Mathematical expression for cross entropy loss is
-y_i*sum(logy_k) but in the cross entropy function it is given as - np.log(y_{hat}[range(len(y_hat)), y]).
You did not multiply with true y label.
I’m stuck on the same thing. But i think the reasoning could be the following:
All the entries in y are 0, except for the true label index which is 1. That is why this is not multiplied for the sum(log y_k) (cause it is 1). All the others are multiplied by 0.
When we need to take from the y_hat the corresponding probability of the predicted label, they all are multiplied by 0, except the true label index.
look at the image. The y array is a one hot vector. Index 2 is the true label.
Yes. Multiplying is not efficient hence we just pick up the corresponding probability. Correct me if i am wrong
This should be the Reasoning.
Maybe someone has a better explanation.