Has anyone tried doing Alexnet's PCA noise augmentation on imagenet data on caffe?

[yxc...@gmail.com]

At first, I am looking for value of the eigenvector and eigenvalues of imagenet 2012 data, so that I can write a python layer to make lighting augmentation. However, when I look closer, caffe stores image value as integers, but one requirement for PCA is to normalize it to between 0 and 1. Therefore, I need to change the data layer as well to add this feature. Can I know why is this feature not implemented on caffe when it can give a boost of 1% on Alexnet. If someone has implemented, can you kindly share? Thanks.

[yxc...@gmail.com]

I also wonder how someone can do PCA on such a large dataset. My computer memory is not sufficient to calculate it so i need the value that someone else has calculated.