motion-blur-SR

Dependenices

• einops == 0.6.0
• ninja == 1.11.1
• pytorch == 2.0.1

Dataset Preparation

1. The GOPRO dataset is available for download at this link: https://drive.google.com/file/d/1H0PIXvJH4c40pk7ou6nAwoxuR4Qh_Sa2/view?usp=sharing.
2. Reference GFN uses the GOPRO dataset to produce training image pairs.
3. Modify the corresponding image path in the training or testing Python file.

Pre-trained Model

Please go to https://drive.google.com/file/d/1pKw_8xxeSqE4lzOPLSMlHCVGHvCLssqg/view?usp=sharing to download the pre-trained models, and put it into "./saved_info/dd_gan/motion-blur-sr/blur_sr_gopro/''.

Train & Evaluate

1. Modify the configuration files in the "./options" folder (such as super resolution, GPU usage, etc.).
2. Run training/evaluation code. The code is for training on 1 GPU.

Train

python3 blur_sr_up.py --exp blur_sr_gopro --num_channels 3 --num_channels_dae 128 --num_timesteps 4 --num_res_blocks 2 --batch_size 16 --num_epoch 75 --ngf 64 --nz 10 --z_emb_dim 256 --n_mlp 4 --embedding_type positional --use_ema --ema_decay 0.9999 --r1_gamma 0.02 --lr_d 1.25e-4 --lr_g 1.6e-4 --lazy_reg 15 --num_process_per_node 1 --ch_mult 1 2 3 4 --save_content --rec_loss --opt ./options/setting1/train/train_setting1_x4.yml

Evaluate

python3 test_ddgan_blur_sr.py --exp blur_sr_gopro --num_channels 3 --num_channels_dae 128 --num_timesteps 4 --num_res_blocks 2 --nz 100 --z_emb_dim 256 --n_mlp 4 --ch_mult 1 2 3 4 --epoch_id 18 --opt ./options/setting1/test/test_setting1_x4.yml

The evaluate results will be saved to the "./Result" folder of the current directory.

Some Results

Acknowledgments

This code is mainly built on SRDiff and DDGAN.