Deployable mixed-precision quantization with co-learning and one-time search.

Abstract

Mixed-precision quantization plays a pivotal role in deploying deep neural networks in resource-constrained environments. However, the task of finding the optimal bit-width configurations for different layers under deployable mixed-precision quantization has barely been explored and remains a challenge. In this work, we present Cobits, an efficient and effective deployable mixed-precision quantization framework based on the relationship between the range of real-valued input and the range of quantized real-valued. It assigns a higher bit-width to the quantizer with a narrower quantized real-valued range and a lower bit-width to the quantizer with a wider quantized real-valued range. Cobits employs a co-learning approach to entangle and learn quantization parameters across various bit-widths, distinguishing between shared and specific parts. The shared part collaborates, while the specific part isolates precision conflicts. Additionally, we upgrade the normal quantizer to dynamic quantizer to mitigate statistical issues in the deployable mixed-precision supernet. Over the trained mixed-precision supernet, we utilize the quantized real-valued ranges to derive quantized-bit-sensitivity, which can serve as importance indicators for efficiently determining bit-width configurations, eliminating the need for iterative validation dataset evaluations. Extensive experiments show that Cobits outperforms previous state-of-the-art quantization methods on the ImageNet and COCO datasets while retaining superior efficiency. We show this approach dynamically adapts to varying bit-width and can generalize to various deployable backends. The code will be made public in https://github.com/sunnyxiaohu/cobits.