Show simple item record

dc.contributor.authorLi, Muyang
dc.contributor.authorLin, Ji
dc.contributor.authorDing, Yaoyao
dc.contributor.authorLiu, Zhijian
dc.contributor.authorHan, Song
dc.date.accessioned2021-01-19T17:04:47Z
dc.date.available2021-01-19T17:04:47Z
dc.date.issued2020-06
dc.identifier.issn1063-6919
dc.identifier.urihttps://hdl.handle.net/1721.1/129446
dc.description.abstractConditional Generative Adversarial Networks (cGANs) have enabled controllable image synthesis for many computer vision and graphics applications. However, recent cGANs are 1-2 orders of magnitude more computationally-intensive than modern recognition CNNs. For example, GauGAN consumes 281G MACs per image, compared to 0.44G MACs for MobileNet-v3, making it difficult for interactive deployment. In this work, we propose a general-purpose compression framework for reducing the inference time and model size of the generator in cGANs. Directly applying existing CNNs compression methods yields poor performance due to the difficulty of GAN training and the differences in generator architectures. We address these challenges in two ways. First, to stabilize the GAN training, we transfer knowledge of multiple intermediate representations of the original model to its compressed model, and unify unpaired and paired learning. Second, instead of reusing existing CNN designs, our method automatically finds efficient architectures via neural architecture search (NAS). To accelerate the search process, we decouple the model training and architecture search via weight sharing. Experiments demonstrate the effectiveness of our method across different supervision settings (paired and unpaired), model architectures, and learning methods (e.g., pix2pix, GauGAN, CycleGAN). Without losing image quality, we reduce the computation of CycleGAN by more than 20x and GauGAN by 9x, paving the way for interactive image synthesis. The code and demo are publicly available.en_US
dc.description.sponsorshipNational Science Foundation (U.S.). Career (Award 1943349)en_US
dc.language.isoen
dc.publisherInstitute of Electrical and Electronics Engineers (IEEE)en_US
dc.relation.isversionof10.1109/CVPR42600.2020.00533en_US
dc.rightsCreative Commons Attribution-Noncommercial-Share Alikeen_US
dc.rights.urihttp://creativecommons.org/licenses/by-nc-sa/4.0/en_US
dc.sourcearXiven_US
dc.titleGAN Compression: Efficient Architectures for Interactive Conditional GANsen_US
dc.typeArticleen_US
dc.identifier.citationLi, Muyang et al. “GAN Compression: Efficient Architectures for Interactive Conditional GANs.” Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2020 (June 2020) © 2020 The Author(s)en_US
dc.contributor.departmentMassachusetts Institute of Technology. Department of Electrical Engineering and Computer Scienceen_US
dc.relation.journalProceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognitionen_US
dc.eprint.versionAuthor's final manuscripten_US
dc.type.urihttp://purl.org/eprint/type/ConferencePaperen_US
eprint.statushttp://purl.org/eprint/status/NonPeerRevieweden_US
dc.date.updated2020-12-17T16:34:00Z
dspace.orderedauthorsLi, M; Lin, J; Ding, Y; Liu, Z; Zhu, J-Y; Han, Sen_US
dspace.date.submission2020-12-17T16:34:07Z
mit.journal.volume2020en_US
mit.licenseOPEN_ACCESS_POLICY


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record