Batch class avg

[garrettwrong]

Stashing WIP before vacation. Still needs a lot of cleanup and testing.

[codecov]

Codecov Report

Attention: Patch coverage is 97.72727% with 2 lines in your changes missing coverage. Please review.

Project coverage is 90.63%. Comparing base (a94bf0d) to head (dea12ca).
Report is 14 commits behind head on develop.

Files with missing lines	Patch %	Lines
src/aspire/classification/averager2d.py	97.56%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1216      +/-   ##
===========================================
- Coverage    90.66%   90.63%   -0.04%     
===========================================
  Files          132      132              
  Lines        13707    13702       -5     
===========================================
- Hits         12428    12419       -9     
- Misses        1279     1283       +4     

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

[garrettwrong]

Passing along for initial review. I believe #1214 should go in first, then this will need to be rebased, maybe resolving some conflicts along the way.

[j-c-c]

Looks good!

[janden]

Looks great! Just two things.

[janden]

(cnt, n_rot) * (cnt, 1) -> (cnt, n_rot)?

[garrettwrong]

Sure :), changed.

[garrettwrong]

I want to discuss/confirm the class hierarchy during our meeting and if no changes come out of that I'll merge this then.

[garrettwrong]

Factored the base image rotation table outside the shift loop like we discussed in the meeting.

[janden]

Factored the base image rotation table outside the shift loop like we discussed in the meeting.

Cool. Any speedup?

[garrettwrong]

Factored the base image rotation table outside the shift loop like we discussed in the meeting.

Cool. Any speedup?

Yes. Using GPU mode for 179px and 50 nbrs this saves about 0.3s per class (44.5 vs 44.8 s). I'll call it 1%. I think relative to the matmuls the vector vector broadcast multiplication is very fast for the GPU. On the host, much larger improvement, more like 10%.

Either way across 3-100k classes I'll definitely take it :). Thanks.