Skip to content

Speed up RetinaNet's postprocess_detections()#2819

Closed
datumbox wants to merge 5 commits intopytorch:masterfrom
datumbox:enhancement/retina_speedup
Closed

Speed up RetinaNet's postprocess_detections()#2819
datumbox wants to merge 5 commits intopytorch:masterfrom
datumbox:enhancement/retina_speedup

Conversation

@datumbox
Copy link
Copy Markdown
Contributor

@datumbox datumbox commented Oct 15, 2020 •
edited
Loading

In relation to #2799, one way we could potentially speed up the RetinaNet's postprocess_detections() method is by applying a pre-filter method to:

  • Eliminate any box whose scores do not surpass the threshold across any class.
  • Remove from the candidate classes those that don't have an associated candidate box.

Benchmark (100 iterations) across different images:

  • Before: 12.9425 sec
  • After: 7.4375 sec

To measure the speed replace:

vision/torchvision/models/detection/retinanet.py

Line 567 in dd76ba9

detections = self.postprocess_detections(head_outputs, anchors, images.image_sizes)

with:

import time
start_time = time.time()
for i in range(100):
    detections = self.postprocess_detections(dict(head_outputs), anchors, images.image_sizes)
print('100 Iter in %.2f' % (time.time() - start_time))

I also used a Python profiler to confirm the results.

@codecov
Copy link
Copy Markdown

codecov bot commented Oct 15, 2020 •
edited
Loading

Codecov Report

Merging #2819 into master will increase coverage by 0.01%.
The diff coverage is 100.00%.

Impacted file tree graph

@@            Coverage Diff             @@
##           master    #2819      +/-   ##
==========================================
+ Coverage   73.26%   73.28%   +0.01%     
==========================================
  Files          99       99              
  Lines        8778     8784       +6     
  Branches     1387     1388       +1     
==========================================
+ Hits         6431     6437       +6     
  Misses       1920     1920              
  Partials      427      427
Impacted Files Coverage Δ
torchvision/models/detection/retinanet.py 73.62% <100.00%> (+0.63%) ⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5e33cc8...9d88383. Read the comment docs.

@datumbox datumbox changed the title Speed up Retina's postprocess_detections() Speed up RetinaNet's postprocess_detections() Oct 15, 2020
@datumbox
Copy link
Copy Markdown
Contributor Author

datumbox commented Oct 16, 2020

I discussed with @fmassa and recommended a different optimisation approach I'm closing this and I'll follow up with a different PR.

@datumbox datumbox closed this Oct 16, 2020
@datumbox datumbox mentioned this pull request Oct 19, 2020
Merged
@datumbox datumbox deleted the enhancement/retina_speedup branch October 20, 2020 09:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@datumbox