Adaptive FSP : Adaptive Architecture Search with Filter Shape Pruning
DeepConvolutionalNeuralNetworks(CNNs)havehighmem- ory footprint and computing power requirements, making their deploy- ment in embedded devices difficult. Network pruning has received at- tention in reducing those requirements of CNNs. Among the pruning methods, Stripe-Wise Pruning (SWP) achieved a further network com- pression than conventional filter pruning methods and can obtain op- timal kernel shapes of filters. However, the model pruned by SWP has filter redundancy because some filters have the same kernel shape. In this paper, we propose the Filter Shape Pruning (FSP) method, which prunes the networks using the kernel shape while maintaining the recep- tive fields. To obtain an architecture that satisfies the target FLOPs with the FSP method, we propose the Adaptive Architecture Search (AAS) framework. The AAS framework adaptively searches for the architec- ture that satisfies the target FLOPs with the layer-wise threshold. The layer-wise threshold is calculated at each iteration using the metric that reflects the filter's influence on accuracy and FLOPs together. Compre- hensive experimental results demonstrate that the FSP can achieve a higher compression ratio with an acceptable reduction in accuracy.