Streamlit app for image classification and model evaluation with load testing on GCP

Motivation

Imagenet is an image database for visual object recognition. It contains more than 14 million hand-annotated images. Various types of models have been developed to with great accuracy including: ResNeXt 152, Densenet 201, Darknet, AlexNet, VGG-16. Here, we have used a computationally light convolutional neural network–MobileNet. After training a model, it’s important to be able to visualize the results. Summary metrics are helpful, but being able to directly examine specific images or categories may help better understand the model–especially in the case on image data.

Streamlit App

The app was buit using streamlit, which is an open source library for building custom web apps. The ability to display data, charts, interactive widgets/sidebars, and other media makes it great for showcasing machine learning/data science.

The app was dockerized and can be pulled and run with the code below.

docker run -p 8501:8501 jtwang1027/streamlit_imagenet:multimodel

App demo on youtube

Screenshots:

Kubernetes Deployment

Pull our docker image from DockerHub:

docker pull jtwang1027/streamlit_imagenet:multimodel

Set the project id and time zone:

export PROJECT_ID=project-id export ZONE=us-central1

Change the name of the image for later Container Registry:

docker tag 24a04cb15635 gcr.io/$PROJECT_ID/sl

Upload the container image to a registry so that GKE can download and run it:

Configure the Docker command-line tool to authenticate to Container Registry:

gcloud auth configure-docker

Upload the image to your Container Registry:

docker push gcr.io/${PROJECT_ID}/sldocker push gcr.io/${PROJECT_ID}/sl

Set the project ID and Compute Engine zone options for the gcloud tool:

gcloud config set project $PROJECT_ID gcloud config set compute/zone $ZONE

Create a two-node cluster named imagenet-cluster:

gcloud container clusters create imagenet-cluster --num-nodes=2

Deploy our application:

kubectl create deployment imagenet-web --image=gcr.io/${PROJECT_ID}/sl:latest

Expose our application to traffic from the internet:

kubectl expose deployment imagenet-web --type=LoadBalancer --port 80 --target-port 8501

Scale up our application:

kubectl scale deployment imagenet-web --replicas=3

Delete the Service: This deallocates the Cloud Load Balancer created for your Service:

kubectl delete service imagenet-web

Delete the cluster: This deletes the resources that make up the cluster, such as the compute instances, disks and network resources:

gcloud container clusters delete imagenet-cluster

Load testing using Apache Benchmark

We decided to use Apache Benchmark for our load testing. Apache Bench is a command-line application for load testing on HTTP servers.

The testing result shows that our server is able to server up to about 850 requests per second, with 99th percentile of respond time falls at about 175 ms.

Repository Link here

Completed with: Nathan Tian, Jimmy Lee