基于风格的生成对抗网络的进化道修剪
Yixia Zhang1, Ferrante Neri1,2, Xilu Wang1
1School of Computer Science and Electronic Engineering, University of Surrey, Guildford, United Kingdom.
International journal of neural systems
|September 30, 2025
概括
我们开发了用于StyleGANs (ECP-StyleGANs) 的进化通道修剪来压缩生成模型. 这种方法显著降低了StyleGANs的计算需求,使资源有限的设备能够高效地进行图像合成.
科学领域:
- 计算机视觉 计算机视觉
- 人工智能的人工智能
- 机器学习 机器学习
背景情况:
- 生成对抗网络 (GAN),特别是StyleGAN和StyleGAN2,在高质量的图像合成方面表现出色.
- 大型模型尺寸和高计算成本 (FLOP) 限制了GAN在边缘设备和移动平台上的部署.
研究的目的:
- 为StyleGANs (ECP-StyleGANs) 提出进化道修剪,这是压缩StyleGAN和StyleGAN2的算法.
- 为了保持具有竞争力的图像质量,同时减少实时应用的模型复杂性.
主要方法:
- 使用进化算法来代地完善二进制面具用于卷积通道修剪.
- 采用健身功能,平衡模型复杂性和生成质量.
- 编码修剪配置并应用选择,交叉和突变操作.
主要成果:
- 实现了 StyleGAN 和 StyleGAN2 模型的 FLOP 和参数大约减少 4 倍.
- 与未修剪的模型相比,Fréchet Inception Distance (FID) 仅略有增加,保持了视觉准确性.
- 证明了ECP-StyleGANs在压缩资源有限的环境中的GANs的有效性.
结论:
- ECP-StyleGANs提供了一种有效的方法来压缩像StyleGAN这样的大型生成模型.
- 该研究将生成人工智能修剪作为一个多目标优化任务,平衡效率和质量.
- 这项研究有助于在边缘设备和资源有限的环境中部署先进的GAN.
相关概念视频
01:20Evolutionary Psychology
939
Evolutionary psychology explores the origins of human behavior and mental processes by framing them within the context of natural selection, a theory famously propounded by Charles Darwin. This field asserts that many behaviors common across human societies — ranging from instinctive fear reactions to complex social interactions — arose as evolutionary adaptations. These adaptations enhanced the survival and reproductive success of our ancestors, thereby becoming embedded in the...
939
02:05Gene Evolution - Fast or Slow?
8.0K
The genomes of eukaryotes are punctuated by long stretches of sequence which do not code for proteins or RNAs. Although some of these regions do contain crucial regulatory sequences, the vast majority of this DNA serves no known function. Typically, these regions of the genome are the ones in which the fastest change, in evolutionary terms, is observed, because there is typically little to no selection pressure acting on these regions to preserve their sequences.
In contrast, regions which code...
In contrast, regions which code...
8.0K
02:05Gene Evolution - Fast or Slow?
3.4K
3.4K
01:18Reducing Line Loss
361
In a three-phase circuit, line loss is an indicator of energy dissipated as heat due to the resistance of transmission lines. To address this, incorporating transformers into the system—a step-up transformer at the source and a step-down transformer at the load—is a strategic solution. Two three-phase transformers are introduced to improve this.
With a step-up transformer at the source, the voltage is increased, thereby reducing the current in the transmission lines since power loss in...
With a step-up transformer at the source, the voltage is increased, thereby reducing the current in the transmission lines since power loss in...
361
01:15Source Transformation
11.1K
Source transformation is a fundamental technique employed in circuit analysis, offering a valuable tool for simplifying complex electrical circuits. This technique involves the replacement of either a voltage source in series with a resistor by a current source in parallel with a resistor, or vice versa. The key concept here is that when the original sources are deactivated (turned off), the equivalent resistance at the circuit's end terminals remains the same.
It is essential to note that when...
It is essential to note that when...
11.1K
03:33Genetic Drift
42.9K
Natural selection—probably the most well-known evolutionary mechanism—increases the prevalence of traits that enhance survival and reproduction. However, evolution does not merely propagate favorable traits, nor does it always benefit populations.
42.9K