high loss in regress human poses from images

[himora...@gmail.com]

Hi everyone,

I'm testing regression of human poses using HDF5 format. I have observed that the training loss is very high, more than 10^7. I have tried normalizing images by image mean subtraction or scaling to [0,1]. I don't scale the labels 

 Input image size: 1x3x336x336

 Input label size: 1x52

I have attached the trainval.prototxt and solver.txt. Hope to hear your analysis on this.

Best regard