03-pytorch_computer_vision: question about calculating test accuracy (timestamp = 16:08:10)

[davidfelipeb]

Hello! Thanks for this videos, they are really helpful! I have a question regarding the calculation of test accuracy in the FashionMNISTModelV0. In the video (see exact timestamp in title), you say that we should do

test_acc += accuracy_fn(y_true = y_test, y_pred = test_pred.argmax(dim = 1))

in order to correctly calculate the accuracy, that is because the variable test_pred contains logits, and not label predictions, and hence we should apply the argmax values on them. My question is: shouldn't we first convert test_pred from logits to prediction probabilities and then use argmax to find the maximum value? Something like this:

test_acc += accuracy_fn(y_true = y_test, y_pred = torch.softmax(test_pred, dim = 1).argmax(dim = 1))

Or am I missing something here? Thanks for any help provided!

[fivefishstudios]

I specifically came here to ask the same question.

Looking at output format of y_pred=Model(X), it seems softmax() and argmax() need to be applied first? y_pred and y are definitely not the same format.

[mrdbourke]

Hey @davidfelipeb,

Yes, you're right that is a valid option: logits -> prediction probabilities (torch.softmax) -> labels (torch.argmax).

But you can actually skip the middle step: logits -> labels (torch.argmax) - if you don't require prediction probabilities.

Confusing, yes.

The first option is probably more right. The second option is more of a shortcut.

However, you'll notice they both give the same results.

You could try experimenting using it and not using it and seeing what happens.