def IMDB(root='.data', split=('train', 'test'), offset=0):
""" Defines raw IMDB datasets.
Create supervised learning dataset: IMDB
Separately returns the raw training and test dataset
Args:
root: Directory where the datasets are saved. Default: ".data"
split: a string or tuple for the returned datasets. Default: ('train', 'test')
By default, both datasets (train, test) are generated. Users could also choose any one or two of them,
for example ('train', 'test') or just a string 'train'.
offset: the number of the starting line. Default: 0
Examples:
>>> train, test = torchtext.experimental.datasets.raw.IMDB()
"""
split_ = check_default_set(split, ('train', 'test'), 'IMDB')
dataset_tar = download_from_url(URLS['IMDB'],
root=root,
hash_value=MD5['IMDB'],
hash_type='md5')
extracted_files = extract_archive(dataset_tar)
return wrap_datasets(
tuple(
RawTextIterableDataset("IMDB",
NUM_LINES["IMDB"][item],
generate_imdb_data(item, extracted_files),
offset=offset) for item in split_), split)
def _setup_datasets(dataset_name, root, split_, offset):
split = check_default_set(split_, ('train', 'test'), dataset_name)
if dataset_name == 'AG_NEWS':
extracted_files = [
download_from_url(URLS[dataset_name][item],
root=root,
hash_value=MD5['AG_NEWS'][item],
hash_type='md5') for item in ('train', 'test')
]
else:
dataset_tar = download_from_url(URLS[dataset_name],
root=root,
hash_value=MD5[dataset_name],
hash_type='md5')
extracted_files = extract_archive(dataset_tar)
cvs_path = {}
for fname in extracted_files:
if fname.endswith('train.csv'):
cvs_path['train'] = fname
if fname.endswith('test.csv'):
cvs_path['test'] = fname
return wrap_datasets(
tuple(
RawTextIterableDataset(dataset_name,
NUM_LINES[dataset_name][item],
_create_data_from_csv(cvs_path[item]),
offset=offset) for item in split), split_)
def _setup_datasets(dataset_name, tokenizer, root, vocab, split_, year, language):
if tokenizer is None:
tokenizer = get_tokenizer('basic_english')
split = check_default_set(split_, ('train', 'test', 'valid'), dataset_name)
if vocab is None:
if 'train' not in split:
raise TypeError("Must pass a vocab if train is not selected.")
if dataset_name == 'WMTNewsCrawl':
raw_train, = raw.DATASETS[dataset_name](root=root, split=('train',), year=year, language=language)
else:
raw_train, = raw.DATASETS[dataset_name](root=root, split=('train',))
logger_.info('Building Vocab based on train data')
vocab = build_vocab(raw_train, tokenizer)
logger_.info('Vocab has %d entries', len(vocab))
def text_transform(line):
return torch.tensor([vocab[token] for token in tokenizer(line)], dtype=torch.long)
if dataset_name == 'WMTNewsCrawl':
raw_datasets = raw.DATASETS[dataset_name](root=root, split=split, year=year, language=language)
else:
raw_datasets = raw.DATASETS[dataset_name](root=root, split=split)
raw_data = {name: list(map(text_transform, raw_dataset)) for name, raw_dataset in zip(split, raw_datasets)}
logger_.info('Building datasets for {}'.format(split))
return wrap_datasets(tuple(LanguageModelingDataset(raw_data[item], vocab, text_transform)
for item in split), split_)
def _setup_datasets(dataset_name, separator, root, split_, offset):
split = check_default_set(split_, ('train', 'valid', 'test'), dataset_name)
extracted_files = []
if isinstance(URLS[dataset_name], dict):
for name, item in URLS[dataset_name].items():
dataset_tar = download_from_url(item,
root=root,
hash_value=MD5[dataset_name][name],
hash_type='md5')
extracted_files.extend(extract_archive(dataset_tar))
elif isinstance(URLS[dataset_name], str):
dataset_tar = download_from_url(URLS[dataset_name],
root=root,
hash_value=MD5[dataset_name],
hash_type='md5')
extracted_files.extend(extract_archive(dataset_tar))
else:
raise ValueError(
"URLS for {} has to be in a form of dictionary or string".format(
dataset_name))
data_filenames = {
"train": _construct_filepath(extracted_files, "train.txt"),
"valid": _construct_filepath(extracted_files, "dev.txt"),
"test": _construct_filepath(extracted_files, "test.txt")
}
return wrap_datasets(
tuple(
RawTextIterableDataset(
dataset_name,
NUM_LINES[dataset_name][item],
_create_data_from_iob(data_filenames[item], separator),
offset=offset) if data_filenames[item] is not None else None
for item in split), split_)
def _setup_datasets(dataset_name, root, ngrams, vocab, tokenizer, split_):
text_transform = []
if tokenizer is None:
tokenizer = get_tokenizer("basic_english")
text_transform = sequential_transforms(tokenizer, ngrams_func(ngrams))
split = check_default_set(split_, ('train', 'test'), dataset_name)
raw_datasets = raw.DATASETS[dataset_name](root=root, split=split)
# Materialize raw text iterable dataset
raw_data = {name: list(raw_dataset) for name, raw_dataset in zip(split, raw_datasets)}
if vocab is None:
if "train" not in split:
raise TypeError("Must pass a vocab if train is not selected.")
logger_.info('Building Vocab based on train data')
vocab = build_vocab(raw_data["train"], text_transform)
logger_.info('Vocab has %d entries', len(vocab))
text_transform = sequential_transforms(
text_transform, vocab_func(vocab), totensor(dtype=torch.long)
)
if dataset_name == 'IMDB':
label_transform = sequential_transforms(lambda x: 1 if x == 'pos' else 0, totensor(dtype=torch.long))
else:
label_transform = sequential_transforms(totensor(dtype=torch.long))
logger_.info('Building datasets for {}'.format(split))
return wrap_datasets(tuple(
TextClassificationDataset(
raw_data[item], vocab, (label_transform, text_transform)
)
for item in split
), split_)
def _setup_datasets(dataset_name, root, split_, offset):
split = check_default_set(split_, ('train', 'dev'), dataset_name)
extracted_files = {
key: download_from_url(URLS[dataset_name][key],
root=root,
hash_value=MD5[dataset_name][key],
hash_type='md5')
for key in split
}
return wrap_datasets(
tuple(
RawTextIterableDataset(dataset_name,
NUM_LINES[dataset_name][item],
_create_data_from_json(
extracted_files[item]),
offset=offset) for item in split), split_)
def _setup_datasets(dataset_name, root, split_, year, language, offset):
if dataset_name == 'WMTNewsCrawl':
split = check_default_set(split_, ('train', ), dataset_name)
else:
split = check_default_set(split_, ('train', 'test', 'valid'),
dataset_name)
if dataset_name == 'PennTreebank':
extracted_files = [
download_from_url(URLS['PennTreebank'][key],
root=root,
hash_value=MD5['PennTreebank'][key],
hash_type='md5') for key in split
]
else:
dataset_tar = download_from_url(URLS[dataset_name],
root=root,
hash_value=MD5[dataset_name],
hash_type='md5')
extracted_files = extract_archive(dataset_tar)
if dataset_name == 'WMTNewsCrawl':
file_name = 'news.{}.{}.shuffled'.format(year, language)
extracted_files = [f for f in extracted_files if file_name in f]
path = {}
for item in split:
for fname in extracted_files:
if item in fname:
path[item] = fname
datasets = []
for item in split:
logging.info('Creating {} data'.format(item))
datasets.append(
RawTextIterableDataset(dataset_name,
NUM_LINES[dataset_name][item],
iter(io.open(path[item], encoding="utf8")),
offset=offset))
return wrap_datasets(tuple(datasets), split_)
def _setup_datasets(dataset_name, root, vocab, tokenizer, split_):
text_transform = []
if tokenizer is None:
tokenizer = get_tokenizer('basic_english')
text_transform = sequential_transforms(tokenizer)
split = check_default_set(split_, ('train', 'dev'), dataset_name)
raw_datasets = raw.DATASETS[dataset_name](root=root, split=split)
raw_data = {
name: list(raw_dataset)
for name, raw_dataset in zip(split, raw_datasets)
}
if vocab is None:
if 'train' not in split:
raise TypeError("Must pass a vocab if train is not selected.")
def apply_transform(data):
for (_context, _question, _answers, _ans_pos) in data:
tok_ans = []
for item in _answers:
tok_ans += text_transform(item)
yield text_transform(_context) + text_transform(
_question) + tok_ans
logger_.info('Building Vocab based on train data')
vocab = build_vocab_from_iterator(apply_transform(raw_data['train']),
len(raw_data['train']))
logger_.info('Vocab has %d entries', len(vocab))
text_transform = sequential_transforms(text_transform, vocab_func(vocab),
totensor(dtype=torch.long))
transforms = {
'context': text_transform,
'question': text_transform,
'answers': text_transform,
'ans_pos': totensor(dtype=torch.long)
}
logger_.info('Building datasets for {}'.format(split))
return wrap_datasets(
tuple(
QuestionAnswerDataset(raw_data[item], vocab, transforms)
for item in split), split_)
def _setup_datasets(dataset_name, root, vocabs, split_):
split = check_default_set(split_, ('train', 'valid', 'test'), dataset_name)
raw_iter_tuple = raw.DATASETS[dataset_name](root=root, split=split)
raw_data = {}
for name, raw_iter in zip(split, raw_iter_tuple):
raw_data[name] = list(raw_iter)
if vocabs is None:
if "train" not in split:
raise TypeError("Must pass a vocab if train is not selected.")
logger_.info('Building Vocab based on train data')
vocabs = build_vocab(raw_data["train"])
else:
if not isinstance(vocabs, list):
raise TypeError("vocabs must be an instance of list")
# Find data that's not None
notnone_data = None
for key in raw_data.keys():
if raw_data[key] is not None:
notnone_data = raw_data[key]
break
if len(vocabs) != len(notnone_data[0]):
raise ValueError(
"Number of vocabs must match the number of columns "
"in the data")
transformers = [
sequential_transforms(vocab_func(vocabs[idx]),
totensor(dtype=torch.long))
for idx in range(len(vocabs))
]
logger_.info('Building datasets for {}'.format(split))
return wrap_datasets(
tuple(
SequenceTaggingDataset(raw_data[item], vocabs, transformers)
for item in split), split_)
def _setup_datasets(dataset_name, train_filenames, valid_filenames,
test_filenames, split_, root, vocab, tokenizer):
split = check_default_set(split_, ('train', 'valid', 'test'), dataset_name)
src_vocab, tgt_vocab = vocab
if tokenizer is None:
src_tokenizer = get_tokenizer("spacy", language='de_core_news_sm')
tgt_tokenizer = get_tokenizer("spacy", language='en_core_web_sm')
elif isinstance(tokenizer, tuple):
if len(tokenizer) == 2:
src_tokenizer, tgt_tokenizer = tokenizer
else:
raise ValueError("tokenizer must have length of two for"
"source and target")
else:
raise ValueError(
"tokenizer must be an instance of tuple with length two"
"or None")
raw_datasets = raw.DATASETS[dataset_name](train_filenames=train_filenames,
valid_filenames=valid_filenames,
test_filenames=test_filenames,
split=split,
root=root)
raw_data = {
name: list(raw_dataset)
for name, raw_dataset in zip(split, raw_datasets)
}
src_text_vocab_transform = sequential_transforms(src_tokenizer)
tgt_text_vocab_transform = sequential_transforms(tgt_tokenizer)
if src_vocab is None:
if 'train' not in split:
raise TypeError("Must pass a vocab if train is not selected.")
logger_.info('Building src Vocab based on train data')
src_vocab = build_vocab(raw_data["train"],
src_text_vocab_transform,
index=0)
else:
if not isinstance(src_vocab, Vocab):
raise TypeError("Passed src vocabulary is not of type Vocab")
logger_.info('src Vocab has %d entries', len(src_vocab))
if tgt_vocab is None:
if 'train' not in split:
raise TypeError("Must pass a vocab if train is not selected.")
logger_.info('Building tgt Vocab based on train data')
tgt_vocab = build_vocab(raw_data["train"],
tgt_text_vocab_transform,
index=1)
else:
if not isinstance(tgt_vocab, Vocab):
raise TypeError("Passed tgt vocabulary is not of type Vocab")
logger_.info('tgt Vocab has %d entries', len(tgt_vocab))
logger_.info('Building datasets for {}'.format(split))
datasets = []
for key in split:
src_text_transform = sequential_transforms(src_text_vocab_transform,
vocab_func(src_vocab),
totensor(dtype=torch.long))
tgt_text_transform = sequential_transforms(tgt_text_vocab_transform,
vocab_func(tgt_vocab),
totensor(dtype=torch.long))
datasets.append(
TranslationDataset(raw_data[key], (src_vocab, tgt_vocab),
(src_text_transform, tgt_text_transform)))
return wrap_datasets(tuple(datasets), split_)
def _setup_datasets(dataset_name, train_filenames, valid_filenames,
test_filenames, split_, root, offset):
split = check_default_set(split_, ('train', 'valid', 'test'), dataset_name)
if not isinstance(train_filenames, tuple) and not isinstance(valid_filenames, tuple) \
and not isinstance(test_filenames, tuple):
raise ValueError("All filenames must be tuples")
src_train, tgt_train = train_filenames
src_eval, tgt_eval = valid_filenames
src_test, tgt_test = test_filenames
extracted_files = [] # list of paths to the extracted files
if isinstance(URLS[dataset_name], list):
for idx, f in enumerate(URLS[dataset_name]):
dataset_tar = download_from_url(f,
root=root,
hash_value=MD5[dataset_name][idx],
hash_type='md5')
extracted_files.extend(extract_archive(dataset_tar))
elif isinstance(URLS[dataset_name], str):
dataset_tar = download_from_url(URLS[dataset_name],
root=root,
hash_value=MD5[dataset_name],
hash_type='md5')
extracted_dataset_tar = extract_archive(dataset_tar)
if dataset_name == 'IWSLT':
# IWSLT dataset's url downloads a multilingual tgz.
# We need to take an extra step to pick out the specific language pair from it.
src_language = train_filenames[0].split(".")[-1]
tgt_language = train_filenames[1].split(".")[-1]
languages = "-".join([src_language, tgt_language])
iwslt_tar = '.data/2016-01/texts/{}/{}/{}.tgz'
iwslt_tar = iwslt_tar.format(src_language, tgt_language, languages)
extracted_dataset_tar = extract_archive(iwslt_tar)
extracted_files.extend(extracted_dataset_tar)
else:
raise ValueError(
"URLS for {} has to be in a form or list or string".format(
dataset_name))
# Clean the xml and tag file in the archives
file_archives = []
for fname in extracted_files:
if 'xml' in fname:
_clean_xml_file(fname)
file_archives.append(os.path.splitext(fname)[0])
elif "tags" in fname:
_clean_tags_file(fname)
file_archives.append(fname.replace('.tags', ''))
else:
file_archives.append(fname)
data_filenames = defaultdict(dict)
data_filenames = {
"train": _construct_filepaths(file_archives, src_train, tgt_train),
"valid": _construct_filepaths(file_archives, src_eval, tgt_eval),
"test": _construct_filepaths(file_archives, src_test, tgt_test)
}
for key in data_filenames.keys():
if len(data_filenames[key]) == 0 or data_filenames[key] is None:
raise FileNotFoundError(
"Files are not found for data type {}".format(key))
datasets = []
for key in split:
src_data_iter = _read_text_iterator(data_filenames[key][0])
tgt_data_iter = _read_text_iterator(data_filenames[key][1])
def _iter(src_data_iter, tgt_data_iter):
for item in zip(src_data_iter, tgt_data_iter):
yield item
datasets.append(
RawTextIterableDataset(dataset_name,
NUM_LINES[dataset_name][key],
_iter(src_data_iter, tgt_data_iter),
offset=offset))
return wrap_datasets(tuple(datasets), split_)