russian_second_level.py

"""Russian Literary Dataset from late 19th century up to early 20th century."""

import json
import os
import warnings
from typing import Dict, List, Tuple

import datasets
import numpy as np
from transformers import PreTrainedTokenizerBase

_DESCRIPTION = """Second level categorization of Russian articles."""

_HOMEPAGE = ""

_LICENSE = ""

_NAMES = [
    "cultural_discourse_subject",
    "cultural_discourse_type",
    "literary_text_type",
]

# Label information as `(num_labels, is_multi_label)` tuples
_LABELS: Dict[str, Tuple[int, bool]] = {
    "cultural_discourse_subject": (7, True),
    "cultural_discourse_type": (7, True),
    "literary_text_type": (5, False),
}


def generate_urls(name: str) -> Dict[str, str]:
    return {
        "train": os.path.join(name, "train.json"),
        "val": os.path.join(name, "val.json"),
        "test": os.path.join(name, "test.json"),
    }


class NonwestlitSecondLevelConfig(datasets.BuilderConfig):
    """BuilderConfig for Dataset."""

    def __init__(
        self, tokenizer: PreTrainedTokenizerBase = None, max_sequence_length: int = None, **kwargs
    ):
        """BuilderConfig for Dataset.

        Args:
            **kwargs: keyword arguments forwarded to super.
        """
        super(NonwestlitSecondLevelConfig, self).__init__(**kwargs)
        self.tokenizer = tokenizer
        self.max_sequence_length = max_sequence_length

    @property
    def features(self):
        if self.name == "literary_text_type":
            labels = datasets.Value("uint8")
        else:
            labels = datasets.Sequence(datasets.Value("uint8"))
        return {
            "labels": labels,
            "input_ids": datasets.Value("string"),
            "title": datasets.Value("string"),
            "iid": datasets.Value("uint32"),
            "chunk_id": datasets.Value("uint32"),
        }


class NonwestlitSecondLevelDataset(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("0.0.1")

    BUILDER_CONFIGS = [
        NonwestlitSecondLevelConfig(name=name, version=version, description=name)
        for name, version in zip(_NAMES, [VERSION] * len(_NAMES))
    ]
    BUILDER_CONFIG_CLASS = NonwestlitSecondLevelConfig
    __current_id = 1
    __current_chunk_id = 1

    @property
    def __next_id(self):
        cid = self.__current_id
        self.__current_id += 1
        return cid

    @property
    def __next_chunk_id(self):
        cid = self.__current_chunk_id
        self.__current_chunk_id += 1
        return cid

    @property
    def label_info(self) -> Tuple[int, bool]:
        return _LABELS[self.config.name]

    def __reset_chunk_id(self):
        self.__current_chunk_id = 1

    def _info(self):
        if self.config.tokenizer is None:
            raise RuntimeError(
                "For HF Datasets and for chunking to be carried out, 'tokenizer' must be given."
            )
        if "llama" in self.config.tokenizer.name_or_path:
            warnings.warn(
                "It is suggested to pass 'max_sequence_length' argument for Llama-2 model family. There "
                "might be errors for the data processing parts as `model_max_len` attributes are set to"
                "MAX_INT64 (?)."
            )
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(self.config.features),
        )

    def _split_generators(self, dl_manager):
        urls = generate_urls(self.config.name)
        data_dir = dl_manager.download_and_extract(urls)
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN, gen_kwargs={"filepath": data_dir["train"]}
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION, gen_kwargs={"filepath": data_dir["val"]}
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST, gen_kwargs={"filepath": data_dir["test"]}
            ),
        ]

    def prepare_articles(self, article: str) -> List[str]:
        tokenizer = self.config.tokenizer
        model_inputs = tokenizer(
            article,
            truncation=True,
            padding=True,
            max_length=self.config.max_sequence_length,
            return_overflowing_tokens=True,
        )
        return tokenizer.batch_decode(model_inputs["input_ids"], skip_special_tokens=True)

    def _to_one_hot(self, labels: List[int], num_labels: int) -> List[int]:
        x = np.zeros(num_labels, dtype=np.float16)
        x[labels] = 1.0
        return x.tolist()

    # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
    def _generate_examples(self, filepath: str):
        with open(filepath, encoding="utf-8") as f:
            dataset = json.load(f)

        num_labels, multi_label = self.label_info
        chunk_id = 0
        for instance in dataset:
            iid = instance.get("id", self.__next_id)
            label = instance.get("label")
            if label is None:
                if not multi_label:
                    continue
                else:
                    label = self._to_one_hot(labels=[], num_labels=num_labels)
            elif isinstance(label, int):
                label = int(label) - 1
            elif isinstance(label, str):
                if multi_label:
                    label = [int(l) - 1 for l in label.split(",")]
                    label = self._to_one_hot(label, num_labels)
                else:
                    label = int(label) - 1

            article = self.prepare_articles(instance["article"])
            self.__reset_chunk_id()
            for chunk in article:
                chunk_inputs = {
                    "iid": iid,
                    "chunk_id": self.__next_chunk_id,
                    "title": instance["title"],
                    "input_ids": chunk,
                    "labels": label,
                }
                yield chunk_id, chunk_inputs
                chunk_id += 1