Skip to content

compute-server.md

Latest commit

 

History

History
75 lines (58 loc) · 3.16 KB

compute-server.md

File metadata and controls

75 lines (58 loc) · 3.16 KB
Raw

The tee-compute server

tessera-eval ships an optional local compute server that lets you run the ML on your own machine (where you have CPU/RAM/GPU) while a hosted TEE server supplies the map UI, tiles, and label sharing. It's the "bring your own compute" companion to the hosted TEE viewer.

pip install "tessera-eval[server]"
tee-compute --hosted https://tee.cl.cam.ac.uk
# open http://localhost:8001 in your browser

Why

The interactive evaluation — uploading a shapefile, running learning curves over millions of pixels, training models, rendering large-area prediction maps — is compute-heavy and data-heavy. Running it locally means:

  • your labelled shapefiles and embeddings stay on your machine;
  • you use your own cores/RAM (and GPU for the U-Net) instead of a shared server;
  • you still get the hosted UI, basemaps, and tile delivery for free.

How it works

            browser ──▶ tee-compute (localhost:8001)
                          │
          ML requests ────┤ handled locally (this package)
          everything else ┴──▶ proxied to --hosted (UI, tiles, labels)

tee-compute is a small Flask app. Requests under /api/evaluation/* are served locally by tessera_eval (loading embeddings via GeoTessera, running run_learning_curve / run_kfold_cv, training and applying models). Every other request is transparently proxied to the --hosted server, so the browser sees a single origin.

Local endpoints (consumed by the hosted UI; not a stable public API):

Endpoint Purpose
POST /api/evaluation/upload-shapefile accept a labelled shapefile/GeoJSON
POST /api/evaluation/run-large-area learning-curve / CV over the labelled area (streams progress)
POST /api/evaluation/train-models fit final models on all labels
GET /api/evaluation/download-model/<name> download a trained model
POST /api/evaluation/create-map render a prediction map for a region
GET /api/evaluation/download-map/<name> download a rendered map
POST /api/evaluation/cancel · clear-shapefiles · finish-classifier session control
GET /health liveness
* /<path> proxied to --hosted

Configuration

Flag Default Meaning
--hosted https://tee.cl.cam.ac.uk hosted TEE server for UI/data/proxy
--port 8001 local port to serve on
--host 127.0.0.1 bind address (keep loopback unless you know you want LAN access)
--debug off Flask debug mode (auto-reload, verbose errors)

In production mode it serves via waitress (4 threads, long channel timeout for big jobs); --debug uses the Flask dev server.

Requirements

The [server] extra pulls Flask, waitress, requests, and geotessera. Tile access (GeoTessera) requires network access to fetch embeddings unless they're cached locally. For the U-Net path, install torch as well (pip install torch).

Programmatic alternative

If you don't need the UI, skip the server entirely and call the library directly — see the tutorial. The server is purely a convenience layer over the same tessera_eval functions.