Homepage model demo

The homepage runs a live, watch-only Pong simulation in the browser. The saved network controls the green/right paddle against the CPU's cyan/left paddle. It does not train, mutate, or select a population, and the checkpoint's performance is not a guarantee of reliability.

The browser uses the Python game's seven normalized observations, tanh/ReLU/ sigmoid layers, and two actions thresholded strictly above 0.5. Both active actions cancel each other; neither active action leaves the paddle still. Simulation uses the original 476 x 500 coordinates and is rotated only for display in the existing CSS court.

Export a new checkpoint

From the repository root, with the project's Python dependencies installed:

python -m scripts.export_web_model --checkpoint assets/model.pt --output assets/model.json

After replacing assets/model.pt, rerun this command and include both the checkpoint and the generated JSON in your change. For a checkpoint elsewhere:

python -m scripts.export_web_model --checkpoint src/elite_model.pt --output assets/model.json

The exporter loads on CPU using PyTorch's weights_only=True. It accepts only the current state dict's six finite float32 tensors with the 7 -> 8 -> 6 -> 2 shapes. Full pickled models, other architectures, and invalid weights fail explicitly. Activations are specified by the current architecture, not stored in the .pt file; changing them requires updating the exporter and browser implementation together.

JSON version 1 contains architecture, layer weights/biases/activations, and the source checkpoint's SHA-256 hash. Export is deterministic and does not round away float32 values. The browser loads assets/model.json, not the .pt file, and needs no external ML runtime. JavaScript and PyTorch arithmetic may differ slightly; raw outputs are compared within tolerance in the parity coverage.

Preview locally

Serve the repository root over HTTP (opening index.html directly as a file: URL will not reliably allow model fetching):

python -m http.server 8000

Open http://localhost:8000/. The demo autoplays, displays actual AI/CPU scores, and re-serves after misses. Pause/resume is available; reduced-motion visitors start paused. Hidden tabs and offscreen courts suspend without accumulating simulation time or discarding a manual pause.

The browser uses 144 logical decisions per second for watchable playback, independent of display refresh rate. This is deliberately slower than the Python training runner's accelerated multi-step updates. No audio is played. Missing or malformed JSON produces an explicit unavailable message and a console error, not fake model-controlled gameplay. Without JavaScript the court remains a static illustration.

GitHub Pages publishes the committed JSON, homepage scripts, and demo GIF. Deployment does not install PyTorch or regenerate weights. Keep the committed JSON in sync using the export command above.

Coverage

python -m pytest tests/test_web_model.py tests/test_web_demo.py covers export validation and browser/PyTorch parity. Browser coverage uses an installed Chromium-family browser (Edge, Chrome, or Chromium), or the executable selected by PONG_TEST_BROWSER; it skips explicitly when none is available. Existing deployment tests exercise asset assembly on Linux with Bash.