Visible to the public Collaborative Distillation for Ultra-Resolution Universal Style Transfer

TitleCollaborative Distillation for Ultra-Resolution Universal Style Transfer
Publication TypeConference Paper
Year of Publication2020
AuthorsWang, H., Li, Y., Wang, Y., Hu, H., Yang, M.-H.
Conference Name2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Date Publishedjun
KeywordsCollaboration, collaborative distillation, compressed models, convolutional filters, convolutional neural nets, data compression, Decoding, deep convolutional neural network models, encoder-decoder based neural style transfer, encoder-decoder pairs, encoding, exclusive collaborative relationship, feature size mismatch, graphics processing units, Image coding, Image reconstruction, Image resolution, knowledge distillation method, Knowledge engineering, learning (artificial intelligence), leverage rich representations, linear embedding loss, neural style transfer, optimisation, Predictive Metrics, pubcrawl, Resiliency, Scalability, student network, style transfer models, Task Analysis, trained models, ultra-resolution images, ultra-resolution universal style transfer, universal style transfer, universal style transfer methods, VGG-19
AbstractUniversal style transfer methods typically leverage rich representations from deep Convolutional Neural Network (CNN) models (e.g., VGG-19) pre-trained on large collections of images. Despite the effectiveness, its application is heavily constrained by the large model size to handle ultra-resolution images given limited memory. In this work, we present a new knowledge distillation method (named Collaborative Distillation) for encoder-decoder based neural style transfer to reduce the convolutional filters. The main idea is underpinned by a finding that the encoder-decoder pairs construct an exclusive collaborative relationship, which is regarded as a new kind of knowledge for style transfer models. Moreover, to overcome the feature size mismatch when applying collaborative distillation, a linear embedding loss is introduced to drive the student network to learn a linear embedding of the teacher's features. Extensive experiments show the effectiveness of our method when applied to different universal style transfer approaches (WCT and AdaIN), even if the model size is reduced by 15.5 times. Especially, on WCT with the compressed models, we achieve ultra-resolution (over 40 megapixels) universal style transfer on a 12GB GPU for the first time. Further experiments on optimization-based stylization scheme show the generality of our algorithm on different stylization paradigms. Our code and trained models are available at https://github.com/mingsun-tse/collaborative-distillation.
DOI10.1109/CVPR42600.2020.00193
Citation Keywang_collaborative_2020