Super-Resolution (SR) Generative Adversarial Networks (GAN) methods for Fluid Dynamics (FD) images.
FD-SR is a project part of a Master's Thesis at FEUP partnered with LIACC (Artificial Intelligence and Computer Science Lab).
The Thesis is named Application of Novel Techniques in Super Resolution GANs for Fluid Dynamics.
The goal is to use super-resolution GAN techniques to enhance the quality of images obtained from Computational Fluid Dynamics (CFD) simulations and Experimental Fluid Dynamics.
Models explored include the ESRGAN, BSRGAN, Real-ESRGAN (Discriminator only), and the A-ESRGAN.
The hyperparameters and training configurations can be adjusted in the respective file. For example for the BSRGAN it is the bsrgan_config.py file.
This configuration file sets various parameters, including:
- Train, Development, Test datasets
- Degradation process parameters
- Random seed for maintaining reproducible results
- The device for training (default: GPU)
- Whether only the Y channel image data should be verified when evaluating the SR model
- Model architecture names
- Discriminator and generator configurations
- MLflow experiment details
- Training parameters like dataset address, batch size, number of epochs, learning rate, etc
- Testing parameters like directory for ground truth images, whether to save images and metrics, etc.
You can find the BSRGAN configuration file here.
To train the models, you need to navigate into the respective model's folder. For example, to train the BSRGAN model, follow these steps:
- Change your current directory to the BSRGAN folder:
cd BSRGAN - Modify the mode to
trainin the configuration script and choose if a pretrained model should be loaded (from a MLFlow run or from a file):python bsrgan_config.py - Run the training script:
python train_bsrgan.py
To test the models, you need to run the test script inside the respective model's folder. For example, to test the BSRGAN model, follow these steps:
- Change your current directory to the BSRGAN folder:
cd BSRGAN - Modify the mode to
testin the configuration script and provide the path to the model:python bsrgan_config.py - Run the testing script:
python test_bsrgan.py - Resulting metrics and images are presented in the respective MLFlow folder:
cd mlruns/{experiment_id}/{run_id}
For detailed instructions and more information about the project, see the comments in the provided scripts.