def test_cached_path(self):
url = 'http://fake.datastore.com/glove.txt.gz'
set_up_glove(url, self.glove_bytes)
# non-existent file
with pytest.raises(FileNotFoundError):
filename = cached_path(
"tests/fixtures/does_not_exist/fake_file.tar.gz")
# unparsable URI
with pytest.raises(ValueError):
filename = cached_path("fakescheme://path/to/fake/file.tar.gz")
# existing file as path
assert cached_path(self.glove_file) == self.glove_file
# caches urls
filename = cached_path(url, cache_dir=self.TEST_DIR)
assert len(responses.calls) == 2
assert filename == os.path.join(self.TEST_DIR,
url_to_filename(url, etag="0"))
with open(filename, 'rb') as cached_file:
assert cached_file.read() == self.glove_bytes
def __init__(self,
file_uri: str,
encoding: str = DEFAULT_ENCODING,
cache_dir: str = None) -> None:
self.uri = file_uri
self._encoding = encoding
self._cache_dir = cache_dir
self._archive_handle: Any = None # only if the file is inside an archive
main_file_uri, path_inside_archive = parse_embeddings_file_uri(
file_uri)
main_file_local_path = cached_path(main_file_uri, cache_dir=cache_dir)
if zipfile.is_zipfile(main_file_local_path): # ZIP archive
self._open_inside_zip(main_file_uri, path_inside_archive)
elif tarfile.is_tarfile(main_file_local_path): # TAR archive
self._open_inside_tar(main_file_uri, path_inside_archive)
else: # all the other supported formats, including uncompressed files
if path_inside_archive:
raise ValueError('Unsupported archive format: %s' +
main_file_uri)
# All the python packages for compressed files share the same interface of io.open
extension = get_file_extension(main_file_uri)
package = {
'.txt': io,
'.vec': io,
'.gz': gzip,
'.bz2': bz2,
'.lzma': lzma,
}.get(extension, None)
if package is None:
warn(
'The embeddings file has an unknown file extension "%s". '
'We will assume the file is an (uncompressed) text file',
extension)
package = io
self._handle = package.open(main_file_local_path,
'rt',
encoding=encoding) # type: ignore
# To use this with tqdm we'd like to know the number of tokens. It's possible that the
# first line of the embeddings file contains this: if it does, we want to start iteration
# from the 2nd line, otherwise we want to start from the 1st.
# Unfortunately, once we read the first line, we cannot move back the file iterator
# because the underlying file may be "not seekable"; we use itertools.chain instead.
first_line = next(self._handle) # this moves the iterator forward
self.num_tokens = EmbeddingsTextFile._get_num_tokens_from_first_line(
first_line)
if self.num_tokens:
# the first line is a header line: start iterating from the 2nd line
self._iterator = self._handle
else:
# the first line is not a header line: start iterating from the 1st line
self._iterator = itertools.chain([first_line], self._handle)
def add_file_to_archive(self, name: str) -> None:
"""
Any class in its ``from_params`` method can request that some of its
input files be added to the archive by calling this method.
For example, if some class ``A`` had an ``input_file`` parameter, it could call
```
params.add_file_to_archive("input_file")
```
which would store the supplied value for ``input_file`` at the key
``previous.history.and.then.input_file``. The ``files_to_archive`` dict
is shared with child instances via the ``_check_is_dict`` method, so that
the final mapping can be retrieved from the top-level ``Params`` object.
NOTE: You must call ``add_file_to_archive`` before you ``pop()``
the parameter, because the ``Params`` instance looks up the value
of the filename inside itself.
If the ``loading_from_archive`` flag is True, this will be a no-op.
"""
if not self.loading_from_archive:
self.files_to_archive[f"{self.history}{name}"] = cached_path(
self.get(name))
def _load_cnn_weights(self):
cnn_options = self._options['char_cnn']
filters = cnn_options['filters']
char_embed_dim = cnn_options['embedding']['dim']
convolutions = []
for i, (width, num) in enumerate(filters):
conv = torch.nn.Conv1d(in_channels=char_embed_dim,
out_channels=num,
kernel_size=width,
bias=True)
# load the weights
with h5py.File(cached_path(self._weight_file), 'r') as fin:
weight = fin['CNN']['W_cnn_{}'.format(i)][...]
bias = fin['CNN']['b_cnn_{}'.format(i)][...]
w_reshaped = numpy.transpose(weight.squeeze(axis=0),
axes=(2, 1, 0))
if w_reshaped.shape != tuple(conv.weight.data.shape):
raise ValueError("Invalid weight file")
conv.weight.data.copy_(torch.FloatTensor(w_reshaped))
conv.bias.data.copy_(torch.FloatTensor(bias))
conv.weight.requires_grad = self.requires_grad
conv.bias.requires_grad = self.requires_grad
convolutions.append(conv)
self.add_module('char_conv_{}'.format(i), conv)
self._convolutions = convolutions
def _read_pretrained_words(embeddings_filename: str) -> Set[str]:
words = set()
with gzip.open(cached_path(embeddings_filename), 'rb') as embeddings_file:
for line in embeddings_file:
fields = line.decode('utf-8').strip().split(' ')
word = fields[0]
words.add(word)
return words
def _load_char_embedding(self):
with h5py.File(cached_path(self._weight_file), 'r') as fin:
char_embed_weights = fin['char_embed'][...]
weights = numpy.zeros(
(char_embed_weights.shape[0] + 1, char_embed_weights.shape[1]),
dtype='float32')
weights[1:, :] = char_embed_weights
self._char_embedding_weights = torch.nn.Parameter(
torch.FloatTensor(weights), requires_grad=self.requires_grad)
def from_file(params_file: str, params_overrides: str = "") -> 'Params':
"""
Load a `Params` object from a configuration file.
"""
# redirect to cache, if necessary
params_file = cached_path(params_file)
file_dict = pyhocon.ConfigFactory.parse_file(params_file)
overrides_dict = pyhocon.ConfigFactory.parse_string(params_overrides)
param_dict = overrides_dict.with_fallback(file_dict)
return Params(param_dict)
def _open_inside_zip(self,
archive_path: str,
member_path: Optional[str] = None) -> None:
cached_archive_path = cached_path(archive_path,
cache_dir=self._cache_dir)
archive = zipfile.ZipFile(cached_archive_path, 'r')
if member_path is None:
members_list = archive.namelist()
member_path = self._get_the_only_file_in_the_archive(
members_list, archive_path)
member_path = cast(str, member_path)
member_file = archive.open(member_path, 'r')
self._handle = io.TextIOWrapper(member_file, encoding=self._encoding)
self._archive_handle = archive
def from_file(params_file: str, params_overrides: str = "") -> 'Params':
"""
Load a `Params` object from a configuration file.
"""
# redirect to cache, if necessary
params_file = cached_path(params_file)
ext_vars = dict(os.environ)
file_dict = json.loads(evaluate_file(params_file, ext_vars=ext_vars))
overrides_dict = parse_overrides(params_overrides)
param_dict = with_fallback(preferred=overrides_dict,
fallback=file_dict)
return Params(param_dict)
def _open_inside_tar(self,
archive_path: str,
member_path: Optional[str] = None) -> None:
cached_archive_path = cached_path(archive_path,
cache_dir=self._cache_dir)
archive = tarfile.open(cached_archive_path, 'r')
if member_path is None:
members_list = archive.getnames()
member_path = self._get_the_only_file_in_the_archive(
members_list, archive_path)
member_path = cast(str, member_path)
member = archive.getmember(
member_path) # raises exception if not present
member_file = cast(IO[bytes], archive.extractfile(member))
self._handle = io.TextIOWrapper(member_file, encoding=self._encoding)
self._archive_handle = archive
def _load_projection(self):
cnn_options = self._options['char_cnn']
filters = cnn_options['filters']
n_filters = sum(f[1] for f in filters)
self._projection = torch.nn.Linear(n_filters,
self.output_dim,
bias=True)
with h5py.File(cached_path(self._weight_file), 'r') as fin:
weight = fin['CNN_proj']['W_proj'][...]
bias = fin['CNN_proj']['b_proj'][...]
self._projection.weight.data.copy_(
torch.FloatTensor(numpy.transpose(weight)))
self._projection.bias.data.copy_(torch.FloatTensor(bias))
self._projection.weight.requires_grad = self.requires_grad
self._projection.bias.requires_grad = self.requires_grad
def __init__(self,
options_file: str,
weight_file: str,
requires_grad: bool = False,
vocab_to_cache: List[str] = None) -> None:
super(_ElmoBiLm, self).__init__()
self._token_embedder = _ElmoCharacterEncoder(
options_file, weight_file, requires_grad=requires_grad)
self._requires_grad = requires_grad
if requires_grad and vocab_to_cache:
warn(
"You are fine tuning ELMo and caching char CNN word vectors. "
"This behaviour is not guaranteed to be well defined, particularly. "
"if not all of your inputs will occur in the vocabulary cache."
)
# This is an embedding, used to look up cached
# word vectors built from character level cnn embeddings.
self._word_embedding = None
self._bos_embedding: torch.Tensor = None
self._eos_embedding: torch.Tensor = None
if vocab_to_cache:
info("Caching character cnn layers for words in vocabulary.")
# This sets 3 attributes, _word_embedding, _bos_embedding and _eos_embedding.
# They are set in the method so they can be accessed from outside the
# constructor.
self.create_cached_cnn_embeddings(vocab_to_cache)
with open(cached_path(options_file), 'r') as fin:
options = json.load(fin)
if not options['lstm'].get('use_skip_connections'):
raise ConfigurationError(
'We only support pretrained biLMs with residual connections')
self._elmo_lstm = ElmoLstm(
input_size=options['lstm']['projection_dim'],
hidden_size=options['lstm']['projection_dim'],
cell_size=options['lstm']['dim'],
num_layers=options['lstm']['n_layers'],
memory_cell_clip_value=options['lstm']['cell_clip'],
state_projection_clip_value=options['lstm']['proj_clip'],
requires_grad=requires_grad)
self._elmo_lstm.load_weights(weight_file)
# Number of representation layers including context independent layer
self.num_layers = options['lstm']['n_layers'] + 1
def __init__(self,
options_file: str,
weight_file: str,
requires_grad: bool = False) -> None:
super(_ElmoCharacterEncoder, self).__init__()
with open(cached_path(options_file), 'r') as fin:
self._options = json.load(fin)
self._weight_file = weight_file
self.output_dim = self._options['lstm']['projection_dim']
self.requires_grad = requires_grad
self._load_weights()
# Cache the arrays for use in forward -- +1 due to masking.
self._beginning_of_sentence_characters = torch.from_numpy(
numpy.array(ELMoCharacterMapper.beginning_of_sentence_characters) +
1)
self._end_of_sentence_characters = torch.from_numpy(
numpy.array(ELMoCharacterMapper.end_of_sentence_characters) + 1)
def _load_highway(self):
# pylint: disable=protected-access
# the highway layers have same dimensionality as the number of cnn filters
cnn_options = self._options['char_cnn']
filters = cnn_options['filters']
n_filters = sum(f[1] for f in filters)
n_highway = cnn_options['n_highway']
# create the layers, and load the weights
self._highways = Highway(n_filters,
n_highway,
activation=torch.nn.functional.relu)
for k in range(n_highway):
# The AllenNLP highway is one matrix multplication with concatenation of
# transform and carry weights.
with h5py.File(cached_path(self._weight_file), 'r') as fin:
# The weights are transposed due to multiplication order assumptions in tf
# vs pytorch (tf.matmul(X, W) vs pytorch.matmul(W, X))
w_transform = numpy.transpose(
fin['CNN_high_{}'.format(k)]['W_transform'][...])
# -1.0 since AllenNLP is g * x + (1 - g) * f(x) but tf is (1 - g) * x + g * f(x)
w_carry = -1.0 * numpy.transpose(
fin['CNN_high_{}'.format(k)]['W_carry'][...])
weight = numpy.concatenate([w_transform, w_carry], axis=0)
self._highways._layers[k].weight.data.copy_(
torch.FloatTensor(weight))
self._highways._layers[
k].weight.requires_grad = self.requires_grad
b_transform = fin['CNN_high_{}'.format(k)]['b_transform'][...]
b_carry = -1.0 * fin['CNN_high_{}'.format(k)]['b_carry'][...]
bias = numpy.concatenate([b_transform, b_carry], axis=0)
self._highways._layers[k].bias.data.copy_(
torch.FloatTensor(bias))
self._highways._layers[
k].bias.requires_grad = self.requires_grad
def load_weights(self, weight_file: str) -> None:
"""
Load the pre-trained weights from the file.
"""
requires_grad = self.requires_grad
with h5py.File(cached_path(weight_file), 'r') as fin:
for i_layer, lstms in enumerate(
zip(self.forward_layers, self.backward_layers)):
for j_direction, lstm in enumerate(lstms):
# lstm is an instance of LSTMCellWithProjection
cell_size = lstm.cell_size
dataset = fin['RNN_%s' %
j_direction]['RNN']['MultiRNNCell'][
'Cell%s' % i_layer]['LSTMCell']
# tensorflow packs together both W and U matrices into one matrix,
# but pytorch maintains individual matrices. In addition, tensorflow
# packs the gates as input, memory, forget, output but pytorch
# uses input, forget, memory, output. So we need to modify the weights.
tf_weights = numpy.transpose(dataset['W_0'][...])
torch_weights = tf_weights.copy()
# split the W from U matrices
input_size = lstm.input_size
input_weights = torch_weights[:, :input_size]
recurrent_weights = torch_weights[:, input_size:]
tf_input_weights = tf_weights[:, :input_size]
tf_recurrent_weights = tf_weights[:, input_size:]
# handle the different gate order convention
for torch_w, tf_w in [[input_weights, tf_input_weights],
[
recurrent_weights,
tf_recurrent_weights
]]:
torch_w[(1 * cell_size):(2 * cell_size), :] = tf_w[(
2 * cell_size):(3 * cell_size), :]
torch_w[(2 * cell_size):(3 * cell_size), :] = tf_w[(
1 * cell_size):(2 * cell_size), :]
lstm.input_linearity.weight.data.copy_(
torch.FloatTensor(input_weights))
lstm.state_linearity.weight.data.copy_(
torch.FloatTensor(recurrent_weights))
lstm.input_linearity.weight.requires_grad = requires_grad
lstm.state_linearity.weight.requires_grad = requires_grad
# the bias weights
tf_bias = dataset['B'][...]
# tensorflow adds 1.0 to forget gate bias instead of modifying the
# parameters...
tf_bias[(2 * cell_size):(3 * cell_size)] += 1
torch_bias = tf_bias.copy()
torch_bias[(1 * cell_size):(2 * cell_size)] = tf_bias[(
2 * cell_size):(3 * cell_size)]
torch_bias[(2 * cell_size):(3 * cell_size)] = tf_bias[(
1 * cell_size):(2 * cell_size)]
lstm.state_linearity.bias.data.copy_(
torch.FloatTensor(torch_bias))
lstm.state_linearity.bias.requires_grad = requires_grad
# the projection weights
proj_weights = numpy.transpose(dataset['W_P_0'][...])
lstm.state_projection.weight.data.copy_(
torch.FloatTensor(proj_weights))
lstm.state_projection.weight.requires_grad = requires_grad
