CenteredWN

KO | EN

gitlite — search

#typescript #javascript #rust #python #css #html #react #android #csharp #nextjs #graphql #ios

★ 21 Open GitHub ↗

This project is the Torch implementation of our ICCV 2017 paper: Centered Weight Normalization in Accelerating Training of Deep Neural Networks

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Centered Weight Normalization
======================================

This project is the Torch implementation of the paper: Centered Weight Normalization  in Accelerating Training of Deep Neural Networks ( ICCV 2017).
* bibtex:
```Bash
@INPROCEEDINGS{Huang2017ICCV,
    author = {Lei Huang and Xianglong Liu and Yang Liu and  Bo Lang and Dacheng Tao},
    title = {Centered Weight Normalization  in Accelerating Training of Deep Neural Networks},
    booktitle = {ICCV},
    year = {2017}}
 ```

### Updates
* Add the [' pytorch module'](./module_pytorch/CWN.py)  of CWN.     2019-12-07.

## Requirements and Dependency
* Install [Torch](http://torch.ch) with CUDA GPU
* Install [cudnn v5](http://torch.ch)
* Install dependent lua packages optnet by run:
luarocks install optnet

## Experiments in the paper

#### 1. MLP architecture over SVHN dataset

* Dataset prepraration:  We can get the preprocessed SVHN dataset for MLP architecture by running:
```Bash
  cd dataset
   th preProcess_div256.lua
 ```

Note that this script is based on the [Torch script for SVHN](https://github.com/torch/tutorials/blob/master/A_datasets/svhn.lua)

*	Execute: 
```Bash
 th exp_MLP.lua 
 ```

*	To reproduce the experimental results, you can run the script below, which include all the information of experimental configurations: 
```Bash
 bash 1_execute_MLP_svhn.sh  
 bash 1_execute_MLP_svhn_adam.sh  
 ```

#### 2. VGG-A architecture over Cifar-10 dataset
* Dataset preparations: the dataset is based on the preprocessed script on: https://github.com/szagoruyko/cifar.torch, and you should put the data file in the directory: './dataset/cifar_provider.t7'

* Execute:
```Bash
 th exp_vggA.lua –dataPath './dataset/cifar_provider.t7'
 ```
 *	To reproduce the experimental results, you can run the script below, which include all the information of experimental configurations: 
```Bash
 bash   2_execute_Conv_CIFAR10_vggA.sh
 ```
 
#### 3. GoogLeNet architecture over Cifar datasets

*	Dataset preparations: The dataset is based on [whitened CIFAR datasets](https://yadi.sk/d/em4b0FMgrnqxy).  
 * Execute: 
```Bash
th exp_GoogleNet_dataWhitening.lua –dataPath './dataset/cifar100_whitened.t7'
 ```
  *	To reproduce the experimental results, you can run the script below, which include all the information of experimental configurations: 
```Bash
 3_execute_Conv_CIFAR100_GoogLeNet.sh 
 ```
 
 
  The GoogLeNet model is based on the project on: https://github.com/soumith/imagenet-multiGPU.torch

#### 4. Residual network architecture over Cifar datasets

*	Dataset preparations: The dataset is based on [original CIFAR datasets](https://yadi.sk/d/eFmOduZyxaBrT), and the data file should be put in the directory: ./dataset/cifar_original.t7.  
 *	Execute: 
```Bash
th exp_res_dataNorm.lua –dataPath './dataset/cifar10_original.t7'
 ```
   *	To reproduce the experimental results, you can run the script below, which include all the information of experimental configurations: 
```Bash
4_execute_Conv_CIFAR10_resnet.sh
 ```
 
  The normlization of Cifar dataset is in the script th exp_res_dataNorm.lua. The residual network model and respective script are based on [facebook ResNet](https://github.com/facebook/fb.resnet.torch).

####  5. GoogLeNet over ImageNet
This experiment is based on the project at: https://github.com/soumith/imagenet-multiGPU.torch.
<br>
The proposed model are in: './models/imagenet/'

## Contact
huanglei@nlsde.buaa.edu.cn, Any discussions and suggestions are welcome!

CenteredWN

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README