Create variables on correct devices and with correct types

[jzakirov]

Related to #288

Proposed Changes

• Propagate input dtype though whole computation pipeline and cast intermediate kernels / filters to it.
• Add tests to verify that for float32 and float64. float16 is not included, because not all operations are defined for CPU.

[codecov]

Codecov Report

Merging #292 (27f9cab) into master (19f0d92) will decrease coverage by 2.33%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #292      +/-   ##
==========================================
- Coverage   93.35%   91.02%   -2.34%     
==========================================
  Files          33       33              
  Lines        2289     2284       -5     
==========================================
- Hits         2137     2079      -58     
- Misses        152      205      +53     

Flag	Coverage Δ
unittests	91.02% <100.00%> (-2.34%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
piq/brisque.py	99.01% <100.00%> (ø)
piq/dss.py	100.00% <100.00%> (+1.03%)	⬆️
piq/fsim.py	100.00% <100.00%> (ø)
piq/functional/filters.py	100.00% <100.00%> (ø)
piq/ms_ssim.py	100.00% <100.00%> (ø)
piq/ssim.py	100.00% <100.00%> (ø)
piq/vsi.py	100.00% <100.00%> (ø)
piq/gs.py	32.96% <0.00%> (-59.35%)	⬇️

[denproc]

Great fix!

• For BRISQUE, '_natural_scene_statistics' uses gaussian kernel (line 178), which is generated as 'torch.FloatTensor' losing the precision for 'torch.DoubleTensor' input. I suggest to propagate type into the kernel generation or use 'torch.float64' as default for generating kernels.
• For DSS, I suggest type assignment in line 90 to avoid implicit type upgrade in following loop. Function '_dct_matrix' and ‘gaussian_fliter’ produce 'torch.FloatTensor's losing the precision if 'torch.DoubleTensor's are used as the input. Therefore, I suggest, to add type propagation inside.
• For FSIM, '_lowpassfilter' amd ‘scharr_filter’ are 'torch.FloatTensor'. Similar suggestion to the previous. Propagate types or generate 'torch.float64' as default with following reduction to 'torch.float32'. Same for ‘_construct_filters’ in lines 154-158.
• For MS-SSIM and SSIM, we use gaussian filter, generated using ‘torch.FloatTensor’ in line 73 (MS-ssim) and in line 64 (SSIM).
• For VSI, ‘scharr_filter’ and ‘_log_gabor’ are generated as ‘torch.FloatTensor’.
• For other functionality, that wasn't changed with presented commits, could contain same issues.

Concluding on the above, ‘functional/filters.py’ could be upgraded to ‘torch.float64’ as default precision, which could be reduces after.

[sonarqubecloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

No Coverage information
0.0% Duplication

[jzakirov]

As discussed with Denis and Sergei, we won't create kernels as torch.DoubleTensor by default.

Reasoning:
It's much more common to have single precision float tensors as inputs. So having kernels in float64 will create a lot of casting back to float32 every time metrics are computed.

This PR adds support for float64 input inference and slight loss of precision due to kernel's dtype is not considered important.

@denproc please review again and merge.

[denproc]

LTGM