Docker deploy (Tile Geo Process)

Publish an oriented-det checkpoint as a Sanic service: base64 JPEG/PNG tiles in, GeoJSON polygons out. The example lives in deploy/example/. Folder README: deploy/example/README.md.

This path still runs PyTorch inside an NVIDIA CUDA 12.1 image. For inference without PyTorch, see ONNX export.

Bake a checkpoint

Weights are gitignored. From the repo root, copy a Hub sidecar + .pth (or a runs/ best checkpoint) into deploy/example/app/, then write description.json:

odet pretrained download oriented_rcnn_dota_le90_1x
mkdir -p deploy/example/app/weights
cp pretrained/oriented_rcnn_*_dota_le90_1x*.json deploy/example/app/config.json
cp pretrained/oriented_rcnn_*_dota_le90_1x*.pth  deploy/example/app/weights/model.pth
python deploy/scripts/generate_description.py \
  --config deploy/example/app/config.json \
  --out deploy/example/app/description.json \
  --deploy-version 0.4.0

Build and run

docker build -f deploy/example/Dockerfile -t odet-example:latest .
docker run --rm -p 8080:8080 --gpus all odet-example:latest

Verb	Path	Role
GET	/api/v1/health	liveness
GET	/api/v1/describe	class list + metadata
GET	/api/v1/openapi	Tile Geo Process YAML
POST	/api/v1/process	{resolution, tiles[]} → FeatureCollection
GET	/swagger/	OpenAPI UI

resolution is meters per pixel (scales length / width on each feature). A second POST while inference is running returns 429.

config.production.* sets the score floor, final NMS, sliding-window overlap, and canvas flags (stick_to_model_canvas default true → 1024 tiles; larger images tile). Walkthrough with overlays: Deploy oriented-det in Docker.