Rank Selection Method of CP Decomposition Based on Deep Deterministic Policy Gradient Algorithm

Abstract

With the popularity of edge computing devices and increasing complexity of convolutional neural network (CNN) models, the need for model compression and acceleration has become increasingly urgent. As an effective model compression technique, CANDECOMP/PARAFAC (CP) decomposition relies heavily on the preset rank for its compression effectiveness. However, no direct algorithm is currently available for determining the optimal tensor rank. Therefore, a novel method for CP rank selection based on deep reinforcement learning is proposed. This method utilizes the DecG single-player game framework based on the Deep Deterministic Policy Gradient (DDPG) algorithm to achieve automation and intelligence in rank selection. In this process, a pre-trained model is introduced, which fuses and reshapes several historical tensors as network inputs. Additionally, a hybrid greedy strategy based on singular value decomposition (SVD) was designed in the exploration phase to enhance the efficiency of finding ideal rank selection results. This method can automatically determine the rank according to the weight tensor characteristics of the convolution layer and optimize the compression efficiency and performance of the model. In addition, a compression efficiency index is developed to visually demonstrate the performance of the various compression methods. Finally, on the CIFAR-10 and CIFAR-100 datasets, CP decomposition experiments are conducted on various convolutional neural network models, and the decomposed models undergo iterative fine-tuning for retraining. The experimental results show that the rank values determined by the DecG method achieve significant optimization and enhancement in the compression efficiency of the models compared to other methods, exhibiting strong robustness.

Keywords

• Convolutional neural network
• deep deterministic policy gradient
• model compression
• low-rank decomposition
• CP rank selection