Skip to content

Latest commit

 

History

History
4 lines (4 loc) · 871 Bytes

RANDOMNESS.md

File metadata and controls

4 lines (4 loc) · 871 Bytes

This file lists some randomness factors which could make the reproduced solution perform slightly different from the winning submission.

  • Environments or hardware: should be negligible
  • YOLOX training: I use official YOLOX repository to train breast ROI detection model. Unfortunately, the code is currently unreproducible between different experiments, even when using same config and seed used. The issue is mentioned here. I did not notice this randomness until the competition ended. So, i think it's pretty hard to reproduce the exact YOLOX model used in the winning submission (we need to guess some random numbers?).
  • Converting model from Torch to TensorRT would introduce small performance changes. But I expect these change to be also negligible