All the python packages required are given in requirements.txt. Run the below code to install them all.
pip install -r requirements.txtThe original dataset is found on Kaggle and can be downloaded (here). It contains 5000 images each of elevation height maps, terrain maps and segmentations maps.
Data preprocessing is done here. Images are deleted if their size is less than 305KB, or if they contain more than 50% pixels of extremely dark color (almost black). There is also code for resizing images to test the SSIM scores. The final dataset used is dataset-1.
The DCGAN code is located in the DCGAN file. It consists of importing the packages, initializing hyperparameters, along with different architectures used, most of which have been commented out, followed by adding loss functions and then training the DCGAN. The models trained are saved using the TensorFlow's save() method.
There are also code cells for loading the model and generating images using the loaded model for performance testing.
The VAE code is located in the VAE file. Follows similar pattern of DCGAN.
The code for perlin noise is located here. perlin_noise library has been installed which is referenced in the disseration report. Multiple images are generated and plotted which are then used to create 3d models of terrain and for some comparative analysis.
The code for SSIM testing can be found here. The generated images from VAE and DCGAN are located in the test folders (in VAE_generated_images and GAN_generated_images) which are compared to the resized images located here from original dataset located here