yolov5-fastapi-demo

This is a demo FastAPI app that allows a user to upload image(s), perform inference using a pretrained YOLOv5 model, and receive results in JSON format. This repo also includes Jinja2 HTML templates, so you can access this interface through a web browser at localhost:8000.

Install Requirements

Python 3.8 or later with all requirements.txt dependencies installed, including torch>=1.7 (per https://github.com/ultralytics/yolov5).

To install run:

pip install -r requirements.txt

Minimal FastAPI Example

See the minimal_client_server_example folder for a minimal client/server wrapper of YOLOv5 with FastAPI and HTML forms.

Again, you can try this out by:

1. Running the server with python server_minimal.py or uvicorn server_minimal:app --reload
2. Test the server with python client_minimal.py, navigating to localhost:8000 in your web browser or localhost:8000/docs -> POST Request -> Try It Out

Inference Methods

You can initialize the server with python server.py (use --help for other args) or uvicorn server:app --reload

You can test the server a couple of ways:

1. Using client.py - this is a basic example of using the Requests library to upload a batch of images + model name to localhost:8000/detect/ and receive JSON inference results.
2. Open localhost:8000 in your web browser, use the web form to upload image(s) and select a model, then click submit. You should see inference results displayed in the web browser shortly.
3. Open http://localhost:8000/drag_and_drop_detect in your web browser, use the drag and drop interface to upload an image, and the image + bounding boxes will be rendered via Javascript.

Models will automatically be downloaded the first time they are used and are cached on disc.

API Documentation

API endpoint documentation is auto-generated in localhost:8000/docs. The general idea is that humans use the "/" route (HTML form + inference results displayed in the browser) and programs use the "/detect/" API route to receive JSON inference results.

Developer Notes

Server.py

Contains the FastAPI server code and helper functions.

Jinja2 Templates (/templates folder)

File	Description
layout.html	Base template with navbar that is common to all pages. home.html and drag_and_drop_detect.html both extend this template.
home.html	Basic web form for uploading images, model selection and inference size to the server. The server gets the YOLO results and renders a bbox image, then returns the results by plugging them into the jinja2 template templates/show_results.html. This is overly fancy, but I wanted to demonstrate how to do this - if you want just JSON results see the minimal client-server example.
drag_and_drop_detect.html	This implements a Drag & Drop interface to upload images. Once dropped onto the dropzone, the image and parameters are sent to the server's /detect endpoint which returns JSON results. The JSON results are then used to render the image + bboxes in the web browser as seen in the Inference Methods section above. The box labels are raised above the box outline such that the labels don't overlap with each other.

Credits

This repository is a wrapper around YOLOv5 from Ultralytics: https://github.com/ultralytics/yolov5

Also modified the results_to_json function from the original here: https://gist.github.com/decent-engineer-decent-datascientist/81e04ad86e102eb083416e28150aa2a1