def __init__(self, config, scaling_factor, dropout):
super(
Deep128NoiseMsDiscS2Alternative.Discriminator.
MultiscaleDisc, self).__init__()
assert scaling_factor > 0
if scaling_factor != 1:
size_x = int(128 * scaling_factor)
size_y = int(128 * scaling_factor)
tf.logging.info(
"Multiscale discriminator operating on resolution: {}x{}"
.format(size_x, size_y))
self.resize = lambda x: tf.image.resize_nearest_neighbor(
x, (size_x, size_y))
else:
tf.logging.info(
"Multiscale discriminator operating on regular resolution"
)
self.resize = lambda x: x
initial_filters = 32 * 1
self.blocks = [
ConvBlock(initial_filters * 1, 4, 2),
ConvBlock(initial_filters * 2, 4, 2),
ConvBlock(initial_filters * 4, 4, 2),
ConvBlock(initial_filters * 8, 4, 2),
]
self.dropout = dropout
self.flatten = Flatten()
# self.pre_fc = Dense(1000, use_bias=False)
self.fc = Dense(config.discriminator_classes, use_bias=False)
def __init__(self, config, scaling_factor, dropout):
super(Deep240pTo480p.Discriminator.MultiscaleDisc,
self).__init__()
assert scaling_factor > 0
if scaling_factor != 1:
size_x = int(640 * scaling_factor)
size_y = int(480 * scaling_factor)
tf.logging.info(
"Multiscale discriminator operating on resolution: {}x{}"
.format(size_x, size_y))
self.resize = lambda x: tf.image.resize_nearest_neighbor(
x, (size_x, size_y))
else:
tf.logging.info(
"Multiscale discriminator operating on regular resolution"
)
self.resize = lambda x: x
initial_filters = 32 // 1
self.blocks = [
ConvBlock(initial_filters * 2, 5, 2),
ConvBlock(initial_filters * 2, 5, 1),
ConvBlock(initial_filters * 4, 5, 2),
ConvBlock(initial_filters * 4, 5, 1),
ConvBlock(initial_filters * 8, 5, 2),
ConvBlock(initial_filters * 8, 5, 1),
ConvBlock(initial_filters * 16, 5, 2),
ConvBlock(initial_filters * 16, 5, 1),
]
self.dropout = dropout
self.flatten = Flatten()
self.fc = Dense(config.discriminator_classes, use_bias=False)
def __init__(self, config):
super(Deep240pTo480p.Generator, self).__init__()
initial_filters = 1024 // 4
self.blocks = [
ConvBlock(initial_filters * 1, 5, 1),
ConvBlock(initial_filters * 2, 5, 1),
DeconvBlock(initial_filters * 4, 5, 2),
ConvBlock(initial_filters * 2, 5, 1),
ConvBlock(initial_filters * 1, 5, 1),
]
self.final_conv = Conv(3 if config.has_colored_target else 1, 5, 1)
self._low_res_generator = None
def __init__(self, config):
super(Deep480NoiseS2.Discriminator, self).__init__()
initial_filters = 32 // 2 * 1
self.blocks = [
ConvBlock(initial_filters * 2, 4, 2),
ConvBlock(initial_filters * 4, 4, 2),
ConvBlock(initial_filters * 8, 4, 2),
ConvBlock(initial_filters * 16, 4, 2),
ConvBlock(initial_filters * 32, 4, 2),
]
self.dropout = Dropout(0.3)
self.flatten = Flatten()
self.fc = Dense(config.discriminator_classes, use_bias=False)
def __init__(self, config):
super(Deep480NoiseS2.Generator, self).__init__()
initial_filters = int(512 / 32) * 1
self.fc = tf.keras.layers.Dense(15 * 15 * 64, use_bias=False)
self.initial_norm = tf.keras.layers.BatchNormalization()
self.blocks = [
DeconvBlock(initial_filters * 32, 5, 2),
ConvBlock(initial_filters * 16, 5, 1),
DeconvBlock(initial_filters * 16, 5, 2),
ConvBlock(initial_filters * 8, 5, 1),
DeconvBlock(initial_filters * 8, 5, 2),
ConvBlock(initial_filters * 4, 5, 1),
DeconvBlock(initial_filters * 4, 5, 2),
ConvBlock(initial_filters * 2, 5, 1),
DeconvBlock(initial_filters * 2, 5, 2),
ConvBlock(initial_filters * 1, 5, 1),
]
self.final_conv = Conv(3 if config.has_colored_target else 1, 5, 1)