jbk3407 / Self-Regularized-Entropy Public

forked from unique-chan/Classwise-Regulated-Entropy

Notifications You must be signed in to change notification settings
Fork 0
Star 1

Robust Image Classification via Class-wise Softmax Suppression (PyTorch)

1 star 1 fork Branches Tags Activity

Notifications

Name		Name	Last commit message	Last commit date
Latest commit History 17 Commits
my_criterion		my_criterion
my_utils		my_utils
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
train.py		train.py

Repository files navigation

Self-Regularized-Entropy

Robust Image Classification via Class-wise Softmax Suppression (PyTorch)

[Core Idea]

Notation
- $N$ : Total number of samples
- $K$ : Total number of classes
- $g$ : Index of the ground-truth class
- $\hat{y}$ : Predicted probability vectors
- $y$ : Label vectors (One-hot encoded)
- $(i)$ : Index of the $i$ th vector
- $[j]$ : Index of the probability mass on class $j$
Cross Entropy
- $H(y, \hat{y})=-\frac{1}{N}\sum_{i=1}^{N}log(\hat{y}^{(i)[g]})$
Self-Regularized Entropy
- $S(\hat{y})=-\frac{1}{N}\sum_{i=1}^{N}\sum_{j=1, j \ne g}^{K}R(\hat{y}^{(i)[j]})$
- $R(\hat{y}^{(i)[j]})=\frac{c(\hat{y}^{(i)[j]})}{c(\hat{y}^{(i)[j]})+\hat{y}^{(i)[j]}}log(\frac{c(\hat{y}^{(i)[j]})}{c(\hat{y}^{(i)[j]}+\hat{y}^{(i)[j]}})+\frac{\hat{y}^{(i)[j]}}{c(\hat{y}^{(i)[j]})+\hat{y}^{(i)[j]}}log(\frac{\hat{y}^{(i)[j]}}{c(\hat{y}^{(i)[j]})+\hat{y}^{(i)[j]}})$ - where $c(\hat{y}^{(i)[j]})$ denotes $\frac{1}{K-1}\hat{y}^{(i)[j]}$ .
Proposed Training Loss Function
- $H(y,\hat{y})-S(\hat{y})$

About

Robust Image Classification via Class-wise Softmax Suppression (PyTorch)

Report repository

Releases

No releases published

Packages

No packages published

Languages

Python 100.0%