5

I have a question regarding "on-the-fly" tokenization. This question was elicited by reading the "How to train a new language model from scratch using Transformers and Tokenizers" here. Towards the end there is this sentence: "If your dataset is very large, you can opt to load and tokenize examples on the fly, rather than as a preprocessing step". I've tried coming up with a solution that would combine both datasets and tokenizers, but did not manage to find a good pattern.

I guess the solution would entail wrapping a dataset into a Pytorch dataset.

As a concrete example from the docs

import torch

class SquadDataset(torch.utils.data.Dataset):
    def __init__(self, encodings):
        # instead of doing this beforehand, I'd like to do tokenization on the fly
        self.encodings = encodings 

    def __getitem__(self, idx):
        return {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}

    def __len__(self):
        return len(self.encodings.input_ids)

train_dataset = SquadDataset(train_encodings)

How would one implement this with "on-the-fly" tokenization exploiting the vectorized capabilities of tokenizers?

Pietro
  • 415
  • 6
  • 16

1 Answers1

3

UPDATE Feb 2021

As of v1.3.0 datasets supports lazy evaluation of functions via the set_transform method. Therefore, you can apply on-the-fly tokenization directly like shown here.

OLD ANSWER

In the end I settled for this solution. I do not like that the batch_size is now controlled at the dataset level. However, it does its job.

In this way we exploit two nice things:

  • fast indexing the HuggingFace datasets

  • vectorization capabilities of the HuggingFace tokenizer

class CustomPytorchDataset(Dataset):
    """
    This class wraps the HuggingFace dataset and allows for 
    batch indexing into the dataset. This allows exploiting
    the capabilities of the tokenizer to work on batches.

    NOTE: now we control batch_size at the Dataset level, not
    in the DataLoader therefore the DataLoader should always be
    used with `batch_size=1`.
    """

    def __init__(self, batch_size: int):
        self.batch_size = batch_size
        self.dataset = train_ds          # HuggingFace dataset
        self.tokenizer = bert_tokenizer  # HuggingFace tokenizer

    def __getitem__(self, batch_idx: List[int]):
        instance = self.dataset[batch_idx]

        # tokenize on-the-fly
        tokenized_instance = self.tokenizer(
            instance[text_col], 
            truncation=True, 
            padding=True
        )
        
        return tokenized_instance

    def __len__(self):
        return len(self.dataset)

    def sampler(self):
        # shuffling can be controlled by the sampler, 
        # without touching the dataset
        return BatchSampler(
            SequentialSampler(self), 
            batch_size=self.batch_size, 
            drop_last=True
        )

    @staticmethod
    def collate_fn(batches: List[Dict[str, int]]):
        return {
            k: torch.tensor(v, dtype=torch.int64) 
            for k, v in batches[0].items()
        }
Pietro
  • 415
  • 6
  • 16