English | 简体中文

[TOC]

一、Introduction

An implementation of the SR models (PULSE) proposed in paper PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models with PaddlePaddle. PULSE traverses the high-resolution natural image manifold, searching for images that downscale to the original LR image.This is formalized through the “down-scaling loss,” which guides exploration through the latent space of a generative model.

Reference Project：

- https://github.com/adamian98/pulse

AI Studio Project：

- notebook task：https://aistudio.baidu.com/aistudio/projectdetail/2255411

二、Accuracy

2.1 Visual Comparison

​

2.2 NIQE

torch：average_NIQE=2.174

paddle：average_NIQE=2.132

pretrain GAN

Address：Baidu cloud disk Extraction code：f35u

三、Dataset

The input picture needs to be placed in the input folder, which is the same as [original reference code]（ https://github.com/adamian98/pulse ）Consistent, the input image is 16 × 16 celabahq face dataset:

https://pan.baidu.com/s/1wGbZ4UxPDpQj2gV_Zq37pQ Extraction code: mo0s

四、Environment

scipy==1.2.1
paddlepaddle==2.1.2
numpy==1.20.1

五、Quick Start

Step1: Clone

# clone this repo
git https://github.com/Martion-z/Paddle-PULSE.git
cd Paddle-PULSE-main

Step2: Load pretrain GAN model

Before running, you need to download the weight of the pre training network in advance. Place the weight file(styleGan.pdparams) in the models/cache folder. The model address:

Baidu cloud disk Extraction code: f35u

Step3: Run

python3 run.py

Just wait for the results. The output result (1024x1024) is stored in the output1024 folder.

六、Code Structure and Explanation

6.1 Code Structure

./Paddle-Pulse
|-- images               
       |--input          #the path of input
       |--output1024		#the path of output
|-- models               
       |--cache 		#store the weight of model
       |--loss			#loss
       |--utils			#tool API
       |--pulse.py		#network of PULSE
       |--stylegan_paddle.py	#stylegan
|-- utils                #public tool API
|-- run.py					#main
|-- README.md            
|-- README_cn.md

6.2 Parameter Explanation

Parameters	Default	Explanation
input_dir	input	The path of the input
output_dir	Output1024	The path of the output
batch_size	1	Batchsize
seed	0	Random seed
eps	2e-3	Optimizer
opt_name	adam	The class of Optimizer
steps	100	The number of iterations to find the best picture

七、Model Infomation

Field	Content
Author	皮蛋瘦肉周
Date	2021.08
Framework version	paddlepaddle 2.1.2
Application scenarios	Image Supper-Resolution
Supported hardware	CPU、GPU