batch.py

import os
from pandas import read_csv

from datasets import GeneratorBasedBuilder, Value, Version, BuilderConfig, Features, DatasetInfo, SplitGenerator, Split, Audio, Sequence

_DESCRIPTION = '''
The dataset contains threads parsed from the /b/ board of 2ch archive
'''

_HOMEPAGE = 'https://huggingface.co/datasets/zeio/batch'

_LICENSE = 'Apache License Version 2.0'

_CLUSTER = '{first_page:04d}-{last_page:04d}'
_URLS = {
    'written': 'https://huggingface.co/datasets/zeio/batch/resolve/main/threads-compressed/{cluster}.tar.xz',
    'spoken': 'https://huggingface.co/datasets/zeio/batch-speech/raw/main/threads-compressed/{cluster}.tar.xz'
}
_INDEX = 'https://huggingface.co/datasets/zeio/batch/resolve/main/index.tsv'

_N_ITEMS = 1750
_N_BATCH = 20


class Batch(GeneratorBasedBuilder):

    VERSION = Version('06.11.2023')

    BUILDER_CONFIGS = [
        BuilderConfig(
            name = 'written',
            version = VERSION,
            description = 'The base modification which contains only text representation of threads, which are divided into topics, which in turn are made of posts'
        ),
        BuilderConfig(
            name = 'spoken',
            version = VERSION,
            description = (
                'An extended configuration of the dataset in which besides text some threads have an associated audio data with speech '
                'generated for text in the respective thread using an alternating speaker pattern'
            )
        )
    ]

    DEFAULT_CONFIG_NAME = 'written'

    def _info(self):
        if self.config.name == 'written':
            features = Features({
                'title': Value('string'),
                'topics': Sequence({
                    'posts': Sequence({
                        'text': Value('string')
                    })
                })
            })
        elif self.config.name == 'spoken':
            features = Features({
                'title': Value('string'),
                'speech': Audio(sampling_rate = 48_000),
                'topics': Sequence({
                    'posts': Sequence({
                        'text': Value('string')
                    })
                })
            })
        else:
            raise ValueError(f'Unknown config: {self.config.name}')

        return DatasetInfo(
            description=_DESCRIPTION,
            features = features,
            homepage=_HOMEPAGE,
            license=_LICENSE
        )

    def _split_generators(self, dl_manager):
        name = self.config.name

        url = _URLS['written']
        spoken_url = _URLS['spoken'] if name == 'spoken' else None

        offset = 0

        written = {}
        spoken = None if spoken_url is None else {}

        while offset < _N_ITEMS:
            cluster = _CLUSTER.format(first_page = offset, last_page = (offset := min(offset + _N_BATCH - 1, _N_ITEMS)))
            written[f'threads/{cluster}'] = dl_manager.download_and_extract(url.format(cluster = cluster))
            if spoken is not None:
                try:
                    spoken[f'threads/{cluster}'] = dl_manager.download_and_extract(spoken_url.format(cluster = cluster))
                except:  # speech for some clusters may be missing
                    pass

        index = dl_manager.download_and_extract(_INDEX)

        # print(clusters)
        # print(index)

        return [
            SplitGenerator(
                name = Split.TRAIN,
                gen_kwargs = {
                    'written': written,
                    'spoken': spoken,
                    'index': index
                }
            )
        ]

    def _generate_examples(self, written: dict, index: str, spoken: dict = None):
        for i, row in read_csv(index, sep = '\t').iterrows():
            # print(row)

            path = os.path.join(written[row['path']], f'{row["thread"]}.txt')

            topics = []
            posts = []

            # def append_topic():
            #     nonlocal posts, topics

            #     if len(posts) > 0:
            #         topics.append({'posts': posts})
            #         posts = []

            with open(path, 'r', encoding = 'utf-8') as file:
                for line in file.read().split('\n'):
                    if line:
                        posts.append({'text': line})
                    # else:
                    #     append_topic()
                    elif len(posts) > 0:
                        topics.append({'posts': posts})
                        posts = []

                # append_topic()

            item = {
                'title': row['title'],
                'topics': topics
            }

            if spoken is not None:
                speech_cluster_path = spoken.get(row['path'])

                if speech_cluster_path is not None:
                    speech_file_path = os.path.join(speech_cluster_path, f'{row["thread"]}.mp3')

                    if os.path.isfile(speech_file_path):
                        item['speech'] = speech_file_path

            yield i, item

            # if sound is None:
            #     yield i, dict(row)
            # else:
            #     data = dict(row)

            #     folder = data['folder']
            #     filename = data['filename']

            #     if folder == folder and filename == filename:  # if folder and filename are not nan
            #         data['sound'] = os.path.join(sound, folder, f'{filename}.ogg')
            #     else:
            #         data['sound'] = NA

            #     data.pop('folder')
            #     data.pop('filename')

            #     yield i, data