Magnitude and Similarity Based Variable Rate Filter Pruning for Efficient Convolution Neural Networks

The superior performance of the recent deep learning models comes at cost a significant increase in computational complexity, memory use, and power consumption. Filter pruning is one effective neural network compression techniques suitable for model deployment modern low-power edge devices. In this paper, we propose loss-aware filter Magnitude Similarity based Variable rate Pruning (MSVFP) technique. We studied several selection criteria on magnitude similarity among filters within convolution layer, assumption that sensitivity each layer throughout different, unlike conventional fixed methods, our algorithm using automatically finds network. addition, proposed adapts two different to remove weak as well redundant score respectively. Finally, iterative retraining approach are used maintain accuracy during its target float point operations (FLOPs) reduction rate. algorithm, small number steps sufficient prevent an abrupt drop Experiments with commonly VGGNet ResNet CIFAR-10 ImageNet benchmark show superiority method over existing methods literature. Notably, VGG-16, ResNet-56, ResNet-110 dataset even improved original more than 50% FLOPs. Additionally, ResNet-50 reduces FLOPs by 42% negligible accuracy.