Skip to content

Add support for WNLI dataset with unit tests #1724

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Jun 1, 2022
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions docs/source/datasets.rst
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,11 @@ STSB

.. autofunction:: STSB

WNLI
~~~~

.. autofunction:: WNLI

YahooAnswers
~~~~~~~~~~~~

Expand Down
84 changes: 84 additions & 0 deletions test/datasets/test_wnli.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,84 @@
import os
import zipfile
from collections import defaultdict
from unittest.mock import patch

from parameterized import parameterized
from torchtext.datasets.wnli import WNLI

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, "WNLI")
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 ("train.tsv", "test.tsv", "dev.tsv"):
txt_file = os.path.join(temp_dataset_dir, file_name)
with open(txt_file, "w", encoding="utf-8") as f:
f.write("index\tsentence1\tsentence2\tlabel\n")
for i in range(5):
label = seed % 2
rand_string_1 = get_random_unicode(seed)
rand_string_2 = get_random_unicode(seed + 1)
if file_name == "test.tsv":
dataset_line = (rand_string_1, rand_string_2)
f.write(f"{i}\t{rand_string_1}\t{rand_string_2}\n")
else:
dataset_line = (label, rand_string_1, rand_string_2)
f.write(f"{i}\t{rand_string_1}\t{rand_string_2}\t{label}\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, "WNLI.zip")
# create zip file from dataset folder
with zipfile.ZipFile(compressed_dataset_path, "w") as zip_file:
for file_name in ("train.tsv", "test.tsv", "dev.tsv"):
txt_file = os.path.join(temp_dataset_dir, file_name)
zip_file.write(txt_file, arcname=os.path.join("WNLI", file_name))

return mocked_data


class TestWNLI(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", "test", "dev"])
def test_wnli(self, split):
dataset = WNLI(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", "test", "dev"])
def test_wnli_split_argument(self, split):
dataset1 = WNLI(root=self.root_dir, split=split)
(dataset2,) = WNLI(root=self.root_dir, split=(split,))

for d1, d2 in zip_equal(dataset1, dataset2):
self.assertEqual(d1, d2)
2 changes: 2 additions & 0 deletions torchtext/datasets/__init__.py
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
from .udpos import UDPOS
from .wikitext103 import WikiText103
from .wikitext2 import WikiText2
from .wnli import WNLI
from .yahooanswers import YahooAnswers
from .yelpreviewfull import YelpReviewFull
from .yelpreviewpolarity import YelpReviewPolarity
Expand Down Expand Up @@ -53,6 +54,7 @@
"UDPOS": UDPOS,
"WikiText103": WikiText103,
"WikiText2": WikiText2,
"WNLI": WNLI,
"YahooAnswers": YahooAnswers,
"YelpReviewFull": YelpReviewFull,
"YelpReviewPolarity": YelpReviewPolarity,
Expand Down
100 changes: 100 additions & 0 deletions torchtext/datasets/wnli.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,100 @@
# Copyright (c) Facebook, Inc. and its affiliates.
import os
from functools import partial

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://dl.fbaipublicfiles.com/glue/data/WNLI.zip"

MD5 = "a1b4bd2861017d302d29e42139657a42"

NUM_LINES = {
"train": 635,
"dev": 71,
"test": 146,
}

_PATH = "WNLI.zip"

DATASET_NAME = "WNLI"

_EXTRACTED_FILES = {
"train": os.path.join("WNLI", "train.tsv"),
"dev": os.path.join("WNLI", "dev.tsv"),
"test": os.path.join("WNLI", "test.tsv"),
}


def _filepath_fn(root, x=None):
return os.path.join(root, os.path.basename(x))


def _extracted_filepath_fn(root, split, _=None):
return os.path.join(root, _EXTRACTED_FILES[split])


def _filter_fn(split, x):
return _EXTRACTED_FILES[split] in x[0]


def _modify_res(split, t):
if split == "test":
return (t[1], t[2])
else:
return (int(t[3]), t[1], t[2])


@_create_dataset_directory(dataset_name=DATASET_NAME)
@_wrap_split_argument(("train", "dev", "test"))
def WNLI(root, split):
"""WNLI Dataset

For additional details refer to https://arxiv.org/pdf/1804.07461v3.pdf

Number of lines per split:
- train: 635
- dev: 71
- test: 146

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`, `test`)

:returns: DataPipe that yields tuple of text and/or label (0 to 1). The `test` split only returns text.
:rtype: Union[(int, str, str), (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`"
)

url_dp = IterableWrapper([URL])
cache_compressed_dp = url_dp.on_disk_cache(
filepath_fn=partial(_filepath_fn, root),
hash_dict={_filepath_fn(root, 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=partial(_extracted_filepath_fn, root, split))
cache_decompressed_dp = (
FileOpener(cache_decompressed_dp, mode="b").load_from_zip().filter(partial(_filter_fn, split))
)
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").map(partial(_modify_res, split))
return parsed_data.shuffle().set_shuffle(False).sharding_filter()