Abstract
Although the demands of deploying deep models to the edge devices are proliferating, it is a challenging problem to compress image-to-image translation models based on Generative Adversarial Networks (GANs). In particular, most of the compression approaches do not apply to the GAN models. In this paper, we revisit weight pruning approaches for GANs and theoretically build a novel perturbation model to analyze the effect of pruning certain weights for the instance normalization(IN)-based image-to-image translation GAN models. Furthermore, we develop a new training framework by imposing perturbation-bound-induced pruning loss. In the experimental analysis, we observe that the former pruning approaches do wrongly prune the channels with high visual impacts. We then depict the effectiveness of the proposed model by conducting both on-training pruning and zero-shot pruning of current state-of-the-art models. Specifically, we compress the CycleGAN[51] model using the horse2zebra dataset. In the on-training pruning task, we achieve x5.31 and x5.44 compression ratio to the original model in terms of FLOPs and Memory consumption, respectively; In the zero-shot pruning tasks, we obtain a decrease of 4.94 in the FID score compared to the best model provided in OMGD[39], both with a negligible decrease in output visual quality.Video
Citation
@inproceedings{Li_2022_BMVC,
author = {Yiyong Li and Zhun Sun and Chao Li},
title = {Are we pruning the correct channels in image-to-image translation models?},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0551.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by