Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks

Yonggan Fu, Wuyang Chen, Haotao Wang, Haoran Li, Yingyan Celine Lin, Zhangyang Wang

TL;DR

The paper tackles the challenge of compressing GAN generators for on-device deployment by introducing AutoGAN-Distiller (AGD), an AutoML framework designed to discover compact generators under explicit resource constraints. AGD combines a customized sequential search space with differentiable NAS and a distillation-guided proxy task that preserves generation quality without requiring trained discriminators. It demonstrates the approach on unpaired image translation (CycleGAN) and image super-resolution (ESRGAN), achieving substantial reductions in model size, FLOPs, and real-device latency while maintaining or improving perceptual metrics such as FID and PSNR. The work offers a practical, generalizable pathway to deploy efficient GANs on mobile devices, with potential impact on a broad range of image translation and enhancement tasks.

Abstract

The compression of Generative Adversarial Networks (GANs) has lately drawn attention, due to the increasing demand for deploying GANs into mobile devices for numerous applications such as image translation, enhancement and editing. However, compared to the substantial efforts to compressing other deep models, the research on compressing GANs (usually the generators) remains at its infancy stage. Existing GAN compression algorithms are limited to handling specific GAN architectures and losses. Inspired by the recent success of AutoML in deep compression, we introduce AutoML to GAN compression and develop an AutoGAN-Distiller (AGD) framework. Starting with a specifically designed efficient search space, AGD performs an end-to-end discovery for new efficient generators, given the target computational resource constraints. The search is guided by the original GAN model via knowledge distillation, therefore fulfilling the compression. AGD is fully automatic, standalone (i.e., needing no trained discriminators), and generically applicable to various GAN models. We evaluate AGD in two representative GAN tasks: image translation and super resolution. Without bells and whistles, AGD yields remarkably lightweight yet more competitive compressed models, that largely outperform existing alternatives. Our codes and pretrained models are available at https://github.com/TAMU-VITA/AGD.

AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks

TL;DR

The paper tackles the challenge of compressing GAN generators for on-device deployment by introducing AutoGAN-Distiller (AGD), an AutoML framework designed to discover compact generators under explicit resource constraints. AGD combines a customized sequential search space with differentiable NAS and a distillation-guided proxy task that preserves generation quality without requiring trained discriminators. It demonstrates the approach on unpaired image translation (CycleGAN) and image super-resolution (ESRGAN), achieving substantial reductions in model size, FLOPs, and real-device latency while maintaining or improving perceptual metrics such as FID and PSNR. The work offers a practical, generalizable pathway to deploy efficient GANs on mobile devices, with potential impact on a broad range of image translation and enhancement tasks.

Abstract

The compression of Generative Adversarial Networks (GANs) has lately drawn attention, due to the increasing demand for deploying GANs into mobile devices for numerous applications such as image translation, enhancement and editing. However, compared to the substantial efforts to compressing other deep models, the research on compressing GANs (usually the generators) remains at its infancy stage. Existing GAN compression algorithms are limited to handling specific GAN architectures and losses. Inspired by the recent success of AutoML in deep compression, we introduce AutoML to GAN compression and develop an AutoGAN-Distiller (AGD) framework. Starting with a specifically designed efficient search space, AGD performs an end-to-end discovery for new efficient generators, given the target computational resource constraints. The search is guided by the original GAN model via knowledge distillation, therefore fulfilling the compression. AGD is fully automatic, standalone (i.e., needing no trained discriminators), and generically applicable to various GAN models. We evaluate AGD in two representative GAN tasks: image translation and super resolution. Without bells and whistles, AGD yields remarkably lightweight yet more competitive compressed models, that largely outperform existing alternatives. Our codes and pretrained models are available at https://github.com/TAMU-VITA/AGD.

Paper Structure

This paper contains 15 sections, 3 equations, 4 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. AutoML: Neural Architecture Search
  4. Model Compression
  5. GANs and GAN Compression
  6. Our AutoGAN-Distiller Framework
  7. The Proposed Customized Search Space
  8. The Proposed Customized Proxy Task
  9. Experiment Results
  10. Considered Tasks & Models.
  11. Evaluation Metrics.
  12. Training details.
  13. AGD for Efficient Unpaired Image Translation
  14. AGD for Efficient Super Resolution
  15. Conclusion

Figures (4)

  • Figure 3: Visualization examples of CycleGAN compression on the horse2zebra task (rows 1 and 2), zebra2horse task (row 3), and summer2winter task (row 4). Columns from left to right: source images, translation results by original CycleGAN, CEC, and AGD, respectively, and the FLOPs and memory (model size) of each method on each task are annotated above the images.
  • Figure 4: Visualization results of CycleGAN compressed by AGD with quantization, where the first two figures are from the horse2zebra task,and the last figure is from the zebra2horse task.
  • Figure 5: Visualization results of different super resolution methods (Better viewed after zooming in). From top to bottom rows are examples from Set5, Set14, BSD100, and Urban100. Columns from left to right: original ESRGAN wang2018esrgan, ESRGAN after structural pruning li2016pruning, SRGAN ledig2017photo, VDSR kim2016accurate, and ESRGAN compressed by AGD. FLOPs (calculated as processing a 256$\times$256 low resolution image with a scale factor of four) and memory (model size) of each method on each task are annotated above the images.
  • Figure : The Proposed AutoGAN-Distiller Framework