Proposed Changes

FSIM: got rid of a duplicate for filter generation _construct_filters
FSIM: removed the duplicate of mesh grid generation performed for _lowpassfilter. The filter is computed explicitly.
FSIM: minor optimisation of tensor generations across the fsim to avoid transfer between devices
base.py: mesh grid creation (get_meshgrid) is performed straight on target device and data type according to input tensor without redundant transfers between devices
base.py: removed redundant to() operation in gradient_map
filters: filters are generated straight on target device and data type according to input tensor without redundant transfers between devices
colour_conversion: moved tensor generation to the target device instead of transfer between devices
Updated .gitignore
minor change in benchmark script to allow more flexible choice for GPU device
minor fixes across the code to match the changes for colour conversion and filter generation

Outcome of the Proposed Changes

Reduced cpu usage with cuda-enabled computations for FSIM and FSIMc (visual comparison is atttached)
CPU time was reduced by 6 sec according to PyTorch profiler

Note

Other places to consider for additional assessment of leaks

piq/base.py, fid.py, perceptual.py and pieapp.py use .to() for a various transfers. They are out of scope for current PR due to a different purpose of the transfer between devices/types.
custom made piq/functional/resize.py could be refactored as well to get rid of redundant transfers. Added to #227
vsi.py has redundant mesh grid creation. Opened #335

Appendix

Visual Comparison of CPU usage

htop following #266

Before

Branch: master

After

Branch: bug/cpu_usage

PyTorch Profiler Results. Top-8 operations.

-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                                   Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls  
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  
                                        model_inference        16.38%        8.624s       100.00%       52.639s       52.639s             1  
                                               aten::to         0.07%      37.231ms        68.84%       36.234s       1.020ms         35525  
                                         aten::_to_copy         0.14%      72.795ms        68.76%       36.197s       4.993ms          7250  
                                            aten::copy_         0.34%     177.512ms        61.16%       32.196s       2.019ms         15950  
                                  cudaStreamSynchronize        40.22%       21.173s        40.22%       21.173s       1.718ms         12325  
                                        cudaMemcpyAsync        20.68%       10.883s        20.68%       10.883s     883.015us         12325  
                                    aten::empty_strided         0.17%      88.161ms         7.78%        4.093s     564.510us          7250  
                                             cudaMalloc         7.62%        4.009s         7.62%        4.009s     129.312ms            31

codecov[bot] commented 1 year ago

Codecov Report

Merging #334 (59fa153) into master (753030f) will increase coverage by 0.00%. The diff coverage is 96.66%.

@@           Coverage Diff           @@
##           master     #334   +/-   ##
=======================================
  Coverage   91.96%   91.97%           
=======================================
  Files          34       34           
  Lines        2490     2493    +3     
=======================================
+ Hits         2290     2293    +3     
  Misses        200      200

Flag	Coverage Δ
unittests	`91.97% <96.66%> (+<0.01%)`	:arrow_up:

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

Impacted Files	Coverage Δ
piq/psnr.py	`80.00% <0.00%> (ø)`
piq/functional/base.py	`92.10% <71.42%> (ø)`
piq/brisque.py	`99.00% <100.00%> (ø)`
piq/dss.py	`100.00% <100.00%> (ø)`
piq/fsim.py	`100.00% <100.00%> (ø)`
piq/functional/colour_conversion.py	`100.00% <100.00%> (ø)`
piq/functional/filters.py	`100.00% <100.00%> (ø)`
piq/gmsd.py	`100.00% <100.00%> (ø)`
piq/haarpsi.py	`100.00% <100.00%> (ø)`
piq/isc.py	`90.90% <100.00%> (ø)`
... and 7 more