Introduction

This repo contains an unofficial implementation of pseudo-sr.

With this code, I've tested unpaired face SR (section 4.3 in the paper). You can get the dataset from here.

I am certain that you can configure the experiment 4.2 (track 4 of the NTIRE 2018 SR challenge) using this repo, provided that you have the NTIRE dataset and implement the unpaired LR-HR data paser.

Usage

First, configure the yaml file which is located at configs/faces.yaml. Set the root folder of face dataset to DATA.FOLDER.

To train:

CUDA_VISIBLE_DEVICES=2,3 python train.py configs/faces.yaml --port 12121

The --port option is only required for multi-gpu training. You can use a number between 49152 and 65535 for the port number.

Reproducibility

TBA

Reference

Maeda, Shunta. "Unpaired Image Super-Resolution using Pseudo-Supervision." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.