Read up on results from competition

[thomashopkins32]

The competition has ended and I should dive into what worked for people and what didn't.

[thomashopkins32]

Notes on the competition:

1st Place

• Found the RTMDet model to train fast and perform well.
  • https://arxiv.org/pdf/2212.07784v2.pdf
• Used Exponential Moving Average (EMA) for parameter updates (?)
• Used MMDetection and MMEngine for training + evaluation
• Fairly simple data augmentations but aggressive
• Trained with a batch size of 8, keeping 3 images from dataset 1 and 5 images from dataset 2 in the same batch
• Best single model config found here: https://github.com/tascj/kaggle-hubmap-hacking-the-human-vasculature/blob/main/configs/r0.py