Skip to content

ryujaehun/EDSR-Mobile

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

73 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EDSR with depthwise separable convolution

About

This repository is the repository that applies the Depthwise Separable convolution to EDSR. The size of the feature map and the size of the residual block are significantly reduced to accommodate the number of parameters. Since the MSE loss is better than the L1 loss from the time when the parameter is reduced, MSE is adopted.

Dependencies

  • Python (Tested with 3.6)
  • PyTorch >= 0.4.0
  • numpy
  • imageio
  • matplotlib
  • tqdm

Code

Clone this repository into any place you want.

git clone https://github.com/ryujaehun/EDSR-PyTorch
cd EDSR-PyTorch

Quick start (Demo)

You can test our super-resolution algorithm with your own images. Place your images in test folder. (like test/<your_image>) We support png and jpeg files.

Run the script in code folder. Before you run the demo, please uncomment the appropriate line in demo.sh that you want to execute.

cd code       # You are now in */EDSR-PyTorch/code
sh demo.sh
Model Scale Dataset Parameters Loss Filter ResBlock **PSNR
EDSR 2 Set5 3.5K MSE 8 2 36.119 dB
EDSR 2 Set14 3.5K MSE 8 2 32.178 dB
EDSR 2 B100 3.5K MSE 8 2 31.084 dB
EDSR 2 Urban100 3.5K MSE 8 2 28.884 dB
EDSR 2 Set5 3.5K L1 8 2 35.222 dB
EDSR 2 Set14 3.5K L1 8 2 31.515 dB
EDSR 2 B100 3.5K L1 8 2 30.853 dB
EDSR 2 Urban100 3.5K L1 8 2 27.945 dB
EDSR 2 Set5 62.1K MSE 32 8 37.19 dB
EDSR 2 Set14 62.1K MSE 32 8 32.873 dB
EDSR 2 B100 62.1K MSE 32 8 31.694 dB
EDSR 2 Urban100 62.1K MSE 32 8 30.353 dB
EDSR 2 Set5 62.1K L1 32 8 37.246 dB
EDSR 2 Set14 62.1K L1 32 8 32.854 dB
EDSR 2 B100 62.1K L1 32 8 31.662 dB
EDSR 2 Urban100 62.1K L1 32 8 30.329 dB

*Baseline models are in experiment/model. Please download our final models from here (542MB) **We measured PSNR using DIV2K 0801 ~ 0900, RGB channels, without self-ensemble. (scale + 2) pixels from the image boundary are ignored.

You can evaluate your models with widely-used benchmark datasets:

Set5 - Bevilacqua et al. BMVC 2012,

Set14 - Zeyde et al. LNCS 2010,

B100 - Martin et al. ICCV 2001,

Urban100 - Huang et al. CVPR 2015.

For these datasets, we first convert the result images to YCbCr color space and evaluate PSNR on the Y channel only. You can download benchmark datasets (250MB). Set --dir_data <where_benchmark_folder_located> to evaluate the EDSR and MDSR with the benchmarks.

How to train EDSR and MDSR

We used DIV2K dataset to train our model. Please download it from here (7.1GB).

Unpack the tar file to any place you want. Then, change the dir_data argument in code/option.py to the place where DIV2K images are located.

We recommend you to pre-process the images before training. This step will decode all png files and save them as binaries. Use --ext sep_reset argument on your first run. You can skip the decoding part and use saved binaries with --ext sep argument.

If you have enough RAM (>= 32GB), you can use --ext bin argument to pack all DIV2K images in one binary file.

You can train EDSR and MDSR by yourself. All scripts are provided in the code/demo.sh. Note that EDSR (x3, x4) requires pre-trained EDSR (x2). You can ignore this constraint by removing --pre_train <x2 model> argument.

cd code       # You are now in */EDSR-PyTorch/code
sh demo.sh

TODO

  • visdom

Update log

  • Aug 19, 2018
  • Application of Depthwise Separable Convolution and Prediction of Trainng Time

Releases

No releases published

Packages

No packages published

Languages

  • Python 98.2%
  • MATLAB 1.6%
  • Shell 0.2%