| --- |
| license: apache-2.0 |
| tags: |
| - pytorch |
| - mechanistic-interpretability |
| - gravitational-lensing |
| - physics-informed |
| - equivariant |
| - computer-vision |
| - d4-symmetry |
| --- |
| |
| # D4LensPINN Weights |
|
|
| This repository contains the trained checkpoints for the D4LensPINN |
| gravitational lens classification model and its ResNet-18 baseline. |
|
|
| ## Checkpoints |
|
|
| - `d4phase2best.pth`: Best D4LensPINN checkpoint. |
| - `resnet18baseline_best.pth`: ResNet-18 baseline checkpoint. |
|
|
| ## Intended Use |
|
|
| These weights are provided for reproducibility and research use in |
| mechanistic interpretability experiments on hybrid physics-ML |
| architectures. |
|
|
| ## Notes |
|
|
| - Inputs are 150×150 grayscale images. |
| - D4LensPINN expects the full physics pipeline described in the paper. |
| - The checkpoint files are not meant to be used as standalone generic |
| image classifiers without the corresponding model code. |
|
|
| ## Loading |
|
|
| Example: |
|
|
| ```python |
| import torch |
| |
| ckpt = torch.load("d4phase2best.pth", map_location="cpu") |
| ``` |
|
|
| The exact model class definitions must match the training code used to |
| create the checkpoints. |
|
|
| ## Reproducibility |
|
|
| If you use these weights, please cite the associated paper. |
|
|
| ## License |
|
|
| Apache-2.0 License. |
