Using tokenizers from 🤗 Tokenizers

The PreTrainedTokenizerFast depends on the 🤗 Tokenizers library. The tokenizers obtained from the 🤗 Tokenizers library can be loaded very simply into 🤗 Transformers.

Before getting in the specifics, let’s first start by creating a dummy tokenizer in a few lines:

>>> from tokenizers import Tokenizer
>>> from tokenizers.models import BPE
>>> from tokenizers.trainers import BpeTrainer
>>> from tokenizers.pre_tokenizers import Whitespace

>>> tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
>>> trainer = BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])

>>> tokenizer.pre_tokenizer = Whitespace()
>>> files = [...]
>>> tokenizer.train(files, trainer)

We now have a tokenizer trained on the files we defined. We can either continue using it in that runtime, or save it to a JSON file for future re-use.

Loading directly from the tokenizer object

Let’s see how to leverage this tokenizer object in the 🤗 Transformers library. The PreTrainedTokenizerFast class allows for easy instantiation, by accepting the instantiated tokenizer object as an argument:

>>> from transformers import PreTrainedTokenizerFast

>>> fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)

This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to the tokenizer page for more information.

Loading from a JSON file

In order to load a tokenizer from a JSON file, let’s first start by saving our tokenizer:

>>> tokenizer.save("tokenizer.json")

The path to which we saved this file can be passed to the PreTrainedTokenizerFast initialization method using the tokenizer_file parameter:

>>> from transformers import PreTrainedTokenizerFast

>>> fast_tokenizer = PreTrainedTokenizerFast(tokenizer_file="tokenizer.json")

This object can now be used with all the methods shared by the 🤗 Transformers tokenizers! Head to the tokenizer page for more information.