This repository hosts Jupyter notebooks dedicated to training and inference of a Conditional Generative Adversarial Network (cGAN) based on the Pix2Pix architecture. The project focuses on generating standardized face-on images of Protoplanetary Disks from source images that are randomly oriented, using a dataset obtained from FARGO3D simulations.
-
Training Notebook:
- Implements a cGAN model with Pix2Pix architecture to transform randomly oriented Protoplanetary Disk images into face-on views.
- Includes attention mechanisms to enhance image generation quality by focusing on crucial features.
- Utilizes TensorFlow for model development and training.
-
Inference Notebook:
- Deploys the trained cGAN model to generate face-on images from new, randomly oriented Protoplanetary Disk images.
- Enables comparison between generated images and corresponding target images for evaluation.
-
Dataset Handling:
- Preprocesses and stores Protoplanetary Disk images in numpy arrays for efficient memory usage during training.
- Includes 21,000 images sourced from FARGO3D simulations, comprising 30 orientations per image.
-
Performance Evaluation:
- Incorporates periodic model checkpoints and performance visualizations using
summarize_performance
function. - Automatically organizes saved models and plots into uniquely named folders for each training session.
- Incorporates periodic model checkpoints and performance visualizations using
Dr. Sayantan Auddy
Dyutiman Santra