我想在TensorFlow中创建一个GAN架构，但它没有学习:(
我已经在代码中进行了一些测试，但我意识到问题出在自动创建架构时：
def createNNfromTopology( topology, activation):
layers = []
out = 0
input = 0
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(topology[1], activation=leaky_relu,input_shape=(topology[0],), kernel_initializer = he_init))
for i in range(2,len(topology)):
input = topology[i-1]
out = topology[i]
if i == len(topology):
model.add(tf.keras.layers.Dense(out, activation=activation,kernel_initializer = he_init))
else:
model.add(tf.keras.layers.Dense(out, activation=leaky_relu,kernel_initializer = he_init))
return model
def CreateGan(gen_topology, dis_topology, gl,dl):
#generator
gen_model = createNNfromTopology( gen_topology, tanh)
gen_model.summary()
#discriminator
dis_model = createNNfromTopology(dis_topology, sigmoid)
dis_model.summary()
#gan
gan = tf.keras.Sequential([
gen_model,
dis_model], name="gan")
gan.summary()
dis_model.compile(optimizer=tf.train.AdamOptimizer(learning_rate=dl), loss='binary_crossentropy')
gan.compile(optimizer=tf.train.AdamOptimizer(learning_rate=gl), loss='binary_crossentropy')
return gan, gen_model, dis_model
g_topology = [sample_size, 128, image_size]
d_topology = [image_size, 128, 1]
# create a GAN, a generator and a discriminator
gan, gen_model, dis_model = CreateGan(g_topology, d_topology,g_learning_rate,d_learning_rate)
输出：
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_43 (Dense) (None, 128) 12928
_________________________________________________________________
dense_44 (Dense) (None, 784) 101136
=================================================================
Total params: 114,064
Trainable params: 114,064
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_45 (Dense) (None, 128) 100480
_________________________________________________________________
dense_46 (Dense) (None, 1) 129
=================================================================
Total params: 100,609
Trainable params: 100,609
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
sequential_18 (Sequential) (None, 784) 114064
_________________________________________________________________
sequential_19 (Sequential) (None, 1) 100609
=================================================================
Total params: 214,673
Trainable params: 214,673
Non-trainable params: 0
Epoch: 1/100 Discriminator Loss: 0.0000 Generator Loss: 16.1181
Epoch: 2/100 Discriminator Loss: 0.0000 Generator Loss: 16.1181
而如果我以这种方式创建GAN，那么它将起作用：
def make_simple_GAN(sample_size,
g_hidden_size,
d_hidden_size,
leaky_alpha,
g_learning_rate,
d_learning_rate):
gen_model = tf.keras.Sequential([
tf.keras.layers.Dense(g_hidden_size, activation=leaky_relu,input_shape=(sample_size,),kernel_initializer = he_init),
tf.keras.layers.Dense(784, activation=tanh,input_shape=(g_hidden_size,),kernel_initializer = he_init),
], name='generator')
gen_model.summary()
dis_model = tf.keras.Sequential([
tf.keras.layers.Dense(d_hidden_size, activation=leaky_relu,input_shape=(784,),kernel_initializer = he_init),
tf.keras.layers.Dense(1, activation=sigmoid,input_shape=(d_hidden_size,),kernel_initializer = he_init),
], name='discriminator')
dis_model.summary()
gan = tf.keras.Sequential([
gen_model,
dis_model
])
gan.summary()
dis_model.compile(optimizer=adamOptimizer(d_learning_rate), loss='binary_crossentropy')
gan.compile(optimizer=adamOptimizer(g_learning_rate), loss='binary_crossentropy')
return gan, gen_model, dis_model
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_55 (Dense) (None, 128) 12928
_________________________________________________________________
dense_56 (Dense) (None, 784) 101136
=================================================================
Total params: 114,064
Trainable params: 114,064
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_57 (Dense) (None, 128) 100480
_________________________________________________________________
dense_58 (Dense) (None, 1) 129
=================================================================
Total params: 100,609
Trainable params: 100,609
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
generator (Sequential) (None, 784) 114064
_________________________________________________________________
discriminator (Sequential) (None, 1) 100609
=================================================================
Total params: 214,673
Trainable params: 214,673
Non-trainable params: 0
_________________________________________________________________
Epoch: 1/100 Discriminator Loss: 0.3978 Generator Loss: 2.9608
Epoch: 2/100 Discriminator Loss: 0.2341 Generator Loss: 2.7517
我以看到here的方式填充模型，也发现了其他方法，但是我的想法更加清晰，因为我没有那么深入。 here和here
任何使此工作正常的提示？