Efficient deep convolutional model compression with an active stepwise pruning approach