ValueError: No gradients provided for any variable in gan network

TensorFlow General Discussion
1

I am a novice, I am building a gan network, but I found that there is no gradient transfer in my network, the following is the reproducible code, I do not know where the problem, can help me?

# Model
def generator():
    model=tf.keras.Sequential()
    model.add(tf.keras.layers.InputLayer(input_shape=(opt.latent_dim,)))
    model.add(tf.keras.layers.Dense(1024, activation="relu"))
    model.add(tf.keras.layers.Dense(5024, activation="relu"))
    model.add(tf.keras.layers.Dense(10 * 10 * 128,activation="relu"))
    model.add(tf.keras.layers.Reshape([10, 10, 128]))
    model.add(tf.keras.layers.Conv2DTranspose(128, kernel_size=4, strides=2, padding="SAME",activation="relu"))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.Conv2DTranspose(128, kernel_size=4, strides=4, padding="SAME",activation="relu"))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.Conv2DTranspose(128, kernel_size=4, strides=2, padding="SAME",activation="relu"))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.Conv2DTranspose(3, kernel_size=4, strides=2, padding="SAME",activation="tanh"))
    return model
Generator = generator()
Generator.summary()

def discriminator():
    input_values_dim = Input(shape=(opt.latent_dim,))
    x1 = Dense(units=8, activation='relu')(input_values_dim)
    x1 = Dense(units=4, activation='relu')(x1)
    mlp = Model(input_values_dim, x1)
    #mlp.summary()
    filters = (16, 32, 64)
    input_images_dim = Input(shape=(opt.img_size*2,opt.img_size*2,opt.channels))
    for (i, f) in enumerate(filters):
        if i == 0:
            x2 = input_images_dim
            x2 = Conv2D(f, (3, 3), padding="same")(x2)
            x2 = Activation("relu")(x2)
            x2 = BatchNormalization()(x2)
            x2 = MaxPooling2D(pool_size=(2, 2))(x2)
            # flatten the volume, then FC => RELU => BN => DROPOUT
            x2 = Flatten()(x2)
            x2 = Dense(16)(x2)
            x2 = Activation("relu")(x2)
            x2 = BatchNormalization()(x2)
            x2 = Dropout(0.5)(x2)
            x2 = Dense(4)(x2)
            x2 = Activation("relu")(x2)
        cnn = Model(input_images_dim, x2)
    #cnn.summary()
    combinedInput = concatenate([mlp.output, cnn.output])
    x3 = Dense(4, activation="relu")(combinedInput)
    x3 = Dense(1, activation="linear")(x3)
    discriminator = Model(inputs=[mlp.input, cnn.input], outputs=x3)
    return discriminator
Discriminator = discriminator()
Discriminator.summary()
Discriminator.trainable = False

My loss is as follows:

def discriminator_loss(real_output, fake_output):
    real_loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)(tf.ones_like(real_output), real_output)
    fake_loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)(tf.zeros_like(fake_output), fake_output)
    total_loss = real_loss + fake_loss
    return total_loss

def generator_loss(fake_output):
    gen_loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)(tf.ones_like(fake_output), fake_output)
    return gen_loss


optimizer = tf.keras.optimizers.Adam(learning_rate=0.00002, beta_1=0.6, beta_2=0.9)
Generator.compile(loss="binary_crossentropy", optimizer=optimizer, metrics=['accuracy'])
Discriminator.compile(loss="binary_crossentropy", optimizer=optimizer, metrics=['accuracy'])

But I printed the following results and found that there are some problems with the gradient.

a = tf.random.normal(shape=(2,10))
fake_images = Generator(a, training=True)
with tf.GradientTape() as gen_tape:
    gen_loss = generator_loss(fake_output)
gradients_of_generator = gen_tape.gradient(gen_loss, Generator.trainable_variables)
print(gradients_of_generator)

b = tf.random.normal(shape=(2,320,320,3))
real_output = Discriminator([a,b], training=True)
fake_output = Discriminator([a,fake_images], training=True)
with tf.GradientTape() as disc_tape:
    disc_loss = discriminator_loss(real_output, fake_output)
gradients_of_discriminator = disc_tape.gradient(disc_loss, Discriminator.trainable_variables)
print(gradients_of_discriminator)

error：

[None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None]
[]
2

I have solved some of the issues and changed the code to:

a = tf.random.normal(shape=(2,10))
a = tf.convert_to_tensor(a,dtype=tf.float32)
with tf.GradientTape() as gen_tape:
fake_images = Generator(a, training=True)
gen_loss = generator_loss(fake_images)
gradients_of_generator = gen_tape.gradient(gen_loss, Generator.trainable_variables)
print(gradients_of_generator )

b = tf.random.normal(shape=(2,320,320,3))
b = tf.convert_to_tensor(b,dtype=tf.float32)
with tf.GradientTape() as disc_tape:
real_output = Discriminator([a,b], training=True)
fake_output = Discriminator([a,fake_images], training=True)
disc_loss = discriminator_loss(real_output, fake_output)
gradients_of_discriminator = disc_tape.gradient(disc_loss,Discriminator.trainable_variables)
print(gradients_of_discriminator)

I successfully printed gradients_of_generator，but my gradients_of_discriminator is：
I guess this may be due to my Discriminator model being a composite model, but I don’t know how to modify it. Please help me, thank you

3

i got it, I deleted this code: Discriminator.trainable = False