MBQuant: A novel multi-branch topology method for arbitrary bit-width network quantization

Yunshan Zhong, Yuyao Zhou, Fei Chao, Rongrong Ji

Abstract: Highlights•We find that existing arbitrary bit-width methods suffer from accumulated quantization errors from switching weight and activations bit-widths.•We introduce MBQuant, which utilizes the multi-branch topology and an amortization strategy to address accumulated quantization errors.•Extensive experiments demonstrate that MBQuant achieves significant performance gains compared to existing arbitrary bit-width methods.