How can I calculate the loss (and get a scalar) between two FloatTensors of different size? For example, let’s assume that one FloatTensor is 115x5 and the other is 5x5. How can one get the loss between those two?

I’m doing:

loss = (y_predicted - y).pow(2).sum()

but of course it gives an error as the dimensions are different.

You need to explain what your ground-truth values and predictions are.
And if why 115 vs 5 (different batch size or other stuff)?

Because normally it is not logical to do sth like to want. You could resize tensor to same size, but then I think than network will be not learning anything.

I have 10 subjects and each subject has a number of 5-dimensional measurements. So,

Subject_1 with 1, …, N measurements
Subject_2 with 1, …, K measurements
.
.
.
Subject_10 with 1, …, L measurements

For example, I’m trying find the loss between all of the N measurements of subject_1 with my latent sample which is 5 dimensional and I assume it’s drawn from a multivariate Gaussian.