def apply(self, is_train, x, c, mask=None, context_mask=None):
x = dropout(x, self.keep_probs, is_train)
c = dropout(c, self.context_keep_probs, is_train)
init = get_keras_initialization(self.init)
x_w = tf.get_variable("merge_x_weights", (x.shape.as_list()[-1], self.output_size), initializer=init)
c_w = tf.get_variable("merge_context_weights", (c.shape.as_list()[-1], self.output_size), initializer=init)
output = tf.tensordot(x, x_w, axes=[[2], [0]]) + tf.expand_dims(tf.matmul(c, c_w), 1)
if self.use_dots:
dots = tf.einsum("aij,aj->aij", x, c)
dot_w = tf.get_variable("dot_weights", (c.shape.as_list()[-1], self.output_size), initializer=init)
output += tf.tensordot(dots, dot_w, axes=[[2], [0]])
bais = tf.get_variable("merge_bias", (1, 1, self.output_size))
output += bais
return get_keras_activation(self.activation)(output)
def apply(self, is_train, x, mask=None):
num_channels = x.get_shape()[3]
filter_ = tf.get_variable("conv1d/filters", shape=[1, self.filter_size, num_channels, self.num_filters], dtype='float')
bias = tf.get_variable("conv1d/bias", shape=[self.num_filters], dtype='float')
strides = [1, 1, 1, 1]
if self.keep_probs < 1.0:
x = dropout(x, self.keep_probs, is_train)
fn = get_keras_activation(self.activation)
return fn(tf.nn.conv2d(x, filter_, strides, "VALID") + bias)
def apply(self, is_train, x, mask=None):
shape = tf.shape(x)
noise_shape = []
for i in range(len(shape)):
if i == self.axis:
noise_shape.append(1)
else:
noise_shape.append(shape[i])
return dropout(x, self.keep_probs, is_train, noise_shape)
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(self.scope):
c, h = state
h = dropout(h, self.keep_recurrent_probs, self.is_train)
mat = _compute_gates(inputs, h, self.num_units, self.forget_bias,
self.kernel_initializer, self.recurrent_initializer, True)
i, j, f, o = tf.split(value=mat, num_or_size_splits=4, axis=1)
new_c = (c * self.recurrent_activation(f) + self.recurrent_activation(i) *
self.activation(j))
new_h = self.activation(new_c) * self.recurrent_activation(o)
new_state = LSTMStateTuple(new_c, new_h)
return new_h, new_state
def apply(self, is_train, x, mask=None):
shape = tf.shape(x)
return dropout(x, self.keep_prob, is_train, [shape[0], 1, shape[2]])
def apply(self, is_train, x, mask=None):
return dropout(x, self.keep_prob, is_train)
def apply(self, is_train, x, mask=None):
shape = tf.shape(x)
#if x.shape[1] == 4 :
#pdb.set_trace()
#return dropout(x, self.keep_prob, is_train, [1, 1, shape[2]])
return dropout(x, self.keep_prob, is_train, [shape[0], 1, shape[2]])
