https://github.com/loseall/ntire2019-competition-test
Test code for NTIRE 2019. Two tracks here: RSR and DN(sRGB)
https://github.com/loseall/ntire2019-competition-test
ntire
Last synced: 2 months ago
JSON representation
Test code for NTIRE 2019. Two tracks here: RSR and DN(sRGB)
- Host: GitHub
- URL: https://github.com/loseall/ntire2019-competition-test
- Owner: LoSealL
- Created: 2019-03-18T14:23:16.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2019-03-24T12:33:16.000Z (about 6 years ago)
- Last Synced: 2025-02-03T12:51:36.677Z (4 months ago)
- Topics: ntire
- Language: Shell
- Size: 13.7 KB
- Stars: 2
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
## One-click script
Make sure you are in the root folder. You should also have`python>=3.6`, `tensorflow(-gpu)>=1.12.0` and `pytorch>=1.0.0` installed manually.
Before running `one_click_ntire19_rsr.sh` for Real Image Super Resolution, set two paths: `RSR_TEST_DIR` for testing images and `RSR_SAVE_DIR` for saving results.
*It's recommended to use absolute path.*
```bash
RSR_TEST_DIR=/bla/bla/bla
RSR_SAVE_DIR=/bli/bli/bli
. one_click_ntire19_rsr.sh
```
Before running `one_click_ntire19_drn.sh` for sRGB Image Denoising, set two paths: `DRN_TEST_MAT` for testing mat file and `DRN_SAVE_DIR` for saving results.
*It's recommended to use absolute path.*
```bash
DRN_TEST_MAT=/bla/bla/bla/BenchmarkNoisyBlocksSrgb.mat
DRN_SAVE_DIR=/bli/bli/bli
. one_click_ntire19_drn.sh
```
You can also do it step-by-step as follows.
## Step by step reproduce instructions
1. Install the whole VSR package and its requirements:
```bash
git clone https://github.com/LoSealL/VideoSuperResolution -b ntire_2019 && cd VideoSuperResolution
pip install -e .
```
Note that you should pre-install `tensorflow` and `pytorch`.
2. Download the pre-trained model:
**make sure you are in the root folder.*
For Real Image Super-Resolution
```bash
python prepare_data.py --filter=rsr -q
```
For sRGB Real Image Denoising (Track #2: sRGB)
```bash
python prepare_data.py --filter=drn -q
```
Model url for manually download:
- [rsr](https://github.com/LoSealL/Model/releases/download/crdn/rsr.zip): https://github.com/LoSealL/Model/releases/download/crdn/rsr.zip
- [drn](https://github.com/LoSealL/Model/releases/download/mldn/drn.zip): https://github.com/LoSealL/Model/releases/download/mldn/drn.zip
3. Prepare testing data:
**make sure you are in the root folder.*
For RSR:
You need to crop images into small patches by:
```bash
python VSR/Tools/DataProcessing/NTIRE19RSR.py --ref_dir=path/to/test/data/folder --patch_size=768 --stride=760 --save_dir=path/to/saving/folder
```
For sRGB Denoising:
You need to convert .MAT file to png images by:
```bash
python VSR/Tools/DataProcessing/NTIRE19Denoise.py --validation=path/to/.MAT --save_dir=path/to/saving/folder
```
4. Predicting
**make sure you are in the root folder.*
For RSR:
Entering VSRTorch folder
```bash
cd VSRTorch
python eval.py rsr --cuda -t=/path/to/divided/test/images/folder --pth=../Results/rsr/save/rsr_ep2000.pth --ensemble
```
The output will be saved in `../Results/rsr/`. To combine them back together:
```bash
cd ..
python VSR/Tools/DataProcessing/NTIRE19RSR.py --ref_dir=path/to/test/data/folder --patch_size=768 --stride=760 --results=Results/rsr// --save_dir=path/to/saving/folder
```
Where `--ref_dir` should keep the same as the folder in step 3, it's a reference to know how to combine patches. `--patch_size` and `--stride` should also keep the same.
For sRGB Denoising:
Entering VSRTorch folder
```bash
cd VSRTorch
python eval.py drn --cuda -t=/path/to/divided/test/images/folder --pth=../Results/drn/save/drn_ep2000.pth --output_index=0 --ensemble
```
The output will be saved in `../Results/drn/`. To pack them into mat file:
```bash
cd ..
python VSR/Tools/DataProcessing/NTIRE19Denoise.py --results=Results/drn/ --save_dir=path/to/saving/folder
```
*If OOM happened, try not to enable `--cuda` flag.