Exploring Generative Image Modeling with Block PixelCNNs

Generative image modeling is one of the central but extremely challenging tasks in the modern unsupervised learning. There are numerous potential applications of probabilistic density models which are interesting both from theoretical and practical point of view. Block Pixel Convolutional Neural Networks, or simply Block PixelCNNs, proposed in this thesis combine approaches of Multiscale PixelCNNs and original PixelCNNs and fill the gap between them with the purpose to explore the impact made by different group structuring of the image pixels. Resulting model demonstrates competitive log-likelihood scores compared to other state-of-the-art generative models on CIFAR-10 and ImageNet datasets. Qualitative performance of different Block PixelCNNs is explored for the upscaling task on CIFAR-10 dataset.

Recommended citation: E. Iakovleva. "Exploring Generative Image Modeling with Block PixelCNNs." MSc Thesis at Grenoble INP, 2018.
Download Paper