Skin Cancer MNIST: HAM10000

Networks (Results)

• resnet34 pretrained + adamW+ wd = 1e-3+ oversampling
  • overfitting like hell initially
• resnet34 pretrained + adamW+ wd = 1e-3+no oversampling
  • overfitting like hell initially too but more than oversampled
• resnet50 pretrained + adamW+ wd = 1e-3+no oversampling
  • even more overfitting
• resnet34 not pretrained + adamW+ wd = 1e-3+ oversampling
  • less overfitting
  • still over fit
• resnet34 not pretrained + adamW+ wd = 1e-1+ oversampling+ mixup
  • no overfitting
  • doesnt do that great tho
• resnet34 pretrained + adamW+ wd = 1e-1+ oversampling+ mixup
  • no overfitting
  • 71%
• vgg16bn pretrained + adamW+ wd = 1e-1+ oversampling+ mixup
  • no overfitting
  • 85.5%
• vgg19bn pretrained + adamW+ wd = 1e-1+ oversampling+ mixup
  • no overfitting
  • 88%

Conclusion

• oversampling helps a lot