This is a reimplementation of the paper On Data-Driven Saak Transform (https://arxiv.org/abs/1710.04176), maintained by Jiali Duan and Yueru Chen.

psnr metric: 104.294772826

psnr metric: 105.637763477

psnr metric: 105.513759179

psnr metric: 106.509034118

• [Dataset] ( Hand-written digits classification)

  • [MNIST] ( train set: 60000, 28x28. We used the same with downloaded from http://yann.lecun.com/exdb/mnist/)

• [Installation] (sklearn and Pytorch)

  • [Sklearn Installation] Refer to http://scikit-learn.org/stable/install.html)
  • [Pytorch Installation] (Refer to http://pytorch.org)
  • [Optional: Jupyter Notebook] (Refer to http://jupyter.org/install.html)

• [How to] (Forward and Inverse Transform

  • Command python multi-stage_saak_v2.py
  • Forward Transform: multi_stage_saak_trans
  • Inverse Transform: toy_recon(outputs,filters)
  • Detailed params:

usage: multi-stage_saak_v2.py [-h] [--loadSize LOADSIZE]
                              [--train_batch_size TRAIN_BATCH_SIZE]
                              [--test_batch_size TEST_BATCH_SIZE]
                              [--size SIZE] [--windsize WINDSIZE]
                              [--stride STRIDE] [--save_path SAVE_PATH]
                              [--recStage RECSTAGE] [--visNum VISNUM]
                              [--use_SP]

optional arguments:
  -h, --help            show this help message and exit
  --loadSize LOADSIZE   Number of samples to be loaded
  --train_batch_size TRAIN_BATCH_SIZE
                        Batch size for loading
  --test_batch_size TEST_BATCH_SIZE
                        Batch size for loading
  --size SIZE           Size of the input
  --windsize WINDSIZE   Size of moving window
  --stride STRIDE       Stride to take during convolution
  --save_path SAVE_PATH
                        Path to save result
  --recStage RECSTAGE   Reconstruction start stage
  --visNum VISNUM       Number of visualizations
  --use_SP              Use S/P conversion

• [To-do list]

  • [x] One-stage Saak Transform
  • [x] Multi-stage Saak Transform
  • [x] Inverse Transform
  • [x] S/P Conversion (multi-stage_saak_v2.py)

   An alternative view about the Cascade of kernel augmentation and ReLU operation
   (5,-3) -> (5, 0, 0, 3)
  

• [Other Code]

  • [notebook] multi-stage_saak_v2.ipynb
  • [dataset I/O] datasets.py, utils.py

• Contact Me

Jiali Duan (Email: jli.duan@gmail.com)

Yueru Chen (Email: yueruche@usc.edu)