[WIP] torch-text
This repository consists of:
- torchtext.data : Generic data loaders, abstractions, and iterators for text
- torchtext.datasets : Pre-built loaders for common NLP datasets
- (maybe) torchtext.models : Model definitions and pre-trained models for popular NLP examples (though the situation is not the same as vision, where people can download a pretrained ImageNet model and immediately make it useful for other tasks -- it might make more sense to leave NLP models in the torch/examples repo)
Data
The data module provides the following:
- Ability to describe declaratively how to load a custom NLP dataset that's in a "normal" format:
pos = data.TabularDataset(
path='data/pos/pos_wsj_train.tsv', format='tsv',
fields=[('text', data.Field()),
('labels', data.Field())])
sentiment = data.TabularDataset(
path='data/sentiment/train.json', format='json',
fields=[{'sentence_tokenized': ('text', data.Field(sequential=True)),
'sentiment_gold': ('labels', data.Field(sequential=False))}])
path='data/pos/pos_wsj_train.tsv', format='tsv',
fields=[('text', data.Field()),
('labels', data.Field())])
sentiment = data.TabularDataset(
path='data/sentiment/train.json', format='json',
fields=[{'sentence_tokenized': ('text', data.Field(sequential=True)),
'sentiment_gold': ('labels', data.Field(sequential=False))}])
- Ability to define a preprocessing pipeline:
src = data.Field(tokenize=my_custom_tokenizer)
trg = data.Field(tokenize=my_custom_tokenizer)
mt_train = datasets.TranslationDataset(
path='data/mt/wmt16-ende.train', exts=('.en', '.de'),
fields=(src, trg))
trg = data.Field(tokenize=my_custom_tokenizer)
mt_train = datasets.TranslationDataset(
path='data/mt/wmt16-ende.train', exts=('.en', '.de'),
fields=(src, trg))
- Batching, padding, and numericalizing (including building a vocabulary object):
>>next(iter(train_iter))
"># continuing from above
mt_dev = data.TranslationDataset(
path='data/mt/newstest2014', exts=('.en', '.de'),
fields=(src, trg))
src.build_vocab(mt_train, max_size=80000)
trg.build_vocab(mt_train, max_size=40000)
# mt_dev shares the fields, so it shares their vocab objects
train_iter = data.BucketIterator(
dataset=mt_train, batch_size=32,
sort_key=lambda x: data.interleave_keys(len(x.src), len(x.trg)))
# usage
>>>next(iter(train_iter))
<data.Batch(batch_size=32, src=[LongTensor (32, 25)], trg=[LongTensor (32, 28)])>
mt_dev = data.TranslationDataset(
path='data/mt/newstest2014', exts=('.en', '.de'),
fields=(src, trg))
src.build_vocab(mt_train, max_size=80000)
trg.build_vocab(mt_train, max_size=40000)
# mt_dev shares the fields, so it shares their vocab objects
train_iter = data.BucketIterator(
dataset=mt_train, batch_size=32,
sort_key=lambda x: data.interleave_keys(len(x.src), len(x.trg)))
# usage
>>>next(iter(train_iter))
<data.Batch(batch_size=32, src=[LongTensor (32, 25)], trg=[LongTensor (32, 28)])>
- Wrapper for dataset splits (train, validation, test):
TEXT = data.Field()
LABELS = data.Field()
train, val, test = data.TabularDataset.splits(
path='/data/pos_wsj/pos_wsj', train='_train.tsv',
validation='_dev.tsv', test='_test.tsv', format='tsv',
fields=[('text', TEXT), ('labels', LABELS)])
train_iter, val_iter, test_iter = data.BucketIterator.splits(
(train, val, test), batch_sizes=(16, 256, 256),
sort_key=lambda x: len(x.text), device=0)
TEXT.build_vocab(train)
LABELS.build_vocab(train)
LABELS = data.Field()
train, val, test = data.TabularDataset.splits(
path='/data/pos_wsj/pos_wsj', train='_train.tsv',
validation='_dev.tsv', test='_test.tsv', format='tsv',
fields=[('text', TEXT), ('labels', LABELS)])
train_iter, val_iter, test_iter = data.BucketIterator.splits(
(train, val, test), batch_sizes=(16, 256, 256),
sort_key=lambda x: len(x.text), device=0)
TEXT.build_vocab(train)
LABELS.build_vocab(train)
Datasets
Some datasets it would be useful to have built in:
- bAbI and successors from FAIR
- SST (done) and IMDb sentiment
- SNLI (done)
- Penn Treebank (for language modeling (done) and parsing)
- WMT and/or IWSLT machine translation
- SQuAD
See the "test" directory for examples of dataset usage.