HAT-L 모델

Reference

• HAT [Paper Link]
• HAT [Code Link]
• RealEsrgan [Code Link]
• HAT-Pretrained Model - check Training section
• AI-HUB Dataset [link]

Environment

• PyTorch >= 1.7
• BasicSR == 1.3.4.9

Installation

1. Clone Repo

git clone https://github.com/toomy0toons/yangjaeSR.git
cd yangjaeSR

2. Install depedencies

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

3. Troubleshooting when using docker container, install CV2 dependencies

apt-get install ffmpeg libsm6 libxext6 -y

Dataset Preperation

Data Preprocessing

1. Dacon Data download and unzip to data folder

data/
|-- README.md
|-- yangjaeSR 
|   |-- sample_submission.zip
|   |-- test
|   |   |-- lr
|   |-- test.csv
|   |-- train
|   |   |-- hr
|   |   `-- lr
|   `-- train.csv

2. Unzip sample_submission.zip under test/hr folder Inference 시 HR PAIR 가 있어야 작동하기 때문에, 더미로 사용합니다. 학습과 무관함.

data/yangjaeSR/test
|-- hr
|   |-- 20000.png
|   |-- 20001.png
|   |-- 20002.png
|   |-- 20003.png
|   |-- 20004.png
|   |-- 20005.png
|   |-- 20006.png
|   |-- 20007.png
|   |-- 20008.png
|   |-- 20009.png
|   |-- 20010.png
|   |-- 20011.png
|   |-- 20012.png
|   |-- 20013.png
|   |-- 20014.png
|   |-- 20015.png
|   |-- 20016.png
|   `-- 20017.png
`-- lr
    |-- 20000.png
    |-- 20001.png
    |-- 20002.png
    |-- 20003.png
    |-- 20004.png
    |-- 20005.png
    |-- 20006.png
    |-- 20007.png
    |-- 20008.png
    |-- 20009.png
    |-- 20010.png
    |-- 20011.png
    |-- 20012.png
    |-- 20013.png
    |-- 20014.png
    |-- 20015.png
    |-- 20016.png
    `-- 20017.png

3. run scripts/resize_image.py

yangjaeSR# python scripts/resize_image.py 
mkdir data/processed/train/hr ...
Extract:  11%|██████████████████▌                

AIHUB DATA[Optional]

AI 허브에서 데이터를 다운로드합니다.

• 야외 실제 촬영 한글 이미지
• 초해상화(Super Resolution 이미지)
• 노후 시설물 이미지

scripts/convert_png.py 를 이용하여 aihub data 를 2048 로 리사이징 한뒤, 잘라냅니다.

data/aihub-list 에서 사용된 데이터 목록을 볼 수 있습니다.

    # HR images
    opt['input_folder'] = 'aihub-sr/노후 시설물 이미지'
    opt['save_folder'] = 'aihub-processed/노후 시설물 이미지'
    opt['crop_size'] = 480
    opt['step'] = 360
    opt['thresh_size'] = 0
    extract_subimages(opt)

How To Test

• Refer to ./options/test for the configuration file of the model to be tested, and prepare the testing data and pretrained model.

• The pretrained models are available at Google Drive

• Download model to experiments/pretrained_models/dacon_submission.pth

• Then run the follwing codes to reproduce submission resulsts

python hat/test.py -opt options/test/yangjaeSR/HAT-L_Dacon_Submission.yml

The testing results will be saved in the ./results folder.

Submission

Training

• refer to options/train for config for each model runs
• download imagenet pretrained HAT-L model at

1. Download Pretrained HAT-L Model to experiments/pretrained_models/HAT-L_SRx4_ImageNet-pretrain.pth

• The pretrained models are available at Google Drive or Baidu Netdisk (access code: qyrl).

2. Configure Opt

# in options/train/train--- yml

batch_size_per_gpu: 4 <--- change this number

• Batch Size 4 needs ~40GB Gpu RAM

3. Run Train code

• Set GPU IDs for distributed training For example, distributed training in 4 GPU

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 hat/train.py \
-opt options/train/yangjaeSR/train_HAT-L_Dacon_Coarse.yml \
--auto_resume --launcher pytorch

• In case of containerized server, set NCCL sockets
• run containeer with net=host and ipc=host

NCCL_IB_DISABLE=1 NCCL_SOCKET_IFNAME=en,eth,em,bond \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 hat/train.py \
-opt options/train/yangjaeSR/train_HAT-L_Dacon_Coarse.yml \
--auto_resume --launcher pytorch

대회 세팅

Coarse -> Pretrain -> Submission 순서대로

1. Coarse 를 225000 iter 까지

NCCL_IB_DISABLE=1 NCCL_SOCKET_IFNAME=en,eth,em,bond \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 hat/train.py \
-opt options/train/yangjaeSR/train_HAT-L_Dacon_Coarse.yml \
--auto_resume --launcher pytorch

2. Pretrain 을 335000 iter 까지

NCCL_IB_DISABLE=1 NCCL_SOCKET_IFNAME=en,eth,em,bond \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 hat/train.py \
-opt options/train/yangjaeSR/train_HAT-L_Dacon_Pretrain.yml \
--auto_resume --launcher pytorch

3. Submission 185000 까지 (AIHUB 데이터 필요)

NCCL_IB_DISABLE=1 NCCL_SOCKET_IFNAME=en,eth,em,bond \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 hat/train.py \
-opt options/train/yangjaeSR/train_HAT-L_Dacon_Submission.yml \
--auto_resume --launcher pytorch

4. The training logs and weights will be saved in the ./experiments folder.

Results

The inference results will be saved under results/modelname/visualization folder

Contact