This work is a modified fork of Erik Lindernoren's SRGAN implementation in the excellent PyTorch-GAN collection. (https://github.com/eriklindernoren/PyTorch-GAN)
This repository was created as part of a research experiment exploring the influence that the depth of the discriminator network has on the performance of the SRGAN model. Weights are included for models trained on the Linnaeus 5 dataset of 256x256 images, with discriminator networks containing B=8, 12, 16, 20 residual blocks.
Download the Linnaeus 5 dataset (Chaladze, G. Kalatozishvili L. 2017.) here: http://chaladze.com/l5/
- You will need to rename the image files and flatted the directory structure so that all images are located in the ./data/Linnaeus 5 256X256_train and ./data/Linnaeus 5 256X256_test/ directories, respectively.
The code has only been tested in a Linux environment. Use other operating systems at your own risk.
-
Ensure your dataset folder is located in the ./data/ directory
-
Launch the srgan.py script within your IDE of choice (eg. VSCode) OR Launch the srgan.py script via the command line with the following steps: 2.1 Navigate to the ./implementation/srgan/ directory 2.2 Execute the following command: python srgan.py
Recommended: Use this command to view a list of optional command line parameters to control program execution:
python srgan.py --help
-
Ensure your images are located in the ./data/testImages/ subdirectory.
-
Copy the desired pre-trained weights and training state data files to the ./saved_models/ directory
-
Launch the processImage.py script within your IDE of choice (eg. VSCode)
OR
Launch the processImage.py script via the command line with the following steps:
3.1 Navigate to the ./implementation/srgan/ directory
3.2 Execute the following command:
python processImage.py\- You may need to supply the location of your test images via command line arguments
Note: Use the following command to view a list of optional command line parameters to control program execution:
python processImage.py --help
| File | Description |
|---|---|
| /data/ | Image data directory |
| /images/ | Automatically created: Images are output here during training |
| /implementation/srgan/dataset.py | Helper functions for loading and using datasets |
| /implementation/srgan/models.py | SRGAN generator and discriminator network implementation |
| /implementation/srgan/processImage.py | Test script: Super-resolve an image using pre-trained weights |
| /implementation/srgan/srgan.py | Network training. This is the main program funcitonality |
| /implementation/srgan/validateGenerator.py | Test script: Evaluate generator loss (ie. content loss) only |
| /implementation/srgan/validateModel.py | Validation testing script. Called by srgan.py |
| /learned weights/8 residual blocks/ | Pre-trained network, for a generator with b=8 residual blocks |
| /learned weights/12 residual blocks/ | Pre-trained network, for a generator with b=12 residual blocks |
| /learned weights/16 residual blocks/ | Pre-trained network, for a generator with b=16 residual blocks |
| /learned weights/20 residual blocks/ | Pre-trained network, for a generator with b=20 residual blocks |
| /LICENCE | Licence file, as provided by Erik Linernoren |
| /README.md | This file |
| /saved_models/ | Automatically created: Saved weights/training states output here |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent