Black-Box Diagnosis and Calibration on GAN Intra-Mode Collapse: A Pilot Study

Generative adversarial networks (GANs) nowadays are capable of producing images incredible realism. Two concerns raised whether the state-of-the-art GAN’s learned distribution still suffers from mode collapse and what to do if so. Existing diversity tests samples GANs usually conducted qualitatively on a small scale and/or depend access original training data as well trained model parameters. This article explores GAN intra-mode calibrates that in novel black-box setting: neither nor parameters is assumed. The new setting practically demanded yet rarely explored significantly more challenging. As first stab, we devise set statistical tools based sampling can visualize, quantify, rectify . We demonstrate effectiveness our proposed diagnosis calibration techniques, via extensive simulations experiments, unconditional image generation (e.g., face vehicle). Our study reveals prevailing problem diagnosable calibratable settings. codes available at https://github.com/VITA-Group/BlackBoxGANCollapse