Can DALI be integrated into HuggingFace Trainer?

[ShyFoo]

Describe the question.

Is it possible to use DALI as the dataloader for HuggingFace Trainer?

Check for duplicates

• [X] I have searched the open bugs/issues and have found no duplicates for this bug report

[JanuszL]

Hi @ShyFoo,

Thank you for reaching out. While we haven't tried this directly yet DALI provides compatibility with PyTorch so you can create a DALI iterator that would behave as PyTorch one. Have you tried this yet? Have you encountered any issues?

[JanuszL]

I just checked and unfortunately, HuggingFace trainer (I checked only one example so far) expects PyTorch DataIterator (checks for the type of passes object). I still believe it is possible to make DALI work there but I may need a bit of adjustments to the internal methods. If you have some spare time we would be more than happy to review a PR with such an example.

[ShyFoo]

I just checked and unfortunately, HuggingFace trainer (I checked only one example so far) expects PyTorch DataIterator (checks for the type of passes object). I still believe it is possible to make DALI work there but I may need a bit of adjustments to the internal methods. If you have some spare time we would be more than happy to review a PR with such an example.

Thanks for your reply. Yeah, I have tried integrating DALI into the HuggingFace trainer. As you mentioned, it seems to expect either torch.utils.data.Dataset or torch.utils.data.IterableDataset as input. It might be possible to customize a data collator, which can preprocess data in a DALI pipeline, for the HuggingFace trainer, but I'm not sure if it will work. If you have any idea for this, please let me know. I'm sure it would be a huge step for the whole community!

[JanuszL]

I did some basic investigation:

• for CV tasks DALI needs to have such wrapper:

  class DALIwrapper(DALIGenericIterator):
    def __init__(self, *args, **kargs):
        super().__init__(*args, **kargs)
  
    def __next__(self):
        batch = super().__next__()
        return batch[0]
  
  pipe_train = GetDaliPipeline(...)
  data_train = DALIwrapper([pipe_train], ["pixel_values", "labels"], reader_name="Reader")

• trainer needs to have these methods replaced:

  trainer.get_train_dataloader = lambda : data_train
  trainer.get_eval_dataloader = lambda x: x

  so DALI dataloader is used as is without wrapping into PyTorch DataIterator

• there is probably another layer of complexity using DALI with a distributed strategy which is handled by accelerate

[ShyFoo]

For the third point, _get_traindataloader returns _self.accelerator.prepare(DataLoader(train_dataset, **dataloaderparams)) for a distributed training or evaluation. So what should I do in this setting? Hope for your help.

[JanuszL]

@ShyFoo - I'm afraid I don't know the answer. My idea was to avoid wrapping DALI into DataLoader as it serves as one already. If you have time I would be more than happy to learn your findings.

[ShyFoo]

e basic investigation:

Got it. I have tried using DALI + DeepSpeed, where I have to define a triaing loop and other practical functions myself, instead of using the Huggingface Trainer for convenience, so I prefer to use DALI, DeepSpeed, and Huggingface Trainer simultaneously if possible.

[JanuszL]

If you have any examples of DALI + DeepSpeed feel free to post them as PR to DALI - we would be more than happy to enrich our documentation and the base of samples.

[ShyFoo]

Well, I'd be happy to.