Co-Evolutionary Compression for Unpaired Image Translation

Abstract

Generative adversarial networks (GANs) have been successfully used for considerable computer vision tasks, especially the image-to-image translation. However, generators in these networks are of complicated architectures with large number of parameters and huge computational complexities. Existing methods are mainly designed for compressing and speeding-up deep neural networks in the classification task, and cannot be directly applied on GANs for image translation, due to their different objectives and training procedures. To this end, we develop a novel co-evolutionary approach for reducing their memory usage and FLOPs simultaneously. In practice, generators for two image domains are encoded as two populations and synergistically optimized for investigating the most important convolution filters iteratively. Fitness of each individual is calculated using the number of parameters, a discriminator-aware regularization, and the cycle consistency. Extensive experiments conducted on benchmark datasets demonstrate the effectiveness of the proposed method for obtaining compact and effective generators.

GAN Compression: Efficient Architectures for Interactive Conditional GANs

“…It can compress state-of-the-art conditional GANs by 5-21×, and reduce the model size by 4-33×, with only negligible degradation in the model performance. Specifically, our proposed method shows a clear advantage of CycleGAN compression compared to the previous Co-Evolution method [60]. We can reduce the computation of CycleGAN generator by 21.2×, which is 5× better compared to the previous CycleGANspecific method [60] while achieving a better FID by more than 30 ‡ .…”

“…The FID difference between the two protocols is small. The FIDs for the original model, Shu et al [60], and our compressed model are 65.48, 96.15, and 69.54 using their protocol.…”

“…We assume no knowledge of the original cGAN learning algorithm. Experiments show that our general-purpose method achieves 21.1× compression ratio (5× better than CycleGAN-specific method [60]) while retaining the FID of original models. Knowledge distillation.…”

“…Liu et al [45] later replaced the reinforcement learning by an evolutionary search algorithm. Recently, Shu et al [60] proposed co-evolutionary pruning for CycleGAN by modifying the original Cycle-GAN algorithm. This method is tailored for a particular algorithm.…”

CPGAN : An Efficient Architecture Designing for Text-to-Image Generative Adversarial Networks Based on Canonical Polyadic Decomposition

“…To this end, we propose an endto-end compression framework based on CPD. Compared to Shu et al [32] and Li et al [33], we do not need to pretrain GAN model. We design and train the compression model from scratch.…”