Python 文本到图像GAN-自定义块和tf.decordent函数

我试图用Keras再现本文中描述的体系结构：。这是一个用于文本到图像生成的GAN。他们提出了一个层来多次融合文本嵌入和图像特征。我已经实现了该层，但它不起作用，并且在尝试开始培训时出现以下错误：ValueError:tf.function-decorded函数尝试在非第一次调用时创建变量

以下是第一个自定义图层的代码：

class affine_T(Layer):
def __init__(self, text, units_out, **kwargs):
    super(affine_T, self).__init__(**kwargs)

    self.text = text
    self.dense_1 = Dense(units=256)
    self.dense_2 = Dense(units=units_out)
    self.dense_3 = Dense(units=256)
    self.dense_4 = Dense(units=units_out)
    self.reshape_1 = Reshape((1, 1, units_out))
    self.reshape_2 = Reshape((1, 1, units_out))
     
def build(self, input_shape):
    self.built = True

def call(self, img):

    gamma = self.dense_1(self.text)
    gamma = ReLU()(gamma)
    gamma = self.dense_2(gamma)
    gamma = self.reshape_1(gamma)

    beta = self.dense_3(self.text)
    beta = ReLU()(beta)
    beta = self.dense_4(beta)
    beta = self.reshape_2(beta)

    self.out = gamma * img + beta 
    
    return self.out

以下是使用第一个自定义图层的第二个自定义图层的代码：

class deep_F(Layer):

def __init__(self, nf_in, nf_out, text_input, **kwargs):
  super(deep_F, self).__init__(**kwargs)
  self.nf_in = nf_in
  self.nf_out = nf_out
  self.text = text_input
  self.units_out = nf_in

  self.affine_1 = affine_T(self.text, self.units_out)
  self.affine_2 = affine_T(self.text, self.units_out)

  self.affine_3 = affine_T(self.text, self.units_out)
  self.affine_4 = affine_T(self.text, self.units_out)
  
  self.conv2d_1 = Conv2D(filters=self.nf_in, kernel_size=(3, 3), strides=1, padding='same')
  self.conv2d_2 = Conv2D(filters=self.nf_out, kernel_size=(3, 3), strides=1, padding='same')

def build(self, input_shape):
    self.built = True

def call(self, inputs):
  image_features = inputs
  
  shortcut = image_features #preserve input information for residual block

  #Fusion Block
  x = self.affine_1(image_features)
  x = LeakyReLU(alpha=0.2)(x)
  x = self.affine_2(x)
  x = LeakyReLU(alpha=0.2)(x)
  x = self.conv2d_1(x)

  #Fusion Block 
  x_1 = self.affine_3(x)
  x_1 = LeakyReLU(alpha=0.2)(x_1)
  x_1 = self.affine_4(x_1)
  x_1 = LeakyReLU(alpha=0.2)(x_1)
  x_1 = self.conv2d_2(x_1)
  
  #check number of channels for residual operation
  if self.nf_in != self.nf_out: 
    shortcut = Conv2D(filters=self.nf_out, kernel_size=(1, 1), strides=1, padding='valid') (shortcut) #!


  return add([x_1, shortcut]) #Residual 

相反，发电机是这样的：

input_z = Input(shape = (100,))

input_text = Input(shape = (256,))
#i = Dense(units=128)(input_text)
#i = LeakyReLU(alpha=0.2)(i)

concat_layer = Concatenate(axis=1)([input_z, input_text])
x = Dense(4*4*8*64)(concat_layer)
x = Reshape((4, 4, 64*8))(x)

#x = affine_T(i, 512)(x)
x = UpSampling2D(size=(2, 2))(x)
x = deep_F(512, 512, input_text)(x)

x = UpSampling2D(size=(2, 2))(x)
x = deep_F(512, 256, input_text)(x)

x = UpSampling2D(size=(2, 2))(x)
x = deep_F(256, 128, input_text)(x)

x = UpSampling2D(size=(2, 2))(x)
x = deep_F(128, 128, input_text)(x)

x = UpSampling2D(size=(2, 2))(x)
x = deep_F(128, 128, input_text)(x)

#x = UpSampling2D(size=(2, 2))(x)
#x = deep_F(512, 512, i)(x)
x = LeakyReLU(alpha=0.2)(x)
x_out = Conv2DTranspose(filters=3, kernel_size=(3, 3), strides=1, padding='same', use_bias=False, activation='tanh')(x)

• 是因为我多次使用文本输入吗
• 我是否在自定义层实现方面做错了什么

---

非常感谢你的帮助

因为我只使用API中的默认层，所以我通过将这些层包装到python函数中解决了这个问题。无论如何，在未来，我将研究如何更好地创建自定义层。