TODO.md

TODO:

• 0.8 dropout after timesformer
• multiple dropouts (like topology)
• Loss = alpha * MSE - beta * Spearman?
  • not differentiable, maybe it is?
  • https://forum.numer.ai/t/differentiable-spearman-in-pytorch-optimize-for-corr-directly/2287
  • https://discuss.pytorch.org/t/spearmans-correlation/91931/5
• image preprocessing
  • random cropping
  • random rotation
  • normalize image to -0.5 to 0.5
• different cropping/scaling techniques?
  • run our architecture on 448x448 models