Single image super resolution using GANs

This is a python framwork for training, evaluating and comparing GAN models with different loss functions and hyperparameters.

The goal of the repo and framework is to be able to reproduce the results and methods presented in SISR papers such as SRGAN and Best-Buddy GAN, and to try out how different loss functions affect a GANs performance.

---

Usage

Initialization

1. Clone the repo

git clone https://github.com/SebastianBitsch/SRGAN-ST.git

2. Install dependencies

pip install -r requirements.txt

3. Pepare data

   1. Training data
      Download a training dataset containing high-res images. We recommend either the DIV2K or Flickr2K dataset for comparable results. Place the dataset in /data/

   2. Evaluation data
      Download one or more validation datasets. Set5, Set14, Urban100 or BSDS100 are all good options. Place the dataset(s) in /data/

   3. Make train dataset
      Create a dataset of smaller equal sized images from the given training dataset by running the following.

       python data-prep/prepare_dataset.py --input_dir=/data/original/ --output_dir=/data/train/ --output_size=96
      

      See prepare_dataset.py -h for more options.

   4. Config
      Update config.py to reflect the locations of the dataset(s). In particular update DATA.TRAIN_GT_IMAGES_DIR, DATA.TEST_GT_IMAGES_DIR and DATA.TEST_LR_IMAGES_DIR

Training

Train RResnet
To train a resnet / warmup the generator in GAN, edit relevant training parameters in the config.py - or keep the default, and run:

python warmup.py

Train GAN
To train a GAN edit any relevant parameters in the config.py file. To initialize the generator with the weights after warmup, set MODEL.G_CONTINUE_FROM_WARMUP to true and set the path to the weights in MODEL.G_WARMUP_WEIGHTS. To start training run:

python train.py

Evaluation

Train RResnet
To evalute a trained model (GAN or resnet), set EXP.NAME in the config to the name of the model you want to evaluate, then run:

python validate.py

See python validate.py -h for more options

---

Acknowledgement

This repo is loosely based on the work of Github user Lornatangs repo for SISR using SRGAN.

Notes

• Moduler: link

  • module load python3/3.10.7
  • module load cuda/11.7

• Launch tensorboard fra ssh i vsc:

  • tensorboard --logdir=tensorboard/ --host localhost --port 3000
  • fuser -k 3000/tcp slå processen ned hvis den allerede kører

• Kopier filer til og fra hpc link

• Find ud hvor meget space er brugt på hpc

  • getquota_zhome.sh
  • getquota_work3.sh
  • du -h --max-depth=1 --apparent $HOME

• How to use venv in notebooks link

• Kopier filer fra hpc til pc

  • scp -r -i /Users/sebastianbitsch/.ssh/gbar s204163@transfer.gbar.dtu.dk:SRGAN-ST/samples/logs /Users/sebastianbitsch/Desktop/

• Scratch ligger på /work3/