Loss function for a classifier with output and target shape: (batch_size, N, M)

nlp
1

Hi,
I have a model with output shape: (batch_size, N, M) and target shape: (batch_size, N, M)
this is a classifier.
so I should use NLL or CrossEntropy as the loss function.
but torch gives me 1D target tensor expected, multi-target not supported error
how can I fix this?

2

Hi Ali!

Well, you should tell us what “output” means, what “target” means,
and what “N” and “M” are. Otherwise, we will have to guess.

But let me guess a little bit, and make the following comment:

If you are performing a multi-class classification problem (but not
a multi-label problem), then the shapes of the input (output of
your model) and target passed to your CrossEntropyLoss loss
function must match up in the following way:

Your input can have shape [batch_size, nClass], and then your
target must have shape [batch_size] and consist of integer (long)
class labels that run from 0 to nClass - 1.

Or, in the “K-dimensional case,” your input could have shape, say,
[batch_size, nClass, height, width, depth], and your target
would have to have shape [batch_size, height, width, depth]
(and consist of values from 0 to nClass - 1).

Good luck.

K. Frank

2 Likes
3

thanks Frank, that was helpful