def encode(self, x=None):
if x is None:
x = CharLSTMEmbeddings.create_placeholder(self.name)
self.x = x
with tf.variable_scope(self.scope, reuse=tf.AUTO_REUSE):
Wch = tf.get_variable(
"Wch",
initializer=tf.constant_initializer(self.weights, dtype=tf.float32, verify_shape=True),
shape=[self.vsz, self.dsz],
trainable=True
)
ech0 = tf.scatter_update(Wch, tf.constant(Offsets.PAD, dtype=tf.int32, shape=[1]), tf.zeros(shape=[1, self.dsz]))
shape = tf.shape(x)
B = shape[0]
T = shape[1]
W = shape[2]
flat_chars = tf.reshape(x, [-1, W])
word_lengths = tf.reduce_sum(tf.cast(tf.equal(flat_chars, Offsets.PAD), tf.int32), axis=1)
with tf.control_dependencies([ech0]):
embed_chars = tf.nn.embedding_lookup(Wch, flat_chars)
fwd_lstm = stacked_lstm(self.lstmsz // 2, self.pdrop, self.layers)
bwd_lstm = stacked_lstm(self.lstmsz // 2, self.pdrop, self.layers)
_, rnn_state = tf.nn.bidirectional_dynamic_rnn(fwd_lstm, bwd_lstm, embed_chars, sequence_length=word_lengths, dtype=tf.float32)
result = tf.concat([rnn_state[0][-1].h, rnn_state[1][-1].h], axis=1)
return tf.reshape(result, [B, T, self.lstmsz])
def pool(self, word_embeddings, dsz, init, **kwargs):
"""LSTM with dropout yielding a final-state as output
:param word_embeddings: The input word embeddings
:param dsz: The input word embedding depth
:param init: The tensorflow initializer to use (currently ignored)
:param kwargs: See below
:Keyword Arguments:
* *hsz* -- (``int``) The number of hidden units (defaults to `100`)
* *cmotsz* -- (``int``) An alias for `hsz`
:return:
"""
hsz = kwargs.get('rnnsz', kwargs.get('hsz', 100))
if type(hsz) is list:
hsz = hsz[0]
rnntype = kwargs.get('rnntype', 'lstm')
nlayers = int(kwargs.get('layers', 1))
if rnntype == 'blstm':
rnnfwd = stacked_lstm(hsz, self.pkeep, nlayers)
rnnbwd = stacked_lstm(hsz, self.pkeep, nlayers)
((_, _), (fw_final_state,
bw_final_state)) = tf.nn.bidirectional_dynamic_rnn(
rnnfwd,
rnnbwd,
word_embeddings,
sequence_length=self.lengths,
dtype=tf.float32)
# The output of the BRNN function needs to be joined on the H axis
output_state = fw_final_state[-1].h + bw_final_state[-1].h
out_hsz = hsz
else:
rnnfwd = stacked_lstm(hsz, self.pkeep, nlayers)
(_,
(output_state)) = tf.nn.dynamic_rnn(rnnfwd,
word_embeddings,
sequence_length=self.lengths,
dtype=tf.float32)
output_state = output_state[-1].h
out_hsz = hsz
combine = tf.reshape(output_state, [-1, out_hsz])
return combine
def pool(self, word_embeddings, dsz, init, **kwargs):
"""LSTM with dropout yielding a final-state as output
:param word_embeddings: The input word embeddings
:param dsz: The input word embedding depth
:param init: The tensorflow initializer to use (currently ignored)
:param kwargs: See below
:Keyword Arguments:
* *rnnsz* -- (``int``) The number of hidden units (defaults to `hsz`)
* *hsz* -- (``int``) backoff for `rnnsz`, typically a result of stacking params. This keeps things simple so
its easy to do things like residual connections between LSTM and post-LSTM stacking layers
:return:
"""
hsz = kwargs.get('rnnsz', kwargs.get('hsz', 100))
vdrop = bool(kwargs.get('variational_dropout', False))
if type(hsz) is list:
hsz = hsz[0]
rnntype = kwargs.get('rnn_type', kwargs.get('rnntype', 'lstm'))
nlayers = int(kwargs.get('layers', 1))
if rnntype == 'blstm':
rnnfwd = stacked_lstm(hsz//2, self.pdrop_value, nlayers, variational=vdrop, training=TRAIN_FLAG())
rnnbwd = stacked_lstm(hsz//2, self.pdrop_value, nlayers, variational=vdrop, training=TRAIN_FLAG())
((_, _), (fw_final_state, bw_final_state)) = tf.nn.bidirectional_dynamic_rnn(rnnfwd,
rnnbwd,
word_embeddings,
sequence_length=self.lengths,
dtype=tf.float32)
# The output of the BRNN function needs to be joined on the H axis
output_state = tf.concat([fw_final_state[-1].h, bw_final_state[-1].h], -1)
out_hsz = hsz
else:
rnnfwd = stacked_lstm(hsz, self.pdrop_value, nlayers, variational=vdrop, training=TRAIN_FLAG())
(_, (output_state)) = tf.nn.dynamic_rnn(rnnfwd, word_embeddings, sequence_length=self.lengths, dtype=tf.float32)
output_state = output_state[-1].h
out_hsz = hsz
combine = tf.reshape(output_state, [-1, out_hsz])
return combine
def pool(self, word_embeddings, dsz, init, **kwargs):
"""LSTM with dropout yielding a final-state as output
:param word_embeddings: The input word embeddings
:param dsz: The input word embedding depth
:param init: The tensorflow initializer to use (currently ignored)
:param kwargs: See below
:Keyword Arguments:
* *rnnsz* -- (``int``) The number of hidden units (defaults to `hsz`)
* *hsz* -- (``int``) backoff for `rnnsz`, typically a result of stacking params. This keeps things simple so
its easy to do things like residual connections between LSTM and post-LSTM stacking layers
:return:
"""
hsz = kwargs.get('rnnsz', kwargs.get('hsz', 100))
vdrop = bool(kwargs.get('variational_dropout', False))
if type(hsz) is list:
hsz = hsz[0]
rnntype = kwargs.get('rnn_type', kwargs.get('rnntype', 'lstm'))
nlayers = int(kwargs.get('layers', 1))
if rnntype == 'blstm':
rnnfwd = stacked_lstm(hsz//2, self.pdrop_value, nlayers, variational=vdrop, training=TRAIN_FLAG())
rnnbwd = stacked_lstm(hsz//2, self.pdrop_value, nlayers, variational=vdrop, training=TRAIN_FLAG())
((_, _), (fw_final_state, bw_final_state)) = tf.nn.bidirectional_dynamic_rnn(rnnfwd,
rnnbwd,
word_embeddings,
sequence_length=self.lengths,
dtype=tf.float32)
# The output of the BRNN function needs to be joined on the H axis
output_state = tf.concat([fw_final_state[-1].h, bw_final_state[-1].h], -1)
out_hsz = hsz
else:
rnnfwd = stacked_lstm(hsz, self.pdrop_value, nlayers, variational=vdrop, training=TRAIN_FLAG())
(_, (output_state)) = tf.nn.dynamic_rnn(rnnfwd, word_embeddings, sequence_length=self.lengths, dtype=tf.float32)
output_state = output_state[-1].h
out_hsz = hsz
combine = tf.reshape(output_state, [-1, out_hsz])
return combine
def encode(self, x=None):
if x is None:
x = CharLSTMEmbeddings.create_placeholder(self.name)
self.x = x
with tf.variable_scope(self.scope):
Wch = tf.get_variable("Wch",
initializer=tf.constant_initializer(
self.weights,
dtype=tf.float32,
verify_shape=True),
shape=[self.vsz, self.dsz],
trainable=True)
ech0 = tf.scatter_update(
Wch, tf.constant(Offsets.PAD, dtype=tf.int32, shape=[1]),
tf.zeros(shape=[1, self.dsz]))
shape = tf.shape(x)
B = shape[0]
T = shape[1]
W = shape[2]
flat_chars = tf.reshape(x, [-1, W])
word_lengths = tf.reduce_sum(tf.cast(
tf.equal(flat_chars, Offsets.PAD), tf.int32),
axis=1)
with tf.control_dependencies([ech0]):
embed_chars = tf.nn.embedding_lookup(Wch, flat_chars)
fwd_lstm = stacked_lstm(self.lstmsz // 2, self.pdrop, self.layers)
bwd_lstm = stacked_lstm(self.lstmsz // 2, self.pdrop, self.layers)
_, rnn_state = tf.nn.bidirectional_dynamic_rnn(
fwd_lstm,
bwd_lstm,
embed_chars,
sequence_length=word_lengths,
dtype=tf.float32)
result = tf.concat([rnn_state[0][-1].h, rnn_state[1][-1].h],
axis=1)
return tf.reshape(result, [B, T, self.lstmsz])
