Topological optimization of the DenseNet with pretrained-weights inheritance and genetic channel selection

Abstract

Convolutional neural networks (CNNs) have been successfully applied in many computer vision applications [1], especially in image classification tasks, where most of the structures have been designed manually. With the aid of skip connection and dense connection, the depths of the models are becoming “deeper” and the filters of layers are getting “wider” in order to tackle the challenge of large-scale datasets. However, large-scale models in convolutional layers become inefficient due to the redundant channels from input feature maps. In this paper, we aim to automatically optimize the topology of the DenseNet, in which unnecessary convolutional kernels are reduced. To achieve this, we present a training pipeline that generates the network structure using a genetic algorithm. We first propose two encoding methods that can represent the structure of the model using a fixed-length binary string. A three-step based evolutionary process consisting of selection, crossover, and mutation is proposed to optimize the structure. We also present a pretrained weight inheritance method which can largely reduce the total time consumption of the genetic process. Experimental results have demonstrated that our proposed model can achieve comparable accuracy to the state-of-the-art models, across a wide range of image recognition and classification datasets, whilst significantly reducing the number of parameters.