Speed up the YOLO detection module

ocean-data-factory-sweden / kso

Notebooks to upload/download marine footage, connect to a citizen science project, train machine learning models and publish marine biological observations.

GNU General Public License v3.0

4 stars 12 forks source link

Speed up the YOLO detection module #342

Closed victor-wildlife closed 4 months ago

victor-wildlife commented 6 months ago

I am not sure what the solution might be but if it will be possible to batch-process the frames maybe we can make the most of the GPU resources in a shorter amount of time (at the moment running the spyfish model in Collab's T4 for +30 mins movies takes ~20 mins and only uses 0.4 GB of the 15GB available from the GPU).

victor-wildlife commented 6 months ago

Maybe this multiprocessing approach for terrestrial camera traps could help as a starting point?

jannesgg commented 4 months ago

Jupyter Notebooks and multiprocessing do not play well together. We could potentially use this speed-up technique on Cloudina, but not for Colab.