Need to adjust initial values for each layer.

[hantek]

Currently for all kinds of layers we are applying the weight initialization method which is actually for Sigmoid. i.e., the uniform distribution from -numpy.sqrt(6. / (n_in + n_out)) to numpy.sqrt(6. / (n_in + n_out)).

Sometimes this causes problems. We need to find good scale and distribution specifically for each kind of layers. That improves the default performance of the models built on top of it.