Loading...

Model pruning in convolutional encoder-decoder network

We use the Network Slimming to prune all the unimportant neurons by enforcing the channel-wise sparsity on all the matching convolutional layers and batch normalization layers within a block. We evaluate the approach by pruning the DPDB-Net for classification which gives a 50% reduction in model size and 5% reduction in computation and on image segmentation task 35% reduction in model size and 12% reduction in computation.