Rethinking Data-Free Quantization as a Zero-Sum Game

Data-free quantization (DFQ) recovers the performance of quantized network (Q) without accessing real data, but generates fake sample via a generator (G) by learning from full-precision (P) instead. However, such generation process is totally independence Q, specialized as failing to consider adaptability generated samples, i.e., beneficial or adversarial, over resulting into non-ignorable loss. Building on this, several crucial questions --- how measure and exploit Q under varied bit-width scenarios? generate samples with desirable benefit network? impel us revisit DFQ. In this paper, we answer above game-theory perspective specialize DFQ zero-sum game between two players network, further propose an Adaptability-aware Sample Generation (AdaSG) method. Technically, AdaSG reformulates dynamic maximization-vs-minimization anchored adaptability. The maximization aims adaptability, reduced minimization after calibrating for recovery. Balance Gap defined guide stationarity maximally Q. theoretical analysis empirical studies verify superiority state-of-the-arts. Our code available at https://github.com/hfutqian/AdaSG.