2

我正在重新组织我的代码以便于阅读，但是在编译时它说： ValueError: No gradients provided for any variable: ['enc_conv_4/kernel:0'.... 我知道我的损失函数是可微的，因为代码在触摸它之前就已经工作了，但是现在我的模型的梯度丢失了。

    @tf.function
    def train_disc(self,real_imgs,gen_imgs):
      with tf.GradientTape() as disc_tape:
        d_loss = self.wasserstein_loss(real_imgs,gen_imgs)
      gradients_d = disc_tape.gradient(d_loss, self.discriminator.trainable_variables)
      self.d_optimizer.apply_gradients(zip(gradients_d, self.discriminator.trainable_variables))
      return d_loss

    @tf.function
    def train_gen(self,real_img,gen_imgs,mask,img_feat,rot_feat_mean):
      with tf.GradientTape() as gen_tape:
        g_loss_param = self.generator_loss(mask,img_feat,rot_feat_mean)
        g_loss = g_loss_param(real_img, gen_imgs)
      gradients_g = gen_tape.gradient(g_loss, self.generator.trainable_variables)
      print(gradients_g)
      self.g_optimizer.apply_gradients(zip(gradients_g, self.generator.trainable_variables))

正如你所看到的，当我对鉴别器和生成器做同样的事情时，生成器给了我一个空的梯度列表。

gen_imgs = self.generator([real_img, mask], training=True)


d_loss = self.train_disc(real_img,gen_imgs[:,:,:,:-1])

if step%self.n_critic == 0:
  masked_images = real_img * mask
  idx = 3  # index of desired layer
  layer_input = Input(shape=(self.img_shape))  #
  x = layer_input
  for layer in self.generator.layers[idx:idx+12]:
      x = layer(x)
  model_feat = Model(inputs=layer_input,outputs=x)
  model_feat.trainable = False
  img_feat = model_feat(masked_images,training=False)
  rot_feat_mean = []
  for i in range(self.batch_size):
      rot = []
      for an in [180, 155, 130, 105, 80, 55, 20, 10]:
          r = tf.keras.preprocessing.image.random_rotation(masked_images[i], an, row_axis=0, col_axis=1,
                                                           channel_axis=2)
          rot.append(r)
      rot = np.array(rot)
      rot_feat_mean.append(np.mean(model_feat(rot,training=False),axis=0))
  rot_feat_mean = np.array(rot_feat_mean)
  g_loss = self.train_gen(real_img,gen_imgs[:,:,:,:-1],mask,img_feat,rot_feat_mean)

最后一段代码的最后一行给了我一个错误。我不知道这个错误是否是由于任何语义错误。

4

0 回答 0