ADAS: Data Engine¶

¶

¶

¶

Business Challenge and Goals ¶

ML Problem Framing ¶

Project Planning, Operational Strategy ¶

Workflows, Team, Roles ¶

End to End MLOPS Testing Strategy ¶

Data Ingestion Workflows ¶

Scene Understanding & Data Mining ¶

Model Training & Experimentation ¶

Packaging, Evaluation & Promotion ¶

Deployment & Serving ¶

Monitoring & Continual Learning ¶

Cost, Lifecycle, Compliance ¶

Reliability, Capacity, Maps ¶

Inference
- PyTorch Quantization
- https://hamel.dev/notes/serving/ ML Serving
- https://openai.com/index/triton/ : Introducing Triton: Open-source GPU programming for neural networks

TODO¶

Scene Understanding & Data Mining¶

OpenSearch Indexing, Semantic Search
- How OpenSearch and GraphQL API works together
Embedding models for images, text and LiDAR
Vector DB indexing/search Algorithms: HNSW, IVF-PQ
Details of Scenes/Triggers/Scenarios
- Full list of cases
- List of 3-5 challenging cases (multi-turn curate -> train -> deply -> monitor -> improve -> curate)

Training¶

Use Ray for Distributed Training ?
Training
- Task balancing for multi-head models (e.g., HydraNet-style): dynamic or curriculum weighting.
Model Testing
- Full list of test cases
- Slices
- Regressions
HPO
- Baseline config (train.yaml) with search space: LR/WD, warmup, aug policy, loss weights, NMS/score thresholds, backbone/neck options, EMA on/off, AMP level, batch size/accum steps.
- Sweep strategy: Bayesian, Hyperband/ASHA, Random, or Population-Based Training (PBT) for long runs.