模型的输出张量必须是模型Api Tensorfow中的TensorFlow`Layer`（因此保留了过去的层元数据）的输出

Question

def generator_model(self):

        input_images = Input(shape=[64,64,1])
        layer1= Conv2D(self.filter_size,self.kernel_size,(2,2),padding='same',use_bias=False,kernel_initializer='random_uniform')(input_images)
        layer1=LeakyReLU(0.2)(layer1)

        layer2= Conv2D(self.filter_size*2,self.kernel_size,(2,2),padding='same',use_bias=False,kernel_initializer='random_uniform')(layer1)
        layer2=BatchNormalization()(layer2)
        layer2=LeakyReLU(0.2)(layer2)

        layer3=Conv2D(self.filter_size*4,self.kernel_size,(2,2),padding='same',use_bias=False,kernel_initializer='random_uniform')(layer2)
        layer3=BatchNormalization()(layer3)
        layer3=LeakyReLU(0.2)(layer3) 

        layer4=Conv2D(self.filter_size*8,self.kernel_size,(2,2),padding='same',use_bias=False,kernel_initializer='random_uniform')(layer3)
        layer4=BatchNormalization()(layer4)
        layer4=LeakyReLU(0.2)(layer4)  

        layer5=Conv2D(self.filter_size*16,self.kernel_size,(2,2),padding='same',use_bias=False,kernel_initializer='random_uniform')(layer4)
        layer5=BatchNormalization()(layer5)
        layer5=LeakyReLU(0.2)(layer5)            

        up_layer5 = Conv2DTranspose(self.filter_size*8,self.kernel_size,strides = (2,2),padding='same',use_bias=False)(layer5)
        up_layer5=BatchNormalization()(up_layer5)
        up_layer5=LeakyReLU(0.2)(up_layer5)
        #shape = 4*4*512
        up_layer5_concat = tf.concat([up_layer5,layer4],0)

        up_layer6 = Conv2DTranspose(self.filter_size*4,self.kernel_size,strides = (2,2),padding='same',use_bias=False)(up_layer5_concat)
        up_layer6 =BatchNormalization()(up_layer6)
        up_layer6 =LeakyReLU(0.2)(up_layer6)
                    up_layer_6_concat = tf.concat([up_layer6,layer3],0)

        up_layer7 = Conv2DTranspose(self.filter_size*2,self.kernel_size,strides = (2,2),padding='same',use_bias=False)(up_layer_6_concat)
        up_layer7 =BatchNormalization()(up_layer7)
        up_layer7 =LeakyReLU(0.2)(up_layer7)
        up_layer_7_concat = tf.concat([up_layer7,layer2],0)

        up_layer8 = Conv2DTranspose(self.filter_size,self.kernel_size,strides = (2,2),padding='same',use_bias=False)(up_layer_7_concat)
        up_layer8 =BatchNormalization()(up_layer8)
        up_layer8 =LeakyReLU(0.2)(up_layer8)
        up_layer_8_concat = tf.concat([up_layer8,layer1],0)    
        output = Conv2D(3,self.kernel_size,strides = (1,1),padding='same',use_bias=False)(up_layer_8_concat)
        final_output = LeakyReLU(0.2)(output)

        model = Model(input_images,output)
        model.summary()
        return model

This is how my generator_model looks like, and I have followed a research paper to make the architecture. But, I am in problem with the error. I have checked the other solutions to given problem here in SO, but none of them worked for me as they are little bit different maybe. My guess, the problem is there with the tf.concat() function which should be put as tensorflow keras layer of Lambda, but I tried that too and of no help. Any help regarding this issue? Bugging me for 2 days now.

Answer 1

When you define a model using the Keras functional API, you must use the Keras Layers to build your model.

Therefore you are right, the problem is in your tf.concat invocation.

In the tf.keras.layers package, however, you can find the Concatenate layer, that uses the functional API too.

Thus, you can replace your concat layers from:

up_layer5_concat = tf.concat([up_layer5,layer4],0)

to

up_layer5_concat = tf.keras.layers.Concatenate()([up_layer5, layer4])

And so on for every other tf.concat invocation in your network