def build(self, input_shape):
#constant layer
self.const_4_4 = self.add_weight(
name='4x4/Const/const',
shape=(1, 512, 4, 4),
initializer=tf.random_normal_initializer(0, 1),
trainable=True)
#early layer 4x4
self.layer_4_4 = SynthesisMainLayer(fmaps=nf(1),
impl=self.impl,
gpu=self.gpu,
name='4x4')
self.torgb_4_4 = ToRgbLayer(impl=self.impl, gpu=self.gpu, name='4x4')
#main layers
for res in range(3, self.resolution_log2 + 1):
res_str = str(2**res)
setattr(
self, 'layer_{}_{}_up'.format(res_str, res_str),
SynthesisMainLayer(fmaps=nf(res - 1),
impl=self.impl,
gpu=self.gpu,
up=True,
name='{}x{}'.format(res_str, res_str)))
setattr(
self, 'layer_{}_{}'.format(res_str, res_str),
SynthesisMainLayer(fmaps=nf(res - 1),
impl=self.impl,
gpu=self.gpu,
name='{}x{}'.format(res_str, res_str)))
setattr(
self, 'torgb_{}_{}'.format(res_str, res_str),
ToRgbLayer(impl=self.impl,
gpu=self.gpu,
name='{}x{}'.format(res_str, res_str)))
def build(self, input_shape):
self.conv2d_0 = Conv2DLayer(fmaps=nf(self.res - 1),
kernel=3,
impl=self.impl,
gpu=self.gpu,
name='Conv0')
self.bias_0 = self.add_weight(name='Conv0/bias',
shape=(nf(self.res - 1), ),
initializer=tf.random_normal_initializer(
0, 1),
trainable=True)
self.conv2d_1_down = Conv2DLayer(fmaps=nf(self.res - 2),
kernel=3,
down=True,
resample_kernel=self.resample_kernel,
impl=self.impl,
gpu=self.gpu,
name='Conv1_down')
self.bias_1_down = self.add_weight(
name='Conv1_down/bias',
shape=(nf(self.res - 2), ),
initializer=tf.random_normal_initializer(0, 1),
trainable=True)
self.conv2d_skip = Conv2DLayer(fmaps=nf(self.res - 2),
kernel=1,
down=True,
resample_kernel=self.resample_kernel,
impl=self.impl,
gpu=self.gpu,
name='Skip')
def build(self, input_shape):
self.mini_btch_std_layer = MinibatchStdLayer()
self.from_rgb = FromRgbLayer(fmaps=nf(self.resolution_log2-1),
name='{}x{}'.format(self.resolution, self.resolution),
impl=self.impl, gpu=self.gpu)
for res in range(self.resolution_log2, 2, -1):
res_str = str(2**res)
setattr(self, 'block_{}_{}'.format(res_str, res_str),
BlockLayer(res=res, name='{}x{}'.format(res_str, res_str),
impl=self.impl, gpu=self.gpu))
#last layers
self.conv_4_4 = Conv2DLayer(fmaps=nf(1), kernel=3, impl=self.impl,
gpu=self.gpu, name='4x4/Conv')
self.conv_4_4_bias = self.add_weight(name='4x4/Conv/bias', shape=(512,),
initializer=tf.random_normal_initializer(0,1), trainable=True)
self.dense_4_4 = DenseLayer(fmaps=512, name='4x4/Dense0')
self.dense_output = DenseLayer(fmaps=1, name='Output')