khyeyoon / Video-Super-Resolution

EDVR in PyTorch

EDVR in PyTorch

paper : https://arxiv.org/pdf/1905.02716.pdf

review : https://khyeyoon.github.io/video%20super-resolution%20paper/EDVR/

• This code was written by modifying the code in the link below

https://github.com/xinntao/BasicSR

Requirements

• Python >= 3.7
• PyTorch >= 1.3
• NVIDIA GPU + CUDA

step

1. pip install -r requirements.txt
2. python setup.py develop

Train

Single GPU Training

PYTHONPATH="./:${PYTHONPATH}" \
CUDA_VISIBLE_DEVICES=0 \
python basicsr/train.py -opt options/train/SRResNet_SRGAN/train_MSRResNet_x4.yml

Distributed Training

You need to modify the yml file to match your data path

8 GPUs

PYTHONPATH="./:${PYTHONPATH}" \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 basicsr/train.py -opt options/train/EDVR/train_video_HD.yml --launcher pytorch

4 GPUs

PYTHONPATH="./:${PYTHONPATH}" \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 basicsr/train.py -opt options/train/EDVR/train_video_HD.yml --launcher pytorch

Self-collection dataset (cartoon, video)

About 30,000 frames were collected through videos in YouTube and used as a dataset

train/validation dataset format

dataroot
├── subfolder1
    ├── frame000
    ├── frame001
    ├── ...
├── subfolder1
    ├── frame000
    ├── frame001
    ├── ...
├── ...

Result of training with Self-collection dataset

You can check the result video through the link below