src/create_dataset.py

import gzip
import json

import numpy as np
import pandas as pd
from transformers import AutoTokenizer

COLLATE_LENGTH = 370


def emit(line_id, nl_str, en_str, nl_l, en_l):
    obj = {
        "id": line_id,
        "translation": {
            "nl": nl_str.strip(),
            "en": en_str.strip(),
        },
        "nl_len": nl_l,
        "en_len": en_l,
    }
    writer.write(str.encode(json.dumps(obj)))
    writer.write("\n".encode("utf-8"))


class TokenLength:
    def __init__(self, tokenizer):
        self.tokenizer = AutoTokenizer.from_pretrained(
            tokenizer, max_length=4096, truncation=False, use_fast=False
        )

    def __call__(self, text: str):
        return len(self.tokenizer.encode(text, max_length=4096, truncation=False))


class Counter:
    def __init__(self, start=0):
        self.count = start

    def __call__(self):
        self.count += 1
        return self.count


class Buffer:
    def __init__(
        self,
        id: int,
        emit_lines: bool,
        max_length: int,
        en_prefix="",
    ):
        self.id = id
        self.emit_lines = emit_lines
        self.max_length = max_length
        self.en_prefix = en_prefix
        self.counter = Counter()
        self.nl_l = None
        self.en_l = None
        self.nl_buf = None
        self.en_buf = None
        self.cur_max_length = None
        self.reset()

    def set_cur_max_length(self):
        """You can check the distribution with the following code:
        %matplotlib notebook
        import numpy as np
        import matplotlib.pyplot as plt

        plt.rcParams['figure.figsize'] = [9.5,6]
        fig, ax = plt.subplots(1, 1)

        r = np.random.beta(20,8,102000)
        ax.hist(r, density=True, histtype='stepfilled', alpha=0.2, bins=200)
        ax.legend(loc='best', frameon=False)
        plt.show()
        """
        self.cur_max_length = int(self.max_length * np.random.beta(20, 8))

    def reset(self):
        self.nl_l = None
        self.en_l = None
        self.nl_buf = None
        self.en_buf = None
        self.set_cur_max_length()

    def add_ok(self, nl_str, en_str, separator="\n"):
        """If the new text fits within the max_length tokens, add it, else return False"""
        nl_new = self.nl_buf + f"{separator}{nl_str}" if self.nl_buf else nl_str
        en_new = self.en_buf + f"{separator}{en_str}" if self.en_buf else en_str
        nl_new_l = token_length(nl_new)
        en_new_l = token_length(en_new)
        # Check if we can add it or if the result would be too long
        if (
            nl_new_l > self.cur_max_length
            or token_length(self.en_prefix + en_new) > self.cur_max_length
        ):
            return False
        else:
            self.nl_buf = nl_new
            self.en_buf = en_new
            self.nl_l = nl_new_l
            self.en_l = en_new_l
            return True

    def emit(self, row, separator):
        nl_str = row.translation["nl"]
        en_str = row.translation["en"]
        nl_id = row.meta["sentenceIds"]["nl"]
        en_id = row.meta["sentenceIds"]["en"]

        # if one of the sentences ends on a . but the other doesn't, add a dot to the other
        if nl_str.endswith(".") and not en_str.endswith("."):
            en_str += "."
        elif en_str.endswith(".") and not nl_str.endswith("."):
            nl_str += "."
        # Strip any leading "- " or "- " from the sentences
        nl_str = nl_str.lstrip("- ")
        en_str = en_str.lstrip("- ")

        nl_len = token_length(nl_str)
        en_len = token_length(en_str)
        if self.emit_lines and nl_len <= COLLATE_LENGTH and en_len <= COLLATE_LENGTH:
            emit(
                line_id=f"{row.tconst}-nl{nl_id}-en{en_id}-l-",
                nl_str=nl_str,
                en_str=en_str,
                nl_l=nl_len,
                en_l=en_len,
            )
        if self.add_ok(nl_str.strip(), en_str.strip(), separator):
            return

        # If buf.add returns false, we've hit the maximum length boundary, so emit the current buffer, if it is not Empty
        if self.nl_buf:
            emit(
                line_id=f"{row.tconst}-b{self.id}-{self.counter()}",
                nl_str=self.nl_buf,
                en_str=self.en_buf,
                nl_l=self.nl_l,
                en_l=self.en_l,
            )
            # After emit of the buffer, we reset the buffer
            self.reset()

        # Add the first line in this new buffer
        result = self.add_ok(nl_str.strip(), en_str.strip())
        if not result:
            self.reset()


if __name__ == "__main__":
    token_length = TokenLength(tokenizer="yhavinga/ul2-base-dutch")
    line_counter = Counter()

    buffers = [
        Buffer(
            id=index, emit_lines=(index == 0), max_length=buf_max_length, en_prefix=""
        )
        for index, buf_max_length in enumerate([0.6 * 370, 370])
    ]

    df = pd.read_json("episode_opensubtitles.json.gz", lines=True)

    with gzip.open("outfile", mode="wb") as writer:
        for row in df.itertuples():
            for buffer in buffers:
                buffer.emit(row, separator="\n")