Learning bounds for risk-sensitive learning (NeurIPS 2020)

This repo contains a PyTorch codebase for the NeurIPS 2020 paper Learning bounds for risk-sensitive learning.

(Disclaimer: It should be straightforward to implement these codes by yourself, perhaps in a more efficient way. Still, these codes may be helpful to beginners? 🤓)

Our codebase is organized as follows. (*) denotes key parts.

• models/ : Contains the model neural network (ResNet18). You can add one if you want to, but make sure to modify __init__.py
• tools/ : Contains some functions.
• tools/loaders.py : Contains dataloaders. You may add additional datasets.
• tools/risks.py : (*) Contains functions to compute (batch)-CVaR and SVP.
• tools/train.py : Ugly-looking train/test modules.
• longtake.py : (*) Our main function. In the file, you would see what kind of arguments you want to use with longtake.py.
• print.py : (*) After the main script is done running, you can use this to take summarize seed-wise results to get the csv file with average/std.

For quick runs, you may try typing:

python longtake.py --cuda 0 --target_risk avg python longtake.py --cuda 0 --target_risk meanstd --stdmult 1.0 python longtake.py --cuda 0 --target_risk cvar --betabar 0.4

Have fun! Jaeho.

P.S. "betabar" is simply an "alpha" for CVaR.