README_test.md

README

NOTE: This document details the inference of our approach on the challenge data for submission. To run our approach on your data, please modify the following scripts.

We provide our data on Dropbox:

• Validation set for Track 1
• Validation set for Track 2
• Validation set for Track 3
• Stage 1 model
• Stage 2 model
• FFmpeg

You can also download the models at Releases.

Data structure:

ntire-submission/
|-- stage1/
`-- stage2/

If you find this work helpful, please cite our workshop paper.

1. Data preparation

# fixme: assume we put the validation zip at /mnt/nfs1/qunliang/data/submission/
cd /mnt/nfs1/qunliang/data/submission/

unzip validation_track1.zip
unzip validation_track2.zip
unzip validation_track3.zip

# fixme: assume the ffmpeg is located at ./ffmpeg-4.3.1-amd64-static/ffmpeg
for idx in `seq -f '%03g' 1 15`; do mkdir Track1/$idx && ./ffmpeg-4.3.1-amd64-static/ffmpeg -i Track1/$idx.mkv Track1/$idx/f%3d.png; done
for idx in `seq -f '%03g' 1 15`; do mkdir Track2/$idx && ./ffmpeg-4.3.1-amd64-static/ffmpeg -i Track2/$idx.mkv Track2/$idx/f%3d.png; done
for idx in `seq -f '%03g' 1 15`; do mkdir Track3/$idx && ./ffmpeg-4.3.1-amd64-static/ffmpeg -i Track3/$idx.mkv Track3/$idx/f%3d.png; done

cd ntire-submission/stage1/
ln -s /mnt/nfs1/qunliang/data/submission data

2. Create environment

# assume we have cuda 10.2
conda create -n mmtest python=3.7 -y && conda activate mmtest
pip install mmcv-full==1.3.14 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.8.0/index.html

cd ntire-submission/ && conda env update -n mmtest --file conda_mmtest.yml

cd ntire-submission/stage1/ && python setup.py develop

The following codes are run with the mmtest environment.

3. Track 1

3.1. Generate x8tta lq data

cd ntire-submission/stage1/toolbox_x8tta/
python generate_x8tta_lq.py -inp_dir ../data/Track1 -out_dir ../data/Track1_x8tta

3.2. Run stage1 model

cd ntire-submission/stage1/toolbox_test/

# fixme: assume we have gpus 0, 1, 2, 3

python test.py -gpu 0 -inp_dir '../data/Track1_x8tta/origin' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/origin' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2 && \
python test.py -gpu 0 -inp_dir '../data/Track1_x8tta/r90' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/r90' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2

python test.py -gpu 1 -inp_dir '../data/Track1_x8tta/r180' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/r180' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2 && \
python test.py -gpu 1 -inp_dir '../data/Track1_x8tta/r270' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/r270' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2

python test.py -gpu 2 -inp_dir '../data/Track1_x8tta/flip' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/flip' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2 && \
python test.py -gpu 2 -inp_dir '../data/Track1_x8tta/flipr90' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/flipr90' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2

python test.py -gpu 3 -inp_dir '../data/Track1_x8tta/flipr180' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/flipr180' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2 && \
python test.py -gpu 3 -inp_dir '../data/Track1_x8tta/flipr270' -out_dir '../work_dirs/track1/results_x8tta_rmdup2/flipr270' -config_path '../configs/restorers/submission/flow5v2r6_ldv_ql_mse1e-6_track1.py' -model_path '../work_dirs/track1/iter_16000.pth' -if_rmdup2

3.3. Check PSNR vs. LQ and replace videos

cd ntire-submission/stage1/toolbox_test/

python check_psnr.py -lq_dir '../data/Track1_x8tta/origin' -enh_dir '../work_dirs/track1/results_x8tta_rmdup2/origin' -tab 'track1_model1'

Check ntire-submission/stage1/toolbox_test/log/track1_model1.log:

1. If this file is empty, go to Section 3.4.
2. If any video appears in this file, run the backup model:

cd ntire-submission/stage1/toolbox_test/

# fixme: assume we have gpus 0, 1, 2, 3

python test.py -gpu 0 -inp_dir '../data/Track1_x8tta/origin' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/origin' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2 && \
python test.py -gpu 0 -inp_dir '../data/Track1_x8tta/r90' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/r90' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2

python test.py -gpu 1 -inp_dir '../data/Track1_x8tta/r180' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/r180' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2 && \
python test.py -gpu 1 -inp_dir '../data/Track1_x8tta/r270' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/r270' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2

python test.py -gpu 2 -inp_dir '../data/Track1_x8tta/flip' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/flip' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2 && \
python test.py -gpu 2 -inp_dir '../data/Track1_x8tta/flipr90' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/flipr90' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2

python test.py -gpu 3 -inp_dir '../data/Track1_x8tta/flipr180' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/flipr180' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2 && \
python test.py -gpu 3 -inp_dir '../data/Track1_x8tta/flipr270' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_backup/flipr270' -config_path '../configs/restorers/submission/s1t1_iter_1000.py' -model_path '../work_dirs/track1/iter_1000.pth' -if_rmdup2

Then replace the videos at list ntire-submission/stage1/toolbox_test/log/track1_model1.log and in ntire-submission/stage1/work_dirs/track1/results_x8tta_rmdup2/ with those in ntire-submission/stage1/work_dirs/track1/results_x8tta_rmdup2_backup.

3.4. Generate image symbolic link dir for stage2

cd ntire-submission/stage2/toolbox_data/

To test full set:

python rename_and_ln.py -inp_dir '../../stage1/work_dirs/track1/results_x8tta_rmdup2' -out_dir '../work_dirs/track1/input_x8tta_rmdup2'

Or test online validation set (f010, f020, ...):

python rename_and_ln.py -inp_dir '../../stage1/work_dirs/track1/results_x8tta_rmdup2' -out_dir '../work_dirs/track1/input_x8tta_rmdup2_online10' -online

3.5. Run stage2 model

To test full set:

# fixme: assume we have gpus 0, 1, 2, 3
cd ntire-submission/stage2/

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/origin' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/origin' --gpus 0 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/r90' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/r90' --gpus 0 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/r180' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/r180' --gpus 1 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/r270' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/r270' --gpus 1 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/flip' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/flip' --gpus 2 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/flipr90' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/flipr90' --gpus 2 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/flipr180' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/flipr180' --gpus 3 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2/flipr270' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy/flipr270' --gpus 3 --save_npy True

Or test online validation set (f010, f020, ...):

# fixme: assume we have gpus 0, 1, 2, 3
cd ntire-submission/stage2/

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/origin' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/origin' --gpus 0 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/r90' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/r90' --gpus 0 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/r180' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/r180' --gpus 1 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/r270' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/r270' --gpus 1 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/flip' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/flip' --gpus 2 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/flipr90' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/flipr90' --gpus 2 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/flipr180' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/flipr180' --gpus 3 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track1/9000_G.pth' --folder_lq 'work_dirs/track1/input_x8tta_rmdup2_online10/flipr270' --folder_output 'work_dirs/track1/results_x8tta_rmdup2_npy_online10/flipr270' --gpus 3 --save_npy True

3.6. Merge x8tta npy files into images in one dir

To test full set:

cd ntire-submission/stage2/toolbox_data/
python merge_npy.py -inp_dir '../work_dirs/track1/results_x8tta_rmdup2_npy' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_npy/merge'

Or test online validation set (f010, f020, ...):

cd ntire-submission/stage2/toolbox_data/
python merge_npy.py -inp_dir '../work_dirs/track1/results_x8tta_rmdup2_npy_online10' -out_dir '../work_dirs/track1/results_x8tta_rmdup2_npy_online10/merge'

3.7. Collect images to each video dir

To test full set:

cd ntire-submission/stage2/toolbox_data/
python collect_and_submit.py -inp_dir '../work_dirs/track1/results_x8tta_rmdup2_npy/merge' -out_dir '../work_dirs/track1/results_final'

The final results are stored at: ntire-submission/stage2/work_dirs/track1/results_final

Or test online validation set (f010, f020, ...):

cd ntire-submission/stage2/toolbox_data/
python collect_and_submit.py -inp_dir '../work_dirs/track1/results_x8tta_rmdup2_npy_online10/merge' -out_dir '../work_dirs/track1/results_final_online10'

The final results are stored at: ntire-submission/stage2/work_dirs/track1/results_final_online10

4. Track 2

4.1. Generate x8tta lq data

cd ntire-submission/stage1/toolbox_x8tta/
python generate_x8tta_lq.py -inp_dir ../data/Track2 -out_dir ../data/Track2_x8tta

4.2. Run stage1 model

cd ntire-submission/stage1/toolbox_test/

# fixme: assume we have gpus 0, 1, 2, 3

python test.py -gpu 0 -inp_dir '../data/Track2_x8tta/origin' -out_dir '../work_dirs/track2/results_x8tta/origin' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth' && \
python test.py -gpu 0 -inp_dir '../data/Track2_x8tta/r90' -out_dir '../work_dirs/track2/results_x8tta/r90' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth'

python test.py -gpu 1 -inp_dir '../data/Track2_x8tta/r180' -out_dir '../work_dirs/track2/results_x8tta/r180' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth' && \
python test.py -gpu 1 -inp_dir '../data/Track2_x8tta/r270' -out_dir '../work_dirs/track2/results_x8tta/r270' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth'

python test.py -gpu 2 -inp_dir '../data/Track2_x8tta/flip' -out_dir '../work_dirs/track2/results_x8tta/flip' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth' && \
python test.py -gpu 2 -inp_dir '../data/Track2_x8tta/flipr90' -out_dir '../work_dirs/track2/results_x8tta/flipr90' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth'

python test.py -gpu 3 -inp_dir '../data/Track2_x8tta/flipr180' -out_dir '../work_dirs/track2/results_x8tta/flipr180' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth' && \
python test.py -gpu 3 -inp_dir '../data/Track2_x8tta/flipr270' -out_dir '../work_dirs/track2/results_x8tta/flipr270' -config_path '../configs/restorers/submission/flow5v2r5_track2_newdata_mse_ldv.py' -model_path '../work_dirs/track2/iter_3000.pth'

4.3. Generate image symbolic link dir for stage2

cd ntire-submission/stage2/toolbox_data/

To test full set:

python rename_and_ln.py -inp_dir '../../stage1/work_dirs/track2/results_x8tta' -out_dir '../work_dirs/track2/input_x8tta'

Or test online validation set (f010, f020, ...):

python rename_and_ln.py -inp_dir '../../stage1/work_dirs/track2/results_x8tta' -out_dir '../work_dirs/track2/input_x8tta_online10' -online

4.4. Run stage2 model

To test full set:

# fixme: assume we have gpus 0, 1, 2, 3
cd ntire-submission/stage2/

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/origin' --folder_output 'work_dirs/track2/results_x8tta_npy/origin' --gpus 0 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/r90' --folder_output 'work_dirs/track2/results_x8tta_npy/r90' --gpus 0 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/r180' --folder_output 'work_dirs/track2/results_x8tta_npy/r180' --gpus 1 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/r270' --folder_output 'work_dirs/track2/results_x8tta_npy/r270' --gpus 1 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/flip' --folder_output 'work_dirs/track2/results_x8tta_npy/flip' --gpus 2 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/flipr90' --folder_output 'work_dirs/track2/results_x8tta_npy/flipr90' --gpus 2 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/flipr180' --folder_output 'work_dirs/track2/results_x8tta_npy/flipr180' --gpus 3 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta/flipr270' --folder_output 'work_dirs/track2/results_x8tta_npy/flipr270' --gpus 3 --save_npy True

Or test online validation set (f010, f020, ...):

# fixme: assume we have gpus 0, 1, 2, 3
cd ntire-submission/stage2/

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/origin' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/origin' --gpus 0 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/r90' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/r90' --gpus 0 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/r180' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/r180' --gpus 1 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/r270' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/r270' --gpus 1 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/flip' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/flip' --gpus 2 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/flipr90' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/flipr90' --gpus 2 --save_npy True

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/flipr180' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/flipr180' --gpus 3 --save_npy True && \
python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track2/15000_G.pth' --folder_lq 'work_dirs/track2/input_x8tta_online10/flipr270' --folder_output 'work_dirs/track2/results_x8tta_npy_online10/flipr270' --gpus 3 --save_npy True

4.5. Merge x8tta npy files into images in one dir

To test full set:

cd ntire-submission/stage2/toolbox_data/
python merge_npy.py -inp_dir '../work_dirs/track2/results_x8tta_npy' -out_dir '../work_dirs/track2/results_x8tta_npy/merge'

Or test online validation set (f010, f020, ...):

cd ntire-submission/stage2/toolbox_data/
python merge_npy.py -inp_dir '../work_dirs/track2/results_x8tta_npy_online10' -out_dir '../work_dirs/track2/results_x8tta_npy_online10/merge'

4.6. Collect images to each video dir

To test full set:

cd ntire-submission/stage2/toolbox_data/
python collect_and_submit.py -inp_dir '../work_dirs/track2/results_x8tta_npy/merge' -out_dir '../work_dirs/track2/results_final'

The final results are stored at: ntire-submission/stage2/work_dirs/track2/results_final

Or test online validation set (f010, f020, ...):

cd ntire-submission/stage2/toolbox_data/
python collect_and_submit.py -inp_dir '../work_dirs/track2/results_x8tta_npy_online10/merge' -out_dir '../work_dirs/track2/results_final_online10'

The final results are stored at: ntire-submission/stage2/work_dirs/track2/results_final_online10

5. Track 3

5.1. Generate x8tta lq data

cd ntire-submission/stage1/toolbox_x8tta/
python generate_x8tta_lq.py -inp_dir ../data/Track3 -out_dir ../data/Track3_x8tta

5.2. Run stage1 modelx2 (model ensemble here)

Model 1:

cd ntire-submission/stage1/toolbox_test/

# fixme: assume we have gpus 0, 1, 2, 3

python test.py -gpu 0 -inp_dir '../data/Track3_x8tta/origin' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/origin' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/origin' -if_rmdup2 && \
python test.py -gpu 0 -inp_dir '../data/Track3_x8tta/r90' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/r90' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/r90' -if_rmdup2

python test.py -gpu 1 -inp_dir '../data/Track3_x8tta/r180' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/r180' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/r180' -if_rmdup2 && \
python test.py -gpu 1 -inp_dir '../data/Track3_x8tta/r270' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/r270' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/r270' -if_rmdup2

python test.py -gpu 2 -inp_dir '../data/Track3_x8tta/flip' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/flip' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/flip' -if_rmdup2 && \
python test.py -gpu 2 -inp_dir '../data/Track3_x8tta/flipr90' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/flipr90' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/flipr90' -if_rmdup2

python test.py -gpu 3 -inp_dir '../data/Track3_x8tta/flipr180' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/flipr180' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/flipr180' -if_rmdup2 && \
python test.py -gpu 3 -inp_dir '../data/Track3_x8tta/flipr270' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model1/flipr270' -config_path '../configs/restorers/submission/c128n25rec3_track3mse.py' -model_path '../work_dirs/track3/iter_11000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1/flipr270' -if_rmdup2

Merge:

python merge_npy.py -inp_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1_merge'

Model 2:

cd ntire-submission/stage1/toolbox_test/

# fixme: assume we have gpus 0, 1, 2, 3

python test.py -gpu 0 -inp_dir '../data/Track3_x8tta/origin' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/origin' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/origin' -if_rmdup2 && \
python test.py -gpu 0 -inp_dir '../data/Track3_x8tta/r90' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/r90' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/r90' -if_rmdup2

python test.py -gpu 1 -inp_dir '../data/Track3_x8tta/r180' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/r180' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/r180' -if_rmdup2 && \
python test.py -gpu 1 -inp_dir '../data/Track3_x8tta/r270' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/r270' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/r270' -if_rmdup2

python test.py -gpu 2 -inp_dir '../data/Track3_x8tta/flip' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/flip' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/flip' -if_rmdup2 && \
python test.py -gpu 2 -inp_dir '../data/Track3_x8tta/flipr90' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/flipr90' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/flipr90' -if_rmdup2

python test.py -gpu 3 -inp_dir '../data/Track3_x8tta/flipr180' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/flipr180' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/flipr180' -if_rmdup2 && \
python test.py -gpu 3 -inp_dir '../data/Track3_x8tta/flipr270' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_model2/flipr270' -config_path '../configs/restorers/submission/flow5v2r6_track3mse.py' -model_path '../work_dirs/track3/iter_8000.pth' -float -out_npy_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2/flipr270' -if_rmdup2

Merge:

python merge_npy.py -inp_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2_merge'

5.3. Model ensemble

cd ntire-submission/stage1/toolbox_test/

python check_psnr.py -lq_dir '../data/Track3_x8tta/origin' -enh_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2_merge/png' -tab 'track3_model1'
python check_psnr.py -lq_dir '../data/Track3_x8tta/origin' -enh_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2_merge/png' -tab 'track3_model2'

python model_ensemble.py --model1_result_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1_merge/npy' --model1_png_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model1_merge/png' --model1_lst_file 'log/track3_model1.log' --model2_result_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2_merge/npy' --model2_png_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_model2_merge/png' --model2_lst_file 'log/track3_model2.log' --save_dir '../work_dirs/track3/results_x8tta_rmdup2_model_ensemble'

5.4. Generate image symbolic link dir for stage2

cd ntire-submission/stage2/toolbox_data/

To test full set:

python rename_and_ln_track3.py -inp_dir '../../stage1/work_dirs/track3/results_x8tta_rmdup2_model_ensemble' -out_dir '../work_dirs/track3/input_x8tta_rmdup2'

Or test online validation set (f010, f020, ...):

python rename_and_ln_track3.py -inp_dir '../../stage1/work_dirs/track3/results_x8tta_rmdup2_model_ensemble' -out_dir '../work_dirs/track3/input_x8tta_rmdup2_online10' -online

5.5. Run stage2 model

To test full set:

cd ntire-submission/stage2/

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track3/15000_G.pth' --folder_lq 'work_dirs/track3/input_x8tta_rmdup2' --folder_output 'work_dirs/track3/results_x8tta_rmdup2_npy' --gpus 0 --save_npy True

Or test online validation set (f010, f020, ...):

# fixme: assume we have gpus 0, 1, 2, 3
cd ntire-submission/stage2/

python main_test_swinir.py --task color_dn --noise 50 --model_path 'work_dirs/track3/15000_G.pth' --folder_lq 'work_dirs/track3/input_x8tta_rmdup2_online10' --folder_output 'work_dirs/track3/results_x8tta_rmdup2_npy_online10' --gpus 0 --save_npy True

5.6. Convert npy files into images in one dir

To test full set:

cd ntire-submission/stage2/toolbox_data/
python merge_npy_track3.py -inp_dir '../work_dirs/track3/results_x8tta_rmdup2_npy' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_npy/merge'

Or test online validation set (f010, f020, ...):

cd ntire-submission/stage2/toolbox_data/
python merge_npy_track3.py -inp_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_online10' -out_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_online10/merge'

5.7. Collect images to each video dir

To test full set:

cd ntire-submission/stage2/toolbox_data/
python collect_and_submit.py -inp_dir '../work_dirs/track3/results_x8tta_rmdup2_npy/merge' -out_dir '../work_dirs/track3/results_final'

The final results are stored at: ntire-submission/stage2/work_dirs/track3/results_final

Or test online validation set (f010, f020, ...):

cd ntire-submission/stage2/toolbox_data/
python collect_and_submit.py -inp_dir '../work_dirs/track3/results_x8tta_rmdup2_npy_online10/merge' -out_dir '../work_dirs/track3/results_final_online10'

The final results are stored at: ntire-submission/stage2/work_dirs/track3/results_final_online10