Skip to content

demo.md

Latest commit

 

History

History
57 lines (45 loc) · 2.79 KB

demo.md

File metadata and controls

57 lines (45 loc) · 2.79 KB
Raw

End-to-end demo

The primary reviewable demo is assets/demo/factoryvision-demo.mp4. It is a 2 minute 30 second walkthrough of the FactoryVision path. A looping GIF fallback is also included:

image upload -> FastAPI /predict -> ONNX mask and score
             -> prediction persistence -> Prometheus/Grafana
             -> MLflow experiment and registered candidate

The video is assembled from real repository evidence. Its API response is generated by calling the real FastAPI /predict route with the ONNX model and an in-memory store. The monitoring frame is a captured local Grafana dashboard, and the MLflow frame uses the actual mini-study results. This keeps the demo reproducible without requiring a cloud account or embedding credentials.

The model still receives its fixed 256 x 640 letterboxed input. For the qualitative prediction image, the padded region is cropped away and the mask is mapped back to the original image dimensions, so the source image appears only once.

Timeline

Time Scene What to explain
0:00–0:12 Opening flow The image moves through serving, persistence, monitoring, and tracking.
0:12–0:36 API upload /predict accepts a multipart image and returns the model response.
0:36–1:06 Segmentation Logits become probabilities, a threshold creates a binary mask, and post-processing derives the score and box.
1:06–1:36 Monitoring Prometheus collects counters/histograms and Grafana displays service and model signals.
1:36–2:06 MLflow Comparable runs are inspected and the baseline is registered as the candidate.
2:06–2:30 Reproduction The README and operational cards provide the commands, evidence, and limitations.

Rebuild the demo

The model artifact must exist at artifacts/models/factoryvision-segmentation.onnx. If necessary, follow the training, registration, and ONNX export instructions in the README first.

.venv\Scripts\python.exe scripts\build_demo_gif.py

The script writes:

  • assets/demo/factoryvision-demo.mp4 — the primary 2:30 demo video;
  • assets/demo/factoryvision-demo.gif — a looping GIF fallback;
  • assets/demo/factoryvision-demo-storyboard.png — a static six-scene review;
  • assets/demo/api-prediction-overlay.png — the actual API prediction overlay;
  • assets/demo/api-response.json — the exact response used in the API scene.

The generated API evidence uses an in-memory persistence adapter, so it does not replace the Docker Compose integration test. For the complete live flow, start the Compose stack using the README instructions, upload an image through the interactive API docs, then refresh Grafana and MLflow while narrating the same sequence.