def __init__(self):
initializer = initializers.HeNormal()
super(Generator, self).__init__(
# num of noise that becomes the seed of Generation is 100
l0z=L.Linear(100, 7 * 7 * 128, initialW=initializer),
ps1=PixelShuffler(2, 128, 256),
ps2=PixelShuffler(2, 64, 4),
bn0=L.BatchNormalization(7 * 7 * 128),
bn1=L.BatchNormalization(64),
)
def wdsr(scale, filters, n_resblocks, res_block):
x_in = Input(shape=(None, None, 3))
m = Conv2D(filters, 3, padding='same')(x_in)
for i in range(n_resblocks):
m = res_block(m, filters)
m = Conv2D(3 * scale**2, 3, padding='same')(m)
m = PixelShuffler(scale)(m)
s = Conv2D(3 * scale**2, 5, padding='same')(x_in)
s = PixelShuffler(scale)(s)
x = Add()([m, s])
return Model(x_in, x)
def encoder():
inputs = Input(shape=IMAGE_SHAPE)
x = inputs
##x = K.constant(inp)
#x = Conv2D( 64, kernel_size=5, strides=2, padding='same' )(x)
#x = LeakyReLU(0.1)(x)
x = Conv2D(128, kernel_size=5, strides=2, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = Conv2D(256, kernel_size=5, strides=2, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = Conv2D(512, kernel_size=5, strides=2, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = Conv2D(1024, kernel_size=5, strides=2, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = Dense(1024)(Flatten()(x))
x = Dense(8 * 8 * 1024)(x)
x = Reshape((8, 8, 1024))(x)
x = PixelShuffler()(x)
x = Conv2D(256 * 4, kernel_size=3, padding='same')(x)
x = LeakyReLU(0.1)(x)
#x = Conv2D( 256*2, kernel_size=3, padding='same' )(x)
#x = LeakyReLU(0.1)(x)
#return x
#arr=K.eval(x)
#return arr
return Model(inputs, x)
def block(x):
x = Conv2D(filters * 4,
kernel_size=5,
use_bias=False,
kernel_initializer=RandomNormal(0, 0.02),
padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
return x
def block(x):
x = Conv2D(out_shape[2] * 4, 3, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
shape = x.get_shape().as_list()[1:]
cx = shape[0] - out_shape[0]
cy = shape[1] - out_shape[1]
if cx != 0 or cy != 0:
x = Cropping2D(((0, cx), (0, cy)))(x)
return x
def block(x, use_instance_norm=use_instance_norm):
x = Conv2D(filters * 4,
kernel_size=3,
use_bias=False,
kernel_initializer=RandomNormal(0, 0.02),
padding='same')(x)
if use_instance_norm:
x = inst_norm()(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
return x
def decoder():
#x = K.constant(inp)
inputs = Input(shape=(8, 8, 512))
x = inputs
x = Conv2D(1024, kernel_size=3, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
x = Conv2D(128 * 4, kernel_size=3, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
x = Conv2D(64 * 4, kernel_size=3, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
x = Conv2D(32 * 4, kernel_size=3, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
#x = Conv2D( 16*4, kernel_size=3, padding='same' )(x)
#x = LeakyReLU(0.1)(x)
#x = PixelShuffler()(x)
x = Conv2D(3, kernel_size=5, padding='same', activation="sigmoid")(x)
return Model(inputs, x)
def residual_decoder(h, w, c=3, k=4, latent_dim=2, dropout_rate=0.1):
inputs_ = Input(shape=(latent_dim, ))
hidden = inputs_
transform = Dense(h * w * 512, kernel_regularizer=l2(0.001))(hidden)
transform = LeakyReLU(0.1)(transform) # more nonlinearity
reshape = Reshape((h, w, 512))(transform)
x = reshape # 2x2@512
x = Dropout(dropout_rate)(x) # prevent overfitting
x = up_bilinear()(x) # 4x4@512
x = Conv2DTranspose(128, k, padding='same')(x) # 4x4@128
x = LeakyReLU(0.2)(x)
x = up_bilinear()(x) # 8x8@128
x = Conv2DTranspose(128, k, padding='same')(x) # 8x8@128
x = LeakyReLU(0.2)(x)
x = up_bilinear()(x) # 16x16@128
x = Conv2DTranspose(64, k, padding='same')(x) # 16x16@64
x = LeakyReLU(0.2)(x)
x = _res_conv(64, k, dropout_rate)(x) # 16x16@64
x = PixelShuffler()(x) # 32x32@16
x = Conv2DTranspose(32, k, padding='same')(x) # 32x32@32
x = LeakyReLU(0.2)(x)
x = _res_conv(32, k, dropout_rate)(x) # 32x32@32
outputs = conv(c, k, 1, act='tanh')(x) # 32x32@c
model = Model([inputs_], [outputs])
return model
def block(x):
x = Conv2D(filters * 4, kernel_size=3, padding='same')(x)
x = LeakyReLU(0.1)(x)
x = PixelShuffler()(x)
return x