https://github.com/wangyongjie-ntu/siamese-network

siamese network
https://github.com/wangyongjie-ntu/siamese-network

Last synced: 3 months ago
JSON representation

siamese network

• Host: GitHub
• URL: https://github.com/wangyongjie-ntu/siamese-network
• Owner: wangyongjie-ntu
• Created: 2019-06-02T05:23:42.000Z (almost 6 years ago)
• Default Branch: master
• Last Pushed: 2019-06-04T13:45:04.000Z (almost 6 years ago)
• Last Synced: 2025-01-09T19:57:44.970Z (4 months ago)
• Language: Python
• Homepage:
• Size: 3.82 MB
• Stars: 0
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# SIAMESE NETWORK

## Introduction

This repo just illustrates the mechanism of contrastive loss function. 

Under this loss, if x_i, x_j are the features from same person, the distance is small. If x_i, x_j are the features from different person, the distance is large. 

![loss](https://github.com/iversonicter/siamese-network/blob/master/data/loss.png)

The network structure can be illustrated in the following image:

![network](https://github.com/iversonicter/siamese-network/blob/master/data/siamese.png)

Download the att face training set, and create your own positive pairs and negative pairs. The ratio should be 1:1 ideally, but in fact, the negative is far more than positive pairs. Hard sample mining is adopted to make learning efficient.

# Training result

positive:negative = 1:1

learning rate = 0.0005

![Training loss](https://github.com/iversonicter/siamese-network/blob/master/log/figure_0.0005_1_1.png)