sigmoid added for faster convergence

L2 regulrization added to the model; gradient clipping of 10 added. I…

…nitializer changed in the prediction/proto head; sigmoid_cross_entropy used for classification loss; binary_cross_entropy without activation used for mask/seg loss.