Deep neural network pruning algorithm based on particle swarm

A powerful neural network consumes a lot of storage space and computing resources, which is unacceptable for mobile devices and embedded devices with limited resources. To solve this problem, PSOPruner was proposed. Firstly, The algorithm randomly initializes a series of particles as the pruned network structure; secondly, calculates the threshold corresponding to each convolutional layer according to the particles, deletes the convolution kernels with corresponding values less than the threshold; then, adapts the pruned network structure evaluation, update the particles and the global optimal fitness particles; finally, retrain the network structure corresponding to the global optimal fitness particles to restore the network accuracy. The experimental results show that the algorithm is applied to VGG16 on the cifar-10 data set. After 100 iterations of the algorithm, the test accuracy is improved by 0.04%, the model size is reduced by 94%, and the running speed is increased by 36% compared with the untrimmed model. The pruned network model is more conducive to deployment in mobile devices and embedded devices.