One‑Stage Training for Generative Adversarial Networks (OSGAN): Methodology and Efficiency Analysis

In recent years, generative adversarial techniques have been applied to many image tasks such as editing, style transfer, caption generation, few‑shot data augmentation, adversarial attacks, and AI font design. However, the training efficiency of GANs remains a bottleneck for large‑scale, iterative development.

The Alibaba Mama search advertising team, in collaboration with Prof. Song Ming‑Li of Zhejiang University, conducted exploratory research and proposed a one‑stage GAN training method (OSGAN, "Training Generative Adversarial Networks in One Stage"). The method achieves a 1.5× speed‑up over the traditional two‑stage training while preserving performance. The work was accepted to CVPR 2021, and the code has been open‑sourced.

Background

Traditional GAN training follows a two‑stage alternating optimization: the generator (G) and discriminator (D) are updated in opposite directions, which introduces redundant computation compared with standard CNN training. This two‑stage scheme limits the overall training efficiency of GANs.

Proposed Solution (OSGAN)

The authors first categorize GANs into symmetric GANs (both G and D share the same loss on fake samples) and asymmetric GANs (different losses on fake samples). For symmetric GANs, the gradient of D with respect to fake samples can be reused to compute the gradient for G, allowing a single‑stage update.

For asymmetric GANs, a combined loss function is designed to avoid gradient cancellation between G and D. By scaling the mixed gradient and decomposing it, the method recovers separate gradients for G and D, enabling one‑stage training for asymmetric GANs as well.

Experimental Analysis

The paper evaluates OSGAN on both symmetric and asymmetric DCGANs. Three aspects are measured: (1) time spent on real vs. generated samples per batch, (2) forward‑pass vs. backward‑pass time, and (3) gradient computation time for network parameters. Results show that the one‑stage approach reduces total training time by approximately 1.6–1.7× compared with the two‑stage baseline.

The acceleration ratio in the worst case reaches around 1.5×, confirming the efficiency gain of the proposed method.

Conclusion and Outlook

OSGAN provides a practical one‑stage training pipeline that speeds up GAN training by about 1.5× without sacrificing quality. The authors also demonstrate that OSGAN can be used for data augmentation in low‑sample supervised CNN tasks, yielding performance improvements in advertising scenarios. Future work will explore even more efficient GAN training strategies and broader applications.