Filter Pruning via Probabilistic Model-based Optimization for Accelerating Deep Convolutional Neural Networks

Accelerating Deep Convolutional Neural Networks(CNNs) has recently received ever-increasing research focus. Among various approaches proposed in the literature, filter pruning has been regarded as a promising solution, which is due to its advantage in significant speedup and memory reduction of both network model and intermediate feature maps. Previous works utilized "smaller-norm-less-important" criterion to prune filters with smaller l(p)-norm values by pruning and retraining alternately. However, they ignore the effects of feedback: most current approaches that prune filters only consider the statistics of the filters (e.g., prune filter with small l(p)-norm values), without considering the performance of the pruned model as an important feedback signal in the next iteration of filter pruning. To solve the problem of non-feedback, we propose a novel filter pruning method, namely Filter Pruning via Probabilistic Model-based Optimization (FPPMO). FPPMO solves the problem of non-feedback by pruning filters in a probabilistic manner. We introduce a pruning probability for each filter, and pruning is guided by sampling from the pruning probability distribution. An optimization method is proposed to update the pruning probability based on the performance of the pruned model in the pruning process. When applied to two image classification benchmarks, the effectiveness of our FPPMO is validated. Notably, on CIFAR-10, our FPPMO reduces more than 57% FLOPs on ResNet-110 with even 0.08% relative accuracy improvement. Moreover, on ILSVRC-2012, our FPPMO reduces more than 50% FLOPs on ResNet-101 without top-5 accuracy drop. Which proving that our FPPMO outperforms the state-of-the-art filter pruning method.