Up to 100x Faster Data-Free Knowledge Distillation

Data-free knowledge distillation (DFKD) has recently been attracting increasing attention from research communities, attributed to its capability to compress a model only using synthetic data. Despite the encouraging results achieved, stateof-the-art DFKD methods still suffer from the inefficiency of data synthesis, making the data-free training process extremely time-consuming and thus inapplicable for large-scale tasks. In this work, we introduce an efficacious scheme, termed as FastDFKD, that allows us to accelerate DFKD by a factor of orders of magnitude. At the heart of our approach is a novel strategy to reuse the shared common features in training data so as to synthesize different data instances. Unlike prior methods that optimize a set of data independently, we propose to learn a meta-synthesizer that seeks common features as the initialization for the fast data synthesis. As a result, FastDFKD achieves data synthesis within only a few steps, significantly enhancing the efficiency of data-free training. Experiments over CIFAR, NYUv2, and ImageNet demonstrate that the proposed FastDFKD achieves 10× and even 100× acceleration while preserving performances on par with state of the art. Code is available at https://github.com/zju-vipa/Fast-Datafree . 2020; Fang et al. 2021a). The conventional setup of KD requires possessing the original training data as input so as to train the student. Unfortunately, due to confidential or copyright reasons, in many cases the original data cannot be released and only the pre-trained models are available to users (Kolesnikov et al. 2020; Shen et al. 2019; Ye et al. 2019) , which, in turn, imposes a major obstacle towards applying KD for a broader domain. To remedy this issue, data-free knowledge distillation (DFKD) approaches have been proposed by assuming that no access to training data is available at all (Lopes, Fenu, and