How can I make input data that processed ZCA whitening?

[Byung-Joon Kwon]

Hello!

I want to test ZCA whitening in caffe. 

- Current procedure

  1. ZCA whitening CIFAR10 in matlab.

       - Rescaling(/255)

       - (X-mean) / (standard deviation) ;each feature

       - ZCA whitening function

       - Save .mat file

 

  2. Save LMDB in python

       - Using caffe tool, make lmdb.

  

  3. Set input LMDB in .prototxt and start training.

Is this wrong way? 

When I print training data, it is uint8, using ipython notebook. But data that processed ZCA whitening is float32. If through data layer, did change data type?

If I save data that processed ZCA whitening as .jpg images in matlab, did my intention changed due to data format change?

I think that zca whitening function is work well. But error rate poorly than original cifar10 data.

Generally, how to input data that preprocessed by some methd into data layer of caffe?

I need your opinion!

[Trần Bằng]

I have the same problem as you. Have you resolved it?

Vào 21:54:30 UTC+7 Thứ Ba, ngày 17 tháng 5 năm 2016, Byungjun Kwon đã viết: