Exploration of Automatic Mixed-Precision Search for Deep Neural Networks

Neural networks have shown great performance in cognitive tasks. When deploying network models on mobile devices with limited computation and storage resources, the weight quantization technique has been widely adopted. In practice, 8-bit or 16-bit quantization is mostly likely to be selected in order to maintain the accuracy at the same level as the models in 32-bit floating-point precision. Binary quantization, on the contrary, aims to obtain the highest compression at the cost of much bigger accuracy drop. Applying different precision in different layers/structures can potentially produce the most efficient model. Seeking for the best precision configuration, however, is difficult. In this work, we proposed an automatic search algorithm to address the challenge. By relaxing the search space of quantization bitwidth from discrete to continuous domain, our algorithm can generate a mixed-precision quantization scheme which achieves the compression rate close to the one from the binary-weighted model while maintaining the testing accuracy similar to the original full-precision model.