Using a custom trained huggingface tokenizer

Question

I’ve trained a custom tokenizer using a custom dataset using this code that’s on the documentation. Is there a method for me to add this tokenizer to the hub and to use it as the other tokenizers by calling the AutoTokenizer.from_pretrained() function? If I can’t do that how can I use the tokenizer to train a custom model from scratch? Thanks for your help!!!

Here's the code below:

from tokenizers import Tokenizer
from tokenizers.models import BPE
tokenizer = Tokenizer(BPE(unk_token="[UNK]"))

from tokenizers.trainers import BpeTrainer
trainer = BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])

from tokenizers.pre_tokenizers import Whitespace
tokenizer.pre_tokenizer = Whitespace()

folder = 'dataset_unicode'
files = [f"/content/drive/MyDrive/{folder}/{split}.txt" for split in ["test", "train", "valid"]]
tokenizer.train(files, trainer)

from tokenizers.processors import TemplateProcessing
tokenizer.post_processor = TemplateProcessing(
    single="[CLS] $A [SEP]",
    pair="[CLS] $A [SEP] $B:1 [SEP]:1",
    special_tokens=[
        ("[CLS]", tokenizer.token_to_id("[CLS]")),
        ("[SEP]", tokenizer.token_to_id("[SEP]")),
    ],
)

# I've tried saving it like this but it doesn't work as I expect it:
tokenizer.save("data/tokenizer-custom.json")

Please follow this [post](https://stackoverflow.com/a/75990055/6664872) to load it. You can afterward save it with `save_pretrained` and upload it to the hub. — cronoik, Apr 19 '23 at 10:47

alvas · Accepted Answer · 2023-04-19T19:16:06.523

The AutoTokenizer expects a few files in the directory:

awesometokenizer/
    tokenizer_config.json
    special_tokens_map.json
    tokenizer.json

But the default tokenizer.Tokenizer.save() function only saves the vocab file in awesometokenizer/tokenizer.json, open up the json file and compare the ['model']['vocab'] keys to your json from data/tokenizer-custom.json.

The simplest way to let AutoTokenizer load .from_pretrained is to follow the answer that @cronoik posted in the comment, using PreTrainedTokenizerFast, i.e. adding a few lines to your existing code:

from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import Whitespace
from tokenizers.processors import TemplateProcessing

from transformers import PreTrainedTokenizerFast  # <---- Add this line.



trainer = BpeTrainer(special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])

tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
tokenizer.pre_tokenizer = Whitespace()

files = ["big.txt"]  # e.g. training with https://norvig.com/big.txt
tokenizer.train(files, trainer)

tokenizer.post_processor = TemplateProcessing(
    single="[CLS] $A [SEP]",
    pair="[CLS] $A [SEP] $B:1 [SEP]:1",
    special_tokens=[
        ("[CLS]", tokenizer.token_to_id("[CLS]")),
        ("[SEP]", tokenizer.token_to_id("[SEP]")),
    ],
)

# Add these lines:
#     |
#     |
#     V
awesome_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)
awesome_tokenizer.save_pretrained("awesome_tokenizer")

Then you can load the trained tokenizer:

from transformers import AutoTokenizer

auto_loaded_tokenizer = AutoTokenizer.from_pretrained(
    "awesome_tokenizer", 
    local_files_only=True
)

Note: tokenizers though can be pip installed, is a library in Rust with Python bindings

Using a custom trained huggingface tokenizer

1 Answers1