Keras 中 GAN 的断开连接图

Question

对于此代码：

     # Initialize generator - feed noise and profile images
noise = random_normal(shape = (-1, 8, 8, z_dim), mean = 0.0, stddev = 1.0, dtype = None, seed = None)

noise      = Input(tensor = noise)
input_data = Input(shape = (128, 128, 3))

generated_img = generator_network(input_data, noise)

# Initialize discriminator - feed frontal faces as ground truth and the generated images as fake
generated_img = Input(tensor = generated_img)

true_score = discriminator_network(input_data)
fake_score = discriminator_network(generated_img)

# Optimizer
Adam_optimizer = Adam(lr = learning_rate, beta_1 = 0.9, beta_2 = 0.999, epsilon = 1e-08, decay = decay_rate)

# Losses
discrim_loss = discriminator_loss(true_score, fake_score)
var_loss     = variation_loss(input_data, generated_img)
pix_loss     = pixel_loss(input_data, generated_img)
cross_loss   = cross_entropy_loss(true_score, fake_score)
gen_loss     = generator_loss(discrim_loss, var_loss, pix_loss, cross_loss)

# Models
discriminator = Model(inputs = generated_img      , outputs = fake_score)
generator     = Model(inputs = [input_data, noise], outputs = generated_img)

# Compilers
discriminator.compile(optimizer = Adam_optimizer, loss = discriminator_loss)
generator.compile(    optimizer = Adam_optimizer, loss = generator_loss)

我收到此错误：

Traceback（最近一次调用最后一次）： 文件“main.py”，第 74 行，在 生成器=模型（输入=[输入数据，噪声]，输出=生成的img） 包装器中的文件“/home/diana/Documents/VirtualNN/local/lib/python2.7/site-packages/keras/legacy/interfaces.py”，第 87 行 返回函数（*args，**kwargs） init 中的文件“/home/diana/Documents/VirtualNN/local/lib/python2.7/site-packages/keras/engine/topology.py”，第 1793 行 str(layers_with_complete_input)) RuntimeError: Graph disconnected: 无法在“input_3”层获得张量 Tensor("conv2d_35/Relu:0", shape=(?, ?, ?, 3), dtype=float32) 的值。访问以下先前层没有问题：[]

有人知道为什么它说我的模型生成器不是连通图吗？据我了解，是有联系的。 但也许还有其他我看不到的东西。

Answer 1

如果您的目的是构建 GAN 模型，您应该将生成器网络和判别器网络包装为另一个网络中的两个连续层。例如：

from keras.models import Sequential

# generator_network() and generator_network() should each have an Input layer 
#   that defines their input shapes respectively
g_network = generator_network()
d_network = discriminator_network()

gan_network = Sequential()
gan_network.add(g_network)
d_network.trainable = False
gan_network.add(d_network)

# Compilers
d_network.compile(optimizer = Adam_optimizer, loss = discriminator_loss)
gan_network.compile(optimizer = Adam_optimizer, loss = generator_loss)