GDS: Gradient Distribution Scaling-based Gradient Quantization for Low-complexity and Hardware-friendly Training of Instance Segmentation Models

Although recent research in the field of instance segmentation has achieved high accuracy, it suffers from the cost of increased parameters and computation due to the increased model size. To address this issue, various compression techniques are being explored to reduce hardware resources. In particular, quantization is gaining attention as it can significantly reduce hardware resources and power. Recently, quantization studies have expanded from the quantization of the forward pass to also include the backward pass. However, in segmentation tasks, gradient quantization is challenging due to the dynamic distribution across channels, leading to significant quantization errors and making it difficult to achieve high accuracy. In this paper, we propose a technique that applies scaling to layers with a dynamic distribution across channels to address this issue. Additionally, for stability in the initial training, we adopt mixed-precision quantization based on quantization errors. Experimental results show that the proposed method achieves higher accuracy in YOLACT than other quantization methods.