Project Structure

The code lives in the PhD-Brown/AstroVision repository (read-only for this documentation). The Python package is src/astrovision/; analyses are notebooks that import it.

AstroVision/
├── src/astrovision/
│   ├── data_loader.py            Galaxy10 HDF5 → DataLoaders (seed-42 split)
│   ├── models.py                 SimpleCNN, EfficientNetGalaxy, build_model
│   ├── trainer.py                Training loop (AdamW, cosine schedule, W&B)
│   ├── gradcam.py                Grad-CAM and overlay grids
│   ├── crossmatcher.py           CDS XMatch with SDSS DR16 and NSA
│   ├── morphometry.py            CAS, Gini, M20, Petrosian radius, Sérsic proxy
│   ├── segmenter.py              Otsu, DINOv2 patch clustering, optional SAM
│   ├── unet.py                   U-Net, Dice + CE loss, training loop
│   ├── fusion.py                 Late-fusion classifier and ablation
│   ├── uncertainty.py            MC Dropout, temperature scaling, conformal
│   ├── attention_morphometry.py  Morphometry on DINOv2 attention masks
│   ├── spectral_viz.py           Morphology × photometry plots
│   └── synthesis.py              Master table and synthesis figures
├── notebooks/                    00 … 10 (see below)
├── data/          (not versioned) Galaxy10 HDF5, caches, features
├── checkpoints/   (not versioned) trained weights (.pt)
└── figures/                      exported figures

Notebooks

#	Notebook	Content	Documented in
00	`00_eda_galaxy10`	Distributions, galleries, pixel PCA	Galaxy10 DECaLS
01	`01_galaxy10_cnn`	SimpleCNN, EfficientNet-B0, Grad-CAM	Baseline CNN, EfficientNet
02	`02_results_analysis`	Confusion matrices, per-class metrics, errors	Benchmarking
03	`03_dinov2_features`	Zero-shot features, UMAP, neighbours, probe	Linear Probe
04	`04_dinov2_finetune`	DINOv2 fine-tuning, attention before/after	Fine-tuned
05	`05_crossmatch_analysis`	SDSS DR16, colours, populations	SDSS Cross-match
06	`06_segmentation`	CAS/Gini/M20, Otsu, DINOv2 clusters, U-Net	Morphometry, U-Net
07	`07_synthesis`	Master table, bulge fraction × colour	Physical Interpretation
08	`08_multimodal_fusion`	Late fusion and ablation	Late-Fusion
09	`09_uncertainty_calibration`	MC Dropout, temperature, conformal	Uncertainty
10	`10_attention_morphometry`	Attention-mask morphometry	Attention & Grad-CAM

Repository README

The upstream README describes eight notebooks and a shorter module list; notebooks 08–10 and the modules fusion.py, uncertainty.py, attention_morphometry.py and spectral_viz.py were added later and are not in its tree. Its build_model('dinov2') example does not match models.py, which registers only simplecnn and efficientnet; the DINOv2 classifier is defined inside notebook 04.

▶Installation

Set up the environment

</>API Reference

Module-level documentation

Notebooks​

Notebooks