AV datahack on deep learning CV
The repo includes most models tried on AV Computer vision hackathon. Pointers:
- Fastai significantly out performs pure pytorch based models
- Around 30 different models were tried mainly from Cadene's pretrained repo, DenseNet, SENET, polynet,Inception variations,ResNets, Resnets with BAM, and some variations of pretrained EfficientNet
- Focal loss works best for training instead of CrossEntropy loss
- Public LB: 17th (so far) Private LB: ??