Normalizing input and role of Bias in Caffe?

[Prabhu]

I have 10k grayscale images, and i am extracting x,y coordinates of features using caffe.
At present i have the input and output normalized between 0 and 1.


1. Does normalizing the input between -1 and 1 make any difference in caffe??
2. Also does setting a non-zero bias to fully connected layers effect the output?

I have used 3 convolution, 2 pooling, 2 IP and 1 euclidean loss output.

[Shahmi Junoh]

Just out of curiosity, how did you do the normalization of input [-1, 1] at prototxt level? I only know how to normalize [0, 1) i.e. by defining scale value of 1/256 in transform_param.

Thank you.

[Przemek D]

Subtracting the mean image does that. Scale by 1/256 gets you a [0;1) image, now if you subtract mean which is another [0;1) image you get (-1;1) output.

[Shahmi]

Thanks a lot Przemek D. Crystal clear now.