Although the deep network model has been widely used in industry, it cannot be applied well to devices with limited memory or limited computing resources, such as mobile phones and satellites. Model compression technology can reduce the size of the model and runs better on devices with memory limitations. In this paper, we proposed a learning-based strategy leveraging reinforcement learning with compression ratio and accuracy exceeding those of current the rule-based policy. The reason why we achieved the great progress of significant performance is that we leverage DDPG of reinforcement learning to provide the model compression strategy based on the pruned model. The method has a higher compression ratio, better retains accuracy and freeing human labor. The proposed method shows that the model achieved more than 3.1% accuracy and more than 6.46X compression ratio compared with the hand-crafted model compression policy for ResNet20 on Ciar10.