Optimize post-processing #128

New Issue

jgeorgi · 2024-03-08T17:52:32Z

jgeorgi commented

2024-03-08 17:52:32 +00:00

Post-processing is where the biggest hits in detection time are, especially for the larger models (e.g., yolov8m based). There must be some places where that can be optimized better.

jgeorgi added the

enhancement

label 2024-03-08 17:52:32 +00:00

jgeorgi commented

2024-03-10 04:05:12 +00:00

The slow step is the transfer of the data from the gpu. Method doesn't matter, but arraySync I think leads to slightly more efficient processing after the fact. But, there's just not much that can be done with that transfer time. I assume that's machine dependent. Right now I'm getting fairly consistently sub 1s total times with the small sized model.

The slow step is the transfer of the data from the gpu. Method doesn't matter, but `arraySync` I think leads to slightly more efficient processing after the fact. But, there's just not much that can be done with that transfer time. I assume that's machine dependent. Right now I'm getting fairly consistently sub 1s total times with the small sized model.

jgeorgi referenced a pull request that will close this issue

2024-03-10 04:06:36 +00:00

Optimize detection steps #129

jgeorgi closed this issue

2024-03-10 04:06:51 +00:00

jgeorgi referenced this issue from a commit

2024-03-10 04:06:53 +00:00

Optimize detection steps (#129)

Sign in to join this conversation.