A Deep Neural Network Pruning Method Based on Gradient L1-norm
Xue Liu, Weijie Xia, Zhimiao Fan · 2020
The deep neural network model usually has a large number of redundant weight parameters. When calculating the deep neural network model, it needs to occupy a large amount of computing resources and storage space, which makes it difficult to deploy on some edge devices and embedded devices. In order to solve this problem, we propose a deep network pruning algorithm based on gradient L1 norm (GLNP). The core idea of GLNP algorithm is to judge the importance of the filter based on the L1-norm of gradient. Then, according to the pruning ratio, we remove the filter and its connection feature diagram layer by layer, and retrain the precision of the depth network model after pruning. We show that the GLNP algorithm can reduce 71% parameters on VGG-16, with guaranteeing the accuracy 90.81%; Similar experiments with ResNet-56 reveal that the GLNP algorithm can reduce 35.2% parameters on ResNet-56, with guaranteeing the accuracy 93.11%. Compared with the current popular PFEC algorithm, the GLNP algorithm compresses of both network model while improving the accuracy, which has better performance of deep network compression and acceleration.