A pytorch implementation of "Fast and Accurate Image Super Resolution by Deep CNN with Skip Connection and Network in Network", a deep learning based Single-Image Super-Resolution (SISR) model. https://arxiv.org/abs/1707.05425
As output of tree -L 3 -I "*.pyc|*cache*|*init*":
├── checkpoints
├── data
│ ├── eval # evaluation data (no augmentation)
│ │ ├── bsd100
│ │ ├── set14
│ │ └── set5
│ └── train # training dataset (no augmentation)
│ ├── bsd200
│ └── yang91
├── dcscn
│ ├── data_utils # data loading and augmentation
│ │ ├── batcher.py
│ │ └── data_loader.py
│ ├── net.py # model definition
│ └── training # training & helpers
│ ├── checkpointer.py
│ ├── metrics.py
│ ├── tf_logger.py
│ └── trainer.py
├── Dockerfile # Dockerfile
├── entrypoint.sh # entrypoint script
├── logs # tensorboard logs
├── README.md
├── requirements.txt
├── setup.cfg
├── tests # python unit tests
│ └── test_checkpointer.py
└── train.py # training entry point
tqdm==4.28.1
matplotlib==2.2.3
numpy==1.13.0
scikit_image==0.13.1
Pillow==6.0.0
ai_utils==1.1.3
coloredlogs==10.0
torch==1.1.0
torchsummary==1.5.1
torchvision==0.2.2.post3
Basic training of a model (default configuration)
python train.py docker build . -f Dockerfile -t dcscndocker run -it \
-v <project-root-dir>/checkpoints:/super-resolution/app/checkpoints \
-v <project-root-dir>/data:/super-resolution/app/data \
-v <project-root-dir>/logs:/super-resolution/app/logs \
dcscn:latest run /bin/bash- Add typing all around the repo when appropiate
- Verify training achieves paper described performance
- Generalise trainer into a better general purpose package
- Populate README
