def __init__(self, embedding, hparams=None):
ModuleBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
if self._hparams.position_embedder_type == 'sinusoids':
self.position_embedder = SinusoidsPositionEmbedder(
self._hparams.position_embedder_hparams)
else:
self.position_embedder = PositionEmbedder(
position_size=self._hparams.position_size,
hparams=self._hparams.position_embedder_hparams)
self._embedding = embedding
self._vocab_size = self._embedding.get_shape().as_list()[0]
self.output_layer = \
self._build_output_layer(shape_list(self._embedding)[-1])
self.multihead_attentions = {'self_att': [], 'encdec_att': []}
self.poswise_networks = []
for i in range(self._hparams.num_blocks):
layer_name = 'layer_{}'.format(i)
with tf.variable_scope(layer_name):
with tf.variable_scope("self_attention"):
multihead_attention = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attentions['self_att'].append(
multihead_attention)
# pylint: disable=protected-access
if self._hparams.dim != \
multihead_attention._hparams.output_dim:
raise ValueError('The output dimenstion of '
'MultiheadEncoder should be equal '
'to the dim of TransformerDecoder')
with tf.variable_scope('encdec_attention'):
multihead_attention = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attentions['encdec_att'].append(
multihead_attention)
if self._hparams.dim != \
multihead_attention._hparams.output_dim:
raise ValueError('The output dimenstion of '
'MultiheadEncoder should be equal '
'to the dim of TransformerDecoder')
poswise_network = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
if self._hparams.dim != \
poswise_network._hparams.layers[-1]['kwargs']['units']:
raise ValueError('The output dimenstion of '
'FeedForwardNetwork should be equal '
'to the dim of TransformerDecoder')
self.poswise_networks.append(poswise_network)
def test_sinusoids_position_embedder(self):
"""Tests :class:`texar.modules.SinusoidsPositionEmbedder`.
"""
position_size = 64
input_size = [100]
hparams = {'dim': 513} # use odd dimension to ensure padding correct
embedder = SinusoidsPositionEmbedder(position_size, hparams=hparams)
inputs = torch.randint(position_size - 1, input_size)
outputs = embedder(inputs)
self.assertEqual(list(outputs.shape), input_size + [hparams['dim']])
def __init__(self, hparams=None):
EncoderBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
self.position_embedder = \
SinusoidsPositionEmbedder(
self._hparams.position_embedder_hparams)
def test_sinusoids_position_embedder(self):
"""Tests :class:`texar.modules.SinusoidsPositionEmbedder`.
"""
position_size = 64
input_size = (23, 18)
hparams = {'dim': 513} # use odd dimension to ensure padding correct
embedder = SinusoidsPositionEmbedder(position_size, hparams=hparams)
inputs = torch.randint(position_size - 1, input_size)
outputs = embedder(inputs)
self.assertEqual(outputs.size(), input_size + (hparams['dim'], ))
embedder_no_cache = SinusoidsPositionEmbedder(
None, hparams={
**hparams, 'cache_embeddings': False
})
wide_inputs = torch.randint(-position_size, position_size * 2,
input_size)
wide_outputs = embedder_no_cache(wide_inputs)
self.assertEqual(wide_outputs.size(), input_size + (hparams['dim'], ))
no_cache_outputs = embedder_no_cache(inputs)
np.testing.assert_array_equal(outputs, no_cache_outputs)
def __init__(self, hparams=None):
EncoderBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
if self._hparams.position_embedder_type == 'sinusoids':
self.position_embedder = SinusoidsPositionEmbedder(
self._hparams.position_embedder_hparams)
else:
self.position_embedder = PositionEmbedder(
position_size=self._hparams.position_size,
hparams=self._hparams.position_embedder_hparams)
# pylint: disable=protected-access
if self._hparams.dim != \
self.position_embedder._hparams.dim:
raise ValueError('"dim" in '
'TransformerEncoder hparams must be equal '
'to "dim" in its '
'position_embedder_hparams.')
self.multihead_attention_list = []
self.poswise_networks = []
for i in range(self._hparams.num_blocks):
with tf.variable_scope("layer_{}".format(i)):
with tf.variable_scope('attention'):
multihead_attention = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attention_list.append(
multihead_attention)
# pylint: disable=protected-access
if self._hparams.dim != \
multihead_attention._hparams.output_dim:
raise ValueError('The "dim" in the hparams of '
'multihead_attention should be equal '
'to the "dim" of TransformerEncoder')
poswise_network = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
# pylint: disable=protected-access
if self._hparams.dim != \
poswise_network._hparams.layers[-1]['kwargs']['units']:
# poswise_network._hparams.layers[-1]['units']:
raise ValueError('The "units" in the "kwargs" of '
'FeedForwardNetwork should be equal '
'to the "dim" of TransformerEncoder')
self.poswise_networks.append(poswise_network)
def __init__(self, embedding, hparams=None):
ModuleBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer( \
layers.get_initializer(self._hparams.initializer))
self.position_embedder = \
SinusoidsPositionEmbedder(
self._hparams.position_embedder_hparams)
self._embedding = embedding
self._vocab_size = self._embedding.get_shape().as_list()[0]
self.output_layer = \
self._build_output_layer(shape_list(self._embedding)[-1])
def __init__(self, hparams=None):
EncoderBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
self.position_embedder = \
SinusoidsPositionEmbedder(
self._hparams.position_embedder_hparams)
self.multihead_attention_list = []
self.poswise_networks = []
for i in range(self._hparams.num_blocks):
with tf.variable_scope("layer_{}".format(i)):
with tf.variable_scope('self_attention'):
multihead_attention = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attention_list.append(
multihead_attention)
# pylint: disable=protected-access
if self._hparams.dim != \
multihead_attention._hparams.output_dim:
raise ValueError('The output dimenstion of'
'MultiheadEncoder should be equal'
'to the dim of TransformerEncoder')
poswise_network = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
# pylint: disable=protected-access
if self._hparams.dim != \
poswise_network._hparams.layers[-1]['kwargs']['units']:
# poswise_network._hparams.layers[-1]['units']:
raise ValueError('The output dimenstion of'
'FeedForwardNetwork should be equal'
'to the dim of TransformerEncoder')
self.poswise_networks.append(poswise_network)
