Thayer-Net is a compact U-Net research testbed for controlled synthetic galaxy deblending with Galaxy10 DECaLS cutouts. It asks whether learned image-to-image models can reconstruct a known target galaxy from synthetic blends more accurately than simple non-learning baselines, and how that behavior changes under harder overlap conditions.
This repository is a controlled synthetic benchmark. It is not a full survey-grade astronomical deblending pipeline.
Current development reference: Thayer-BR v0.2 Moderate Grouped Retrain. It uses exact-pixel and exact-coordinate group-disjoint train, validation, and development-test sources.
| Grouped development suite | Identity affected MSE | Grouped retrain affected MSE | Lower affected MSE vs identity |
|---|---|---|---|
| Normal | 0.066814 | 0.002319 | 28.81x |
| Hard stress | 0.072531 | 0.004590 | 15.80x |
| Compact bright | 0.080147 | 0.008728 | 9.18x |
| High core obstruction | 0.077871 | 0.004917 | 15.84x |
The normal 28.81x figure is an MSE ratio, corresponding to about 5.37x
lower RMSE. It comes from one grouped training seed and is not a final-paper
estimate.
| Evidence category | Status |
|---|---|
| Original development split | Historical only: 32.3x normal and 19.6x hard lower affected MSE than identity. The row split was not duplicate-safe. |
| Historical checkpoint on grouped manifests | Diagnostic only: it scores better than the grouped retrain, but 54.575% of rows expose an old train/validation source group. |
| Grouped retrain | Current source-group-disjoint development reference: 28.81x, 15.80x, 9.18x, and 15.84x across the four suites above. |
| Future final-paper test | Not yet run. It requires a fresh untouched group-disjoint source pool after protocol freeze. |
The historical evaluation-seed audit (32.02 +/- 1.21x normal and
19.55 +/- 0.30x hard) varied blend-generation/evaluation seeds, not
independent training seeds.
Affected-region metrics are emphasized because most pixels are unchanged in each synthetic blend. Whole-image scores are still useful, but affected-region MSE better isolates the thresholded blend-change region. Because target-only blur, post-composite noise, and clipping can also contribute to that region, it must not be interpreted as a pure contaminant-flux mask.
The correctness audit found 57 implicated sources (0.321% of the dataset).
Removing implicated historical evaluation rows changed reported ratios by at
most about 0.31%: a minor measured aggregate effect but a major protocol
defect. The grouped correction is complete for development work. A fresh
untouched final source pool is still required because the earlier provisional
pool was later reused by grouped training/validation.
Thayer-Net refers to the overall project and model family.
The evaluated model variants are:
- Thayer-Direct: direct reconstruction U-Net, trained to map
blended -> target. - Thayer-Residual: residual prediction U-Net, trained to map
blended -> residual, with reconstruction computed asblended - predicted_residual. - Thayer-BR v0.1: previous balanced hard-case residual U-Net trained on 8,000 synthetic blends with a 50/30/20 mix of normal, high-overlap/core-obstruction, and brightness/size-stress cases.
- Thayer-BR v0.2 Moderate: balanced residual U-Net with moderate affected/core-weighted loss; the grouped retrain is the current development reference and the original checkpoint is historical.
- Thayer-BR v0.2 Strong: stronger weighted-loss ablation, not the current best model.
- Thayer-BR v0.3 Delta: color/perceptual tradeoff ablation; it improves some compact/color diagnostics but does not replace v0.2 Moderate.
- Thayer-ResUNet v0.4: small residual-architecture ablation with targeted compact/halo gains and aggregate tradeoffs; it is not the current reference.
The 2026-07-08 evaluation audit reran evaluation only; it did not train, retrain, or modify checkpoints. The later correctness campaign separately built grouped manifests and trained one authorized grouped v0.2 checkpoint.
- Historical multi-seed normal evaluation:
32.02 +/- 1.21xlower affected MSE vs identity. - Historical multi-seed stress evaluation:
19.55 +/- 0.30xlower affected MSE vs identity. - The grouped v0.2 Moderate retrain is the current source-group-disjoint development reference, not a final-paper result.
- Thayer-BR v0.1 remained best among the audited methods across affected-mask
thresholds
0.005--0.04and dilation radii0--9; v0.2 was not part of that earlier mask-robustness audit. - Checkpoint integrity logs confirmed that the Thayer-Direct, Thayer-Residual, Thayer-BR v0.1, and Thayer-BR v0.2 Moderate checkpoints were unchanged before and after evaluation/audit runs.
- Affected-mask thresholds and dilation radii for the v0.1-era comparison; v0.2 was not included.
- Multi-seed normal and hard stress blend generation/evaluation; these are not independent training-seed replications.
- Residual reconstruction logic and sign convention.
- Visual blend, mask, residual, and model-output diagnostics.
- Apparent-size, centrality, halo-band, and visual-vs-metric diagnostics.
- Checkpoint paths, file sizes, and modified times before and after evaluation.
- Split-before-blending logic, duplicate/source leakage, and same-runtime sample comparability. Splitting indices before blending does not eliminate duplicate objects already present in the source file.
The current results are for controlled synthetic blends with known targets. They should not be interpreted as validated real-survey performance.
Galaxy10 DECaLS images are RGB display cutouts, not calibrated FITS flux images. Results should be interpreted as synthetic deblending of Galaxy10 RGB cutouts, not survey-grade source separation. Identity and threshold baselines are sanity checks, not strong astronomical deblenders.
Remaining limitations include ambiguous source overlap, target-core obstruction, target-detail loss, over-smoothing, simplified sky/noise modeling, missing PSF variation, and synthetic foreground-extraction assumptions.
Thayer-BR v0.2 Moderate is not universally best on every individual example. Thayer-Direct and Thayer-Residual still win on some samples, especially where their inductive biases preserve a particular target structure better. Thayer-BR v0.1 also wins on some v0.2 counterexamples.
The size/visual audit found broad apparent size variation and selected halo-like artifacts in individual v0.2 outputs. Aggregate halo-band error still improved relative to Thayer-BR v0.1, but future work should include a size-normalized benchmark.
The audit supports a development-benchmark ranking; it does not prove performance on real crowded survey scenes. Exact-pixel and exact-coordinate grouping is complete, but exhaustive near-duplicate identity is not proven and a fresh untouched final partition must be frozen before final-paper evaluation.
- Methodology: blend generation, masking, metrics, and residual reconstruction.
- Current best model: short v0.2 Moderate summary.
- Release summary: research checkpoint announcement for Thayer-BR v0.2 Moderate.
- Model card: technical model details, intended use, limitations, and metrics.
- Evaluation audit summary: concise audit trail for the headline result.
- Source-leakage audit and final-test protocol: duplicate findings and why the earlier provisional pool is superseded.
- Research correctness audit and grouped retrain results: grouped infrastructure, metrics, and claim status.
- Advisor update: concise audit and grouped-retrain summary.
- Preservation/clipping audits and limitations: null tests, output-range diagnostics, and next scientific corrections.
- Checkpoint summary: experiment history, checkpoint-level results, and audit summary.
- Results interpretation: how to read the affected-region metrics, threshold baseline, core obstruction, and caveats.
- Paper plan: recommended paper framing and claims to avoid.
Can a compact convolutional model recover the target galaxy from controlled synthetic blends more accurately than simple image-processing baselines, and how does performance change with overlap, contaminant brightness, blur, noise, and apparent source size?
This repository does not include the dataset. Download Galaxy10 DECaLS separately and place the HDF5 file at:
data/Galaxy10_DECals.h5
The data/ directory is kept in the repository with data/.gitkeep, while
dataset files are ignored by git.
- Exact-pixel and exact-coordinate source groups are assigned wholly to train, validation, or grouped development-test before blends are generated.
- Synthetic blends add only extracted contaminant foreground light to the target, reducing rectangular cutout/background artifacts.
- Halo-aware masks preserve diffuse contaminant outskirts while tapering before cutout edges.
- Sanity baselines include identity reconstruction and a simple threshold/connected-component method; neither is a competitive astronomical deblender.
- Thayer-Direct, Thayer-Residual, Thayer-BR v0.1, and Thayer-BR v0.2 Moderate are evaluated with whole-image and affected-region metrics.
- Hard stress testing uses smaller shifts, bright contaminants, similar-size sources where possible, blur/noise perturbations, and a minimum affected mask fraction.
- Thayer-BR v0.2 Moderate adds affected/core-weighted residual loss while keeping the residual prediction formulation.
For implementation details, see docs/methodology.md. For a concise project summary, see docs/checkpoint_summary.md.
thayernet/
├── configs/ # Portable experiment defaults
├── data/ # Local dataset location; dataset files ignored
├── docs/ # Methodology, experiment logs, paper planning
├── notebooks/ # Main experiment notebook
├── reports/ # Public-safe figures and paper skeleton
├── scripts/ # Reproducible training/evaluation scripts
├── src/ # Reusable data, blending, model, training code
├── LICENSE
├── README.md
├── pyproject.toml
└── requirements.txt
Python 3.11 or 3.12 is recommended because scientific Python and PyTorch wheels can lag newer Python releases.
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtPlace the dataset at data/Galaxy10_DECals.h5, then start JupyterLab:
jupyter labOpen notebooks/galaxy_deblending.ipynb to inspect the notebook workflow.
Larger formal experiments are captured by scripts under scripts/.
- Historical source indices were split by row. The grouped development manifests instead keep exact-pixel and exact-coordinate groups within one partition and record both source and group IDs.
- Synthetic blend generation accepts a NumPy random generator for fixed-seed experiments.
- Generated outputs, saved model checkpoints, cached files, and the Galaxy10 DECaLS HDF5 file are intentionally excluded from version control.
- Existing blend objects in a live notebook session do not update after editing
src/blend.py; regenerate blends after restarting or reloading.
- Freeze a fresh untouched group-disjoint final source pool after the model, generator, metrics, and analysis protocol are fixed.
- Add clustered uncertainty and an independent grouped training seed; one seed does not establish training-seed robustness.
- Finalize paper figures from
reports/figures/and the latest reviewed output figures. - Write the LaTeX report.
- Improve preprocessing diagnostics and foreground extraction checks.
- Run a size-normalized held-out benchmark before making stronger claims about size-invariant deblending.
- Add more realistic sky, PSF, noise, and background simulation.
This project is licensed under the Apache License 2.0. See LICENSE for
details.
