Pytorch Object Detector - FastAPI App

Takes an image as input and returns metadata (JSON), extracted from the image

Changelog

• 1.0.0 First release.

Development

1. python3 -m venv .venv
2. activate venv via: source .venv/bin/activate
3. pip install -r requirements-dev.txt
4. enjoy!

Tests

1. activate venv via: source .venv/bin/activate
2. Run python3 -m pytest into project-root.

This will trigger unit tests and a simple Demo that:

1. Uses test image data from tests/input folder
2. calls in memory endpoint for prediction making
3. print the desired metadata results.

Start FastAPI app locally

uvicorn app.main:app

Build Docker container

docker build -t objectdetector

Start app via Docker container

docker run: docker run -v /tmp/output:/tmp/output -p 5000:5000 -e APP_PORT=5000 -e APP_HOST=0.0.0.0 -e OUTPUT_DIR=/tmp/output -d

OR

docker compose: docker-compose up --build -d

Configuration

Environment Variables

• APP_PORT: it enables custom port for deployment by overriding it. Default 8000 .
• APP_HOST: it enables custom host for deployment by overriding it. Default 0.0.0.0 .
• OUTPUT_DIR: it enables custom output dir by overriding it. Default /tmp/output. Note: it may be required to previously create folder on host filesystem, depending on your OS.

These variables are already configured both in above docker run command and docker-compose.

Json Configuration

Default App configuration is located under config/config.json. It would be possibile to change configuration path by overriding env variable CONFIG_PATH.

In configuration, it is possibile to change parameter related to:

• confidence: decision threshold on prediction. Predictions which score is less than the threshold are filtered out.
• model: you can choice betweet 3 different models at the moment ("frcnn-resnet", "frcnn-mobilenet", "retinanet")

Demo

The demo script works as the following:

1. Lookups for APP_HOST and APP_PORT variables. If not defined, they assume default values of (0.0.0.0, 5000).
2. start docker service at the specified address (logged in console).
3. Waits 20 seconds for service startup / model downloading (second time can be faster, SO PLEASE ADJUST SLEEP ACCORDINGLY).
4. performs a prediction request for every file comprised in ./tests/input folder.
5. Outputs filename, extracted metadata and time to perform the request.
6. Check $OUTPUT_DIR folder for annotated images.

Bonus tip: Openapi

Extract OpenApi 3.0 json definition by calling http://$HOST:$PORT/openapi.json

Bonus tip: Swagger

Access SwaggerUI by calling http://$HOST:$PORT/docs