AMAM Benchmark Website

Interactive dataset website for the Annotated Metallic Alloys Microstructures (AMAM) benchmark.

Features

• Dataset overview, creation workflow, and statistics
• Section-by-section explorer for all AMAM subsets
• Responsive image gallery with zoom/lightbox navigation
• Per-image metadata and quick property view
• Category-level and global download controls
• Metadata export (amam-dataset-manifest.json)

Local preview

# from repository root
python3 -m http.server 4177
# open http://127.0.0.1:4177

Files

• index.html: page structure
• assets/css/styles.css: design and responsive layout
• assets/js/app.js: rendering, filters, lightbox, downloads
• assets/data/amam-dataset.json: dataset metadata + links
• assets/images/*: representative microstructure samples
• repro/*: full reproducibility package (benchmark code, outputs, and figures)

Reproduce benchmark results

# from repository root
python3 -m venv .venv
.venv/bin/pip install -r repro/requirements.txt
bash repro/benchmark/run_all_repro.sh

This full workflow runs the complete deep survey five times with explicit seeds 17--21, writes one result directory per seed, aggregates the five runs, and promotes the validated seed-17 detail files for compatibility consumers. Inspect the complete command plan without running models or changing result artifacts with:

REPRO_DRY_RUN=1 DEVICE=cuda:0 bash repro/benchmark/run_all_repro.sh

After an interruption, RESUME=1 reuses only deep seed directories that already contain a complete clean 29-model sweep using the current release configuration. Any incomplete, resumed, reordered, or protocol-mismatched deep seed restarts from the beginning with --no-resume so model-order random-number consumption remains canonical.

The complete 45-method run also needs TextureSAM, which is not tracked in this repository. The runner validates it before any model stage, so set it up first:

mkdir -p repro/external
git clone https://github.com/Scientific-Computing-Lab/TextureSAM repro/external/TextureSAM
mkdir -p repro/external/TextureSAM_Datasets/checkpoints
# then download sam2.1_hiera_small_0.3.pt from
# https://drive.google.com/drive/folders/1pUJLa898WYEcb4Y_sOaXsSVe-CsPkwRv
# to repro/external/TextureSAM_Datasets/checkpoints/sam2.1_hiera_small_0.3.pt

To check the environment and those assets without starting a model run:

PREFLIGHT_ONLY=1 bash repro/benchmark/run_all_repro.sh

Use SKIP_FOUNDATION=1 to run without TextureSAM.

The website reports the 29 supervised deep configurations using five-run means and sample standard deviations for mIoU, Dice, and Pixel Accuracy over seeds 17--21. Classical and foundation/edge values are seed-17 point estimates without a published cross-seed uncertainty estimate. See "Scope of reproducibility" in repro/benchmark/README.md before comparing results.

Detailed instructions are in repro/README.md. Execution details for every model family are in repro/benchmark/README.md. Per-model checkpoint/source traceability is in repro/results/model_provenance_manifest.csv. The 45-method internal consistency audit — model counts, cross-file agreement and artifact hashes for the published artifact bundle — is in repro/results/reproducibility_audit_45_models.json. It does not compare separate runs.

Deployment notes

• A .nojekyll file is included so GitHub Pages serves static assets directly.
• For private-only GitHub Pages publication, GitHub requires enterprise access-control support.
• On personal GitHub Free, GitHub Pages is not private-access controlled.