def lstm_reshape(self,
inputs,
name_prefix,
index,
reshaped_inputs=None,
initial=False):
name_prefix = "{0}_{1}_{2}".format(self.controller_network_name,
name_prefix, index)
cell = LSTMCell(
self.lstm_cell_units,
kernel_initializer=get_weight_initializer(initializer="lstm"),
recurrent_initializer=get_weight_initializer(initializer="lstm"))
if initial:
x = RNN(cell,
return_state=True,
name="{0}_{1}".format(name_prefix, "lstm"))(inputs)
else:
x = RNN(cell,
return_state=True,
name="{0}_{1}".format(name_prefix,
"lstm"))(reshaped_inputs,
initial_state=inputs[1:])
rx = Reshape((-1, self.lstm_cell_units),
name="{0}_{1}".format(name_prefix, "reshape"))(x[0])
return x, rx
def reduce_output_size(self, inputs, name_prefix, rep,
half_current_filters):
x_0 = AveragePooling2D(pool_size=(1,1),
strides=(2,2),
padding="valid",
name="{0}_avepool2d_{1}a_".format(name_prefix, rep))(inputs)
x_0 = Conv2D(filters=half_current_filters,
kernel_size=(1,1),
strides=(1,1),
padding="valid",
kernel_initializer=get_weight_initializer(),
kernel_regularizer=get_weight_regularizer(),
name="{0}_conv2d_{1}a_".format(name_prefix, rep))(x_0)
x_1 = ZeroPadding2D(padding=((0,1),(0,1)),
name="{0}_zeropad2d_{1}b_".format(name_prefix, rep))(inputs)
x_1 = Cropping2D(cropping=((1,0),(1,0)),
name="{0}_crop2d_{1}b_".format(name_prefix, rep))(x_1)
x_1 = AveragePooling2D(pool_size=(1,1),
strides=(2,2),
padding="valid",
name="{0}_avepool2d_{1}b_".format(name_prefix, rep))(x_1)
x_1 = Conv2D(filters=half_current_filters,
kernel_size=(1,1),
strides=(1,1),
padding="valid",
kernel_initializer=get_weight_initializer(),
kernel_regularizer=get_weight_regularizer(),
name="{0}_conv2d_{1}b_".format(name_prefix, rep))(x_1)
x = Concatenate(name="{0}_concat_{1}_".format(name_prefix, rep))([x_0, x_1])
x = BatchNormalization(name="{0}_bn_{1}_".format(name_prefix, rep))(x)
return x
def classification_layer(self, inputs, name_prefix, classes):
x = GlobalAveragePooling2D(
name="{0}_gap2d_".format(name_prefix))(inputs)
x = Dense(classes,
kernel_initializer=get_weight_initializer(),
kernel_regularizer=get_weight_regularizer(),
name="{0}_dense_".format(name_prefix))(x)
x = Activation("softmax", name="{0}_softmax_".format(name_prefix))(x)
return x
def relu_sepconv2d_bn(self, inputs, name_prefix, rep,
kernel_size, current_filters, strides=(1,1)):
x = Activation("relu",
name="{0}_relu_{1}_".format(name_prefix,
rep))(inputs)
x = SeparableConv2D(filters=current_filters,
kernel_size=kernel_size,
strides=strides,
padding="same",
depthwise_initializer=get_weight_initializer(),
pointwise_initializer=get_weight_initializer(),
depthwise_regularizer=get_weight_regularizer(),
pointwise_regularizer=get_weight_regularizer(),
name="{0}_sepconv2d_{1}_".format(name_prefix,
rep))(x)
x = BatchNormalization(name="{0}_bn_{1}_".format(name_prefix,
rep))(x)
return x
def adjust_output_depth(self, inputs, name_prefix, current_filters):
x = Activation("relu", name="{0}_relu_".format(name_prefix))(inputs)
x = Conv2D(filters=current_filters,
kernel_size=(1, 1),
strides=(1, 1),
padding="same",
kernel_initializer=get_weight_initializer(),
kernel_regularizer=get_weight_regularizer(),
name="{0}_conv2d_".format(name_prefix))(x)
x = BatchNormalization(name="{0}_bn_".format(name_prefix))(x)
return x