RADiff: Controllable Diffusion Models for Radio Astronomical Map Generation

Renato Sortino, Thomas Cecconello, Andrea DeMarco, Giuseppe Fiameni, Andrea Pilzer, Andrew M. Hopkins, Daniel Magro, Simone Riggi, Eva Sciacca, Adriano Ingallinera, Cristobal Bordiu, Filomena Bufano, Concetto Spampinato

Table of Contents

• Overview
• Requirements
• Dataset
• Pretrained Models
• Usage
• Results
• References

Overview

This repository contains the official PyTorch implementation of the paper titled "RADiff: Controllable Diffusion Models for Radio Astronomical Maps Generation". In the paper, we propose an approach based on diffusion models to augment small datasets for training segmentation methods in radio astronomy.

Requirements

1. Clone the repository:

git clone https://github.com/SKA-INAF/radiff.git
cd radiff

2. Create a conda environment and install the required dependencies:

conda env create -f environment.yaml
conda activate radiff

Dataset

The data used to run the experiments for this paper is under privacy constraints and we are not allowed to publish it. However, the model can be trained on any collection of radio astronomical images in FITS format if it presents the folder structure described in this section.

Data should be put in the data folder. This folder should contain all the images and annotations, and a text file where each line defines the path of the images, one for train (train.txt) and one for validation (val.txt). The structure of the folder should be as follows:

data
├── train.txt
├── val.txt
├── images
│   ├── img0001.fits
│   ├── img0002.fits
│   ├── img0003.fits
│   └── ...
└── annotations
    ├── mask_img0001.json
    ├── mask_img0001_obj1.fits
    ├── mask_img0002.json
    ├── mask_img0002_obj1.fits
    ├── mask_img0002_obj2.fits
    ├── mask_img0002_obj3.fits
    ├── mask_img0003.json
    ├── mask_img0003_obj1.fits
    ├── mask_img0003_obj2.fits
    ├── ...
    └──

The images folder contains the 128x128 images to be used for training the model while the annotations folder contains information about each image in JSON format (class, bbox coordinates, flux intensity). Additionally, each FITS file contains the segmentation mask of each object. Note that this folder structure is adapted to the DataLoader in this implementation but this can be adapted to another file structure.

Pretrained Models

Datset	Task	Model	{METRIC1}	{METRIC2}	Link	Comments
{DATASET_NAME}	{TASK_NAME}	{MODEL_NAME}	5.11	3.29	{WEIGHTS_LINK}

Usage

Inference

The implementation supports two inference modes:

1. CLI inference, allowing to iterate over a folder of masks and generate an image for each one
2. Interactive inference, more user-friendly but supporting only one mask at a time

CLI inference

TODO

Interactive Interface

Run the interface using Gradio

python gradio_inference.py

Train on custom datasets

Train the Autoencoder

The first step is to train the autoencoder to reconstruct the images and prepare the latent space for the diffusion model. To do so, run the following command and specify the configuration file for its architecture:

python scripts/train_ae.py \ 
--ae-config vae-f4.yaml \
--run-dir {OUTPUT_DIR} \ 
--run-name {RUN_NAME} \
--on_wandb

Train the diffusion model

Once the autoencoder is capable of reconstructing the images, we can train the diffusion model. To do so, run the following command by specifying, as done earlier, the configuration file:

python scripts/train_ldm.py \ 
--ae-config vae-f4.yaml \
--ae-ckpt weights/autoencoder/vae-f4.pt \
--unet-config unet-cldm-mask.yaml \
--dataset train.txt \
--run-dir {OUTPUT_DIR} \ 
--run-name {RUN_NAME} \
--on_wandb

Results

TODO

BibTeX

@article{sortino2023radiff,
  title={RADiff: Controllable Diffusion Models for Radio Astronomical Maps Generation},
  author={Sortino, Renato and Cecconello, Thomas and DeMarco, Andrea and Fiameni, Giuseppe and Pilzer, Andrea and Hopkins, Andrew M and Magro, Daniel and Riggi, Simone and Sciacca, Eva and Ingallinera, Adriano and others},
  journal={arXiv preprint arXiv:2307.02392},
  year={2023}
}