A Hardware-friendly Quantization and Model Fine Tuning with STEBC for Object Detection

Chia-Chi Tsai

Abstract: This paper present a hardware-friendly quantization and model training method using dynamic fixed-point and straight through estimator with boundary constraint(STEBC). By the proposed quantization and fine-tune methods, with all bit-width of layers are set to less than 8-bit including the first layer, quantizing a model from floating-point without fine-tuning can achieve less than 5% drop in accuracy. After fine-tune using the straight-through estimator with boundary constraint, we can recover the accuracy to less than 1% drop compared to its floating-point counterpart. We have proven our method works on various detectors and various backbone, including one-stage and two-stage detector.