nocotan / density_fixing

Implementation of the density-fixing training method

This repo contains demo implementation of the density-fixing training code in PyTorch based on the following paper:

Masanari Mimura, Ryohei Izawa. Density Fixing: Simple yet Effective Regularization Method based on the Class Prior https://arxiv.org/abs/2007.03899

Training

The following table shows the mean test errors w/ and w/o density-fixing regularization.

Model	Top 1 Error	Top 5 Error
ResNet-18	12.72%	0.812%
ResNet-18 + density fixing (gamma=1)	12.23%	0.779%

Acknowledgement

The CIFAR-10 reimplementation of density-fixing is adapted from the pytorch-cifar repository by kuangliu.