def __init__(self, latent_size_prev, latent_size):
super(StyleBlock, self).__init__()
self.lrelu = lambda x: tf.keras.activations.relu(x, alpha=0.2)
self.std_dense = layers.ScaledDense(
latent_size_prev, bias_initializer=tf.ones_initializer())
self.latent_size = latent_size
self.conv_outsize = latent_size
if USE_POINTWISE_DENSES:
self.D1 = layers.ScaledDense(latent_size * 4,
activation=self.lrelu)
self.D2 = layers.ScaledDense(latent_size, activation=self.lrelu)
def __init__(self, latent_size, n_denses):
super(LatentMapping, self).__init__()
self.lrelu = lambda x: tf.keras.activations.relu(x, alpha=0.2)
self.denses = [
layers.ScaledDense(latent_size, activation=self.lrelu)
for _ in range(n_denses)
]
def __init__(self):
super(GAN_d, self).__init__()
self.lrelu = lambda x: tf.keras.activations.relu(x, alpha=0.2)
self.cstart = layers.ScaledDense(
LATENT_SIZES[-1], activation=self.lrelu) #equivalent to 1x1 conv
self.blocks = [
D_block(LATENT_SIZES[-1], True, minibatch_stddev=True),
D_block(LATENT_SIZES[-2], True, minibatch_stddev=True),
D_block(LATENT_SIZES[-3], True, minibatch_stddev=True),
D_block(LATENT_SIZES[-4], True, minibatch_stddev=True),
D_block(LATENT_SIZES[-5], True, minibatch_stddev=True),
D_block(LATENT_SIZES[-6], False, minibatch_stddev=True)
]
self.cend = layers.ScaledDense(
LATENT_SIZES[-6], activation=self.lrelu
) #the last layer is basically a dense over the whole thing.
self.cend2 = layers.ScaledDense(1, use_bias=False)
def __init__(self, latent_size, downsample, minibatch_stddev=False):
super(D_block, self).__init__()
self.lrelu = lambda x: tf.keras.activations.relu(x, alpha=0.2)
self.conv1 = layers.ScaledConv2D(latent_size,
activation=self.lrelu,
filter_size=3)
self.conv2 = layers.ScaledConv2D(latent_size,
activation=self.lrelu,
filter_size=3)
self.conv_residual = layers.ScaledDense(
latent_size, activation=self.lrelu) #equivalent to 1x1 conv
if downsample:
self.resizer = tf.keras.layers.AveragePooling2D()
else:
self.resizer = None
self.minibatch_stddev = minibatch_stddev
self.normalisation_coef = tf.math.rsqrt(2.0)
def __init__(self, image_channels, latent_size):
super(PicOutBlock, self).__init__()
self.std_dense = layers.ScaledDense(
latent_size, bias_initializer=tf.ones_initializer())
self.pic_out = layers.ScaledDense(image_channels, use_bias=False)