Project aims to find and analyse galaxy clusters using AI in microwave and IK range

We assume Python 3.10+ is installed. Dependencies can be managed via Poetry or pip.

1. Navigate to the project directory:

   cd galaxyHackers

1. Enter the Poetry environment:

   poetry env activate

2. Install the project dependencies:

   poetry install

If Poetry doesn't suit your needs or it fails to fetch all required libraries, you can use pip to install all dependencies.

1. (Optional) Set up and activate the virtual environment:

   python3.10 -m venv venv
   source ./venv/bin/activate

2. Install the necessary packages using pip:

   pip install torch torchvision timm torch_optimizer tqdm
   pip install numpy pandas matplotlib scikit-learn Pillow
   pip install astropy astroquery pixell dynaconf wget
   pip install comet_ml h5py ultralytics mlflow

Once dependencies are installed, you can start training the models.

Train a model by running the main.py script:

python3 -m galaxy.main --models MODEL_NAME --epochs NUM_EPOCH --data DATASET

Available flags:

In case script fails to download any dataset, it will address Legacy Survey website to download data of the infrared bands W1, W2 (for Russia available with VPN only).

Learning rate is found automatically using LR range test, so it is not passed using command-line.

For more information see A disciplined approach to neural network hyper-parameters: Part 1 -- learning rate, batch size, momentum, and weight decay

You can run several models simultaneously with the same optimizer and learning rate.

Single model training:

python3 -m galaxy.main --models YOLO12n --epochs 5 --data ACT --segment

Multiple models training:

python3 -m galaxy.main --models AlexNet_VGG ResNet18 --epochs 20 --data WISE --segment

python3 -m galaxy.main --models --epochs 20 --data WISE --segment

Full-sky ACT segmentation is no longer launched from galaxy.main. Use the standalone galaxy.full_sky module instead.

Start a new block-based full-sky run from an existing checkpoint:

python3 -m galaxy.full_sky start --checkpoint /abs/path/to/checkpoint.pth --model YOLO12n --optimizer AdamW --data ACT

Start a new run with a custom prediction-center step:

python3 -m galaxy.full_sky start --checkpoint /abs/path/to/checkpoint.pth --model YOLO12n --optimizer AdamW --data ACT --center-step 3

Resume an existing run:

python3 -m galaxy.full_sky resume --run /abs/path/to/storage/segmentation/full_sky/ACT/YOLO12n_AdamW_seed_1

Assemble a partial full-canvas FITS from completed blocks:

python3 -m galaxy.full_sky assemble-partial --run /abs/path/to/storage/segmentation/full_sky/ACT/YOLO12n_AdamW_seed_1

Render a PNG preview from the assembled partial FITS:

python3 -m galaxy.full_sky preview-partial --run /abs/path/to/storage/segmentation/full_sky/ACT/YOLO12n_AdamW_seed_1

1. Train a model with galaxy.main.
2. Start a long full-sky ACT run with galaxy.full_sky start.
3. If needed, continue it later with galaxy.full_sky resume.
4. At any point, inspect the already computed part with:
   • galaxy.full_sky assemble-partial
   • galaxy.full_sky preview-partial

The script supports Comet.ml and MLflow for experiment tracking.

1. Rename .example.secrets.toml to .secrets.toml
2. In .secrets.toml pass your Comet API to the varible COMET_API_KEY and the name of the workspace into COMET_WORKSPACE
3. Enable --comet flag

1. Enable --mlflow flag

Full documentation can be viewed here.