EvoPrunerPool: An Evolutionary Pruner using Pruner Pool for Compressing Convolutional Neural Networks

Abstract

This paper proposes EvoPrunerPool - an Evolutionary Pruner using Pruner Pool for Compressing Convolutional Neural Networks. EvoPrunerPool formulates filter pruning as a search problem for identifying the right set of pruners from a pool of off-the-shelf filter pruners and applying them in appropriate sequence to incrementally sparsify a given Convolutional Neural Network. The efficacy of EvoPrunerPool has been demonstrated on LeNet model using MNIST data as well as on VGG-19 deep model using CIFAR-10 data and its performance has been benchmarked against state-of-the-art model compression approaches. Experiments demonstrate a very competitive and effective performance of the proposed Evolutionary Pruner. Since EvoPrunerPool employs the native representation of a popular machine learning framework and filter pruners from a well-known AutoML toolkit the proposed approach is both extensible and generic. Consequently, a typical practitioner can use EvoPrunerPool without any in-depth understanding of filter pruning in specific and model compression in general.