Nowadays, deep learning methods such as neural network models are highly effective for various tasks including image classification and natural language processing, but the increasingly high computational cost restricts the deployment of these network models in many kinds of scenarios whose resources are usually limited. Among all kinds of methods to solve these difficulties, quantization is a plausible way to reduce the storage size of these network models and accelerate their inference process by replacing the parameters such as weights with low-bit fixed numbers during the training process. This problem can be viewed as a discrete constrained optimization problem. In this work, we use Alternative Direction Methods of Multipliers (ADMM) to decouple the continuous parameters from the discrete constraints so that the original hard optimization problem is separated into several subproblems. In addition, structure-aligned quantization is also achieved, which is usually more friendly for edge computing devices to execute and accelerate. With extensive experiments on ImageNet and CIFAR10 dataset, models represented by low-bit fixed-point numbers with acceptable accuracy loss compared with original full precision models can be acquired. Compared with some previous quantization works, the quantization models obtained in this work have little classification accuracy drop compared with the original pre-trained full precision model, and a kind of hardware-friendly structure that makes the neural network easier to deploy is achieved.
|