NormProjection

This project is the Torch implementation of the paper: Projection Based Weight Normalization for Deep Neural Networks (arXiv:1710.02338)

• bibtex:

@article{Huang_2017_arxiv,
    author = {Lei Huang and Xianglong Liu and  Bo Lang  and Bo Li},
    title = {:Projection Based Weight Normalization for Deep Neural Networks},
   journal   = {CoRR},
  volume    = {abs/1710.02338},
  year      = {2017}}

Requirements and Dependency

• Install Torch with CUDA GPU
• Install cudnn v5
• Install dependent lua packages optnet by run: luarocks install optnet

Experiments in the paper

1. Reproduce the results on MLP architecture over MNIST dataset:

• Execute:

  bash 1_execute_MLP.sh   
  bash 1_execute_MLP_UpdateT.sh

2. Reproduce the results on Incption, VGG and Residual network over CIFAR datsets:

• Dataset preparations: you should download the CIFAR-10 and CIFAR-100 datasets, and put the data file in the directory: './dataset/'

• To reproduce the experimental results, you can run the script below, which include all the information of experimental configurations:

  bash 2_execute_Conv_Inception.sh  
  bash 3_execute_Conv_VGG.sh 
  bash 4_execute_Conv_resnet.sh  

The Inception model is based on the project on: https://github.com/soumith/imagenet-multiGPU.torch.

The residual network model is based on the facebook torch project: https://github.com/facebook/fb.resnet.torch

3. Run the experiment on imageNet dataset.

• (1) You should clone the facebook residual network project from:https://github.com/facebook/fb.resnet.torch
• (2) You should download imageNet dataset and put it on: '/tmp/dataset/imageNet/' directory (you also can use other path, which can be set in 'opts_imageNet.lua')
• (3) Copy 'opts_imageNet.lua', 'exp_Conv_imageNet_expDecay.lua', 'train_expDecay.lua', 'module' and 'models' to the project's root path.
• (4) Execute:

th exp_Conv_imageNet_expDecay.lua -model imagenet/preresnet_BN -LR 0.05

You can training other respective models by using the parameter '-model'

4. Semi-supervised learning experiments on Ladder networks

The semi-supervised tasks based on Ladder network can be find in this project: https://github.com/huangleiBuaa/Ladder_deepSSL_NP

Contact

[email protected], Any discussions and suggestions are welcome!