Repo for training and experiments with skin-cancer dataset from kaggle
I was able to experiments only with CNN that fits in 4GB of RAM on my GPU.
Diffrent architectures that was use:
* Custom CNN - my CNN with 3 conv layers and 2 fc layers
* MobileNetV2 - pretrained wieghts from ImageNetV2, 3.5MLN PARM
* RegNet X 400MF - pretrained weights from ImageNetV2, 5.5MLN PARM
* EfficientNetB0 - pretrained weights from ImageNetV1, 5.3MLN PARM
The best was EfficientNetB0, with early stopping it achived 92% accuracy on test set, bigger EfficientNet like B1 or B2 probably will give even better results that this.
All the experiments was documented using Weights&Biases.
After experiments I wrote training script using PyTorch Lightning.