First use the TMT dataset to fine-tune the DATUM model 40k iter. Please put the data at data/TMT/

CUDA_VISIBLE_DEVICES=0,1 python train_DATUM_static.py \
--train_path ../../data/TMT/train_turb/ \
--val_path ../../data/TMT/test_turb/ -f ../DATUM_dynamic.pth \
--log_path ../experiments/turb_static_dataset/ --start_over \
--iters 40000 --batch-size 2 --patch-size 512 --learning-rate 0.0001 \
--num_frames 50 --num_gpus 2 --other_mod blur

Then, the trained weights are used for two inferences. Use 90 frames patch-size 512 at a time, use 100 frames patch-size 360 at a time.

python test_DATUM_static.py --loaded_frames 90 --valid_frames 1 \
--load  ../experiments/turb_static_dataset/checkpoints/model_40000.pth \
--val_path ../../data/final_data/ \
--result_path ../../data/result/experiment_turb_static_dataset_iter40k_fs5to95_out1/ \
--start_frame 5 --total_frames 100

python test_DATUM_static.py --loaded_frames 100 --valid_frames 1 \
--load  ../experiments/turb_static_dataset/checkpoints/model_40000.pth \
--val_path ../../data/final_data/ \
--result_path ../../data/result/experiment_turb_static_dataset_iter40k_ps360_fs100_out1/ \
--start_frame 0 --total_frames 100 --patch-size 360

Then use the multi_model_fusion.py script to perform a 7:3 weighted average of 100 frames at 360ps and 90 frames at 512ps.

python multi_model_fusion.py --img_folders data/result/experiment_turb_static_dataset_iter40k_ps360_fs100_out1/ \
data/result/experiment_turb_static_dataset_iter40k_fs5to95_out1/ \
--weights 0.7 0.3 --save_path data/result/model_fusion_73/

Finally, use MPRNet's deblurring to deblur the weighted average result.

python demo.py --input_dir ../data/resul/model_fusion_73/ \
--result_dir ../data/result/model_fusion_73_MPRNet_deblur/ --task Deblurring