This folder contains script regarding the strata-aware optimzation procedure. The code required you to first generate Bayesian optimal bins following procedure in the folder named binning.
In order to combat the high variance of errors issue, we enable data-distribution aware optimization. If the predicted count value lies outside the range of bin, then a linear penality is imposed, where as if the predicted count value lies inside the range of bin a logarithmic penality is imposed.
You can imbibe the strata-aware optimization in two ways:
- Using bin aware optimzation along with network-specific loss function (your loss function).
- Solely using bin aware optimization.
The file log_loss.py provides the pyTorch based implementation of the loss. Keep in mind that when using the loss function along side a network-specific loss function a hyper-parameter should be used for the log loss function (λ1 in Sec 3.4). This hyperparameter can be tuned for better results.