iwslt2016.py

import os

from torchtext._internal.module_utils import is_module_available
from torchtext.data.datasets_utils import (
    _wrap_split_argument,
    _clean_files,
    _create_dataset_directory,
    _generate_iwslt_files_for_lang_and_split,
)

if is_module_available("torchdata"):
    from torchdata.datapipes.iter import FileOpener, GDriveReader, IterableWrapper


URL = "https://drive.google.com/uc?id=1l5y6Giag9aRPwGtuZHswh3w5v3qEz8D8"

_PATH = "2016-01.tgz"

MD5 = "c393ed3fc2a1b0f004b3331043f615ae"

SUPPORTED_DATASETS = {
    "valid_test": ["dev2010", "tst2010", "tst2011", "tst2012", "tst2013", "tst2014"],
    "language_pair": {
        "en": ["ar", "de", "fr", "cs"],
        "ar": ["en"],
        "fr": ["en"],
        "de": ["en"],
        "cs": ["en"],
    },
    "year": 16,
}

NUM_LINES = {
    "train": {
        "train": {
            ("ar", "en"): 224126,
            ("de", "en"): 196884,
            ("en", "fr"): 220400,
            ("cs", "en"): 114390,
        }
    },
    "valid": {
        "dev2010": {
            ("ar", "en"): 887,
            ("de", "en"): 887,
            ("en", "fr"): 887,
            ("cs", "en"): 480,
        },
        "tst2010": {
            ("ar", "en"): 1569,
            ("de", "en"): 1565,
            ("en", "fr"): 1664,
            ("cs", "en"): 1511,
        },
        "tst2011": {
            ("ar", "en"): 1199,
            ("de", "en"): 1433,
            ("en", "fr"): 818,
            ("cs", "en"): 1013,
        },
        "tst2012": {
            ("ar", "en"): 1702,
            ("de", "en"): 1700,
            ("en", "fr"): 1124,
            ("cs", "en"): 1385,
        },
        "tst2013": {
            ("ar", "en"): 1169,
            ("de", "en"): 993,
            ("en", "fr"): 1026,
            ("cs", "en"): 1327,
        },
        "tst2014": {("ar", "en"): 1107, ("de", "en"): 1305, ("en", "fr"): 1305},
    },
    "test": {
        "dev2010": {
            ("ar", "en"): 887,
            ("de", "en"): 887,
            ("en", "fr"): 887,
            ("cs", "en"): 480,
        },
        "tst2010": {
            ("ar", "en"): 1569,
            ("de", "en"): 1565,
            ("en", "fr"): 1664,
            ("cs", "en"): 1511,
        },
        "tst2011": {
            ("ar", "en"): 1199,
            ("de", "en"): 1433,
            ("en", "fr"): 818,
            ("cs", "en"): 1013,
        },
        "tst2012": {
            ("ar", "en"): 1702,
            ("de", "en"): 1700,
            ("en", "fr"): 1124,
            ("cs", "en"): 1385,
        },
        "tst2013": {
            ("ar", "en"): 1169,
            ("de", "en"): 993,
            ("en", "fr"): 1026,
            ("cs", "en"): 1327,
        },
        "tst2014": {("ar", "en"): 1107, ("de", "en"): 1305, ("en", "fr"): 1305},
    },
}

SET_NOT_EXISTS = {
    ("en", "ar"): [],
    ("en", "de"): [],
    ("en", "fr"): [],
    ("en", "cs"): ["tst2014"],
    ("ar", "en"): [],
    ("fr", "en"): [],
    ("de", "en"): [],
    ("cs", "en"): ["tst2014"],
}

DATASET_NAME = "IWSLT2016"


# TODO: migrate this to dataset_utils.py once torchdata is a hard dependency to
# avoid additional conditional imports.
def _filter_clean_cache(cache_decompressed_dp, full_filepath, uncleaned_filename):
    cache_inner_decompressed_dp = cache_decompressed_dp.on_disk_cache(
        filepath_fn=lambda x: full_filepath
    )
    cache_inner_decompressed_dp = FileOpener(
        cache_inner_decompressed_dp, mode="b"
    ).read_from_tar()
    cache_inner_decompressed_dp = cache_inner_decompressed_dp.filter(
        lambda x: os.path.basename(uncleaned_filename) in x[0]
    )
    cache_inner_decompressed_dp = cache_inner_decompressed_dp.map(
        lambda x: _clean_files(full_filepath, x[0], x[1])
    )
    cache_inner_decompressed_dp = cache_inner_decompressed_dp.end_caching(
        mode="wb", same_filepath_fn=True
    )
    return cache_inner_decompressed_dp


@_create_dataset_directory(dataset_name=DATASET_NAME)
@_wrap_split_argument(("train", "valid", "test"))
def IWSLT2016(
    root=".data",
    split=("train", "valid", "test"),
    language_pair=("de", "en"),
    valid_set="tst2013",
    test_set="tst2014",
):
    """IWSLT2016 dataset

    For additional details refer to https://wit3.fbk.eu/2016-01

    The available datasets include following:

    **Language pairs**:

    +-----+-----+-----+-----+-----+-----+
    |     |"en" |"fr" |"de" |"cs" |"ar" |
    +-----+-----+-----+-----+-----+-----+
    |"en" |     |   x |  x  |  x  |  x  |
    +-----+-----+-----+-----+-----+-----+
    |"fr" |  x  |     |     |     |     |
    +-----+-----+-----+-----+-----+-----+
    |"de" |  x  |     |     |     |     |
    +-----+-----+-----+-----+-----+-----+
    |"cs" |  x  |     |     |     |     |
    +-----+-----+-----+-----+-----+-----+
    |"ar" |  x  |     |     |     |     |
    +-----+-----+-----+-----+-----+-----+

    **valid/test sets**: ["dev2010", "tst2010", "tst2011", "tst2012", "tst2013", "tst2014"]


    Args:
        root: Directory where the datasets are saved. Default: os.path.expanduser('~/.torchtext/cache')
        split: split or splits to be returned. Can be a string or tuple of strings. Default: (‘train’, ‘valid’, ‘test’)
        language_pair: tuple or list containing src and tgt language
        valid_set: a string to identify validation set.
        test_set: a string to identify test set.

    :return: DataPipe that yields tuple of source and target sentences
    :rtype: (str, str)

    Examples:
        >>> from torchtext.datasets import IWSLT2016
        >>> train_iter, valid_iter, test_iter = IWSLT2016()
        >>> src_sentence, tgt_sentence = next(iter(train_iter))

    """
    if not is_module_available("torchdata"):
        raise ModuleNotFoundError(
            "Package `torchdata` not found. Please install following instructions at `https://github.com/pytorch/data`"
        )

    if not isinstance(language_pair, list) and not isinstance(language_pair, tuple):
        raise ValueError(
            "language_pair must be list or tuple but got {} instead".format(
                type(language_pair)
            )
        )

    assert (
        len(language_pair) == 2
    ), "language_pair must contain only 2 elements: src and tgt language respectively"

    src_language, tgt_language = language_pair[0], language_pair[1]

    if src_language not in SUPPORTED_DATASETS["language_pair"]:
        raise ValueError(
            "src_language '{}' is not valid. Supported source languages are {}".format(
                src_language, list(SUPPORTED_DATASETS["language_pair"])
            )
        )

    if tgt_language not in SUPPORTED_DATASETS["language_pair"][src_language]:
        raise ValueError(
            "tgt_language '{}' is not valid for give src_language '{}'. Supported target language are {}".format(
                tgt_language,
                src_language,
                SUPPORTED_DATASETS["language_pair"][src_language],
            )
        )

    if (
        valid_set not in SUPPORTED_DATASETS["valid_test"]
        or valid_set in SET_NOT_EXISTS[language_pair]
    ):
        raise ValueError(
            "valid_set '{}' is not valid for given language pair {}. Supported validation sets are {}".format(
                valid_set,
                language_pair,
                [
                    s
                    for s in SUPPORTED_DATASETS["valid_test"]
                    if s not in SET_NOT_EXISTS[language_pair]
                ],
            )
        )

    if (
        test_set not in SUPPORTED_DATASETS["valid_test"]
        or test_set in SET_NOT_EXISTS[language_pair]
    ):
        raise ValueError(
            "test_set '{}' is not valid for give language pair {}. Supported test sets are {}".format(
                valid_set,
                language_pair,
                [
                    s
                    for s in SUPPORTED_DATASETS["valid_test"]
                    if s not in SET_NOT_EXISTS[language_pair]
                ],
            )
        )

    (
        file_path_by_lang_and_split,
        uncleaned_filenames_by_lang_and_split,
    ) = _generate_iwslt_files_for_lang_and_split(
        SUPPORTED_DATASETS["year"], src_language, tgt_language, valid_set, test_set
    )

    url_dp = IterableWrapper([URL])
    cache_compressed_dp = url_dp.on_disk_cache(
        filepath_fn=lambda x: os.path.join(root, _PATH),
        hash_dict={os.path.join(root, _PATH): MD5},
        hash_type="md5",
    )
    cache_compressed_dp = GDriveReader(cache_compressed_dp)
    cache_compressed_dp = cache_compressed_dp.end_caching(
        mode="wb", same_filepath_fn=True
    )

    languages = "-".join([src_language, tgt_language])

    # We create the whole filepath here, but only check for the literal filename in the filter
    # because we're lazily extracting from the outer tarfile. Thus,
    # /root/2016-01/texts/.../src-tgt.tgz will never be in /root/2016-01.tgz/texts/.../src-tgt.tgz
    inner_iwslt_tar = (
        os.path.join(
            root,
            os.path.splitext(_PATH)[0],
            "texts",
            src_language,
            tgt_language,
            languages,
        )
        + ".tgz"
    )

    cache_decompressed_dp = cache_compressed_dp.on_disk_cache(
        filepath_fn=lambda x: inner_iwslt_tar
    )
    cache_decompressed_dp = (
        FileOpener(cache_decompressed_dp, mode="b")
        .read_from_tar()
        .filter(lambda x: os.path.basename(inner_iwslt_tar) in x[0])
    )
    cache_decompressed_dp = cache_decompressed_dp.end_caching(
        mode="wb", same_filepath_fn=True
    )

    src_filename = file_path_by_lang_and_split[src_language][split]
    uncleaned_src_filename = uncleaned_filenames_by_lang_and_split[src_language][split]

    # We create the whole filepath here, but only check for the literal filename in the filter
    # because we're lazily extracting from the outer tarfile.
    full_src_filepath = os.path.join(
        root, "2016-01/texts/", src_language, tgt_language, languages, src_filename
    )

    cache_inner_src_decompressed_dp = _filter_clean_cache(
        cache_decompressed_dp, full_src_filepath, uncleaned_src_filename
    )

    tgt_filename = file_path_by_lang_and_split[tgt_language][split]
    uncleaned_tgt_filename = uncleaned_filenames_by_lang_and_split[tgt_language][split]

    # We create the whole filepath here, but only check for the literal filename in the filter
    # because we're lazily extracting from the outer tarfile.
    full_tgt_filepath = os.path.join(
        root, "2016-01/texts/", src_language, tgt_language, languages, tgt_filename
    )

    cache_inner_tgt_decompressed_dp = _filter_clean_cache(
        cache_decompressed_dp, full_tgt_filepath, uncleaned_tgt_filename
    )

    tgt_data_dp = FileOpener(cache_inner_tgt_decompressed_dp, mode="r")
    src_data_dp = FileOpener(cache_inner_src_decompressed_dp, mode="r")

    src_lines = src_data_dp.readlines(return_path=False, strip_newline=False)
    tgt_lines = tgt_data_dp.readlines(return_path=False, strip_newline=False)

    return src_lines.zip(tgt_lines)