A Computation-Efficient Solution for Acceleration of Generative Adversarial Network

Nowadays, Generative Adversarial Network (GAN) has been developed rapidly and has found variety of applications. Convolutional (CONV) and deconvolutional (DeCONV) layers are typical components of GANs. There existe a few researches on acceleration of deconvolution implementations. However, the previous works on accelerating DeCONV layers generally incur several issues, such as resource under-utilization, large memory requirements, and complex data access. In this paper, we propose an efficient method, namely fast transformation algorithm, to accelerate DeCONV layers. The algorithm improves the computational efficiency significantly and solves the problems of resource under-utilization and large memory requirements. Based on the algorithm, a reconfigurable computing core (RCC) for GANs is developed, which can support both convolutions and deconvolutions. Additionally, a reconfigurable architecture based on RCC is proposed to support both CONV and DeCONV layers. Moreover, a typical generative neural network, CycleGAN, is used to verify our design. The experimental results show that our design, when performing deconvolutions and convolutions in CycleGAN, can reach 352.50 GOPS under 200MHz working frequency on Intel Arria 10SX. In brief, the proposed design outperforms existing works significantly, particularly in terms of hardware efficiency.