Skip to content

Latest commit

 

History

History
24 lines (19 loc) · 2.51 KB

README.md

File metadata and controls

24 lines (19 loc) · 2.51 KB
Raw

Repository for the paper Radio Galaxy Zoo: Building a multi-purpose foundation model for radio astronomy with self-supervised learning

Instructions for use

Installation

  • Clone this repo.
  • Checkout the reproduce branch. (IMPORTANT)
  • Create a new python venv with python 3.6+
  • Activate your virtual environment.
  • Install packages by running pip install -r requirements.txt (make sure you are in the root of the project).
  • Navigate to the parent directory of this repo and locally install this repo as a package by running pip install -e byol.
  • Clone the main branch of the AstroAugmentations repository repository and install locally by using pip install -e AstroAugmentations.
  • Make sure you have a gpu available with enough memory to load a ResNet-18 model. Most modern laptops with a dedicated graphics card should be OK. You may need to install some drivers to access the card.

Fine-tuning

  • We provide a pre-trained checkpoint which you can download here.
  • Place byol.ckpt into the main directory of the project (same directory as finetuning.py).
  • This checkpoint is the model with optimized hyper-parameters which achieves an average of ~98% accuracy when fine-tuned on all MiraBest Confident training data and evaluated on the MiraBest Confident test set. To reproduce this benchmark, simply run the finetuning.py script. Please note that the "val acc" in this case is the accuracy on the training set as there is no validation set for this final experiment (see paper for details). If you would like to test the model with a validation curve, change the parameter val_size in the finetune.yml file to a non zero value.
  • If you would like to reproduce a different result or test other settings, specify hyper-parameters/settings in finetune.yml, making sure that run_id: 'none' and preset: 'none'.

Pre-training

The RGZ DR1 data-set is currently proprietary, but will be released in due course, at which point we can release the data-set used to pre-train. All the code required for pre-training is available to view in this repository in train.py and models.py. The representation learned by the model can be queried and explored interactively using our webapp. We give an example below of using the similarity search feature to find similar galaxies to an extended source.