AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks (2006.08198v2)

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.

• The paper introduces the AGD framework that employs AutoML and knowledge distillation to automatically compress GAN generators.
• It achieves competitive results in image translation and super-resolution by significantly reducing FLOPs and inference latency.
• The method offers a scalable, architecture-agnostic solution that facilitates practical GAN deployment on resource-constrained devices.

An Expert Analysis of "AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks"

The paper "AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks" addresses the nascent area of compressing Generative Adversarial Networks (GANs) using an AutoML approach. The significant contribution of this research is the AutoGAN-Distiller (AGD) framework that performs neural architecture search to create efficient compressed models of GAN generators, optimizing for both performance and resource constraints.

Motivation and Approach

The increasing demand for deploying GANs on resource-constrained devices has propelled interest in this area. GAN generators, like other deep models, are computationally intensive, which limits their deployment on mobile devices. Despite the potential benefits, GAN compression remains underexplored compared to classification and segmentation models. Existing GAN compression techniques often target specific architectures and loss functions, limiting their general applicability.

AutoGAN-Distiller leverages AutoML, driven by the success in deep model compression, to automate the search and compression of GANs. AGD operates independently of the GAN type or availability of trained discriminators. This framework uses a novel search space and knowledge distillation to guide the search process.

Technical Framework

AGD's architecture search spans across operator type and layer width within a sequential GAN generator architecture, balancing model size and performance. The devised search space includes essential building blocks such as convolution layers, residual blocks, and depthwise separable convolutions. The framework employs differentiable architecture search to identify optimally compressed GAN architectures. Furthermore, AGD incorporates knowledge distillation as a proxy task, eschewing reliance on trained discriminators.

AGD extends its applicability across diverse GAN tasks, exhibiting favorable outcomes in image translation and super-resolution tasks. The tasks evaluated include unpaired image translation with CycleGAN and super-resolution with ESRGAN. AGD's versatility and automation could potentially facilitate broader applications in mobile-based GAN deployments, emphasizing practical utility.

Experimental Outcomes

This framework was evaluated on tasks such as image translation (CycleGAN) and super resolution (ESRGAN). The numerical results are compelling, demonstrating that AGD-derived models achieve significant reductions in FLOPs and real-device inference latency, while maintaining competitive metrics such as FID and PSNR. For instance, in the unpaired image translation tasks, AGD outperformed state-of-the-art methods like CEC with lower resource utilization.

The visual comparisons align with quantitative findings, where compressed models produce outputs comparable to the original GAN models, retaining both fidelity and texture detail.

Implications and Future Directions

AutoGAN-Distiller underscores the potential of AutoML frameworks in achieving efficient GAN compression. This approach promises to bridge the gap between high-performance GAN models and their practical deployment on edge devices. The proposed method provides a scalable, automated solution, adaptable to various GAN architectures and tasks without requiring manual intervention.

Future research directions could expand on integrating more advanced operator types, improving search efficiency, and exploring joint optimization strategies with other model aspects like GAN discriminators. Additionally, real-world application testing on diverse mobile hardware could further validate and refine AGD’s capabilities, fostering new advancements in deploying AI on constrained environments.

In conclusion, "AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks" provides a significant contribution to the efficient application of GANs in practical scenarios, paving the way for further innovations in model compression and architecture search within machine learning.