Caffe yolo 9000 problem

[snowm...@gmail.com]

I implement the yolo 9000 in caffe, but I get some problem 

in the darknet, the gradient is 

float tw = log(truth.w*w / biases[2*n]); // the darknet version

delta[index + 2*stride] = scale * (tw - x[index + 2*stride]);

but i notice the "log" will let gradient to nan when w is small

so i change the gradient function to

 ( I forgot -1) 

diff[index + 2 * SIZE] = scale * (pred_box.w - truth_box.w) * pred_box.w;

can i do that?

here is my github

https://github.com/karta0807913/caffe-yolo-9000

P.S. I am sorry for my bad English :'(