tianshuichen / wavelet-like-auto-encoder

Wavelet-like Auto-Encoder

Introduction

This repository contains the caffe prototxt and trained model described in the paper "Learning a Wavelet-like Auto-Encoder to Accelerate Deep Neural Networks".

For more details, please visit our project page: WAE project page.

Model

224x224 center crop validation accuracy on ImageNet, evaluated with a C++ implementation on Intel i7 CPU (3.50GHz) and Nvidia GeForce GTX TITAN-X GPU.

Top-1	Top-5	CPU (ms)	GPU (ms)
67.88%	88.27%	411.63	2.37

Note: The model is retrained in the same way as desribed in the paper, and the accuracy is slightly better than that reported in the paper.

The trained model can be download from google drive or baidu cloud.

Citation

If you find this work useful for your research, please cite:

@inproceedings{{chen2018learning,
	author = {Chen, Tianshui and Lin, Liang and Zuo, Wangmeng and Luo, Xiaonan and Zhang, Lei},
	title = {Learning a Wavelet-like Auto-Encoder to Accelerate Deep Neural Networks},
	booktitle = {AAAI},
	year = {2018}
}

Concact

Feel free to contact me if you have any question (Tianshui Chen [email protected])

wavelet-like-auto-encoder's Issues

数据集和测试代码

陈老师好，我刚拜读了您的论文，对您的实验很感兴趣。但这里的caffe代码不太好读，您能提供多种版本的实现代码吗？
论文中的GPU、CPU的测试我也是第一次听说，很好奇，您能上传c++的测试代码，到WAE中来吗？

最后，您的损失函数设计让我想起了Context Encoder: Feature Learning By Inpainting 这篇文章。这篇文章虽然没有将输入分成两部分，但在损失函数设计时，也考虑了reconstruction loss 和 adversarial loss， 两部分，我觉得你们的设计有异曲同工之妙，但还没有思考出内在的联系。

我会尝试用tensorflow在小一些的数据集上，复现老师您的论文的实验，但如果您能提供更多帮助文档和资料，我将不胜感激。谢谢！

a more radical resize rate?

Have you ever tried a more radical resize rate, such as 76x76 or 56x56?

What's the structure of decoding layer

I could extract the encoding layer on stage 1, but there isn't enough information about decoding layer structure. I've tried the symmetrical structure of encoding layer which contains only one BN after deconv. However, I'm not sure if it is correct. The image reconstruction after 1 epoch shows that high-freq image is likely all-dark, and loss decreases to 131.449 from 10000, which are confusing.

About use this method on detection

Hi,
Maybe we use 256 image size, can detection tiny object, when we use this method( the each input become 128) can influence the detection of small objects?
Thanks.

could you please share the code training the network

Thanks for your nice works! @tianshuichen
I am very interested in your wonderful work. However, I only find the testing code, could you please share the code training the network?