I am calculating my accuracy like
(outputs.round() == targets).all(dim=2).all(dim=1).sum().item() / outputs.shape[0]
with outputs and targets being of shape NxAxB.
N is the batch size.
The remaining part is the prediction/truth value, where I want to see if they are identical.
Currently I am using .all(dim=2).all(dim=1).
The problem now is if I have a different model, the shapes will differ. They will be NxA, so my current approach doesn't work because dim=2 is not present.
(outputs.round() == targets).all(dim=1).sum().item() / outputs.shape[0]
, would work, but then again only for the second model.
Ideally I want to apply .all on everything BUT the fist dimension (batch dimension).
How would I do this?
To generalize to any number of dimensions, you could flatten the boolean tensor from the dim=1 outwards using torch.flatten, then apply all, and mean:
>>> (outputs.round() == targets).flatten(1).all(1).float().mean()
Note: torch.flatten(dim=1) will flatten the tensor from dim=1 to dim=-1.