PEDCC-Loss

The article is available in A New Loss Function for CNN Classifier Based on Pre-defined Evenly-Distributed Class Centroids

Requirements

• Python >= 3.6
• Pytorch >= 1.0.0
• torchvision
• yacs (Yet Another Configuration System)

Data

CIFAR 100 (only)

Note: This is just a demo to show how we generate and use the PEDCC weights in Classification Task and Metric Learning.

Generate PEDCC weight

Firstly, we should change the num_classes and dim to generate PEDCC as you want to use in your task. And in this code, we use CIFAR100 to train a model and the feature's dimention is 512.

python utils\PEDCC.py

Then we get the file named 100_512.pkl, this is the PEDCC weights

Train Net

python tools\train_net.py

Visualize the PEDCC in 2-D and 3-D

Reference

The designed architecture follows this guide PyTorch-Project-Template if you think this is useful to you, please cite our technical report

@misc{zhu2019new,
    title={A New Loss Function for CNN Classifier Based on Pre-defined Evenly-Distributed Class Centroids},
    author={Qiuyu Zhu and Pengju Zhang and Xin Ye},
    year={2019},
    eprint={1904.06008},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}