Train BPE with fastBPE, and load to Huggingface Tokenizer.

The BPETrainer of Huggingface consumes a lot of memory when I am training on a large corpus (e.g. 50000 merges on 20GB corpus). And I got a memory error.

So I use fastBPE (implemented with C) instead, which returns a list of merge operations.

However, I still want to use the huggingface Tokenizer API. So I write a simple convertor for generating the json file for Huggingface Tokenizer.

Train BPE:

cd fastBPE
./fast learnbpe [merges, e.g. 50000] [train.txt] > allvocab

Convert to json:

python convertjs.py

This tokenizer does not indicate the start of a token.

E.g. BPE result for "I am" and "Iam" may be the same. Please split the sentence by space before you use it.

    words = "I am".split()
    for word in words:
        subs = tokenizer.tokenize(word)
        subs[0] = "<begin>" + subs[0]

This results in ["<begin>I", "am"] and ["<begin>I", "<begin>am"] for "Iam" and "I am".