Bug suspected in MS-SSIM\SSIM metric

[amichayfeldman]

MS-SSIM was used to measure two image tensors (BxCxHxW),
set reduction flag to None to get the losses without any reduction.

When I executed it, I got an error of size mismatch for this line:
https://github.com/Lightning-AI/metrics/blob/19355a9d2c51b3b39b311694d7b8e6856a73eae6/src/torchmetrics/functional/image/ssim.py#L407

When debugging the code, I realized that the betas repetition one line before is incorrect.
I believe that instead of
betas = torch.tensor(betas).unsqueeze(1).repeat(1, sim_stack.shape[0])
it should be
betas = torch.tensor(betas).unsqueeze(1).repeat(1, sim_stack.shape[1])

The shape of betas after this command should be (len(origin_betas), batch_size), and this is not what actually happens
(betas is already a column tensor, each element is power for each scale, so we want to repeat it with the amount of the batch size).

In addition, the next two lines raise warnings:
https://github.com/Lightning-AI/metrics/blob/19355a9d2c51b3b39b311694d7b8e6856a73eae6/src/torchmetrics/functional/image/ssim.py#L408
https://github.com/Lightning-AI/metrics/blob/19355a9d2c51b3b39b311694d7b8e6856a73eae6/src/torchmetrics/functional/image/ssim.py#L409

UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor)

The fix can be:
sim_stack = sim_stack ** betas.to(device=sim_stack.device); cs_stack = cs_stack ** betas.to(device=cs_stack.device)
(since betas is already a tensor)

[github-actions]

Hi! thanks for your contribution!, great first issue!

[SkafteNicki]

Hi @amichayfeldman,
Could you provide an minimal example to reproduce the shape error you are encountering?

[Queuecumber]

I can confirm this bug, I fixed it as part of #1231