Docker image for doing inference on vision models trained with fastai

This repo is an adaptation of https://github.com/gdoteof/fastai-vision-uvicorn-gunicorn-starlette-docker. I tried his default code and the Dockerfile is a bit old on fast.ai version so I updated it. Also added Semantic UI for better UI on the webapp part.

This docker image is a wrapper around https://github.com/tiangolo/uvicorn-gunicorn-starlette-docker, adding the dependencies to run fastai models, as well as a minimal "webapp" to make requests.

Known Problems:

• Fastai vision models are not necessarily compatible version to version (TODO: tags for each fastai version)

Installation

Recommended usage is to build the docker image with name <your_docker_name> yourself, and overwrite the export.pkl at runtime, and provide environment variables and substitute <your_docker_name> with the docker image name you created.

	docker run -p 80:80 \
		-v ./local/path/to/export.pkl:/app/export.pkl \
		-e TITLE="Chinese Calligraphy Classifier" \
		-e SUBTITLE="Can disambiguate Chinese calligraphy styles like KaiShu, LiShu, XiaoZhuan" 
		<your_docker_name>

Usage

API

Starlette backend provides a /classify endpoint and separate callbacks for GET and POST.

GET requests expect a query param url which indicates a URL to an image on the internet which should be classified

Example:

GET /classify?url=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fen%2Fd%2Fd4%2FMickey_Mouse.png

Response:

{"predictions":[["mickey",0.9781272411346436],["computer",0.021808452904224396],["animal",6.434295937651768e-05]]}

POST requests expect the body to be a Byte array of an image to be classified.

Webapp

Starlette also serves a simple html/javascript app which can serve as a UI to interact with the backend, and display the prediction of a photo.