The article discusses a new method, Robustness-Guided Image Synthesis (RIS), to improve the semantics and diversity of synthetic images used in data-free quantization. Data-free quantization is a model compression technique that avoids privacy concerns by using synthesized images instead of real training data. Existing methods ensure the reliability of these synthetic images using classification loss. However, these images often suffer from low semantics and homogenization issues. The proposed method, RIS, introduces perturbations on input and model weight, defines inconsistency metrics at feature and prediction levels, and designs a robustness optimization objective. It also makes the approach diversity-aware by forcing the generator to synthesize images with small correlations in the label space. The method achieves state-of-the-art performance for various settings on data-free quantization.
Publication date: 6 Oct 2023
Project Page: https://anonymous.4open.science/r/RIS-00DF/
Paper: https://arxiv.org/pdf/2310.03661
