Initialize network with 'Convolution Aware' weights. This may help to achieve a higher accuracy model, but consumes time at first run.