vidprocnet / gp-dip Goto Github PK

This repository contains the source code for the CVPR 2019 paper A Bayesian Perspective on the Deep Image Prior.

[Paper] [Supplementary] [arXiv] [Project page]

Our implementation is based on the code from Deep Image Prior [Ulyanov et al. CVPR 2018]. Refer to their project page for installation. (Dependencies: python = 3.6; pytorch = 0.4; numpy; scipy; matplotlib; scikit-image; jupyter; seaborn; _pickle)

Recommanded way to install:

conda create -n GP-DIP python=3.6 anaconda
source activate GP-DIP
cat /usr/local/cuda/version.txt # check out the CUDA version
conda install pytorch=0.4.1 cuda80 -c pytorch
conda install -c pytorch torchvision 

To run the GP_RBF_Inpainting.ipynb, gpytorch is required (Dependencies: python >= 3.6; pyTorch >= 1.0).

• 1D_toy_example.ipynb: priors and posterior with 1D convolutional networks.

• Image_Denoising.ipynb: compare the SGD variants and SGLD on image denoising task

• Image_Inpainting.ipynb: compare the SGD variants and SGLD on image inpainting task

• GP_RBF_Inpainting.ipynb: the Gaussian Process with RBF kernel for image inpainting.

• GP_DIP_Inpainting.ipynb: the Gaussian Process with DIP kernel for image inpainting.

Download the dataset here for the large-hole image inpainting experiments presented in our supplementary.

@inproceedings{Cheng_2019_CVPR,
	author = {Cheng, Zezhou and Gadelha, Matheus and Maji, Subhransu and Sheldon, Daniel},
	title = {A Bayesian Perspective on the Deep Image Prior},
	booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	year = {2019}
}