Question

我正在使用具有原始权重的StyleGANv2（https://github.com/moono/stylegan2-tf-2.x）。我的TF版本是2.2.0。默认的训练过程和推理工作正常。

但是我正在尝试进行有关输入的学习步骤：

rnd = np.random.RandomState(seed)
latents = rnd.randn(1, generator.z_dim).astype(np.float32)
labels = rnd.randn(1, generator.labels_dim).astype(np.float32)

initial_input = [tf.convert_to_tensor(latents), tf.convert_to_tensor(labels)]
input_tensor = tf.Variable([initial_input[0]])

for step in range(10):
    with tf.GradientTape() as tape:

        tape.watch(initial_input[0])
        if step == 0:
            generated_image = generator(initial_input)
        else:
            generated_image = generator([input_tensor[0,:,:], initial_input[1]])
        
        # target_image - the image I'm trying to learn latents for
        loss_value = mae_loss(generated_image, target_image)
        
    plt.imshow(postprocess_images(generated_image).numpy()[0])
    plt.show()
    
    print("Loss: ", loss_value)
    print("Tensor: ", input_tensor)
    print("Initial Input: ", [initial_input[0]])
    
    gradients = tape.gradient(loss_value, [initial_input[0]])

    print("Gradients: ", gradients)
    
    optimizer.apply_gradients(zip([gradients], [input_tensor]))

第一步成功完成，但是在第二步，我得到一个错误： ValueError: Attempt to convert a value (None) with an unsupported type (<class 'NoneType'>) to a Tensor.

第二步，gradients = tape.gradient(loss_value, [initial_input[0]])返回None。

那么你知道可能是什么问题吗？

Answer 1

现在可以使用。感谢@klapeyron。我用input_tensor中的值更新了tf.Variable值：

optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)

initial_input = [tf.convert_to_tensor(latents), tf.convert_to_tensor(labels)]
input_tensor = tf.Variable([initial_input[0]])

for step in range(100):
    with tf.GradientTape() as tape:
        tape.watch(initial_input[0])

        initial_input[0] = input_tensor[0,:,:]
        generated_image = generator(initial_input)initial_input[1]])
        
        loss_value = mse_loss(generated_image, target_image)
        
    plt.imshow(postprocess_images(generated_image).numpy()[0])
    plt.show()
    
    gradients = tape.gradient(loss_value, [initial_input[0]])
    optimizer.apply_gradients(zip([gradients], [input_tensor]))

尝试对输入应用渐变时出现错误

1 个答案: