Fast and Effective: A Novel Sequential Single-Path Search for Mixed-Precision-Quantized Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Model quantization reduces model size and computational latency, crucial for mobile and embedded devices.
• Mixed-precision quantization optimizes performance by assigning different bit precisions based on layer sensitivity.
• Determining optimal mixed-precision settings under constraints (hardware, energy, latency) is challenging.

Purpose of the Study:

• To propose a novel sequential single-path search (SSPS) method for efficient mixed-precision model quantization.
• To guide the search process using predefined constraints for practical applications.
• To accelerate the search for optimal quantization configurations.

Main Methods:

• Introduced a sequential single-path search (SSPS) method incorporating constraints.
• Developed a single-path search cell within a fully differentiable supernet for gradient-based optimization.
• Sequentially determined candidate precisions based on selection certainties to reduce search space.

Main Results:

• SSPS efficiently searched for mixed-precision models across diverse architectures (ResNet, MobileNet-V2) and datasets (CIFAR-10, ImageNet, COCO).
• The method effectively handled constraints related to hardware resources, energy consumption, model size, and computational latency.
• Experimental results demonstrated that SSPS significantly outperformed uniform-precision quantization methods.

Conclusions:

• The proposed SSPS method offers an efficient approach to finding optimal mixed-precision quantization configurations.
• SSPS successfully balances model performance with practical deployment constraints.
• This technique provides a significant advancement over traditional uniform-precision quantization strategies.