Add support for MNLI dataset with unit tests

docs/source/datasets.rst

@@ -57,6 +57,11 @@ IMDb
 
 .. autofunction:: IMDB
 
+MNLI
+~~~~
+
+.. autofunction:: MNLI
+
 MRPC
 ~~~~
 

test/datasets/test_mnli.py

@@ -0,0 +1,83 @@
+import os
+import zipfile
+from collections import defaultdict
+from unittest.mock import patch
+
+from parameterized import parameterized
+from torchtext.datasets.mnli import MNLI
+
+from ..common.case_utils import TempDirMixin, zip_equal, get_random_unicode
+from ..common.torchtext_test_case import TorchtextTestCase
+
+
+def _get_mock_dataset(root_dir):
+    """
+    root_dir: directory to the mocked dataset
+    """
+    base_dir = os.path.join(root_dir, "MNLI")
+    temp_dataset_dir = os.path.join(base_dir, "temp_dataset_dir")
+    os.makedirs(temp_dataset_dir, exist_ok=True)
+
+    seed = 1
+    mocked_data = defaultdict(list)
+    for file_name in ["multinli_1.0_train.txt", "multinli_1.0_dev_matched.txt", "multinli_1.0_dev_mismatched.txt"]:
+        txt_file = os.path.join(temp_dataset_dir, file_name)
+        with open(txt_file, "w", encoding="utf-8") as f:
+            f.write(
+                "gold_label\tsentence1_binary_parse\tsentence2_binary_parse\tsentence1_parse\tsentence2_parse\tsentence1\tsentence2\tpromptID\tpairID\tgenre\tlabel1\tlabel2\tlabel3\tlabel4\tlabel5"
+            )
+            for i in range(5):
+                label = seed % 3
+                rand_string = get_random_unicode(seed)
+                dataset_line = (label, rand_string, rand_string)
+                f.write(
+                    f"{label}\t{rand_string}\t{rand_string}\t{rand_string}\t{rand_string}\t{rand_string}\t{rand_string}\t{i}\t{i}\t{i}\t{i}\t{i}\t{i}\t{i}\t{i}\n"
+                )
+
+                # append line to correct dataset split
+                mocked_data[os.path.splitext(file_name)[0]].append(dataset_line)
+                seed += 1
+
+    compressed_dataset_path = os.path.join(base_dir, "multinli_1.0.zip")
+    # create zip file from dataset folder
+    with zipfile.ZipFile(compressed_dataset_path, "w") as zip_file:
+        for file_name in ("multinli_1.0_train.txt", "multinli_1.0_dev_matched.txt", "multinli_1.0_dev_mismatched.txt"):
+            txt_file = os.path.join(temp_dataset_dir, file_name)
+            zip_file.write(txt_file, arcname=os.path.join("multinli_1.0", file_name))
+
+    return mocked_data
+
+
+class TestMNLI(TempDirMixin, TorchtextTestCase):
+    root_dir = None
+    samples = []
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+        cls.root_dir = cls.get_base_temp_dir()
+        cls.samples = _get_mock_dataset(cls.root_dir)
+        cls.patcher = patch("torchdata.datapipes.iter.util.cacheholder._hash_check", return_value=True)
+        cls.patcher.start()
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.patcher.stop()
+        super().tearDownClass()
+
+    @parameterized.expand(["train", "dev_matched", "dev_mismatched"])
+    def test_mnli(self, split):
+        dataset = MNLI(root=self.root_dir, split=split)
+
+        samples = list(dataset)
+        expected_samples = self.samples[split]
+        for sample, expected_sample in zip_equal(samples, expected_samples):
+            self.assertEqual(sample, expected_sample)
+
+    @parameterized.expand(["train", "dev_matched", "dev_mismatched"])
+    def test_sst2_split_argument(self, split):
+        dataset1 = MNLI(root=self.root_dir, split=split)
+        (dataset2,) = MNLI(root=self.root_dir, split=(split,))
+
+        for d1, d2 in zip_equal(dataset1, dataset2):
+            self.assertEqual(d1, d2)

torchtext/datasets/__init__.py

@@ -11,6 +11,7 @@
 from .imdb import IMDB
 from .iwslt2016 import IWSLT2016
 from .iwslt2017 import IWSLT2017
+from .mnli import MNLI
 from .mrpc import MRPC
 from .multi30k import Multi30k
 from .penntreebank import PennTreebank
@@ -38,6 +39,7 @@
     "IMDB": IMDB,
     "IWSLT2016": IWSLT2016,
     "IWSLT2017": IWSLT2017,
+    "MNLI": MNLI,
     "MRPC": MRPC,
     "Multi30k": Multi30k,
     "PennTreebank": PennTreebank,

torchtext/datasets/mnli.py

@@ -0,0 +1,98 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+import csv
+import os
+
+from torchtext._internal.module_utils import is_module_available
+from torchtext.data.datasets_utils import (
+    _create_dataset_directory,
+    _wrap_split_argument,
+)
+
+if is_module_available("torchdata"):
+    from torchdata.datapipes.iter import FileOpener, IterableWrapper
+
+    # we import HttpReader from _download_hooks so we can swap out public URLs
+    # with interal URLs when the dataset is used within Facebook
+    from torchtext._download_hooks import HttpReader
+
+
+URL = "https://cims.nyu.edu/~sbowman/multinli/multinli_1.0.zip"
+
+MD5 = "0f70aaf66293b3c088a864891db51353"
+
+NUM_LINES = {
+    "train": 392702,
+    "dev_matched": 9815,
+    "dev_mismatched": 9832,
+}
+
+_PATH = "multinli_1.0.zip"
+
+DATASET_NAME = "MNLI"
+
+_EXTRACTED_FILES = {
+    "train": "multinli_1.0_train.txt",
+    "dev_matched": "multinli_1.0_dev_matched.txt",
+    "dev_mismatched": "multinli_1.0_dev_mismatched.txt",
+}
+
+LABEL_TO_INT = {"entailment": 0, "neutral": 1, "contradiction": 2}
+
+
+@_create_dataset_directory(dataset_name=DATASET_NAME)
+@_wrap_split_argument(("train", "dev_matched", "dev_mismatched"))
+def MNLI(root, split):
+    """MNLI Dataset
+
+    For additional details refer to https://cims.nyu.edu/~sbowman/multinli/
+
+    Number of lines per split:
+        - train: 392702
+        - dev_matched: 9815
+        - dev_mismatched: 9832
+
+    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`, `dev_matched`, `dev_mismatched`)
+
+    :returns: DataPipe that yields tuple of text and label (0 to 2).
+    :rtype: Tuple[int, str, str]
+    """
+    # TODO Remove this after removing conditional dependency
+    if not is_module_available("torchdata"):
+        raise ModuleNotFoundError(
+            "Package `torchdata` not found. Please install following instructions at `https://github.com/pytorch/data`"
+        )
+
+    def _filepath_fn(x=None):
+        return os.path.join(root, os.path.basename(x))
+
+    def _extracted_filepath_fn(_=None):
+        return os.path.join(root, _EXTRACTED_FILES[split])
+
+    def _filter_fn(x):
+        return _EXTRACTED_FILES[split] in x[0]
+
+    def _filter_res(x):
+        return x[0] in LABEL_TO_INT
+
+    def _modify_res(x):
+        return (LABEL_TO_INT[x[0]], x[5], x[6])
+
+    url_dp = IterableWrapper([URL])
+    cache_compressed_dp = url_dp.on_disk_cache(
+        filepath_fn=_filepath_fn,
+        hash_dict={_filepath_fn(URL): MD5},
+        hash_type="md5",
+    )
+    cache_compressed_dp = HttpReader(cache_compressed_dp).end_caching(mode="wb", same_filepath_fn=True)
+
+    cache_decompressed_dp = cache_compressed_dp.on_disk_cache(filepath_fn=_extracted_filepath_fn)
+    cache_decompressed_dp = FileOpener(cache_decompressed_dp, mode="b").read_from_zip().filter(_filter_fn)
+    cache_decompressed_dp = cache_decompressed_dp.end_caching(mode="wb", same_filepath_fn=True)
+
+    data_dp = FileOpener(cache_decompressed_dp, encoding="utf-8")
+    parsed_data = (
+        data_dp.parse_csv(skip_lines=1, delimiter="\t", quoting=csv.QUOTE_NONE).filter(_filter_res).map(_modify_res)
+    )
+    return parsed_data.shuffle().set_shuffle(False).sharding_filter()