Skip to content

Latest commit

 

History

History
106 lines (95 loc) · 3.33 KB

README.md

File metadata and controls

106 lines (95 loc) · 3.33 KB
Raw

Efficient Deep Neural Network for Photo-realistic Image Super-Resolution

Namhyuk Ahn, Byungkon Kang, Kyung-Ah Sohn. [arXiv]

Requirements

Dataset

We use the same protocols of CARN, our prior work. Please see the details on this repo.

Test models on given directory

To test on given image directory,

$ python pcarn/inference_dir.py \
    --model pcarn \
    --ckpt ./checkpoints/<path>.pth \
    --data_root <dataset_root> \
    --scale [2|3|4] \
    --save_root <sample_dir_root>

More argument details are on the below section.

Test models on benchmark dataset

We provide the pretrained models in the checkpoints directory. To test the PCARN on benchmark dataset:

# For PCARN and PCARN (L1)
$ python pcarn/inference.py \
    --model pcarn \
    --ckpt ./checkpoints/<path>.pth \
    --data ./dataset/<dataset> \
    --scale [2|3|4] \
    --sample_dir <sample_dir>

# For PCARN-M and PCARN-M (L1)
$ python pcarn/inference.py \
    --model pcarn \
    --ckpt ./checkpoints/<path>.pth \
    --data ./dataset/<dataset> \
    --scale [2|3|4] \
    --sample_dir <sample_dir> \
    --mobile --group 4

We provide our results on four benchmark dataset (Set5, Set14, B100 and Urban100). Google Drive

Training models

Before train the PCARN(-M), models have to be pretrained with L1 loss.

# For PCARN (L1)
python pcarn/main.py \
    --model pcarn \
    --ckpt_dir ./checkpoints/<save_directory> \
    --batch_size 64 --patch_size 48 \
    --scale 0 --max_steps 600000 --decay 400000 \
    --memo <message_shown_in_logfile>

# For PCARN-M (L1)
python pcarn/main.py \
    --model pcarn \
    --ckpt_dir ./checkpoints/<save_directory> \
    --mobile --group 4 \
    --batch_size 64 --patch_size 48 \
    --scale 0 --max_steps 600000 --decay 400000 \
    --memo <message_shown_in_logfile>

Train the PCARN(-M) using below commands. Note that PerceptualSimilarity has to be ready to evaluate the model performance during training.

# For PCARN
python pcarn/main.py \
    --model pcarn \
    --ckpt_dir ./checkpoints/<save_directory> \
    --perceptual --msd \
    --pretrained_ckpt <pretrained_model_path> \
    --batch_size 32 --patch_size 48 \
    --scale 0 --max_steps 600000 --decay 400000 \
    --memo <message_shown_in_logfile>
    
# For PCARN-M
python pcarn/main.py \
    --model pcarn \
    --ckpt_dir ./checkpoints/<save_directory> \
    --perceptual --msd \
    --pretrained_ckpt <pretrained_model_path> \
    --mobile --group 4 \
    --batch_size 32 --patch_size 48 \
    --scale 0 --max_steps 600000 --decay 400000 \
    --memo <message_shown_in_logfile>

Results

Citation

@article{ahn2019efficient,
  title={Efficient Deep Neural Network for Photo-realistic Image Super-Resolution},
  author={Ahn, Namhyuk and Kang, Byungkon and Sohn, Kyung-Ah},
  journal={arXiv preprint arXiv:1903.02240},
  year={2019}
}