Hierarchical classification at multiple operating points

https://arxiv.org/abs/2210.10929

This repo contains the code for the NeurIPS 2022 paper.

For now, it's in a raw state. I will refactor the code and remove legacy parts when I have time!

The main functionality can be found in:

• hier.py: Representation of class hierarchy and basic functionality.
• hier_torch.py: Prediction and loss functions written in pytorch.
• infer.py: Inference functions that map likelihoods to labels.
• metrics.py: Functions to evaluate metrics.

Code to run experiments and generate plots:

• main.py
• eval_inat21mini.ipynb
• eval_inat21_errorbars.ipynb
• eval_inat21_split.ipynb
• eval_imagenet.ipynb