Drone Detector MLOps

A drone detection framework with extra MLOps sauce

Cool features

• Automatically delete branches after PR is merged
• make pr uses Claude Code with the gh-pull-requests skill to create a PR
• make test runs tests with coverage
• Use the uv run train command to run the training script
• Set up pre-commit hooks for linting and formatting
• Github Actions check to make sure PR branch is up to date with main before running other workflows to minimize runner usage
• Load environment variables from Github Secrets

Cloud Infrastructure

Region Architecture

We initially started only on europe-north2, but found GPU computing to be an issue here, so we expanded to europe-west4, thus giving us the setup:

• Storage (GCS) & Registry: europe-north2 (Finland) - Minimal latency for local uploads.
• Compute (Vertex AI): europe-west4 (Netherlands) - Selected for NVIDIA T4 availability.

Storage Context Manager

We use a storage abstraction layer (storage.py) that handles local vs cloud paths transparently. Set MODE=local or MODE=cloud in settings - the same code works in both environments.

GCS Buckets

• gs://drone-detection-mlops-data/ - Training data
  • /structured/ - Direct training access (drone/, bird/, splits/)
  • /dvc-store/ - DVC versioning history (not used during training)
• gs://drone-detection-mlops-models/ - Trained models
  • /checkpoints/ - Production model storage

Update structured data - for example if we get new data

gsutil -m rsync -r -d data/drone gs://drone-detection-mlops-data/structured/drone

gsutil -m rsync -r -d data/bird gs://drone-detection-mlops-data/structured/bird

gsutil -m rsync -r -d data/splits gs://drone-detection-mlops-data/structured/splits

Cloud Training Pipeline

1. Build the Container

We use Google Cloud Build to offload the construction of the PyTorch/CUDA environment.

invoke cloud-build-train

Artifact Registry: europe-north2-docker.pkg.dev/drone-detection-mlops/ml-containers/train.

Every build updates the :latest tag used for production runs.

2. Submit to Vertex AI

Training jobs run on dedicated NVIDIA T4 GPUs.

# Default
invoke cloud-train --epochs 50

# Custom configuration
invoke cloud-train --epochs 50 --batch-size 64

API Deployment

The API uses ONNX Runtime for optimized CPU inference (much smaller container than PyTorch).

Build and Deploy

# Build API container
uv run invoke cloud-build-api

# Deploy to Cloud Run
uv run invoke deploy-api

Live endpoint: https://drone-detector-api-66108710596.europe-north2.run.app

Models are automatically loaded from GCS (model-latest.onnx).