perf: improved performance in various methods in Image and ImageList

[Marsmaennchen221]

Summary of Changes

1. Improved memory usage and runtime in:

• Image
  • convert_to_grayscale
  • adjust_brightness
  • add_noise
  • adjust_contrast
  • adjust_color_balance
  • find_edges
• ImageList
  • from_images
  • add_image
  • add_images
  • remove_image_by_index
  • remove_duplicate_images
  • convert_to_grayscale
  • resize
  • crop
  • adjust_brightness
  • add_noise
  • adjust_contrast
  • adjust_color_balance
  • find_edges

2. Changed blur algorithm in Image and ImageList from Gaussian blur to box blur
3. Fixed a bug in blur and sharpen, that they could not work with Tensors of size greater than 2**31

Details to the performance upgrades:

These details will explain the performance upgrades in the ImageList. All performance upgrades and changes in Image are made according to the changes in ImageList

Early stopping

convert_to_grayscale returns self when it has only one channel (in this case, the ImageList is already in grayscale)
remove_duplicate_images returns self if the unique image tensor has the same size as the original (in this case, there are no duplicates)
adjust_brightness returns ImageList with complete black images if factor is 0
adjust_color_balance returns ImageList.convert_to_grayscale if factor is 0

General changes

If a float Tensor is used during the computation, it will be float16 instead of float32
Improved the order of Tensor allocations, so that there will be fewer problems with tensors not being completely on the VRAM

Benchmark

Only the transformation methods have benchmarks for the runtime. Their benchmark includes only changes over/under a change factor of 0.25, rounded to one decimal point, as runtime depends on multiple factors and fluctuates heavily with changes of only a few milliseconds in most cases.
All differences are measured as a factor compared to the original results. That means a factor below 1 is worse, while a factor above 1 is a better result. For readability, all factors equal to 1 are not included.
Due to the bug fix mentioned above for blur and sharpen the performance of these methods decreased in most cases.

Benchmark with RGB images of size 250×250

method	result size difference	runtime difference	max memory allocation during runtime difference
from_images	8		7,95
remove_image_by_index			1,5
adjust_brightness		1,6	1,5
add_noise	4	42,5	2,25
adjust_contrast		1,4	1,69
adjust_color_balance	4	6,3	3
blur		0,7	0,9
sharpen		0,6	0,91
find_edges		1,3

Benchmark with RGBA images (RGB images with transparent layer) of size 256×256

method	result size difference	runtime difference	max memory allocation during runtime difference
from_images	8		6,29
remove_image_by_index			1,5
adjust_brightness		1,7	1,36
add_noise	4	24,5	2,25
adjust_contrast			1,53
adjust_color_balance	4	1,7	2,6
blur		1,6	0,9
sharpen			0,89

Benchmark with RGB images of multiple different sizes

method	result size difference	runtime difference	max memory allocation during runtime difference
from_images	7,7		6,3
add_images	8		2,38
remove_image_by_index			1,28
resize		1,3
crop		1,3	1,09
adjust_brightness		1,3	1,09
add_noise	4	44,3	1,47
adjust_contrast		2	2,17
adjust_color_balance	4	6,7	1,65
blur			2,87
sharpen		0,7	0,9
find_edges		1,9	0,91

[github-actions]

🦙 MegaLinter status: ✅ SUCCESS

Descriptor	Linter	Files	Fixed	Errors	Elapsed time
✅ PYTHON	black	7	0	0	1.95s
✅ PYTHON	mypy	7		0	3.16s
✅ PYTHON	ruff	7	0	0	0.4s
✅ REPOSITORY	git_diff	yes		no	0.68s

See detailed report in MegaLinter reports
Set VALIDATE_ALL_CODEBASE: true in mega-linter.yml to validate all sources, not only the diff

MegaLinter is graciously provided by

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.71%. Comparing base (b99a760) to head (4d23fe0).
Report is 38 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #879      +/-   ##
==========================================
+ Coverage   97.67%   97.71%   +0.04%     
==========================================
  Files         120      120              
  Lines        6234     6478     +244     
==========================================
+ Hits         6089     6330     +241     
- Misses        145      148       +3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[lars-reimann]

Do you have benchmarks to quantify the performance improvement?

[Marsmaennchen221]

Do you have benchmarks to quantify the performance improvement?

@lars-reimann I included the benchmarks in the description of this pr

[lars-reimann]

Nice performance improvements, great work.

[lars-reimann]

🎉 This PR is included in version 0.27.0 🎉

The release is available on:

• v0.27.0
• GitHub release

Your semantic-release bot 📦🚀