def test_read_embedding_file_inside_archive(self):
token2vec = {
"think": torch.Tensor([0.143, 0.189, 0.555, 0.361, 0.472]),
"make": torch.Tensor([0.878, 0.651, 0.044, 0.264, 0.872]),
"difference": torch.Tensor([0.053, 0.162, 0.671, 0.110, 0.259]),
"àèìòù": torch.Tensor([1.0, 2.0, 3.0, 4.0, 5.0])
}
vocab = Vocabulary()
for token in token2vec:
vocab.add_token_to_namespace(token)
params = Params({
'pretrained_file':
str(self.FIXTURES_ROOT / 'embeddings/multi-file-archive.zip'),
'embedding_dim':
5
})
with pytest.raises(
ValueError,
message=
"No ValueError when pretrained_file is a multi-file archive"):
Embedding.from_params(vocab, params)
for ext in ['.zip', '.tar.gz']:
archive_path = str(
self.FIXTURES_ROOT / 'embeddings/multi-file-archive') + ext
file_uri = format_embeddings_file_uri(
archive_path, 'folder/fake_embeddings.5d.txt')
params = Params({'pretrained_file': file_uri, 'embedding_dim': 5})
embeddings = Embedding.from_params(vocab, params).weight.data
for tok, vec in token2vec.items():
i = vocab.get_token_index(tok)
assert torch.equal(embeddings[i],
vec), 'Problem with format ' + archive_path
def test_embedding_layer_actually_initializes_word_vectors_correctly(self):
vocab = Vocabulary()
vocab.add_token_to_namespace("word")
vocab.add_token_to_namespace("word2")
unicode_space = "\xa0\xa0\xa0\xa0\xa0\xa0\xa0\xa0"
vocab.add_token_to_namespace(unicode_space)
embeddings_filename = str(self.TEST_DIR / "embeddings.gz")
with gzip.open(embeddings_filename, 'wb') as embeddings_file:
embeddings_file.write("word 1.0 2.3 -1.0\n".encode('utf-8'))
embeddings_file.write(
f"{unicode_space} 3.4 3.3 5.0\n".encode('utf-8'))
params = Params({
'pretrained_file': embeddings_filename,
'embedding_dim': 3,
})
embedding_layer = Embedding.from_params(vocab, params)
word_vector = embedding_layer.weight.data[vocab.get_token_index(
"word")]
assert numpy.allclose(word_vector.numpy(),
numpy.array([1.0, 2.3, -1.0]))
word_vector = embedding_layer.weight.data[vocab.get_token_index(
unicode_space)]
assert numpy.allclose(word_vector.numpy(), numpy.array([3.4, 3.3,
5.0]))
word_vector = embedding_layer.weight.data[vocab.get_token_index(
"word2")]
assert not numpy.allclose(word_vector.numpy(),
numpy.array([1.0, 2.3, -1.0]))
def from_params(cls, vocab: Vocabulary, params: Params) -> 'TokenCharactersEncoder': # type: ignore
# pylint: disable=arguments-differ
embedding_params: Params = params.pop("embedding")
# Embedding.from_params() uses "tokens" as the default namespace, but we need to change
# that to be "token_characters" by default.
embedding_params.setdefault("vocab_namespace", "token_characters")
embedding = Embedding.from_params(vocab, embedding_params)
encoder_params: Params = params.pop("encoder")
encoder = Seq2VecEncoder.from_params(encoder_params)
dropout = params.pop_float("dropout", 0.0)
params.assert_empty(cls.__name__)
return cls(embedding, encoder, dropout)
def test_get_embedding_layer_initializes_unseen_words_randomly_not_zero(
self):
vocab = Vocabulary()
vocab.add_token_to_namespace("word")
vocab.add_token_to_namespace("word2")
embeddings_filename = str(self.TEST_DIR / "embeddings.gz")
with gzip.open(embeddings_filename, 'wb') as embeddings_file:
embeddings_file.write("word 1.0 2.3 -1.0\n".encode('utf-8'))
params = Params({
'pretrained_file': embeddings_filename,
'embedding_dim': 3,
})
embedding_layer = Embedding.from_params(vocab, params)
word_vector = embedding_layer.weight.data[vocab.get_token_index(
"word2")]
assert not numpy.allclose(word_vector.numpy(),
numpy.array([0.0, 0.0, 0.0]))
def test_read_hdf5_raises_on_invalid_shape(self):
vocab = Vocabulary()
vocab.add_token_to_namespace("word")
embeddings_filename = str(self.TEST_DIR / "embeddings.hdf5")
embeddings = numpy.random.rand(vocab.get_vocab_size(), 10)
with h5py.File(embeddings_filename, 'w') as fout:
_ = fout.create_dataset('embedding',
embeddings.shape,
dtype='float32',
data=embeddings)
params = Params({
'pretrained_file': embeddings_filename,
'embedding_dim': 5,
})
with pytest.raises(ConfigurationError):
_ = Embedding.from_params(vocab, params)
def test_read_hdf5_format_file(self):
vocab = Vocabulary()
vocab.add_token_to_namespace("word")
vocab.add_token_to_namespace("word2")
embeddings_filename = str(self.TEST_DIR / "embeddings.hdf5")
embeddings = numpy.random.rand(vocab.get_vocab_size(), 5)
with h5py.File(embeddings_filename, 'w') as fout:
_ = fout.create_dataset('embedding',
embeddings.shape,
dtype='float32',
data=embeddings)
params = Params({
'pretrained_file': embeddings_filename,
'embedding_dim': 5,
})
embedding_layer = Embedding.from_params(vocab, params)
assert numpy.allclose(embedding_layer.weight.data.numpy(), embeddings)
def test_forward_works_with_projection_layer(self):
vocab = Vocabulary()
vocab.add_token_to_namespace('the')
vocab.add_token_to_namespace('a')
params = Params({
'pretrained_file':
str(self.FIXTURES_ROOT / 'embeddings/glove.6B.300d.sample.txt.gz'),
'embedding_dim':
300,
'projection_dim':
20
})
embedding_layer = Embedding.from_params(vocab, params)
input_tensor = torch.LongTensor([[3, 2, 1, 0]])
embedded = embedding_layer(input_tensor).data.numpy()
assert embedded.shape == (1, 4, 20)
input_tensor = torch.LongTensor([[[3, 2, 1, 0]]])
embedded = embedding_layer(input_tensor).data.numpy()
assert embedded.shape == (1, 1, 4, 20)