Jihao-Li / PCANet_PyTorch

This is a PyTorch implementation of PCANet.

PCANet_PyTorch

This is a PyTorch implementation of PCANet. Details are described in the original paper.
Unlike other implementations, the number of stages in PCANet can be set arbitrarily, rather than two. So the structure is more flexible.

Requirements

• Python 3.5
• PyTorch==1.0.0
• sklearn, tensorboardX, numpy

Usage

train

python train.py

eval

python eval.py --pretrained_path <path to trained PCANet model and SVM>

Results on MNIST

use 70% of training data

convolution kernel in stage 0

convolution kernel in stage 1

feature maps of an image in stage 0

feature maps of an image in stage 1

the accuracy rate in total testing data is 93.42%

use 100% of training data

convolution kernel in stage 0

convolution kernel in stage 1

feature maps of an image in stage 0

feature maps of an image in stage 1

the accuracy rate in total testing data is 93.48%

Results on cifar10

in progress

Citation

Chan T H , Jia K , Gao S , et al. PCANet: A Simple Deep Learning Baseline for Image Classification?[J]. IEEE Transactions on Image Processing, 2015, 24(12):5017-5032.