CrossEntropyLoss - Why is it possible to have more columns in input than there are classes in the target?

[ptrblck]

I assume the discussion is targeting the original post before the edit as the current code snippet uses 8 valid classes so I refer to:

@SuryaKiran

Why does the following code work -

input = torch.cat([torch.randn(3,5),torch.randn(3,3)],dim=1)
target = torch.empty(3,dtype=torch.long).random_(5)
F.cross_entropy(input,target)

This code snippet is working, as it meets the requirements.
The logits have a shape of [batch_size=3, nb_classes=8] while the targets have a shape of [batch_size=3] and contain values in [0, nb_classes-1]. In this case you are limiting the target values to [0, 4] which is inside the valid range of [0, 7].