Preprocessing of the target image

[ruguevara]

In some applications there is a need to repeatedly compare generated images with only one target image, and if all needed independent precalculations on this image could be done beforehand, this could save a lot of gpu/cpu calculation time.

How can this be implemented, for example, in subclass of some metrica? Or pytorch can cache this calculations by himself?

[francois-rozet]

Hello @ruguevara,

Comparing sequentially several images with a single target is indeed wasteful.

z = []
for i in range(len(x)):
    z.append(metric(x[i], y))

In PyTorch (and NumPy) there is something called broadcasting. For most operations (+, *, /, ...), if a tensor has a shape (..., N, ...) and the other (..., 1, ...), the backend will cast the shape of the latter to the shape of the former while performing the operation.

As a result, if the metric is parallelized (meaning that it does not split the batch dimension during the processing), it should accept a batch x of input images of shape (N, C, ...) and a batch y of target images of shape (1, C, ...) and still perform the comparison correctly.

z = metric(x, y)

In this case, the "preprocessing" of the target image is shared among the N input images. I believe all metrics of piqa support parallelization. Tell me if this suits your needs!