment-Pipeline/resolve/main/config.json --------------------------------------------------------------------------- HTTPError Traceback (most recent call last) ~\sentiment_pipeline\lib\site-packages\transformers\configuration_utils.py in get_config_dict(cls, pretrained_model_name_or_path, **kwargs) 423 local_files_only=local_files_only, --> 424 use_auth_token=use_auth_token, 425 ) ~\sentiment_pipeline\lib\site-packages\transformers\file_utils.py in cached_path(url_or_filename, cache_dir, force_download, proxies, resume_download, user_agent, extract_compressed_file, force_extract, use_auth_token, local_files_only) 1085 use_auth_token=use_auth_token, -> 1086 local_files_only=local_files_only, 1087 ) ~\sentiment_pipeline\lib\site-packages\transformers\file_utils.py in get_from_cache(url, cache_dir, force_download, proxies, etag_timeout, resume_download, user_agent, use_auth_token, local_files_only) 1215 r = requests.head(url, headers=headers, allow_redirects=False, proxies=proxies, timeout=etag_timeout) -> 1216 r.raise_for_status() 1217 etag = r.headers.get("X-Linked-Etag") or r.headers.get("ETag") ~\sentiment_pipeline\lib\site-packages\requests\models.py in raise_for_status(self) 942 if http_error_msg: --> 943 raise HTTPError(http_error_msg, response=self) 944 HTTPError: 404 Client Error: Not Found for url: https://huggingface.co/%5CHuggingface-Sentiment-Pipeline/resolve/main/config.json During handling of the above exception, another exception occurred: OSError Traceback (most recent call last) <ipython-input-7-5074b39a82b6> in <module> ----> 1 classifier = pipeline(task='sentiment-analysis', model=model, tokenizer=model, from_pt=True) ~\sentiment_pipeline\lib\site-packages\transformers\pipelines\__init__.py in pipeline(task, model, config, tokenizer, framework, revision, use_fast, **kwargs) 338 model = get_default_model(targeted_task, framework, task_options) 339 --> 340 framework = framework or get_framework(model) 341 342 task_class, model_class = targeted_task["impl"], targeted_task[framework] ~\sentiment_pipeline\lib\site-packages\transformers\pipelines\base.py in get_framework(model, revision) 64 if isinstance(model, str): 65 if is_torch_available() and not is_tf_available(): ---> 66 model = AutoModel.from_pretrained(model, revision=revision) 67 elif is_tf_available() and not is_torch_available(): 68 model = TFAutoModel.from_pretrained(model, revision=revision) ~\sentiment_pipeline\lib\site-packages\transformers\models\auto\modeling_auto.py in from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs) 768 if not isinstance(config, PretrainedConfig): 769 config, kwargs = AutoConfig.from_pretrained( --> 770 pretrained_model_name_or_path, return_unused_kwargs=True, **kwargs 771 ) 772 ~\sentiment_pipeline\lib\site-packages\transformers\models\auto\configuration_auto.py in from_pretrained(cls, pretrained_model_name_or_path, **kwargs) 366 {'foo': False} 367 """ --> 368 config_dict, _ = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs) 369 370 if "model_type" in config_dict: ~\sentiment_pipeline\lib\site-packages\transformers\configuration_utils.py in get_config_dict(cls, pretrained_model_name_or_path, **kwargs) 434 f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a {CONFIG_NAME} file\n\n" 435 ) --> 436 raise EnvironmentError(msg) 437 438 except json.JSONDecodeError: OSError: Can't load config for '\Huggingface-Sentiment-Pipeline'. Make sure that: - '\Huggingface-Sentiment-Pipeline' is a correct model identifier listed on 'https://huggingface.co/models' - or '\Huggingface-Sentiment-Pipeline' is the correct path to a directory containing a config.json file ```

https://api.github.com/repos/huggingface/transformers/issues/11034

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11034/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11034/comments

https://api.github.com/repos/huggingface/transformers/issues/11034/events

https://github.com/huggingface/transformers/issues/11034

MDU6SXNzdWU4NDg5MzkzMTA=

GPT-2 example is broken?

open

NONE

## Environment info <!-- You can run the command `transformers-cli env` and copy-and-paste its output below. Don't forget to fill out the missing fields in that output! --> - `transformers` version: I have had this issue with both 4.3.0 and 4.4.2 (and probably other versions as well) - Python version: 3.7.6 - PyTorch version (GPU?): 1.7.0 - Using GPU in script?: No, I just tested it on the CPU, but it would probably also happen on the GPU - Using distributed or parallel set-up in script?: No ### Who can help <!-- Your issue will be replied to more quickly if you can figure out the right person to tag with @ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**. Please tag fewer than 3 people. --> - gpt2: @patrickvonplaten, @LysandreJik - Documentation: @sgugger ## Information Model I am using (Bert, XLNet ...): gpt2 The problem arises when using: * [ x] the official example scripts: (give details below) * [ ] my own modified scripts: (give details below) ## To reproduce Hello, I am trying to run this example here: https://huggingface.co/transformers/task_summary.html#causal-language-modeling. When I run that code, exactly the same as it is on that page, I get strange/very bad results. Even when I change the input text, it still gives weird results (e.g., predicting empty spaces or strange characters). I also asked my coworker to try it on her computer, and she also got strange results. I am planning to fine-tune GPT-2 for a different purpose later, but was a bit concerned because I couldn't even get this simple example demo to work. Thanks for your help! Steps to reproduce the behavior: 1. Just run the exact example code that I linked above <!-- If you have code snippets, error messages, stack traces please provide them here as well. Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting Do not use screenshots, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.-->

https://api.github.com/repos/huggingface/transformers/issues/11033

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11033/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11033/comments

https://api.github.com/repos/huggingface/transformers/issues/11033/events

https://github.com/huggingface/transformers/issues/11033

MDU6SXNzdWU4NDg5MzY1NzM=

RuntimeError: The size of tensor a (1024) must match the size of tensor b (1025) at non-singleton dimension 3

open

NONE

Here I try to use gpt2 to generation the text under the prompt text. I have several datasets, some of them, such as AG_NEWS and POP_NEWS, are made of short sentences while when I use YAHOO_NEWS, consisting of longer sentences, the error came out. Anything to modify for my codes? Thanks. ``` from transformers import ( CTRLLMHeadModel, CTRLTokenizer, GPT2LMHeadModel, GPT2Tokenizer, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer, TransfoXLLMHeadModel, TransfoXLTokenizer, XLMTokenizer, XLMWithLMHeadModel, XLNetLMHeadModel, XLNetTokenizer, ) class generation(): def __init__(self, model_name='gpt2',num_return_sequences=1): self.model_name = model_name self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") self.MODEL_CLASSES = { "gpt2": (GPT2LMHeadModel, GPT2Tokenizer), "ctrl": (CTRLLMHeadModel, CTRLTokenizer), "openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer), "xlnet-base-cased": (XLNetLMHeadModel, XLNetTokenizer), "transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer), "xlm": (XLMWithLMHeadModel, XLMTokenizer), } self.length = 100 self.k = 0 self.p = 0.9 self.num_return_sequences = num_return_sequences self.model_class, self.tokenizer_class = self.MODEL_CLASSES[self.model_name] self.tokenizer = self.tokenizer_class.from_pretrained(self.model_name) self.model = self.model_class.from_pretrained(self.model_name) self.model.to(self.device) if self.model_name == "xlnet-base-cased": self.p=0.95 self.k=60 self.length = self.adjust_length_to_model(self.length, max_sequence_length=self.model.config.max_position_embeddings) if self.model_name == 'ctrl': self.temperature = 0.3 self.repetition_penalty = 1.2 else: self.temperature = 1.0 self.repetition_penalty = 1.0 def adjust_length_to_model(self, length, max_sequence_length): if length < 0 and max_sequence_length > 0: length = max_sequence_length elif 0 < max_sequence_length < length: length = max_sequence_length # No generation bigger than model size elif length < 0: length = 1000 # avoid infinite loop return length def ctrl_label2prefix(self, label): # https://github.com/salesforce/ctrl/blob/master/control_codes.py ''' 'Pregnancy Christianity Explain Fitness Saving Ask Ass Joke Questions Thoughts Retail Feminism Writing Atheism Netflix Computing Opinion Alone Funny Gaming Human India Joker Diet Legal Norman Tip Weight Movies Running Science Horror Confession Finance Politics Scary Support Technologies Teenage Event Learned Notion Wikipedia Books Extract Confessions Conspiracy Links Narcissus Relationship Relationships Reviews News Translation multilingual' ''' return 'News' if label in ('Sci/Tech', 'tech'): return 'Technologies' elif label in ('politics'): return 'Politics' elif label in ('Sports', 'sport'): return 'Fitness' else: return 'News' def augment(self, prompt_text): if self.model_name == 'ctrl': prefix = 'News ' else: prefix = '' encoded_prompt = self.tokenizer.encode(prefix + prompt_text, add_special_tokens=False, return_tensors="pt") encoded_prompt = encoded_prompt.to(self.device) if encoded_prompt.size()[-1] == 0: input_ids = None else: input_ids = encoded_prompt output_sequences = self.model.generate( input_ids=input_ids, max_length= self.length + len(encoded_prompt[0]), temperature=self.temperature, top_k=self.k, top_p=self.p, repetition_penalty=self.repetition_penalty, do_sample=True, num_return_sequences=self.num_return_sequences, ) # Decode text text_generated = self.tokenizer.decode(output_sequences[0][len(encoded_prompt[0]):], clean_up_tokenization_spaces=True) return text_generated # unit test ''' augmentor = generation('gpt2') prompt_text = "Microsoft has said it will replace more than 14 million power cables for its Xbox consoles due to safety concerns." prompt_text = "Versace art portfolio up for sale The art collection of murdered fashion designer Gianni Versace could fetch \ up to £9m ($17m) when it is auctioned in New York and \ London later this year. <eod> </s> <eos>" augmentor.augment(prompt_text) ''' ``` ERROR information: > File "baseline_classifier.py", line 45, in run_benchmark > ds.df_train['content_aug'] = ds.df_train['content'].map(lambda x: augmentor.augment(x)) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/pandas/core/series.py", line 3382, in map > arg, na_action=na_action) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/pandas/core/base.py", line 1218, in _map_values > new_values = map_f(values, mapper) > File "pandas/_libs/lib.pyx", line 2217, in pandas._libs.lib.map_infer > File "baseline_classifier.py", line 45, in <lambda> > ds.df_train['content_aug'] = ds.df_train['content'].map(lambda x: augmentor.augment(x)) > File "/workspace/user-workspace/topic_classification_augmentation/aug_generation.py", line 110, in augment > num_return_sequences=self.num_return_sequences, > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line 15, in decorate_context > return func(*args, **kwargs) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/generation_utils.py", line 1019, in generate > **model_kwargs, > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/generation_utils.py", line 1486, in sample > output_hidden_states=output_hidden_states, > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__ > result = self.forward(*input, **kwargs) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/models/gpt2/modeling_gpt2.py", line 917, in forward > return_dict=return_dict, > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__ > result = self.forward(*input, **kwargs) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/models/gpt2/modeling_gpt2.py", line 760, in forward > output_attentions=output_attentions, > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__ > result = self.forward(*input, **kwargs) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/models/gpt2/modeling_gpt2.py", line 296, in forward > output_attentions=output_attentions, > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__ > result = self.forward(*input, **kwargs) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/models/gpt2/modeling_gpt2.py", line 241, in forward > attn_outputs = self._attn(query, key, value, attention_mask, head_mask, output_attentions) > File "/workspace/.conda/miniconda/lib/python3.7/site-packages/transformers/models/gpt2/modeling_gpt2.py", line 176, in _attn > w = torch.where(mask.bool(), w, self.masked_bias.to(w.dtype)) > RuntimeError: The size of tensor a (1024) must match the size of tensor b (1025) at non-singleton dimension 3

https://api.github.com/repos/huggingface/transformers/issues/11032

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11032/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11032/comments

https://api.github.com/repos/huggingface/transformers/issues/11032/events

https://github.com/huggingface/transformers/issues/11032

MDU6SXNzdWU4NDg5MjE5ODI=

How to get masked word prediction for other languages

open

NONE

Hello, I trying to get masked words predictions for languages except English with Roberta or XLM Roberta. ``` from transformers import pipeline nlp = pipeline("fill-mask", model="roberta-base") template = f"That woman is {nlp.tokenizer.mask_token}." output = nlp(template) nlp4 = pipeline("fill-mask", model="roberta-base") nlp4(f"Женщины работают {nlp4.tokenizer.mask_token}.") ``` The output for English example is quite good, while for Russian one does not make sense at all: `[{'sequence': 'Женщины работаюта.', 'score': 0.2504434883594513, 'token': 26161, 'token_str': 'а'}, {'sequence': 'Женщины работають.', 'score': 0.24665573239326477, 'token': 47015, 'token_str': 'ь'}, {'sequence': 'Женщины работаюты.', 'score': 0.1454654186964035, 'token': 46800, 'token_str': 'ы'}, {'sequence': 'Женщины работаюте.', 'score': 0.07919821888208389, 'token': 25482, 'token_str': 'е'}, {'sequence': 'Женщины работаюти.', 'score': 0.07401203364133835, 'token': 35328, 'token_str': 'и'}]` Neither "roberta-base" nor "xlm-roberta-base" work for Russian language example. Maybe I am doing it wrong, but how would one use masked word prediction for other languages? Thanks!

https://api.github.com/repos/huggingface/transformers/issues/11030

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11030/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11030/comments

https://api.github.com/repos/huggingface/transformers/issues/11030/events

https://github.com/huggingface/transformers/issues/11030

MDU6SXNzdWU4NDg4MjM3MDI=

pipeline.from_pretrained

open

CONTRIBUTOR

# 🚀 Feature request Nearly everyone who is using the transformers library is aware of the `from_pretrained()` and `save_pretrained()` concept. The [Pipeline class](https://huggingface.co/transformers/main_classes/pipelines.html#parent-class-pipeline) is currently only providing the `save_pretrained()` method which can cause confusion for some users as saving and loading of the pipeline needs to be done like this: ``` from transformers import pipeline TASK = 'something' DIRECTORY='something' classifier = pipeline(TASK) classifier.save_pretrained(DIRECTORY) c2 = pipeline(task = TASK, model=DIRECTORY, tokenizer=DIRECTORY) ``` This is probably not that obvious for people who just read the documentation and not the code. I suggest implementing the `from_pretrained()` method for the pipelines to make the library even more intuitive. ## Your contribution There is actually not much to do since the tokenizer and model are loaded by the corresponding Auto classes. The only information that is missing to construct the saved pipeline with `pipeline.from_pretraind(PATH)` is the task identifier. This information should be stored in a new separate file called `pipeline_config.json`. I can provide a PR if you think this is a useful enhancement for the library @LysandreJik

https://api.github.com/repos/huggingface/transformers/issues/11029

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11029/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11029/comments

https://api.github.com/repos/huggingface/transformers/issues/11029/events

https://github.com/huggingface/transformers/pull/11029

MDExOlB1bGxSZXF1ZXN0NjA3Njc4Nzg3

Documentation about loading a fast tokenizer within Transformers

open

MEMBER

This PR does two things: - Allows to load a fast tokenizer from an instantiated `tokenizers` object - Adds a page to document how to use these tokenizers within `transformers` See [here](https://190138-155220641-gh.circle-artifacts.com/0/docs/_build/html/fast_tokenizers.html) for the generated docs

https://api.github.com/repos/huggingface/transformers/issues/11028

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11028/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11028/comments

https://api.github.com/repos/huggingface/transformers/issues/11028/events

https://github.com/huggingface/transformers/issues/11028

MDU6SXNzdWU4NDg3NjkwNjE=

Fine Tune GPT-NEO 2.7B

open

NONE

Hello to everyone, is there a script to fine tune this new model? Thanks

https://api.github.com/repos/huggingface/transformers/issues/11027

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11027/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11027/comments

https://api.github.com/repos/huggingface/transformers/issues/11027/events

https://github.com/huggingface/transformers/pull/11027

MDExOlB1bGxSZXF1ZXN0NjA3NjUzMTAy

[WIP] Refactor AutoModel classes and add Flax Auto classes

open

MEMBER

# What does this PR do? This PR refactors the logic behind all the Auto model classes in one function that automatically builds those classes from a template. In passing, it uses this new function to build the auto classes for FLAX (at least the ones that have at least one model implemented).

https://api.github.com/repos/huggingface/transformers/issues/11026

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11026/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11026/comments

https://api.github.com/repos/huggingface/transformers/issues/11026/events

https://github.com/huggingface/transformers/pull/11026

MDExOlB1bGxSZXF1ZXN0NjA3NjQyMjM1

Add `examples/language_modeling/run_clm_no_trainer.py`

open

CONTRIBUTOR

# What does this PR do? <!-- Congratulations! You've made it this far! You're not quite done yet though. Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution. Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change. Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost. --> <!-- Remove if not applicable --> This PR adds an example of finetuning a Causal Language Model (without using Trainer) to show the functionalities of the new accelerate library. ## Who can review? @sgugger <!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**. Please tag fewer than 3 people. Models: - albert, bert, xlm: @LysandreJik - blenderbot, bart, marian, pegasus, encoderdecoder, t5: @patrickvonplaten, @patil-suraj - longformer, reformer, transfoxl, xlnet: @patrickvonplaten - fsmt: @stas00 - funnel: @sgugger - gpt2: @patrickvonplaten, @LysandreJik - rag: @patrickvonplaten, @lhoestq - tensorflow: @LysandreJik Library: - benchmarks: @patrickvonplaten - deepspeed: @stas00 - ray/raytune: @richardliaw, @amogkam - text generation: @patrickvonplaten - tokenizers: @n1t0, @LysandreJik - trainer: @sgugger - pipelines: @LysandreJik Documentation: @sgugger HF projects: - datasets: [different repo](https://github.com/huggingface/datasets) - rust tokenizers: [different repo](https://github.com/huggingface/tokenizers) Examples: - maintained examples (not research project or legacy): @sgugger, @patil-suraj - research_projects/bert-loses-patience: @JetRunner - research_projects/distillation: @VictorSanh -->

https://api.github.com/repos/huggingface/transformers/issues/11024

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11024/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11024/comments

https://api.github.com/repos/huggingface/transformers/issues/11024/events

https://github.com/huggingface/transformers/pull/11024

MDExOlB1bGxSZXF1ZXN0NjA3NjEwNjQ1

Add a script to check inits are consistent

open

MEMBER

# What does this PR do? Most inits in the project define the same objects twice (once in `_import_structure` and once in TYPE_CHECKING) to have a fast import so objects are only grabbed when actually needed. The problem is that those two halves have a tendency to diverge as contributors do not always pay attention to have them exactly match. Well not anymore. Introducing `utils/check_inits.py`. This script will parse all the inits with two halves and return an error with a delightful and informative message telling the user what they did wrong. It is enforced in the CI and added to `make fixup` and `make quality`.

https://api.github.com/repos/huggingface/transformers/issues/11023

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11023/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11023/comments

https://api.github.com/repos/huggingface/transformers/issues/11023/events

https://github.com/huggingface/transformers/issues/11023

MDU6SXNzdWU4NDg2ODAxNjg=

Strange ValueError with GPT-2

open

NONE

## Environment info <!-- You can run the command `transformers-cli env` and copy-and-paste its output below. Don't forget to fill out the missing fields in that output! --> - `transformers` version: 4.4.2 - Platform: macOS-10.15.7-x86_64-i386-64bit - Python version: 3.8.6 - PyTorch version (GPU?): 1.7.1 (False) - Tensorflow version (GPU?): 2.4.0 (False) - Using GPU in script?: No. - Using distributed or parallel set-up in script?: No. ### Who can help Models: - gpt2: @patrickvonplaten, @LysandreJik ## Information Model I am using (Bert, XLNet ...): GPT-2 The problem arises when using: * [ ] the official example scripts: (give details below) * [X] my own modified scripts: (give details below) The tasks I am working on is: * [ ] an official GLUE/SQUaD task: (give the name) * [X] my own task or dataset: (give details below) ## To reproduce Create and train GPT-2 with 8 heads and batch size 32. I get an error message when training GPT-2. When number of heads and batch size are the same it works. Looks like a shape check is wronk. See error message below. ``` File "train.py", line 66, in <module> history = model.fit( File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py", line 1100, in fit tmp_logs = self.train_function(iterator) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 828, in __call__ result = self._call(*args, **kwds) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 871, in _call self._initialize(args, kwds, add_initializers_to=initializers) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 725, in _initialize self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 2969, in _get_concrete_function_internal_garbage_collected graph_function, _ = self._maybe_define_function(args, kwargs) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 3361, in _maybe_define_function graph_function = self._create_graph_function(args, kwargs) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/function.py", line 3196, in _create_graph_function func_graph_module.func_graph_from_py_func( File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py", line 990, in func_graph_from_py_func func_outputs = python_func(*func_args, **func_kwargs) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/eager/def_function.py", line 634, in wrapped_fn out = weak_wrapped_fn().__wrapped__(*args, **kwds) File "/Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py", line 977, in wrapper raise e.ag_error_metadata.to_exception(e) ValueError: in user code: /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py:805 train_function * return step_function(self, iterator) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py:795 step_function ** outputs = model.distribute_strategy.run(run_step, args=(data,)) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/distribute/distribute_lib.py:1259 run return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/distribute/distribute_lib.py:2730 call_for_each_replica return self._call_for_each_replica(fn, args, kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/distribute/distribute_lib.py:3417 _call_for_each_replica return fn(*args, **kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py:788 run_step ** outputs = model.train_step(data) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/engine/training.py:758 train_step self.compiled_metrics.update_state(y, y_pred, sample_weight) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/engine/compile_utils.py:408 update_state metric_obj.update_state(y_t, y_p, sample_weight=mask) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/utils/metrics_utils.py:90 decorated update_op = update_state_fn(*args, **kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/metrics.py:177 update_state_fn return ag_update_state(*args, **kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/metrics.py:618 update_state ** matches = ag_fn(y_true, y_pred, **self._fn_kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py:201 wrapper return target(*args, **kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/keras/metrics.py:3315 sparse_categorical_accuracy return math_ops.cast(math_ops.equal(y_true, y_pred), K.floatx()) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/util/dispatch.py:201 wrapper return target(*args, **kwargs) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/ops/math_ops.py:1679 equal return gen_math_ops.equal(x, y, name=name) /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/ops/gen_math_ops.py:3177 equal _, _, _op, _outputs = _op_def_library._apply_op_helper( /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/op_def_library.py:748 _apply_op_helper op = g._create_op_internal(op_type_name, inputs, dtypes=None, /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/func_graph.py:590 _create_op_internal return super(FuncGraph, self)._create_op_internal( # pylint: disable=protected-access /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/ops.py:3528 _create_op_internal ret = Operation( /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/ops.py:2015 __init__ self._c_op = _create_c_op(self._graph, node_def, inputs, /Users/tristanbehrens/Development/python-venvs/tf2-p38/lib/python3.8/site-packages/tensorflow/python/framework/ops.py:1856 _create_c_op raise ValueError(str(e)) ValueError: Dimensions must be equal, but are 32 and 8 for '{{node Equal_1}} = Equal[T=DT_FLOAT, incompatible_shape_error=true](Cast_6, Cast_7)' with input shapes: [32,301], [2,32,8,301]. ```

https://api.github.com/repos/huggingface/transformers/issues/11022

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11022/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11022/comments

https://api.github.com/repos/huggingface/transformers/issues/11022/events

https://github.com/huggingface/transformers/issues/11022

MDU6SXNzdWU4NDg2NzkxNzQ=

cannot import name 'AutoModelForSequenceClassification' from 'transformers'

open

NONE

``` from transformers import pipeline classifier = pipeline('sentiment-analysis') #This code will download the pipeline classifier('We are very happy to show you the 🤗 Transformers library.') classifier.save_pretrained('/some/directory') ``` I'm trying to save the model and trying to perform the sentiment-analysis operation offline ``` from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline t = AutoTokenizer.from_pretrained('/some/directory') m = AutoModelForSequenceClassification.from_pretrained('/some/directory') c2 = pipeline(task = 'sentiment-analysis', model=m, tokenizer=t) I'm facing import error for jupyter notebook as given below `**cannot import name 'AutoModelForSequenceClassification' from 'transformers'**` ```

https://api.github.com/repos/huggingface/transformers/issues/11021

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11021/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11021/comments

https://api.github.com/repos/huggingface/transformers/issues/11021/events

https://github.com/huggingface/transformers/issues/11021

MDU6SXNzdWU4NDg2NTE0MzQ=

Module Not found: datasets_modules.datasets.output

open

NONE

## Environment info - `transformers` version: 4.5.0.dev0 - Platform: Linux-3.10.0-1160.15.2.el7.x86_64-x86_64-with-glibc2.10 - Python version: 3.8.5 - PyTorch version (GPU?): 1.8.1 (False) - Tensorflow version (GPU?): not installed (NA) - Using GPU in script?: not sure - Using distributed or parallel set-up in script?: <fill in> ? ### Who can help @patil-suraj ## Information Model I am using (Bert, XLNet ...): BART seq2seq The problem arises when using: * [x] the official example scripts: (give details below) * [ ] my own modified scripts: (give details below) The tasks I am working on is: * [ ] an official GLUE/SQUaD task: (give the name) * [x] my own task or dataset: (give details below) ## To reproduce Steps to reproduce the behavior: 1. "Install from source" method 2. ran this command, where "data/output.jsonl" is my dataset: ``` python examples/seq2seq/run_translation.py \ --model_name_or_path t5-small \ --do_train \ --do_eval \ --source_lang en \ --target_lang de \ --source_prefix "Translate English to Logical Forms: " \ --dataset_name data/output.jsonl \ --output_dir /tmp/tst-translation \ --per_device_train_batch_size=4 \ --per_device_eval_batch_size=4 \ --overwrite_output_dir \ --predict_with_generate ``` 3. Got the following error: ``` ModuleNotFoundError: No module named 'datasets_modules.datasets.output' ``` At first it told me that "datasets" was not installed, so I did ```pip install datasets``` and that worked fine. Then I got this error and haven't been able to figure out what it means or how to fix it.

https://api.github.com/repos/huggingface/transformers/issues/11020

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11020/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11020/comments

https://api.github.com/repos/huggingface/transformers/issues/11020/events

https://github.com/huggingface/transformers/issues/11020

MDU6SXNzdWU4NDg1NjY2NjY=

Trainer API crashes GPUs

open

NONE

## Environment info - `transformers` version: 4.5.0.dev0 - Platform: Ubuntu 20.04.2 LTS - Python version: Python 3.8.5 - PyTorch version (GPU?): 1.7.1 - Tensorflow version (GPU?): 2.4.1 - Using GPU in script?: Yes - Using distributed or parallel set-up in script?: Yes My scripts that use Trainer API crash GPUs on a Linux server that has 4 Quadro RTX 8000 GPUs (NVIDIA-SMI 460.39, Driver Version: 460.39, CUDA Version: 11.2). In order to understand if this is my problem or not, I installed Huggingface examples as described in https://huggingface.co/transformers/examples.html. I then run python3 examples/seq2seq/run_summarization.py \ > --model_name_or_path t5-large \ > --do_train \ > --do_eval \ > --dataset_name cnn_dailymail \ > --dataset_config "3.0.0" \ > --source_prefix "summarize: " \ > --output_dir /tmp/tst-summarization \ > --per_device_train_batch_size=2 \ > --per_device_eval_batch_size=2 \ > --overwrite_output_dir \ > --predict_with_generate After this script runs for a few minutes (and I can see that the GPUs are being utilized when I run nvidia-smi), all GPUs crash with the following error: Traceback (most recent call last): File "examples/seq2seq/run_summarization.py", line 591, in <module> main() File "examples/seq2seq/run_summarization.py", line 529, in main train_result = trainer.train(resume_from_checkpoint=checkpoint) File "/home/dima/.local/lib/python3.8/site-packages/transformers/trainer.py", line 1120, in train tr_loss += self.training_step(model, inputs) File "/home/dima/.local/lib/python3.8/site-packages/transformers/trainer.py", line 1542, in training_step loss.backward() File "/usr/lib/python3/dist-packages/torch/tensor.py", line 221, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "/usr/lib/python3/dist-packages/torch/autograd/__init__.py", line 130, in backward Variable._execution_engine.run_backward( RuntimeError: CUDA error: unspecified launch failure When I run nvidia-smi, I get: Unable to determine the device handle for GPU 0000:40:00.0: Unknown Error Rebooting the server helps to restore the GPUs, but the same problem happens again if I try to run the example script above. Please help! :)

https://api.github.com/repos/huggingface/transformers/issues/11019

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11019/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11019/comments

https://api.github.com/repos/huggingface/transformers/issues/11019/events

https://github.com/huggingface/transformers/issues/11019

MDU6SXNzdWU4NDg1NDM0NjI=

Enable multiple `eval_dataset` in `Trainer` API

open

NONE

# 🚀 Feature request Allow for two or more (equally long) validation sets to be passed to the `Trainer` API which are evaluated sequentially each `eval_steps`. ## Motivation You can find my motivation in this [thread](https://discuss.huggingface.co/t/use-trainer-api-with-two-valiation-sets/5212) and the referenced paper. My idea would be to evaluate language model pre-training on an overlapping validation set (coming from the same data distribution as the training set) and a non-overlapping validation set (sampled from future periods or another domain). Ideally, I would like to track and log the validation loss during pre-training for both validation sets.

https://api.github.com/repos/huggingface/transformers/issues/11018

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11018/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11018/comments

https://api.github.com/repos/huggingface/transformers/issues/11018/events

https://github.com/huggingface/transformers/issues/11018

MDU6SXNzdWU4NDg1MzcyNDA=

T5 documentation for computing pretraining loss seems to have a mistake

open

NONE

Dear @patrickvonplaten The documentation of T5 for computing loss of pretraining seems to have a mistake, where it talks on the loss formulation: https://huggingface.co/transformers/model_doc/t5.html?highlight=decoder_input_ids ``` input_ids = tokenizer('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt').input_ids labels = tokenizer('<extra_id_0> cute dog <extra_id_1> the <extra_id_2>', return_tensors='pt').input_ids # the forward function automatically creates the correct decoder_input_ids loss = model(input_ids=input_ids, labels=labels).loss ``` 1) the loss as per mt5 paper "T5 is pre-trained on a masked language modeling “span-corruption” objective, where consecutive spans of input tokens are replaced with a mask token and the model is trained to **reconstruct the masked-out tokens.** " https://arxiv.org/pdf/2010.11934.pdf So I believe you need to set the labels of masked tokens in labels to -100 then compute the loss in this example for pretraining. 2) Based on other examples in the repository, I think one need to do shift_right to compute the correct ` decoder_input_ids` before modifying the labels for the loss (before replacing -100 for the masked tokens) and pass it to the model as well, please correct me if mistaken. I greatly appreciate updating the documentations to show the correct procedure. thanks.

https://api.github.com/repos/huggingface/transformers/issues/11016

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11016/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11016/comments

https://api.github.com/repos/huggingface/transformers/issues/11016/events

https://github.com/huggingface/transformers/issues/11016

MDU6SXNzdWU4NDg0OTAwNjA=

Add new CANINE model

open

COLLABORATOR

# 🌟 New model addition ## Model description Google recently proposed a new **C**haracter **A**rchitecture with **N**o tokenization **I**n **N**eural **E**ncoders architecture (CANINE). Not only the title is exciting: > Pipelined NLP systems have largely been superseded by end-to-end neural modeling, yet nearly all commonly-used models still require an explicit tokenization step. While recent tokenization approaches based on data-derived subword lexicons are less brittle than manually engineered tokenizers, these techniques are not equally suited to all languages, and the use of any fixed vocabulary may limit a model's ability to adapt. In this paper, we present CANINE, a neural encoder that operates directly on character sequences, without explicit tokenization or vocabulary, and a pre-training strategy that operates either directly on characters or optionally uses subwords as a soft inductive bias. To use its finer-grained input effectively and efficiently, CANINE combines downsampling, which reduces the input sequence length, with a deep transformer stack, which encodes context. CANINE outperforms a comparable mBERT model by 2.8 F1 on TyDi QA, a challenging multilingual benchmark, despite having 28% fewer model parameters. Overview of the architecture:  Paper is available [here](https://arxiv.org/abs/2103.06874). We heavily need this architecture in Transformers (RIP subword tokenization)! The first author (Jonathan Clark) said on [Twitter](https://twitter.com/JonClarkSeattle/status/1377505048029134856) that the model and code will be released in April :partying_face: ## Open source status * [ ] the model implementation is available: soon [here](https://caninemodel.page.link/code) * [ ] the model weights are available: soon [here](https://caninemodel.page.link/code) * [x] who are the authors: @jhclark-google (not sure), @dhgarrette, @jwieting (not sure)

https://api.github.com/repos/huggingface/transformers/issues/11014

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11014/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11014/comments

https://api.github.com/repos/huggingface/transformers/issues/11014/events

https://github.com/huggingface/transformers/issues/11014

MDU6SXNzdWU4NDgzNzUxMTk=

OSError: Can't load config for '/content/wav2vec2-large-xlsr-asr-demo'. Make sure that:

open

NONE

I'm using pip install transformers==4.4.2 After completing the training process of ASR I can not read the trained file from my local storage. Although the path is right. But can read from hugging face model = Wav2Vec2ForCTC.from_pretrained("/content/wav2vec2-large-xlsr-asr-demo").to("cuda") The error: OSError: Can't load config for '/content/wav2vec2-large-xlsr-asr-demo'. Make sure that: - '/content/wav2vec2-large-xlsr-asr-demo' is a correct model identifier listed on 'https://huggingface.co/models' - or '/content/wav2vec2-large-xlsr-asr-demo' is the correct path to a directory containing a config.json file

https://api.github.com/repos/huggingface/transformers/issues/11013

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11013/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11013/comments

https://api.github.com/repos/huggingface/transformers/issues/11013/events

https://github.com/huggingface/transformers/issues/11013

MDU6SXNzdWU4NDgzNDk0NTM=

use `BaseModelOutput` as common interface for all different `BaseModelOutputWith*`?

open

NONE

Hello team, I have been taking a look at the `different` output models from your models, and I wonder if it would make sense to inherit all the `BaseModelOutputWithPool` and all the other flavours of modeling output, instead of using `ModelOutput`. https://github.com/huggingface/transformers/blob/c301c26370dfa48f6a6d0408b5bb9eb70ca831b3/src/transformers/modeling_outputs.py#L24 We are trying to build a wrapper around many of the public models hosted on hugging face, and it would be useful to know if we can assume that all the potential `outputs` of the models will contain `hidden_states`. Since now they all only inherit from `ModelOutput` it seems a little confusing. Am I missing something? Is it not something that can be assumed?

https://api.github.com/repos/huggingface/transformers/issues/11012

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11012/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11012/comments

https://api.github.com/repos/huggingface/transformers/issues/11012/events

https://github.com/huggingface/transformers/pull/11012

MDExOlB1bGxSZXF1ZXN0NjA3MjM3OTQ4

Add multi-class, multi-label and regression to transformers

open

MEMBER

https://api.github.com/repos/huggingface/transformers/issues/11050

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11050/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11050/comments

https://api.github.com/repos/huggingface/transformers/issues/11050/events

https://github.com/huggingface/transformers/pull/11050

MDExOlB1bGxSZXF1ZXN0NjA4NTM5Nzgw

accelerate scripts for question answering with no trainer

open

NONE

# What does this PR do? <!-- Congratulations! You've made it this far! You're not quite done yet though. Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution. Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change. Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost. --> <!-- Remove if not applicable --> Fixes # (issue) ## Before submitting - [ ] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case). - [x] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#start-contributing-pull-requests), Pull Request section? - [ ] Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link to it if that's the case. - [ ] Did you make sure to update the documentation with your changes? Here are the [documentation guidelines](https://github.com/huggingface/transformers/tree/master/docs), and [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/master/docs#writing-source-documentation). - [ ] Did you write any new necessary tests? ## Who can review? Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors which may be interested in your PR. @sgugger <!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**. Please tag fewer than 3 people. Models: - albert, bert, xlm: @LysandreJik - blenderbot, bart, marian, pegasus, encoderdecoder, t5: @patrickvonplaten, @patil-suraj - longformer, reformer, transfoxl, xlnet: @patrickvonplaten - fsmt: @stas00 - funnel: @sgugger - gpt2: @patrickvonplaten, @LysandreJik - rag: @patrickvonplaten, @lhoestq - tensorflow: @LysandreJik Library: - benchmarks: @patrickvonplaten - deepspeed: @stas00 - ray/raytune: @richardliaw, @amogkam - text generation: @patrickvonplaten - tokenizers: @n1t0, @LysandreJik - trainer: @sgugger - pipelines: @LysandreJik Documentation: @sgugger HF projects: - datasets: [different repo](https://github.com/huggingface/datasets) - rust tokenizers: [different repo](https://github.com/huggingface/tokenizers) Examples: - maintained examples (not research project or legacy): @sgugger, @patil-suraj - research_projects/bert-loses-patience: @JetRunner - research_projects/distillation: @VictorSanh -->

https://api.github.com/repos/huggingface/transformers/issues/11049

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11049/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11049/comments

https://api.github.com/repos/huggingface/transformers/issues/11049/events

https://github.com/huggingface/transformers/pull/11049

MDExOlB1bGxSZXF1ZXN0NjA4NDQyMTM1

[docs] fix xref to `PreTrainedModel.generate`

open

COLLABORATOR

This PR partially resolves the issue raised in https://github.com/huggingface/transformers/issues/9202 I spent quite some time to try to figure out how to get sphinx to figure out the inheritance so that it could cross-reference inherited methods, but it can't even handle mixins it seems. i.e. it can't resolve: `transformers.PreTrainedModel.generate` So one has to explicitly specify the fully qualified original method for xref to work :( Bottom line - if currently you want a cross-reference link to `transformers.PreTrainedModel.generate` or `T5ForConditionalGeneration.generate` to work, you have to use `~transformers.generation_utils.GenerationMixin.generate` So I did: ``` find . -type d -name ".git" -prune -o -type f -exec perl -pi -e 's|transformers\.\w+\.generate|transformers.generation_utils.GenerationMixin.generate|g' {} \; ``` Fixes: https://github.com/huggingface/transformers/issues/9202 @sgugger

https://api.github.com/repos/huggingface/transformers/issues/11048

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11048/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11048/comments

https://api.github.com/repos/huggingface/transformers/issues/11048/events

https://github.com/huggingface/transformers/pull/11048

MDExOlB1bGxSZXF1ZXN0NjA4NDQwMjYz

fix incorrect case for s|Pretrained|PreTrained|

open

COLLABORATOR

This PR fixes incorrect `Pretrained` case for 2 cases: ``` git-replace PretrainedTokenizer PreTrainedTokenizer git-replace transformers.PretrainedModel transformers.PreTrainedModel ``` there might be other cases to fix, but these stood out. @sgugger

https://api.github.com/repos/huggingface/transformers/issues/11047

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11047/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11047/comments

https://api.github.com/repos/huggingface/transformers/issues/11047/events

https://github.com/huggingface/transformers/issues/11047

MDU6SXNzdWU4NDk2MDQ3OTE=

Use Bert model without pretrained weights

closed

NONE

Hi, I wanted to train a Bert classifier from scratch without any pretrained weights. It has to be randomly initialized and trained. Example: ``` bert_base_model = BertForSequenceClassification() trainer = Trainer(model=bert_base_model, args=training_args, train_dataset=train_loader, eval_dataset=test_loader, compute_metrics=compute_metrics ) ```

https://api.github.com/repos/huggingface/transformers/issues/11046

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11046/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11046/comments

https://api.github.com/repos/huggingface/transformers/issues/11046/events

https://github.com/huggingface/transformers/issues/11046

MDU6SXNzdWU4NDk1Njg0NTk=

Potential incorrect application of layer norm in BlenderbotSmallDecoder

open

NONE

In BlenderbotSmallDecoder, layer norm is applied only on the token embeddings, and not on the hidden_states, whereas in the BlenderbotSmallEncoder, layer norm is applied after adding the input_embeds and positional embeds BlenderbotSmallEncoder: `hidden_states = inputs_embeds + embed_pos` `hidden_states = self.layernorm_embedding(hidden_states)` BlenderbotSmallDecoder: `inputs_embeds = self.layernorm_embedding(inputs_embeds)` `hidden_states = inputs_embeds + positions`

https://api.github.com/repos/huggingface/transformers/issues/11045

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11045/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11045/comments

https://api.github.com/repos/huggingface/transformers/issues/11045/events

https://github.com/huggingface/transformers/issues/11045

MDU6SXNzdWU4NDk1NDQzNzQ=

Multi-GPU seq2seq example evaluation significantly slower than legacy example evaluation

open

NONE

### Who can help @patil-suraj @sgugger Models: T5 ## Information I've been doing multi-GPU evaluation for some weeks using a Transformers pull from Feb 12th, just using the example scripts for training/evaluating custom datasets (specifically `run_distributed_eval.py` , though that seq2seq example is now legacy: https://github.com/huggingface/transformers/tree/master/examples/legacy/seq2seq ) Today I grabbed a fresh pull and migrated the data over to the JSON lines format for the new seq2seq example `run_summarization.py` : https://github.com/huggingface/transformers/blob/master/examples/seq2seq/run_summarization.py run_summarization.py appears to use all visible GPUs to do the evaluation (great!), but it also appears significantly slower than the old run_distributed_eval.py . When examining GPU utilization using `nvtop`, it appears that it allocates GPU memory from all devices (much more from device 0), but only uses device 0 for processing:  ## Script In case it's my issue and I'm not invoking it correctly (I know the legacy one required being invoked with `torch.distributed.launch` for multi-GPU evaluation), the runscript I'm using is: ``` #/bin/bash python run_summarization.py \ --model_name_or_path mymodel-debug1000 \ --do_predict \ --train_file mydata/train.json \ --validation_file mydata/val.json \ --test_file mydata/val.json \ --max_source_length 256 \ --max_target_length 512 \ --num_beams 8 \ --source_prefix "" \ --output_dir tst-debug \ --overwrite_output_dir \ --per_device_train_batch_size=4 \ --per_device_eval_batch_size=4 \ --predict_with_generate ```

https://api.github.com/repos/huggingface/transformers/issues/11044

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11044/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11044/comments

https://api.github.com/repos/huggingface/transformers/issues/11044/events

https://github.com/huggingface/transformers/issues/11044

MDU6SXNzdWU4NDk1Mjk3NjE=

[DeepSpeed] ZeRO stage 3 integration: getting started and issues

open

COLLABORATOR

**[This is not yet alive, preparing for the release, so please ignore for now]** While we are waiting for deespeed to make a new release and then merge the PR, you can try `pip install -e .` in these 2 branches: https://github.com/stas00/DeepSpeed/tree/zero3-everything https://github.com/stas00/transformers/tree/ds-zero-3 and see if it works for you. ------------------------------- The DeepSpeed ZeRO-3 has been integrated into HF `transformers`. While I tried to write tests for a wide range of situations I'm sure I've missed some scenarios so if you run into any problems please file a separate issue. I'm going to use this issue to track progress on individual ZeRO3 issues. # Why would you want ZeRO-3 In a few words, while ZeRO-2 was very limited scability-wise - if `model.half()` couldn't fit onto a single gpu, adding more gpus won't have helped so if you had a 24GB GPU you couldn't train a model larger than about 5B params. Since with ZeRO-3 the model weights are partitioned across multiple GPUs plus offloaded to CPU, the upper limit on model size has increased by about 2 orders of magnitude. That is ZeRO-3 allows you to scale to huge models with Trillions of parameters assuming you have enough GPUs and general RAM to support this. ZeRO-3 can benefit a lot from general RAM if you have it. If not that's OK too. ZeRO-3 combines all your GPUs memory and general RAM into a vast pool of memory. If you don't have many GPUs but just a single one but have a lot of general RAM ZeRO-3 will allow you to fit larger models. Of course, if you run in an environment like the free google colab, while you can use run Deepspeed there, you get so little general RAM it's very hard to make something out of nothing. Some users (or some sessions) one gets 12GB of RAM which is impossible to work with - you want at least 24GB instances. Setting is up might be tricky too, please see this notebook for an example: https://github.com/stas00/porting/blob/master/transformers/deepspeed/DeepSpeed_on_colab_CLI.ipynb # Getting started Install the latest deepspeed version: ``` pip install deepspeed ``` You will want to be on a transformers master branch, if you want to run a quick test: ``` git clone https://github.com/huggingface/transformers cd transformers BS=4; PYTHONPATH=src USE_TF=0 deepspeed examples/seq2seq/run_translation.py \ --model_name_or_path t5-small --output_dir /tmp/zero3 --overwrite_output_dir --max_train_samples 64 \ --max_val_samples 64 --max_source_length 128 --max_target_length 128 --val_max_target_length 128 \ --do_train --num_train_epochs 1 --per_device_train_batch_size $BS --per_device_eval_batch_size $BS \ --learning_rate 3e-3 --warmup_steps 500 --predict_with_generate --logging_steps 0 --save_steps 0 \ --eval_steps 1 --group_by_length --adafactor --dataset_name wmt16 --dataset_config ro-en --source_lang en \ --target_lang ro --source_prefix "translate English to Romanian: " \ --deepspeed examples/tests/deepspeed/ds_config_zero3.json ``` You will find a very detailed configuration here: https://huggingface.co/transformers/master/main_classes/trainer.html#deepspeed Your new config file will look like this: ```json { "fp16": { "enabled": true, "loss_scale": 0, "loss_scale_window": 1000, "initial_scale_power": 16, "hysteresis": 2, "min_loss_scale": 1 }, "zero_optimization": { "stage": 3, "cpu_offload": true, "cpu_offload_params": true, "cpu_offload_use_pin_memory" : true, "overlap_comm": true, "contiguous_gradients": true, "stage3_max_live_parameters": 1e9, "stage3_max_reuse_distance": 1e9, "stage3_prefetch_bucket_size": 0.94e6, "stage3_param_persistence_threshold": 1e4, "reduce_bucket_size": 1e6, "prefetch_bucket_size": 3e6, "sub_group_size": 1e14, "stage3_gather_fp16_weights_on_model_save": true }, "optimizer": { "type": "AdamW", "params": { "lr": 3e-5, "betas": [0.8, 0.999], "eps": 1e-8, "weight_decay": 3e-7 } }, "scheduler": { "type": "WarmupLR", "params": { "warmup_min_lr": 0, "warmup_max_lr": 3e-5, "warmup_num_steps": 500 } }, "steps_per_print": 2000, "wall_clock_breakdown": false } ``` So if you were already using ZeRO-2 it's only the `zero_optimization` stage that has changed. One of the biggest nuances of ZeRO-3 is that the model weights aren't inside `model.state_dict`, as they are spread out through multiple gpus. The Trainer has been modified to support this but you will notice a slow model saving - as it has to consolidate weights from all the gpus. I'm planning to do more performance improvements in the future PRs, but for now let's focus on making things work. # Issues / Questions If you have any general questions or something is unclear/missing in the docs please don't hesitate to ask in this thread. But for any bugs or problems please open a new Issue and tag me there. You don't need to tag anybody else. Thank you!

https://api.github.com/repos/huggingface/transformers/issues/11043

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11043/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11043/comments

https://api.github.com/repos/huggingface/transformers/issues/11043/events

https://github.com/huggingface/transformers/issues/11043

MDU6SXNzdWU4NDk0OTk3MzQ=

Can't load model to estimater

open

NONE

I was trying to follow the Sagemaker instructions [here](https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model.html) to load the model I just trained and test an estimation. I get the error message: NotImplementedError: Creating model with HuggingFace training job is not supported. Can someone share some sample code to run to do this? Here is the basic thing I am trying to do: ``` from sagemaker.estimator import Estimator # job which is going to be attached to the estimator old_training_job_name='huggingface-sdk-extension-2021-04-02-19-10-00-242' # attach old training job huggingface_estimator_loaded = Estimator.attach(old_training_job_name) # get model output s3 from training job testModel = huggingface_estimator_loaded.model_data ner_classifier = huggingface_estimator.deploy(initial_instance_count=1, instance_type='ml.m4.xlarge') ``` I also tried some things with .deploy() and endpoints but didn't have any luck there either.

https://api.github.com/repos/huggingface/transformers/issues/11042

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11042/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11042/comments

https://api.github.com/repos/huggingface/transformers/issues/11042/events

https://github.com/huggingface/transformers/issues/11042

MDU6SXNzdWU4NDkyNzQzNjI=

[LXMERT] Unclear what img_tensorize does with color spaces

open

NONE

## Environment info - `transformers` version: Not using transformers directly, I'm loading a model "unc-nlp/frcnn-vg-finetuned" - Platform: MacOS - Python version: 3.8 - PyTorch version (GPU?): 1.6.0, no GPU - Tensorflow version (GPU?): don't have - Using GPU in script?: no - Using distributed or parallel set-up in script?: no ### Who can help @eltoto1219 probably Models: "LXMERT": "unc-nlp/frcnn-vg-finetuned" Library: https://github.com/huggingface/transformers/tree/master/examples/research_projects/lxmert ## Information Model I am using (Bert, XLNet ...): "LXMERT": "unc-nlp/frcnn-vg-finetuned" The problem arises when using: * [ ] the official example scripts: (give details below) * [x ] my own modified scripts: (give details below) The tasks I am working on is: * [ ] an official GLUE/SQUaD task: (give the name) * [x ] my own task or dataset: (give details below) ## Problem I don't get what img_tensorize in utils.py is doing with color spaces. I run the following code to load the model. ``` # load models and model components frcnn_cfg = Config.from_pretrained("unc-nlp/frcnn-vg-finetuned") frcnn = GeneralizedRCNN.from_pretrained("unc-nlp/frcnn-vg-finetuned", config=frcnn_cfg) image_preprocess = Preprocess(frcnn_cfg) ``` Turns out that frcnn_cfg.input.format is "BGR" so I wanted to know what exactly is going on? Here is where the image is loaded (utils.img_tensorize) ``` def img_tensorize(im, input_format="RGB"): assert isinstance(im, str) if os.path.isfile(im): img = cv2.imread(im) else: img = get_image_from_url(im) assert img is not None, f"could not connect to: {im}" img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) if input_format == "RGB": img = img[:, :, ::-1] return img ``` See, we seem to be converting the images to RGB, then if it's "RGB" format we flip the blue (?) channel? Is the image ever converted to "BGR"?

https://api.github.com/repos/huggingface/transformers/issues/11041

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11041/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11041/comments

https://api.github.com/repos/huggingface/transformers/issues/11041/events

https://github.com/huggingface/transformers/pull/11041

MDExOlB1bGxSZXF1ZXN0NjA4MDcxNjc1

wav2vec2 converter: create the proper vocab.json while converting fairseq wav2vec2 finetuned model

open

CONTRIBUTOR

# What does this PR do? While converting a finetuned wav2vec2 model we also need to convert the related dictionary `dict.ltr.txt` to hugging face `vocab.json` format. If a `dict_path` is specified: - Creates&saves the necessary vocab.json file - Modifies config file special token ids and vocab size accordingly - Creates a processor with the right special tokens and saves the processor `preprocessor_config.json` ## Before submitting - [x] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#start-contributing-pull-requests), Pull Request section? - [ ] Did you make sure to update the documentation with your changes? -> not sure if there are any docs related to this. - [ ] Did you write any new necessary tests? -> not sure if there are tests related to this. ## Who can review? Models: - wav2vec2: @patrickvonplaten @LysandreJik

https://api.github.com/repos/huggingface/transformers/issues/11040

https://api.github.com/repos/huggingface/transformers

https://api.github.com/repos/huggingface/transformers/issues/11040/labels{/name}

https://api.github.com/repos/huggingface/transformers/issues/11040/comments

https://api.github.com/repos/huggingface/transformers/issues/11040/events

https://github.com/huggingface/transformers/issues/11040

MDU6SXNzdWU4NDkyNjU2MTU=

max_length in beam_search() and group_beam_search() does not consider beam_scorer.max_length

open

CONTRIBUTOR

## Environment info <!-- You can run the command `transformers-cli env` and copy-and-paste its output below. Don't forget to fill out the missing fields in that output! --> - `transformers` version: - Platform: 4.3.2 - Python version: 3.8.5 - PyTorch version (GPU?): 1.8.0 - Using GPU in script?: No - Using distributed or parallel set-up in script?: No ### Who can help <!-- Your issue will be replied to more quickly if you can figure out the right person to tag with @ If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**. Please tag fewer than 3 people. Models: - albert, bert, xlm: @LysandreJik - blenderbot, bart, marian, pegasus, encoderdecoder, t5: @patrickvonplaten, @patil-suraj - longformer, reformer, transfoxl, xlnet: @patrickvonplaten - fsmt: @stas00 - funnel: @sgugger - gpt2: @patrickvonplaten, @LysandreJik - rag: @patrickvonplaten, @lhoestq - tensorflow: @LysandreJik Library: - benchmarks: @patrickvonplaten - deepspeed: @stas00 - ray/raytune: @richardliaw, @amogkam - text generation: @patrickvonplaten - tokenizers: @LysandreJik - trainer: @sgugger - pipelines: @LysandreJik Documentation: @sgugger Model hub: - for issues with a model report at https://discuss.huggingface.co/ and tag the model's creator. HF projects: - datasets: [different repo](https://github.com/huggingface/datasets) - rust tokenizers: [different repo](https://github.com/huggingface/tokenizers) Examples: - maintained examples (not research project or legacy): @sgugger, @patil-suraj - research_projects/bert-loses-patience: @JetRunner - research_projects/distillation: @VictorSanh --> ## Information I am using BART model in particular, but this problem exists for all the other models, using `beam_search()` and `group_beam_search()` for decoding the generated text. The `max_length` variable set using `BeamSearchScorer` is not used by `beam_search()` or `group_beam_search()` function in `generation_utils.py` script. Thus using a smaller `max_length` while initializing the object of class, for example: ``` beam_scorer = BeamSearchScorer( batch_size=1, max_length=5, num_beams=num_beams, device=model.device, ) ``` instead of ``` beam_scorer = BeamSearchScorer( batch_size=1, max_length=model.config.max_length, num_beams=num_beams, device=model.device, ) ``` in the example given [here](https://huggingface.co/transformers/main_classes/model.html#transformers.generation_utils.GenerationMixin.beam_search) results in an error: ``` File "temp.py", line 34, in <module> outputs = model.beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs) File "<conda_env_dir>/lib/python3.8/site-packages/transformers/generation_utils.py", line 1680, in beam_search sequence_outputs = beam_scorer.finalize( File "<conda_env_dir>/lib/python3.8/site-packages/transformers/generation_beam_search.py", line 328, in finalize decoded[i, : sent_lengths[i]] = hypo RuntimeError: The expanded size of the tensor (5) must match the existing size (6) at non-singleton dimension 0. Target sizes: [5]. Tensor sizes: [6] ``` The problem arises when using: * [x] the official example scripts: - [beam_scorer_example](https://huggingface.co/transformers/main_classes/model.html#transformers.generation_utils.GenerationMixin.beam_search) * [x] my own modified scripts: - Also, using `max_length` higher than `model.config.max_length` while initializing object (`beam_scorer`) of type `BeamSearchScorer` does not help in generating longer sequences, as `beam_scorer.max_length` is not used by `beam_search()` or `group_beam_search()` ## To reproduce Steps to reproduce the behavior: 1. The above mentioned modification in the [example](https://huggingface.co/transformers/main_classes/model.html#transformers.generation_utils.GenerationMixin.beam_search) <!-- If you have code snippets, error messages, stack traces please provide them here as well. Important! Use code tags to correctly format your code. See https://help.github.com/en/github/writing-on-github/creating-and-highlighting-code-blocks#syntax-highlighting Do not use screenshots, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.--> ## Expected behavior 1. The program to run without any errors when lower `max_length` is set for object of type `BeamSearchScorer` 2. Generate longer length sequences (longer than `model.config.max_length`) when higher `max_length` is set for object of type `BeamSearchScorer` <!-- A clear and concise description of what you would expect to happen. -->