def call(self, x, mask=None):
x0 = x
if mask is not None:
mask = K.cast(mask, K.floatx())
mask = K.expand_dims(mask, 2)
# x = x0 * mask if mask is not None else x0
x0 = Lambda(lambda x_: x_, output_shape=lambda s: s)(x0) # drop mask so do not put mask to conv1d
x = self.conv1d(x0)
x, g = x[:, :, :self.o_dim], x[:, :, self.o_dim:]
if self.dropout_rate is not None:
g = K.in_train_phase(K.dropout(g, self.dropout_rate), g)
g = K.sigmoid(g)
# mask is none
mask = mask if mask is not None else K.ones_like(x)
if self.skip_connection:
if K.int_shape(x0)[-1] != self.o_dim:
x0 = self.conv1d_1x1(x0)
return (x0 * (1 - g) + x * g) * mask
return x * g * mask
def call(self, inputs):
clf, x_pre, x_next = inputs
uncertain = normal_shannon_entropy(clf, num_classes)
cond = K.greater(self.speed, uncertain)
x = K.switch(cond, x_pre, x_next)
return K.in_train_phase(x_next, x)
def call(self, inputs):
source, target = inputs
source = source * self.proportion
target = target * (1 - self.proportion)
output = (source + target) / 2
return K.in_train_phase(output, target)
def call(self, inputs):
source, target = inputs
mask = K.random_binomial(shape=[1], p=0.5)
output = mask * source + (1 - mask) * target
return K.in_train_phase(output, target)
