train_wordpiece_tokenizer.py

# Copyright 2024 EPFL and Apple Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import argparse
from fourm.utils.tokenizer import train_unified_wordpiece_tokenizer
from fourm.utils.tokenizer import generate_sentinel_tokens, generate_coord_tokens, generate_object_class_tokens


def get_args():
    parser = argparse.ArgumentParser('Train unified WordPiece tokenizer', add_help=False)
    parser.add_argument('--text_files', type=str, default='/datasets/imagenet_multitask/metadata/all_captions_BLIP.txt',
                        help="Files to train the tokenizer on, separated by a double dash '--'")
    parser.add_argument('--save_file', type=str, default="utils/tokenizer/trained/default_tokenizer.json",
                        help="Path to the saved tokenizer. Can then be loaded using Tokenizer.from_file(path).")
    parser.add_argument('--vocab_size', type=int, default=30_000,
                        help="Vocabulary size")
    parser.add_argument('--num_sentinels', type=int, default=200, help="Number of sentinel tokens")
    parser.add_argument('--coord_bins', type=int, default=1000, help="Number of coordinate bins (for detection)")
    parser.add_argument('--object_classes', type=str, default='coco', choices=['none', 'coco'],
                        help="Special tokens for detection instances (e.g., instance class names from the COCO dataset)")
    parser.add_argument('--lowercase', action='store_true')
    parser.add_argument('--no_lowercase', action='store_false', dest='lowercase')
    parser.set_defaults(lowercase=True)
    return parser.parse_args()


def train_tokenizer(args):

    files = args.text_files.split("--")
    # Get special tokens
    sentinel_tokens = generate_sentinel_tokens(num=args.num_sentinels)
    coord_tokens = generate_coord_tokens(bins=args.coord_bins)
    if args.object_classes == 'none':
        object_class_tokens = None
    else:
        object_class_tokens = generate_object_class_tokens(args.object_classes)

    print(f"Training tokenizer on files: {files}")

    # Train tokenizer
    tokenizer = train_unified_wordpiece_tokenizer(
        files=files,
        vocab_size=args.vocab_size,
        sentinel_tokens=sentinel_tokens,
        coord_tokens=coord_tokens,
        object_class_tokens=object_class_tokens,
        lowercase=args.lowercase,
    )

    # Create directory of target file if it doesn't exist
    os.makedirs(os.path.dirname(args.save_file), exist_ok=True)
    tokenizer.save(path=args.save_file)

    print(f"Tokenizer saved to: {args.save_file}!")


if __name__ == "__main__":
    args = get_args()
    train_tokenizer(args)