Wrong outputs from torch.sub?-CodePudding

I’m currently using torch.sub alongside torch.div to obtain the MAPE between my predicted and true labels for my neural network although I’m not getting the answers I’m expecting. According to the example in the documentation, I should be getting a 4x1 tensor, not 4x4. Could anyone clear this up for me?

print('y_true ', y_true)
y_true tensor([[ 46],
[262],
[ 33],
[ 35]], device=‘cuda:0’, dtype=torch.int16)

print('y_pred ', y_pred)
y_pred tensor([[[308.5075]],
[[375.8983]],
[[389.4587]],
[[406.4957]]], device=‘cuda:0’, grad_fn=)

print('torch.sub ', torch.sub(y_true, y_pred))
torch.sub tensor([[[-262.5075],
[ -46.5075],
[-275.5075],
[-273.5075]],

    [[-329.8983],
     [-113.8983],
     [-342.8983],
     [-340.8983]],

    [[-343.4587],
     [-127.4587],
     [-356.4587],
     [-354.4587]],

    [[-360.4957],
     [-144.4957],
     [-373.4957],
     [-371.4957]]], device='cuda:0', grad_fn=<SubBackward0>)

CodePudding user response：

That is because y_pred has an extra dimension which means the y_true tensor probably gets broadcasted to the correct dimension.

If you remove the extra last dimension you get the desired result:

>>> torch.sub(y_true, y_pred[...,0]).shape
torch.Size([4, 1])