ModSSC-Dashboard

The official visualization dashboard for the ModSSC benchmark project. It allows researchers to explore, analyze, and inspect the results of Semi-Supervised Learning (SSL) experiments.

Project Structure

• dashboard/: React-based visualization frontend (Vite + Tailwind + Recharts).
• runs/: Experiment logs and configuration files from ModSSC.
• src/etl.py: ETL pipeline to parse logs (extracting metrics, configs, and logs) into dashboard/public/data.
• .github/workflows/deploy.yml: CI/CD for automated ETL and GitHub Pages deployment.

Getting Started

Prerequisites

• Node.js (v18+)
• Python (v3.10+)

Development

1. Process Data:

   pip install pandas pyyaml
   python src/etl.py

2. Run Dashboard:

   cd dashboard
   npm install
   npm run dev

Deployment

The project is configured to automatically deploy to GitHub Pages on every push to the main branch. The deployment process includes:

1. Running the ETL script to refresh data.
2. Building the React application.
3. Deploying the static files to the gh-pages branch.

Analytics Tabs

The dashboard exposes the leaderboard matrix at / and additional analytics routes:

• /coverage: modality/regime coverage for successful visible runs.
• /families: family-level normalized rank across regimes.
• /variability: seed variability box plots from multi-seed sweeps.
• /performance: runtime and hardware diagnostics, including family-level runtime charts, hardware filters, and mismatch review.
• /case-studies: representative dataset trajectories across regimes.
• /pipelines: canonical preprocessing and augmentation protocol summaries.
• /variability-design: non-method design axes such as sampling, preprocessing, evaluation, and class filtering.
• /modality: modality-specific method explorer with per-family/per-method charts and cross-modality comparison.

All analytics pages reuse the public data/results-manifest.json chunks, exclude hidden datasets, and ignore crashed runs. Method families are encoded in dashboard/src/lib/methodFamilies.js; pipeline summaries combine compact protocol fields extracted by src/etl.py with documented canonical ModSSC presets.

Runtime And Hardware Metadata

The ModSSC benchmark runner writes, and the dashboard ETL expects, each future run.json to expose:

{
  "run_info": {
    "run_time_seconds": 123.456,
    "gpu_device": "NVIDIA A100-SXM4-80GB",
    "hardware_profile": "a100",
    "hardware_mismatch": false
  }
}

run_time_seconds is wall-clock runtime in seconds. gpu_device is the actual detected device name, or CPU for CPU runs. hardware_profile mirrors config.limits.profile so the expected profile can be compared with the actual device. hardware_mismatch is set when a specific profile does not match the detected GPU, or when a method resolves to CUDA but the run reports CPU.

For historical artefacts that do not yet contain run_info, src/etl.py derives runtime from run.started_at and run.finished_at, parses GPU names from bench logs when available, and falls back to Unknown when CUDA was resolved but no model name was recorded. Runtime summaries are exported to dashboard/public/data/runtime-summary.{csv,json} and analysis/output/runtime_summary.{csv,json}.