def create_g2(self):
depth = 64
dim = 16
alpha = 0.2
channel = 3
self.G.add(Dense(depth * dim * dim, input_dim=self.noise_size))
self.G.add(Reshape((dim, dim, depth)))
MyDCGAN.add_cbl(self.G, depth, 5, 1, 0.2)
MyDCGAN.add_dbr(self.G, depth, 11, 4)
MyDCGAN.add_cbl(self.G, int(depth / 2), 4, 1, 0.2)
MyDCGAN.add_dbr(self.G, int(depth / 2), 7, 4)
MyDCGAN.add_cbl(self.G, int(depth / 4), 5, 1, 0.2)
self.G.add(Conv2D(channel, 7, strides=1, padding='same'))
self.G.summary()
def create_g1(self):
depth = 64 * int(np.power(2, 4))
channel = 3
dim = 8
momentum = 0.5
self.G.add(
Dense(dim * dim * channel * depth, input_dim=self.noise_size))
self.G.add(BatchNormalization(momentum=momentum))
self.G.add(Reshape((dim, dim, channel)))
for i in range(1, 4):
MyDCGAN.add_dbr(self.G, int(depth / np.power(2, i)), 4, 2,
momentum)
self.G.add(
Conv2DTranspose(int(depth / np.power(2, 4)), 4, 2, padding='same'))
self.G.add(Activation('tanh'))
self.G.summary()
def create_d5(self):
depth = 64
dim = 4
channel = 3
momentum = 0.9
dropout = 0.4
alpha = 0.2
self.D.add(
Conv2D(depth,
11,
strides=4,
input_shape=self.input_shape,
padding='same'))
self.D.add(LeakyReLU(alpha=alpha))
self.D.add(Dropout(dropout))
MyDCGAN.add_cbl(self.D, depth * 2, 11, 4, alpha)
self.D.add(Dropout(dropout))
MyDCGAN.add_cbl(self.D, depth * 4, 11, 4, alpha)
self.D.add(Dropout(dropout))
MyDCGAN.add_cbl(self.D, depth * 8, 5, 2, alpha)
self.D.add(Flatten())
self.D.add(Dropout(dropout))
self.D.add(Dense(1))
self.D.add(Activation('sigmoid'))
self.D.summary()
def create_d1(self):
depth = 64
alpha = 0.2
channel = 3
self.D.add(
Conv2D(depth * channel,
4,
strides=2,
input_shape=(256, 256, 3),
padding='same'))
self.D.add(LeakyReLU(alpha=alpha))
for i in range(1, 5):
MyDCGAN.add_cbl(self.D,
64 * int(np.power(2, i)) * channel,
4,
2,
alpha=0.2,
bn=True)
self.D.add(Flatten())
self.D.add(Dense(1))
self.D.add(Activation('sigmoid'))
self.D.summary()
def __init__(self,
input_shape,
model=1,
noise_size=100,
kdsfromzip=False,
batchsize=32,
print_term=32,
epoch=100,
D_optimizer=Adam(lr=0.001),
G_optimizer=Adam(lr=0.0005),
folder='kianap'):
self.input_shape = input_shape
self.D = Sequential()
self.G = Sequential()
self.noise_size = noise_size
if model == 1:
self.create_d1()
self.create_g1()
if model == 2:
self.create_d2()
self.create_g2()
if model == 3:
self.create_d3()
self.create_g3()
if model == 4:
self.create_d4()
self.create_g4()
if model == 5:
self.create_d5()
self.create_g5()
self.kdata = KianaDataSet(img_size=input_shape[0],
load_from_zip=kdsfromzip,
folder=folder)
self.gan = MyDCGAN(self.noise_size,
D=self.D,
G=self.G,
batchsize=batchsize,
print_term=print_term,
Dfname='k_d_w.h5',
Gfname='k_g_w.h5',
epoch=epoch,
D_optimizer=D_optimizer,
G_optimizer=D_optimizer,
img_size=input_shape[0])
self.gan.img_data = self.kdata.normalized
def create_g5(self):
depth = 64 * 8
dim = 4
channel = 3
momentum = 0.9
dropout = 0.05
alpha = 0.2
self.G.add(Dense(dim * dim * depth, input_dim=self.noise_size))
self.G.add(Reshape((dim, dim, depth)))
MyDCGAN.add_dbr(self.G, int(depth), 11, 4)
MyDCGAN.add_dbr(self.G, int(depth / 2), 11, 4)
MyDCGAN.add_dbr(self.G, int(depth / 4), 5, 2)
MyDCGAN.add_dbr(self.G, int(depth / 8), 5, 2)
self.G.add(Conv2D(channel, 4, strides=1, padding='same'))
self.G.add(Activation('tanh'))
self.G.summary()
def create_g3(self):
depth = 64 * 4
alpha = 0.2
channel = 3
dim = 16
momentum = 0.8
self.G.add(Dense(depth * dim * dim, input_dim=self.noise_size))
self.G.add(BatchNormalization(momentum=momentum))
self.G.add(Reshape((dim, dim, depth)))
MyDCGAN.add_dbr(self.G, depth, 7, 2, bn_momentum=momentum)
MyDCGAN.add_dbr(self.G, int(depth / 2), 7, 2, bn_momentum=momentum)
MyDCGAN.add_dbr(self.G, int(depth / 4), 7, 2, bn_momentum=momentum)
self.G.add(Conv2D(3, 2, padding='same'))
self.G.add(Activation('tanh'))
self.G.summary()
def create_d3(self):
depth = 64
alpha = 0.2
channel = 3
dropout = 0.4
self.D.add(
Conv2D(depth, 3, input_shape=self.input_shape, padding='same'))
self.D.add(LeakyReLU(alpha=alpha))
MyDCGAN.add_cbl(self.D, depth * 2, 4, 2, 0.2)
MyDCGAN.add_cbl(self.D, depth * 4, 4, 2, 0.2)
MyDCGAN.add_cbl(self.D, depth * 8, 4, 2, 0.2)
self.D.add(Flatten())
self.D.add(Dropout(dropout))
self.D.add(Dense(1))
self.D.add(Activation('sigmoid'))
self.D.summary()
def create_g4(self):
depth = 32 + 32 + 32 + 32
dim = 8
channel = 3
momentum = 0.9
dropout = 0.3
alpha = 0.2
self.G.add(Dense(dim * dim * depth, input_dim=self.noise_size))
self.G.add(Reshape((dim, dim, depth)))
MyDCGAN.add_cbl(self.G, depth, 5, 1, alpha=alpha)
self.G.add(UpSampling2D())
MyDCGAN.add_dbr(self.G, int(depth), 4, 2)
self.G.add(UpSampling2D())
MyDCGAN.add_dbr(self.G, int(depth / 2), 4, 2)
self.G.add(Conv2D(channel, 7, strides=1, padding='same'))
self.G.add(Activation('tanh'))
self.G.summary()