Conditional pseudo-supervised contrast for data-Free knowledge distillation

•We propose a learning paradigm to improve DFKD with a conditional generative adversarial network, which is able to synthesize category-specific images and promote student learning.•We introduce a categorical feature embedding block to effectively distinguish the distribution of different categories of samples, which connect the features and categories embedding in the middle layers.•To our knowledge, we are the first to attempt to utilize the condition annotations to supervise the contrast of features representation of teacher and student in DFKD, which aims to optimize the diversity of synthesized images and improve the effect of distillation.•Massive experiments are conducted on three mainstream benchmark datasets, i.e., CIFAR-10, CIFAR-100, Tiny-ImageNet. The results demonstrate the effectiveness of the proposed CPSCDFKD in both improving the student and generator.