Mean Directional Accuracy criterion loss function

1

Hello, can someone help me create a mean directional accuracy loss function for pytorch?
in numpy it should be this: https://gist.github.com/bshishov/5dc237f59f019b26145648e2124ca1c9
def mda(actual: np.ndarray, predicted: np.ndarray):
“”" Mean Directional Accuracy “”"
return np.mean((np.sign(actual[1:] - actual[:-1]) == np.sign(predicted[1:] - predicted[:-1])).astype(int))

but i get errors if use it as loss criterion in pytorch

2

You cannot have sign in the loss function and expect good results and int won’t harmonize with autograd either.

3

the result of the function there is the mean as float (not an int)

You cannot have sign in the loss function and expect good results

i think i understand the reason but could that be “added” to the result of a normal mse to give importance also to the direction?

4

Probably multiplying would work better.

And I might add, it could be a good idea to take, say 3-5 elements of your prediction and target vector and visualize the loss you’re getting.