def _construct_decoder(self):
"""
CNN decoder.
"""
z = Input(shape=(self.latent_dim,))
z0 = Dense(self.hidden_dim)(z)
z_reshp = Reshape((1, 1, self.hidden_dim))(z0)
z1 = bn_deconv_layer(z_reshp, 256, 4, 4)
z2 = bn_deconv_layer(z1, 128, 4, 2)
z3 = bn_deconv_layer(z2, 64, 4, 2)
z4 = bn_deconv_layer(z3, 32, 4, 2)
gen_img = bn_deconv_layer(z4, self.img_dim[-1], 4, 2,
activation='sigmoid',
batchnorm=False)
decoder = Model(z, gen_img)
return decoder
def _construct_generator(self):
z = Input(shape=(self.in_dim, ))
# z0 = bn_dense(z, self.g_hidden * 64 * 4 * 4, 'selu')
z0 = Reshape((1, 1, self.in_dim))(z)
z1 = bn_deconv_layer(z0,
16 * self.g_hidden,
4,
4,
activation='selu',
batchnorm=False)
# z2 = bn_deconv_layer(
# z1, 32 * self.g_hidden, 4, 2, activation='selu', batchnorm=False)
# z3 = bn_deconv_layer(
# z1, 16 * self.g_hidden, 4, 2, activation='selu', batchnorm=False)
z4 = bn_deconv_layer(z1,
8 * self.g_hidden,
4,
2,
activation='selu',
batchnorm=False)
z5 = bn_deconv_layer(z4,
4 * self.g_hidden,
4,
2,
activation='selu',
batchnorm=False)
z6 = bn_deconv_layer(z5,
2 * self.g_hidden,
4,
2,
activation='selu',
batchnorm=False)
gen_img = bn_deconv_layer(z6,
self.img_dim[-1],
4,
2,
activation='sigmoid',
batchnorm=False)
generator = Model(z, gen_img)
# generator.compile(optimizer=self.critic_opt(lr=self.critic_lr),
# loss='binary_crossentropy')
return generator
def _construct_decoder(self):
"""
CNN decoder.
"""
z = Input(shape=(self.latent_dim, ))
z0 = Dense(self.hidden_dim)(z)
deconv_block = deconvnet(z0, self.img_dim, self.dec_param, bias=False)
gen_img = bn_deconv_layer(deconv_block,
self.img_dim[-1],
4,
2,
activation='sigmoid',
batchnorm=False)
decoder = Model(z, gen_img)
return decoder
def _construct_generator(self):
z = Input(shape=(self.in_dim, ))
z0 = Reshape((1, 1, self.in_dim))(z)
deconv_block = deconvnet(z0, self.img_dim, self.g_hidden, bias=False)
gen_img = bn_deconv_layer(deconv_block,
self.img_dim[-1],
4,
2,
activation='sigmoid',
batchnorm=False,
use_bias=False)
generator = Model(z, gen_img)
generator.compile(
optimizer=self.critic_opt(lr=self.critic_lr),
# loss='binary_crossentropy')
loss='mse')
return generator
def _construct_imggen(self):
"""
CNN decoder/generator.
The decoder and generator share structure but do NOT share weights.
"""
z = Input(shape=(self.latent_dim, ))
z0 = Dense(512)(z)
deconv_block = deconvnet(z0,
self.img_dim,
self.dec_param,
activation='elu')
gen_img = bn_deconv_layer(deconv_block,
self.img_dim[-1],
4,
2,
activation='sigmoid',
batchnorm=False)
generator = Model(z, gen_img)
return generator
def _construct_generator(self):
z = Input(shape=(self.in_dim, ))
z1 = bn_dense(z, 512)
z_reshp = Reshape((1, 1, 512))(z1)
deconv_block = deconvnet(z_reshp,
self.img_dim,
self.g_hidden,
activation='selu',
batchnorm=True,
bias=False)
gen_img = bn_deconv_layer(deconv_block,
self.img_dim[-1],
4,
2,
activation='tanh',
batchnorm=False,
use_bias=False)
generator = Model(z, gen_img)
# generator.compile(optimizer=self.critic_opt(lr=self.critic_lr),
# loss='mse')
return generator