[Paper Review] AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks

Yonggan Fu, Wuyang Chen+4 more|arXiv (Cornell University)|Jun 15, 2020

Generative Adversarial Networks and Image Synthesis48 citations

TL;DR

AutoGAN-Distiller (AGD) searches and distills GAN architectures to compress CycleGAN and ESRGAN, achieving much smaller models with competitive or improved quality.

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.

Motivation & Objective

• Motivate the need for compressing generative adversarial networks without sacrificing translation quality.
• Propose an architecture search framework to discover efficient generator blocks for various unpaired image translation tasks.
• Demonstrate AGD’s effectiveness on CycleGAN variants and ESRGAN, including quantized configurations.
• Provide visualization and quantitative comparisons to state-of-the-art baselines.

Proposed method

• Independent search of generator architectures within a unified search space for each task.
• Utilization of varied operators and widths to form task-specific blocks and stems.
• Evaluation on unpaired image translation tasks (horse2zebra, zebra2horse, summer2winter, winter2summer) and on ESRGAN-based super-resolution.
• Application of 8-bit quantization to searched architectures to assess memory and performance trade-offs.
• Quantitative benchmarking against CEC and PSNR-oriented SR models (VDSR, ESRGAN).
• Visualization of results to illustrate qualitative improvements in edge sharpness and detail.

Experimental results

Research questions

• RQ1 How effectively can AGD compress CycleGAN architectures for unpaired image translation tasks while maintaining visual quality?
• RQ2 Can AGD-derived PSNR-oriented ESRGAN architectures achieve competitive PSNR with substantially reduced FLOPs and memory?
• RQ3 What are the benefits and trade-offs of applying 8-bit quantization to AGD-discovered architectures?

Key findings

• AGD discovers diverse operator sets across tasks, with larger widths appearing early in stems and headers.
• AGD-compressed CycleGAN architectures yield higher visual quality (sharper edges and more details) compared to baselines.
• Quantized AGD architectures maintain comparable visualization quality while delivering substantial memory reductions (8.1x–10.1x relative to CEC).
• On PSNR-oriented ESRGAN, AGD achieves up to 0.44 PSNR improvement across datasets with ~84.1% fewer FLOPs and ~32.6% smaller model size compared to ESRGAN baselines.
• In SR benchmarks, AGD achieves 110.9 GFLOPs and 1.8 MB memory with PSNRs of 31.79 (Set5), 28.36 (Set14), 27.41 (BSD100), and 25.55 (Urban100).