src/truncate.py

import argparse
import os
import shutil
from transformers import T5Tokenizer
from tqdm import tqdm


def parse():
    parser = argparse.ArgumentParser()

    parser.add_argument("--input-dir", type=str)
    parser.add_argument("--output-dir", type=str)
    parser.add_argument("--max-src-length", type=int, default=512)
    parser.add_argument("--max-dialogue-history-len", type=int, default=256)
    parser.add_argument("--tokenizer-path", type=str)
    parser.add_argument("--special-tokens-file", type=str, default=None)
    parser.add_argument(
        "--truncate-side", type=str, default="left", choices=["left", "right"]
    )

    return parser.parse_args()


def truncate(args):
    left_tokenizer = T5Tokenizer.from_pretrained(
        args.tokenizer_path, truncate_side="left"
    )
    right_tokenizer = T5Tokenizer.from_pretrained(
        args.tokenizer_path, truncate_side="right"
    )
    tokenizer = T5Tokenizer.from_pretrained(args.tokenizer_path)

    if args.special_tokens_file is not None:
        with open(args.special_tokens_file, "r") as reader:
            special_tokens_dict = {
                "additional_special_tokens": [
                    token.strip() for token in reader.readlines()
                ]
            }

        left_tokenizer.add_special_tokens(special_tokens_dict)
        right_tokenizer.add_special_tokens(special_tokens_dict)
        tokenizer.add_special_tokens(special_tokens_dict)

    def normalize(x):
        return tokenizer.decode(tokenizer(x).input_ids[:-1])

    def divide_chunks(src):
        prefix, postfix = src.split("]", 1)
        prefix = prefix + "]"

        knowledge_start_index = postfix.index("[EK]")
        dialogue = postfix[: knowledge_start_index - 1]
        knowledge_and_instruction = postfix[knowledge_start_index - 1 :]

        instruction_start_index = knowledge_and_instruction.rfind("[C]")
        knowledge = knowledge_and_instruction[: instruction_start_index - 1]
        instruction = knowledge_and_instruction[instruction_start_index - 1 :]

        return prefix, dialogue, knowledge, instruction

    def token_num(x):
        return len(tokenizer.tokenize(x))

    min_knowledge_len = token_num(" [EK] None")

    if not os.path.exists(args.output_dir):
        os.makedirs(args.output_dir)

    print(f" {os.path.basename(args.input_dir)} ".center(70, "="))

    for filename in os.listdir(args.input_dir):
        if not filename.endswith(".src"):
            filepath = os.path.join(args.input_dir, filename)
            if not os.path.exists(os.path.join(args.output_dir, filename)):
                if os.path.isfile(filepath):
                    shutil.copyfile(
                        os.path.join(args.input_dir, filename),
                        os.path.join(args.output_dir, filename),
                    )
                else:
                    shutil.copytree(
                        os.path.join(args.input_dir, filename),
                        os.path.join(args.output_dir, filename),
                    )
        else:
            dialogue_cut_num = 0
            knowledge_cut_num = 0
            cut_token_num = 0
            # Truncate the source file
            with open(os.path.join(args.input_dir, filename), "r") as reader, open(
                os.path.join(args.output_dir, filename), "w"
            ) as writer:
                for line in tqdm(reader.readlines()):
                    src = line.strip()
                    src = normalize(src)

                    prefix, dialogue, knowledge, instruction = divide_chunks(src)

                    prefix_token_num = token_num(prefix)
                    dialogue_token_num = token_num(dialogue)
                    knowledge_token_num = token_num(knowledge)
                    instruction_token_num = token_num(instruction)

                    assert (
                        args.max_src_length >= prefix_token_num + instruction_token_num
                    )

                    origin_src_token_num = (
                        prefix_token_num
                        + dialogue_token_num
                        + knowledge_token_num
                        + instruction_token_num
                    )
                    # origin_src_token_num = token_num(src)

                    # assert (
                    #     prefix_token_num
                    #     + dialogue_token_num
                    #     + knowledge_token_num
                    #     + instruction_token_num
                    #     == origin_src_token_num
                    # )

                    if origin_src_token_num > args.max_src_length:
                        left_token_num = (
                            args.max_src_length
                            - prefix_token_num
                            - instruction_token_num
                        )
                        max_dialogue_token_num = min(
                            max(
                                args.max_dialogue_history_len,
                                left_token_num - knowledge_token_num,
                            ),
                            left_token_num - min_knowledge_len,
                        )

                        # The dialogue is out of the maximum token number
                        if dialogue_token_num > max_dialogue_token_num:
                            # Truncate the dialogue from left or right (For DDRel)
                            truncate_tokenizer = (
                                left_tokenizer
                                if args.truncate_side == "left"
                                else right_tokenizer
                            )
                            dialogue_ids = truncate_tokenizer(
                                dialogue,
                                max_length=max_dialogue_token_num
                                + 1,  # +1 is for the eos
                                truncation=True,
                            ).input_ids

                            dialogue = tokenizer.decode(dialogue_ids[:-1])
                            dialogue_token_num = max_dialogue_token_num
                            dialogue_cut_num += 1

                            # assert token_num(dialogue) <= dialogue_token_num

                        if knowledge_token_num > left_token_num - dialogue_token_num:
                            # Truncate the knowledge from right
                            knowledge_ids = right_tokenizer(
                                knowledge,
                                max_length=left_token_num - dialogue_token_num + 1,
                                truncation=True,
                            ).input_ids

                            knowledge = tokenizer.decode(knowledge_ids[:-1])

                            knowledge = " " + knowledge

                            knowledge_token_num = left_token_num - dialogue_token_num
                            knowledge_cut_num += 1

                            # assert (
                            #     token_num(knowledge) <= knowledge_token_num
                            # ), f"{knowledge_token_num}, {token_num(knowledge)}, {tokenizer.convert_ids_to_tokens(knowledge_ids)}, {knowledge_ids}"

                        src = (
                            prefix.strip()
                            + " "
                            + dialogue.strip()
                            + " "
                            + knowledge.strip()
                            + " "
                            + instruction.strip()
                        )

                        src_token_num = token_num(src)

                        # assert src_token_num <= args.max_src_length

                        cut_token_num += origin_src_token_num - src_token_num

                    prefix, dialogue, knowledge, instruction = divide_chunks(src)

                    prefix_token_num = token_num(prefix)
                    dialogue_token_num = token_num(dialogue)
                    knowledge_token_num = token_num(knowledge)
                    instruction_token_num = token_num(instruction)

                    writer.write(src + "\n")

            print(f" {filename} ".center(40, "-"))
            print(f"dialogue cut num: {dialogue_cut_num}")
            print(f"knowledge cut num: {knowledge_cut_num}")
            print(f"token cut num: {cut_token_num}")


if __name__ == "__main__":
    truncate(parse())