Video Classification Deployment FastAPI Kubernetes

Author: Martine Toering

The project serves as a machine learning (ML) deployment tutorial and MLOps practice. It containerizes and deploys a video classification model trained on Kinetics-400 as a simple FastAPI app to Kubernetes with minikube.

Features:

Video classification model in ONNX format
Simple fastapi app containerized with docker
Deployment to Kubernetes with minikube
Simple UI with streamlit
Basic Github Actions workflow for Kubernetes deployment

Get started

Prerequisites

Download the ONNX model files here. Next, put the model_onnx folder in app.

Poetry is used in this project. To install the version of poetry used, do:

pip install -r requirements.txt

The pyproject.toml file contains the environment with the packages used in this project. To install these packages you can for instance use:

poetry install --without dev

FastAPI

FastAPI is a framework for building RESTful APIs in python. It is suitable for machine learning model serving.

The fastapi_app folder contains a file main.py that implements a simple FastAPI app. The endpoint /predict accepts POST requests with input data and returns predictions from the model. The model is loaded in once at startup to avoid performance issues.

Test the API locally by cd fastapi_app and the following command:

uvicorn main:app --reload

After which you can send a request with e.g.

curl -X POST -F video_file=@example_data/k400tiny_videos/Z4tTlrIX0IQ.mp4 localhost:8000/predict/

See example_data for a few example video files.

Streamlit UI

While the FastAPI app is running, go to the streamlit folder, install necessary packages and run the streamlit UI with:

streamlit run app.py

Docker

Next, we are going to containerize the app with Docker, to create a consistent and portable deployment environment. See Dockerfile for how we do this; see below for reference:

FROM python:3.11

WORKDIR /app

# Install poetry version
COPY ./requirements.txt /app/requirements.txt
RUN pip install --no-cache-dir -r /app/requirements.txt

# Install poetry dependencies
COPY ./pyproject.toml /app/pyproject.toml
COPY ./poetry.lock /app/poetry.lock
RUN poetry install --no-root

# Copy code
COPY ./fastapi_app /app

RUN poetry install

EXPOSE 8000

CMD ["poetry", "run", "uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]

We make sure to install dependencies before copying code. This ensures that when the code is modified, the dependencies do not have to be reinstalled because of docker layer caching.

We can test the container locally by executing some commands below.

For building the docker, we can use:

docker build -t <image_registry>/<image_name> .`

In my example, we can do:

docker build -t martinetoering/videopredict_deploy .

To run the container we can then do the following:

docker run -p 8000:8000 martinetoering/videopredict_deploy

after which we can test the exposed API endpoint against predictions:

curl -X POST -F video_file=@example_data/k400tiny_videos/alYgwCPj4wQ.mp4 http://127.0.0.1:8000/predict/

In a deployment setting, we would have to push the docker container to an image registry so that we can make use of it in deployment later, with e.g.:

docker login
docker push martinetoering/videopredict_deploy

In our case of local development, this is not needed.

Kubernetes deployment manifest (YAML file)

The kubernetes YAML file will describe our type of deployment and the resources that we would like to use in the Kubernetes cluster.

Our kubernetes.yaml has two parts: a Deployment and a Service. The Deployment part will configure the Kubernetes pods and resources; we will keep it simple and use a single Pod. The Service will use the Kubernetes NodePort type which is a way to expose the Pod to the outside world. Another common type is LoadBalancer which serves as a way to distribute traffic among pods.

We will use imagePullPolicy: Never in the manifest for local development purposes, so that we do not have to push and pull the image. However, in order for minikube to be able to use your local docker build later, we will run this first:

eval $(minikube docker-env)

And then build the docker container again:

docker build -t martinetoering/videopredict_deploy .

This ensures that minikube can use your locally build docker container. NOTE: This command is only valid for the current terminal. Link for reference.

Deploy to Kubernetes (with Minikube)

Minikube is a Kubernetes engine for creating a kubernetes cluster on a local machine. Minikube is useful for developers testing simple kubernetes environments locally.

Check whether minikube is installed:

minikube version

Initialize minikube with:

minikube start --driver=docker

And deploy to your cluster:

kubectl apply -f kubernetes.yaml

Check your Kubernetes deployment, service and pod with the following commands:

minikube service --all
kubectl get deployments
kubectl get service
kubectl get pods

It may take a few moments before the service has started up.

Access the API

Check whether you can access the service with:

minikube service videopredict-service

You should get the welcome message from the API. The containerized FastAPI is running in the minikube cluster and you can send requests to it with the following:

curl -X POST -F video_file=@example_data/k400tiny_videos/Z4tTlrIX0IQ.mp4 <external_url_ip>/predict/

Github Actions workflow

We included a simple workflow in .github/workflows named deployment.yaml as a build and test step following this article at Minikube.

ycseong07 / ml-demo-fastapi-k8s