LenSiam

LenSiam is the self-supervised learning architecture of SimSiam plus a novel domain-specific augmentation for strong gravitational lens images.

To create a conda env, run:

. create_conda_env.sh

To train Limsiam models, run

python  main.py  --data_dir  path/to/your/dataset  -c  configs/train.yaml

To calculate UMAP embeddings, run

python  calc_umap.py  -c  configs/umap.yaml

The notebook contains the code we used to make graphs in our papers in NeurIPS 2023 workshops.

Note:

• One of the datasets that are used for UMAP calculation is the HST real images. The dataset can be created by running the code in the repo of kuanweih/lensed_quasar_database_scraper.

• We thank Patrick Hua for their open-source code. The code of SimSiam used in this work was adapted from their SimSiam repository.

Key Takeaways

Figure 1

(a) The LenSiam architecture for this work.
We generate positive pairs of lens images through a domain-specific augmentation approach to learn the representation of strong gravitational lens images.

(b) Example of two different source galaxies (top) and their lens images with the identical foreground lens model (bottom). The bottom images represent a positive lens image pair for our LenSiam models.
Our lens augmentation takes into account the domain knowledge of gravitational lensing. This allows LenSiam to learn consistent representations of foreground lens properties.

(c) Example of applying the default augmentation to a lens image. The bottom augmented images represent a positive lens image pair for our baseline SimSiam models.
The commonly used random augmentation methods are problematic here as the lens properties will be easily changed. For example, enlarging a lens image will directly change the Einstein radius.

Figure 2

The UMAPs are color-coded by the Einstein radius $\theta_{\rm E}$, the ellipticity $e_1$, and the radial power-law slope $\gamma$ from the left to right columns. The top row is the UMAPs for LenSiam while the bottom row is the UMAPs for the baseline SimSiam.
The nonuniform distributions on the LenSiam UMAPs indicate that its backbone ResNet101 trained by the LenSiam SSL process does learn some key parameters such as $\theta_{\rm E}$, $e_1$, and $\gamma$, even though it has NEVER seen the true parameters during the entire training process.

Downstream task:

Pretrained ResNet101 Models	Framework	$R^2$ (Einstein radius)
ImageNet-1k (baseline)	Supervised	0.360
SimSiam (baseline)	Unsupervised	0.426
LenSiam (this work)	Unsupervised	0.586

As an exploration, we experiment both our LenSiam and SimSiam learned representations with a downstream regression task as a proof of concept. We finetune the model to estimate the Einstein radius $\theta_{\rm E}$ with the Lens challenge dataset, which simulated Euclid-like observations for strong lensing. To simulate the scarcity of real strong lensing data, we select a sub-sample of 1,000 images as the training set and 1,000 images as the test set. With LenSiam pre-train models, we reach $0.586$ in $R^2$ compared with baseline SimSiam models $0.426$ and supervised-only models (the ResNet101 models pre-trained on ImageNet-1k) $0.360$ on Einstein radius.
We find that the LenSiam pretraining does help downstream regression task.