自定义损失 function 带渐变

Question

I have created a custom loss function which also computes the mean squared error between the gradient of true and predicted labels. The function is given below. However, while debugging the code I found that uxp and uxt are the list and not tensors. Am I doing any mistake in computing the gradients?

def custom_mean_squared_error(y_true, y_pred):
    mse = K.mean(K.square(y_pred - y_true), axis=-1)

    # gradient
    xs = tf.ones_like(y_pred)
    uxp = tf.gradients(y_pred, xs)
    uxt = tf.gradients(y_true, xs)
    grad_mse = K.mean(K.square(uxp - uxt), axis=-1)

    mse1 = mse + grad_mse
    return mse1

Thank you.

Answer 1

Yes, the lists cannot be subtracted in Python (try [1,2,3] - [-1,-2,-3]). And yes, tf.gradients should return the list of tensors (see documentation https://www.tensorflow.org/api_docs/python/tf/gradients ). You can try list comprehension to implement subtraction.