datamodule.py

import abc
from typing import (
    Any,
    Callable,
    Dict,
    Generic,
    List,
    Optional,
    Sequence,
    Tuple,
    Type,
    TypeVar,
    Union,
)

import datasets
import pytorch_lightning as pl
from datasets import ClassLabel, Dataset, DatasetDict
from datasets import Sequence as HFSequence
from torch.utils.data import DataLoader
from torch.utils.data.dataset import Dataset as TorchDataset
from transformers import AutoTokenizer, BatchEncoding

from embeddings.data import dataset as embeddings_dataset
from embeddings.data.data_collator import CustomDataCollatorForTokenClassification
from embeddings.data.data_loader import (
    HuggingFaceDataLoader,
    HuggingFaceLocalDataLoader,
    get_hf_dataloader,
)
from embeddings.data.dataset import LightingDataLoaders, LightingDataModuleSubset
from embeddings.data.io import T_path
from embeddings.utils.loggers import get_logger
from embeddings.utils.utils import initialize_kwargs

Data = TypeVar("Data")
HuggingFaceDataset = Type[Dataset]
_logger = get_logger(__name__)


class TorchFromHuggingFaceDataset(TorchDataset[HuggingFaceDataset]):
    def __init__(self, dataset: Dataset):
        self.dataset = dataset

    def __getitem__(self, index: int) -> Any:
        return self.dataset[index]

    def __len__(self) -> int:
        return len(self.dataset)


class BaseDataModule(abc.ABC, pl.LightningDataModule, Generic[Data]):
    dataset: Data

    def __init__(self, **kwargs: Any) -> None:
        super().__init__()
        self.save_hyperparameters()


class HuggingFaceDataModule(BaseDataModule[DatasetDict]):
    LOADER_COLUMNS = [
        "datasets_idx",
        "input_ids",
        "token_type_ids",
        "attention_mask",
        "start_positions",
        "end_positions",
        "labels",
    ]

    def __init__(
        self,
        dataset_name_or_path: T_path,
        tokenizer_name_or_path: T_path,
        target_field: str,
        max_seq_length: Optional[int],
        train_batch_size: int,
        eval_batch_size: int,
        processing_batch_size: Optional[int] = None,
        tokenizer_kwargs: Optional[Dict[str, Any]] = None,
        load_dataset_kwargs: Optional[Dict[str, Any]] = None,
        dataloader_kwargs: Optional[Dict[str, Any]] = None,
        seed: int = 441,
    ) -> None:
        self.has_setup = False
        self.dataset_name_or_path = dataset_name_or_path
        self.tokenizer_name_or_path = tokenizer_name_or_path
        self.target_field = target_field
        self.max_seq_length = max_seq_length
        self.train_batch_size = train_batch_size
        self.eval_batch_size = eval_batch_size
        self.processing_batch_size = processing_batch_size
        self.tokenizer = AutoTokenizer.from_pretrained(
            self.tokenizer_name_or_path,
            **tokenizer_kwargs if tokenizer_kwargs else {},
        )
        self.load_dataset_kwargs = load_dataset_kwargs if load_dataset_kwargs else {}
        self.dataloader_kwargs = dataloader_kwargs if dataloader_kwargs else {}
        self.seed = seed
        self.setup()
        dataset_info, dataset_version = self._parse_dataset_info()
        super().__init__(
            dataset_info=dataset_info,
            dataset_version=dataset_version,
        )

    def _parse_dataset_info(self) -> Tuple[str, str]:
        assert isinstance(self.dataset, datasets.DatasetDict)
        keys = list(self.dataset.keys())
        dataset_info = self.dataset[keys[0]].info
        dataset_version = self.dataset[keys[0]].info.version.version_str

        if not dataset_info:
            dataset_info = ""
        if not dataset_version:
            dataset_version = ""

        return dataset_info, dataset_version

    def predict_dataloader(
        self,
    ) -> Union[DataLoader[HuggingFaceDataset], Sequence[DataLoader[HuggingFaceDataset]]]:
        return [
            DataLoader(
                dataset=TorchFromHuggingFaceDataset(self.dataset[split]),
                batch_size=self.eval_batch_size,
                collate_fn=self.collate_fn,
                shuffle=False,
                **self.dataloader_kwargs,
            )
            for split in self.splits
        ]

    @abc.abstractmethod
    def prepare_labels(self) -> None:
        pass

    @abc.abstractmethod
    def _class_encode_column(self, column_name: str) -> None:
        pass

    @abc.abstractmethod
    def convert_to_features(
        self, example_batch: Dict[str, Any], indices: Optional[List[int]] = None
    ) -> BatchEncoding:
        pass

    def prepare_data(self) -> None:
        AutoTokenizer.from_pretrained(self.tokenizer_name_or_path)

    def setup(self, stage: Optional[str] = None) -> None:
        if not self.has_setup:
            self.dataset = self.load_dataset()
            self.prepare_labels()
            self.process_data()
            self.has_setup = True
        assert all(hasattr(self, attr) for attr in ["num_classes", "target_names", "dataset"])

    def load_dataset(self, preparation_step: bool = False) -> DatasetDict:
        dataset = embeddings_dataset.Dataset(
            str(self.dataset_name_or_path), **self.load_dataset_kwargs
        )
        loader: Union[HuggingFaceDataLoader, HuggingFaceLocalDataLoader] = get_hf_dataloader(
            dataset
        )

        if isinstance(loader, HuggingFaceLocalDataLoader) and preparation_step:
            return datasets.DatasetDict()

        return loader.load(dataset)

    def process_data(self) -> None:
        columns = [c for c in self.dataset["train"].column_names if c not in self.LOADER_COLUMNS]
        self.dataset = self.dataset.map(
            self.convert_to_features,
            batched=True,
            batch_size=self.processing_batch_size,
            remove_columns=columns,
        )
        self._class_encode_column("labels")
        self.dataset.set_format(type="torch")

    def train_dataloader(self) -> DataLoader[HuggingFaceDataset]:
        return DataLoader(
            dataset=self.dataset["train"],
            batch_size=self.train_batch_size,
            collate_fn=self.collate_fn,
            shuffle=True,
            **self.dataloader_kwargs,
        )

    def val_dataloader(self) -> Union[DataLoader[HuggingFaceDataset], List[Any]]:
        if "validation" in self.dataset:
            return DataLoader(
                dataset=self.dataset["validation"],
                batch_size=self.eval_batch_size,
                collate_fn=self.collate_fn,
                shuffle=False,
                **self.dataloader_kwargs,
            )
        else:
            return []

    def test_dataloader(self) -> DataLoader[HuggingFaceDataset]:
        return DataLoader(
            dataset=self.dataset["test"],
            batch_size=self.eval_batch_size,
            collate_fn=self.collate_fn,
            shuffle=False,
            **self.dataloader_kwargs,
        )

    def get_subset(
        self, subset: Union[str, LightingDataModuleSubset]
    ) -> Union[LightingDataLoaders, None]:
        if subset == "train":
            return self.train_dataloader()
        elif subset in ("dev", "validation"):
            return self.val_dataloader()
        elif subset == "test":
            return self.test_dataloader()
        elif subset == "predict":
            raise NotImplementedError("Predict subset not available in HuggingFaceDataModule")
        else:
            raise ValueError("Unrecognized LightingDataModuleSubset")

    @property
    def collate_fn(self) -> Optional[Callable[[Any], Any]]:
        return None


class TextClassificationDataModule(HuggingFaceDataModule):
    DEFAULT_BATCH_ENCODING_KWARGS = {
        "padding": True,
        "truncation": True,
    }

    def __init__(
        self,
        dataset_name_or_path: T_path,
        tokenizer_name_or_path: T_path,
        text_fields: Union[str, Sequence[str]],
        target_field: str,
        max_seq_length: Optional[int],
        train_batch_size: int,
        eval_batch_size: int,
        tokenizer_kwargs: Optional[Dict[str, Any]] = None,
        batch_encoding_kwargs: Optional[Dict[str, Any]] = None,
        load_dataset_kwargs: Optional[Dict[str, Any]] = None,
        dataloader_kwargs: Optional[Dict[str, Any]] = None,
        **kwargs: Any,
    ):
        if isinstance(text_fields, str):
            text_fields = [text_fields]
        if len(text_fields) > 2:
            raise ValueError("Too many fields given in text_fields attribute")
        self.text_fields = text_fields
        self.batch_encoding_kwargs = initialize_kwargs(
            self.DEFAULT_BATCH_ENCODING_KWARGS, batch_encoding_kwargs
        )
        super().__init__(
            dataset_name_or_path=dataset_name_or_path,
            tokenizer_name_or_path=tokenizer_name_or_path,
            target_field=target_field,
            max_seq_length=max_seq_length,
            train_batch_size=train_batch_size,
            eval_batch_size=eval_batch_size,
            tokenizer_kwargs=tokenizer_kwargs,
            load_dataset_kwargs=load_dataset_kwargs,
            dataloader_kwargs=dataloader_kwargs,
            **kwargs,
        )

    def prepare_labels(self) -> None:
        assert isinstance(self.dataset, DatasetDict)
        if not isinstance(self.dataset["train"].features[self.target_field], ClassLabel):
            self._class_encode_column(self.target_field)
        self.num_classes = self.dataset["train"].features[self.target_field].num_classes
        self.target_names = self.dataset["train"].features[self.target_field].names

    def _class_encode_column(self, column_name: str) -> None:
        if not hasattr(self, "num_classes") or not hasattr(self, "target_names"):
            self.dataset = self.dataset.class_encode_column(column_name)
        else:
            new_features = self.dataset["train"].features.copy()
            new_features[column_name] = ClassLabel(names=self.target_names)
            self.dataset = self.dataset.cast(new_features)

    def convert_to_features(
        self, example_batch: Dict[str, Any], indices: Optional[List[int]] = None
    ) -> BatchEncoding:
        """Encodes either single sentence or sentence pairs."""
        if len(self.text_fields) == 2:
            texts_or_text_pairs = list(
                zip(example_batch[self.text_fields[0]], example_batch[self.text_fields[1]])
            )
        elif len(self.text_fields) == 1:
            texts_or_text_pairs = example_batch[self.text_fields[0]]
        else:
            raise ValueError("Inappropriate length of text_fields attribute")

        features = self.tokenizer(
            texts_or_text_pairs,
            max_length=self.max_seq_length,
            **self.batch_encoding_kwargs,
        )

        features["labels"] = example_batch[self.target_field]

        return features


class SequenceLabelingDataModule(HuggingFaceDataModule):
    IGNORE_INDEX = -100
    DEFAULT_BATCH_ENCODING_KWARGS = {
        "padding": True,
        "truncation": True,
        "is_split_into_words": True,
        "return_offsets_mapping": True,
    }

    def __init__(
        self,
        dataset_name_or_path: T_path,
        tokenizer_name_or_path: T_path,
        text_field: str,
        target_field: str,
        max_seq_length: Optional[int],
        train_batch_size: int,
        eval_batch_size: int,
        label_all_tokens: bool = False,
        tokenizer_kwargs: Optional[Dict[str, Any]] = None,
        batch_encoding_kwargs: Optional[Dict[str, Any]] = None,
        load_dataset_kwargs: Optional[Dict[str, Any]] = None,
        dataloader_kwargs: Optional[Dict[str, Any]] = None,
        **kwargs: Any,
    ):
        self.text_field = text_field
        self.label_all_tokens = label_all_tokens
        self.batch_encoding_kwargs = initialize_kwargs(
            self.DEFAULT_BATCH_ENCODING_KWARGS, batch_encoding_kwargs
        )
        super().__init__(
            dataset_name_or_path=dataset_name_or_path,
            tokenizer_name_or_path=tokenizer_name_or_path,
            target_field=target_field,
            max_seq_length=max_seq_length,
            train_batch_size=train_batch_size,
            eval_batch_size=eval_batch_size,
            tokenizer_kwargs=tokenizer_kwargs,
            load_dataset_kwargs=load_dataset_kwargs,
            dataloader_kwargs=dataloader_kwargs,
            **kwargs,
        )

    def prepare_labels(self) -> None:
        if not isinstance(self.dataset["train"].features[self.target_field].feature, ClassLabel):
            raise TypeError(
                "Target field has inappropiate type; datasets.Sequence(datasets.features.ClassLabel(...)) is required for Sequence Labeling task"
            )
        else:
            self.num_classes = self.dataset["train"].features[self.target_field].feature.num_classes
            self.target_names = self.dataset["train"].features[self.target_field].feature.names

    def convert_to_features(
        self, example_batch: Dict[str, Any], indices: Optional[List[int]] = None
    ) -> BatchEncoding:
        texts = example_batch[self.text_field]
        features = self.tokenizer.batch_encode_plus(
            texts, max_length=self.max_seq_length, **self.batch_encoding_kwargs
        )
        labels = self.encode_tags(labels=example_batch[self.target_field], encodings=features)
        features["labels"] = labels
        features.pop("offset_mapping")
        return features

    def encode_tags(self, labels: List[List[int]], encodings: BatchEncoding) -> List[List[int]]:
        """Encode tags to fix mismatch caused by token split into multiple subtokens by tokenizer.
        Special tokens have a word id that is None.
        We set the label to -100 so they are automatically ignored in the loss function.
        We set the label for the first token of each word.
        For the other tokens in a word, we set the label to either the current label or -100,
        depending on the label_all_tokens flag.
        Source: https://github.com/PyTorchLightning/lightning-transformers/blob/fc4703498a057476205dd4e518f8fcd09654c31b/lightning_transformers/task/nlp/token_classification/data.py
        """
        encoded_labels = []
        for i, label in enumerate(labels):
            word_ids = encodings.word_ids(batch_index=i)
            previous_word_idx = None
            label_ids = []
            for word_idx in word_ids:
                if word_idx is None:  # special token = -100 (IGNORE_INDEX)
                    label_ids.append(self.IGNORE_INDEX)
                elif word_idx != previous_word_idx:  # first token of the word = label
                    label_ids.append(label[word_idx])
                else:  # other tokens = label or -100 depending on the label_all_tokens flag
                    label_ids.append(
                        label[word_idx] if self.label_all_tokens else self.IGNORE_INDEX
                    )
                previous_word_idx = word_idx
            encoded_labels.append(label_ids)
        return encoded_labels

    def _class_encode_column(self, column_name: str) -> None:
        new_features = self.dataset["train"].features.copy()
        new_features[column_name] = HFSequence(feature=ClassLabel(names=self.target_names))
        self.dataset = self.dataset.cast(new_features)

    @property
    def collate_fn(self) -> Optional[Callable[[Any], Any]]:
        if self.processing_batch_size and self.processing_batch_size > 0:
            data_collator = CustomDataCollatorForTokenClassification(tokenizer=self.tokenizer)
            assert callable(data_collator)
            return data_collator
        return None

    def id2str(self, int_: int) -> str:
        if self.dataset is None:
            raise AttributeError("Dataset has not been setup")
        str_ = self.dataset["train"].features["labels"].feature.int2str(int_)
        assert isinstance(str_, str)
        return str_

    def str2id(self, str_: str) -> int:
        if self.dataset is None:
            raise AttributeError("Dataset has not been setup")
        int_ = self.dataset["train"].features["labels"].feature.str2int(str_)
        assert isinstance(int_, int)
        return int_