Skip to content
/ sfess Public

[ICLR 2025] SFESS: Score Function Estimators for k-Subset Sampling

Notifications You must be signed in to change notification settings

klaswijk/sfess

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SFESS: Score Function Estimators for $k$-Subset Sampling
Official PyTorch implementation of the ICLR 2025 paper

SFESS: Score Function Estimators for $k$-Subset Sampling
Klas Wijk, Ricardo Vinuesa & Hossein Azizpour
https://openreview.net/forum?id=q87GUkdQBm

Overview

Implementations of gradient estimators for subset distributions:

  • Gumbel softmax top-$k$ (GS) (Xie and Ermon 2019) [arXiv]
  • Straight-through Gumbel softmax top-$k$ (STGS) (Xie and Ermon 2019) [arXiv]
  • Implicit maximum likelihood estimation (I-MLE) (Niepert et al. 2021) [arXiv]
  • SIMPLE (Ahmed et al. 2023) [arXiv]
  • Score function estimators for $k$-subset sampling (SFESS) [this paper]

Multiple experiments:

  • Feature selection
  • Learning to explain (L2X)
  • Subset VAE
  • Stochastic k-nearest neighbors

Overview of experiments.

Requirements

numpy
matplotlib
seaborn
pytorch
torchvision
lightning
torchmetrics

Usage

To see the list of parameters for an experiment, run:

python main.py [task] --help

where [task] is one of {l2x,vae,knn}.

The toy experiment is found in /notebooks.

Acknowledgements

This implementation extends code from: