Skip to content

Use Tambua, the web application

For: someone who wants a point-and-click interface — for themselves, or for people who will never open a terminal — rather than Python scripts. Assumes AdaptShot is installed; Tambua is a separate, small install on top.

What Tambua is

Tambua is Swahili for identify. It is a general few-shot classification application built on AdaptShot: you point it at a folder of example photographs, it learns them, and it identifies new ones in a browser page — with the prediction set, the confidence, and the "ask a person" signal shown rather than hidden. The domain comes from a configuration file, not from the code: the same application runs maize disease and solar-panel inspection from two YAML files, and yours from a third.

Install and start

pip install tambua
tambua --list-configs
tambua --config maize

Then open the address it prints (by default http://127.0.0.1:7860) in a browser. --port and --host change where it listens; --share creates a temporary public link through Gradio's tunnel — which needs the network, and which you should not use for anything you would not put on the internet.

Tambua is installed from the repository during 0.3.0 (pip install ./apps/tambua); the PyPI name is reserved for it.

The two bundled configurations

config domain classes
maize maize leaf disease healthy, gray leaf spot, northern leaf blight
solar_panel photovoltaic module inspection as named in the file

Each is a YAML file under tambua/configs/. A configuration names the domain, the classes with a local-language label and an advice line for each, the backbone and α, and where photographs live.

Write your own configuration

Copy maize.yaml, change the class keys and their labels, and point paths at your folders. Tambua validates the file before starting and reports the exact line of any mistake:

tambua/configs/mine.yaml:14: unknown key 'diseases' -- did you mean 'classes'?

Renamed keys from earlier versions (crops, diseases, swahili) are recognised and named in the error. The Tambua reference lists every key.

What the page shows

  • The prediction and its calibrated confidence.
  • The prediction set, and whether it is calibrated — before the learner has enough photographs the page says no guarantee yet rather than showing a set that has none.
  • The action: accept, or ask a person, with the reason.
  • A correction control that feeds correct() — the same loop as tutorial 5.

Offline

Tambua needs the network for pip install and for nothing else. The model is bundled; the page is served from your own machine. Switching the wifi off after install changes nothing except that --share stops working.