[optimizing] vectorize, cache and precompute

[derVedro]

Vectorized importance map: The get_importance() method now generates a composite weighting map using numpy, replacing the slow per-pixel computations with efficient array operations.

Caching for repeated calculations: Importance maps are cached per crop dimensions, avoiding redundant computations across multiple crop candidates.

Precomputed features: The new precompute_features() method consolidates scaling, biasing, and weighting operations into a single, vectorized step. This eliminates repetitive calculations during scoring.

Other small things:

• added docstrings
• vectorize debug_crop()

[coveralls]

Pull Request Test Coverage Report for Build 21993256552

Details

• 47 of 60 (78.33%) changed or added relevant lines in 2 files are covered.
• 16 unchanged lines in 1 file lost coverage.
• Overall coverage increased (+0.6%) to 80.392%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
smartcrop/library.py	45	58	77.59%

Files with Coverage Reduction	New Missed Lines	%
library.py	16	90.64%

Totals
Change from base Build 21250558729:	0.6%
Covered Lines:	328
Relevant Lines:	408

---

💛 - Coveralls

[derVedro]

@davidfischer-ch so, that part is done, what do you think about it?

[davidfischer-ch]

This looks great,

I am interested in getting some speed, CPU and memory usage metrics.

For example you may run this and previous version using /usr/bin/time -v (...) on some pictures and show the results?

Thank you.

[derVedro]

For example you may run this and previous version using /usr/bin/time -v (...) on some pictures and show the results?

I already use some kind of benchmark on project's test pix. Here a gist with the code. If you want to run it, just put it into the project directory, create an 'othercrops' directory with directories of other versions of SmartCrop (or just import somehow another SmartCrop from somewhere and fix the import).

Here are the timings

Task                             [b] base        mine   Δ time  speedup
-----------------------------------------------------------------------
business-work-1.jpg (100x400)   1.062696s   0.048699s  -95.42%  +21.82x
business-work-1.jpg (300x300)   0.310685s   0.038552s  -87.59%   +8.06x
business-work-1.jpg (350x50)    0.163094s   0.031437s  -80.72%   +5.19x
nature-1.jpg (100x400)          1.174951s   0.175214s  -85.09%   +6.71x
nature-1.jpg (300x300)          0.414470s   0.166441s  -59.84%   +2.49x
nature-1.jpg (350x50)           0.286055s   0.154219s  -46.09%   +1.85x
orientation.jpg (100x400)       0.919778s   0.127922s  -86.09%   +7.19x
orientation.jpg (300x300)       0.292235s   0.114656s  -60.77%   +2.55x
orientation.jpg (350x50)        0.277473s   0.110509s  -60.17%   +2.51x
travel-1.jpg (100x400)          1.067218s   0.067895s  -93.64%  +15.72x
travel-1.jpg (300x300)          0.305897s   0.055855s  -81.74%   +5.48x
travel-1.jpg (350x50)           0.175847s   0.048551s  -72.39%   +3.62x

average times of 5 shuffeled runs
3 cases: 100x400, 300x300, 350x50 crops
[b] base - current version of SmartCrop from github
mine - this PR version
Δ time means "mine took less time in respect to base"

If you're interested in memory usage, I haven't measured anything yet, but my gut feeling is that there aren't that many differences. Nothing crazy happens in the code, a few importance maps are cached, but it's not a thing at all.