Created by: Sand3r-

Purpose

Currently, the networks which use Convolution with bias, divide the process into two steps (or operations): conv_op->elementwise_add->next_op. In order to increase the performance, the elementwise_add is removed and instead an mkldnn biased convolution is used.

Changes introduced

Convolution now accepts an optional parameter called "Bias" of dimension = 1, which size of is equal to the number of output channels produced by conv operation. Inference transpiler has been updated to replace conv and successing elementwise_add with a single conv op.