def discriminator(self, input_, target, stride=2, layer_count=4):
"""
Using the PatchGAN as a discriminator
"""
input_ = tf.concat([input_, target], axis=3, name='Concat')
layer_specs = self.num_discriminator_filters * np.array([1, 2, 4, 8])
for i, output_channels in enumerate(layer_specs, 1):
with tf.variable_scope('Layer' + str(i)) as scope:
if i != 1:
input_ = BatchNorm(input_, isTrain=self.isTrain)
if i == layer_count:
stride = 1
input_ = LeakyReLU(input_)
input_ = Conv(input_,
output_channels=output_channels,
kernel_size=4,
stride=stride,
padding='VALID',
mode='discriminator')
with tf.variable_scope('Final_Layer') as scope:
output = Conv(input_,
output_channels=1,
kernel_size=4,
stride=1,
padding='VALID',
mode='discriminator')
return tf.sigmoid(output)
def generator_i(self, input_):
with tf.variable_scope("NEW", reuse=tf.AUTO_REUSE):
"""
54 Layer Tiramisu
"""
with tf.variable_scope('InputConv') as scope:
input_ = Conv(input_,
kernel_size=3,
stride=1,
output_channels=self.growth_rate * 4)
collect_conv = []
for i in range(1, 6):
input_ = self.DenseBlock(input_,
name='Encoder' + str(i),
layers=self.layers)
collect_conv.append(input_)
input_ = self.TransitionDown(input_, name='TD' + str(i))
input_ = self.DenseBlock(input_, name='BottleNeck', layers=15)
for i in range(1, 6):
input_ = self.TransitionUp(input_,
output_channels=self.growth_rate *
4,
name='TU' + str(6 - i))
input_ = tf.concat([input_, collect_conv[6 - i - 1]],
axis=3,
name='Decoder' + str(6 - i) + '/Concat')
input_ = self.DenseBlock(input_,
name='Decoder' + str(6 - i),
layers=self.layers)
with tf.variable_scope('OutputConv') as scope:
output = Conv(input_,
kernel_size=1,
stride=1,
output_channels=3)
return tf.nn.tanh(output)
def Layer(self, input_):
"""
This function creates the components inside a composite layer
of a Dense Block.
"""
with tf.variable_scope("Composite"):
next_layer = BatchNorm(input_, isTrain=self.isTrain)
next_layer = ReLU(next_layer)
next_layer = Conv(next_layer,
kernel_size=3,
stride=1,
output_channels=self.growth_rate)
next_layer = DropOut(next_layer, isTrain=self.isTrain, rate=0.2)
return next_layer
def TransitionDown(self, input_, name):
with tf.variable_scope(name):
reduction = 0.5
reduced_output_size = int(int(input_.get_shape()[-1]) * reduction)
next_layer = BatchNorm(input_,
isTrain=self.isTrain,
decay=self.decay)
next_layer = Conv(next_layer,
kernel_size=1,
stride=1,
output_channels=reduced_output_size)
next_layer = DropOut(next_layer, isTrain=self.isTrain, rate=0.2)
next_layer = AvgPool(next_layer)
return next_layer
def op(node_id, op_id, x_shape, channels, strides): # x means feature maps
br_op = None
x_id_fact_reduce = None
x_stride = strides if node_id in [
0, 1
] else 1 ## ??? why set strides=1 when x_id not in [0, 1]
if op_id == 0:
br_op = SepConv(C_in=channels,
C_out=channels,
kernel_size=3,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 1:
br_op = SepConv(C_in=channels,
C_out=channels,
kernel_size=5,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 2:
br_op = layers.AveragePooling2D(pool_size=3,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, x_shape[-1]]
elif op_id == 3:
br_op = layers.MaxPool2D(pool_size=3, strides=x_stride, padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, x_shape[-1]]
elif op_id == 4:
br_op = Identity()
if x_stride > 1:
assert x_stride == 2
x_id_fact_reduce = FactorizedReduce(C_in=x_shape[-1],
C_out=channels)
x_shape = [
x_shape[0] // x_stride, x_shape[1] // x_stride, channels
]
elif op_id == 5:
br_op = Identity()
if x_stride > 1:
assert x_stride == 2
x_id_fact_reduce = FactorizedReduce(C_in=x_shape[-1],
C_out=channels)
x_shape = [
x_shape[0] // x_stride, x_shape[1] // x_stride, channels
]
elif op_id == 6:
br_op = Conv(C_in=channels,
C_out=channels,
kernel_size=1,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 7:
br_op = Conv(C_in=channels,
C_out=channels,
kernel_size=3,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 8:
br_op = Conv(C_in=channels,
C_out=channels,
kernel_size=(1, 3),
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 9:
br_op = Conv(C_in=channels,
C_out=channels,
kernel_size=(1, 7),
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 10:
br_op = layers.MaxPool2D(pool_size=2, strides=x_stride, padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 11:
br_op = layers.MaxPool2D(pool_size=3, strides=x_stride, padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 12:
br_op = layers.MaxPool2D(pool_size=5, strides=x_stride, padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 13:
br_op = layers.AveragePooling2D(pool_size=2,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 14:
br_op = layers.AveragePooling2D(pool_size=3,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
elif op_id == 15:
br_op = layers.AveragePooling2D(pool_size=5,
strides=x_stride,
padding='same')
x_shape = [x_shape[0] // x_stride, x_shape[1] // x_stride, channels]
return br_op, x_shape, x_id_fact_reduce