Channelwise-Saab-Transform

Feature extraction (Module 1) packages for PixelHop/PixelHop++.

Introduction

This is an implementation by Yijing Yang for the feature extraction part in the paper by Chen et.al. PixelHop++: A Small Successive-Subspace-Learning-Based (SSL-based) Model for Image Classification.

It is modified based on Chengyao Wang's implementation (ObjectOriented / Numpy version), with lower memory cost.

Note that this is not the official implementation.

Installation

This code has been tested with Python 3.7 and 3.8. Other dependent packages include: numpy, scikit-image, numba, scikit-learn, xgboost, tensorflow.keras.

Contents

• saab.py: Saab transform.

• cwSaab.py: Channel-wise Saab transform. Use energy threshold TH1 and TH2 to choose intermediate nodes and leaf nodes, respectively. Set 'cw' to 'False' in order to turn off the channel-wise structure.

• pixelhop.py: Built upon cwSaab.py with additional functions of saving models, loading models, and concatenation operation across Hops.

• Example of usage can be found at the bottom of each file.

• main.py: some main structure of PixelHop and PixelHop++. Several TODOs are left blank for students to fill in.

  Note: All the images or data that are fed into these functions should be in the channel last format.

Memory usage

In PixelHop++, for MNIST dataset, using 10000 Training images for CWSaab will take approximately 4GB memory, 60000 Training images for CWSaab will take approximately 11G memory.

In PixelHop, using same dataset, using 10000 Training images for Saab will take approximately 7GB memory, 60000 Training images for Saab will take approximately 14G memory. Note that different threshold value will give different memory usage.

Note: Google Colab is recommended if there is not enough memory on your machine.