Open talenz opened 1 year ago
Thanks for your interest. As mentioned in the supplementary material, we fine-tune the quantized model with additional learnable layer-wise offsets for activations and knowledge distillation following [1][2]. Reference: [1] Nonuniform-to-Uniform Quantization: Towards Accurate Quantization via Generalized Straight-Through Estimation. CVPR 2022. [2] PROFIT: A Novel Training Method for sub-4-bit MobileNet Models. ECCV 2020.
Will you provide the script to reproduce this result?
Great job! In this paper, the top1 of 4bit mobilenetv2 is 72.0 and this is beyond SOTA. Is there a way that can reproduce this result?