How to Deploy YOLOv8 Using Intel's OpenVINO and Amazon SageMaker

Introduction

In this post we will walk through the process of deploying a YOLOv8 model (ONNX format) to an Amazon SageMaker endpoint for serving inference requests, leveraging OpenVino as the ONNX execution provider. We will start by setting up an Amazon SageMaker Studio domain and user profile, followed by a step-by-step notebook walkthrough.

By the end of this tutorial, you will have a better understanding of how to utilize Amazon SageMaker for deploying your ONNX models.

Full notebook here. (opens in a new tab) Let's get started!

Set up Amazon StageMaker Studio

Before we dive into the notebook walkthrough, ensure that you have completed the following steps:

1) Set up a SageMaker Studio domain and user profile

A. Domain overview (opens in a new tab)
B. Domain quick setup (opens in a new tab)
C. Add/remove domain user profile (opens in a new tab)
D. Launch Amazon SageMaker Studio (opens in a new tab)

2) Start SageMaker Studio

First, go to your user profile and click on the launch button to open the studio. Then, upload this notebook to the studio documented here. (opens in a new tab)

Next, select a SageMaker kernel as the configuration mentioned below:

• Image: Base Python 3.0
• Kernel: Python 3
• Instance type: ml.t3.medium
• Start-up script: No script

YOLOv8 Model Conversion and AWS S3 Upload

Next, we're going to prepare our model for inference in SageMaker Studio Labs.

In the code below, we download the yolov8 Pytorch model from the Roboflow GitHub repo (opens in a new tab) and export ONNX weights using the ultralytics YOLO CLI. Then, we zips (as ".tar.gz" file) for use in uploading to S3, the Amazon storage product that we'll use to interact with our weights in SageMaker Studio Labs. The model is then uploaded to S3 for use in SageMaker Studio Labs.

!mkdir -p model
%cd model
!wget https://github.com/roboflow/sagemaker-yolov8/tree/main/yolov8-weights/runs/detect/train -O yolov8n.pt

# Convert pytorch version of YOLOV8 model file to onnx format
if not os.path.isfile('/home/ec2-user/SageMaker/model/yolov8n.onnx'):
    !yolo mode=export model=yolov8n.pt format=onnx dynamic=True # To get a static model chnage dynamic flag to False
else:
    print("ONNX file alreday exist. Please re-run the cell after removing it.")
    
# Model needs to be zipped before storing into a S3 bucket
import tarfile
import os

source_files = ['/root/model/yolov8n.onnx']
output_zip_file = '/root/model/model.tar.gz'

if not os.path.isfile(output_zip_file):
    with tarfile.open(output_zip_file, mode="w:gz") as t:
        for sf in source_files:
            t.add(sf, arcname=os.path.basename(sf))
        print(f"ZIP file is created at {output_zip_file} location.")
else:
    print(f"ZIP file '{output_zip_file}' already exist. Please re-run the cell after removing it.")
 	print("<bucket>: ", bucket)
 
# Upload the model to default S3 bucket of your sagemaker domain
model_location = session.upload_data(path=output_zip_file, key_prefix='deployment_model')
print(f"Zip file is uploaded to {model_location} location of '{bucket}' S3 bucket.")

Build Docker Image on AWS Elastic Container Registry

Next, we need to create a custom service script, container entry point script, and a Docker file to deploy the ONNX model to a SageMaker inference endpoint.

Run cells 12,13, and 14 in this full notebook (opens in a new tab) to write the Docker files to a directory called "docker-inference" which will create a DockerFile, dockerd-entrypoint.py and model-script.py.

# Creating an ECR uri to inference with the build docker image
ecr_repository = 'yolov8-deployment'
tag = ':latest'
uri_suffix = 'amazonaws.com'
inference_repository_uri = '{}.dkr.ecr.{}.{}/{}'.format(account_id, t/model/model.tar.gz'

%cd ..
!mkdir -p docker-inference

This step needs to be manually carried out in an AWS EC2 instance. To build the image, download the "docker-inference" folder from "File Browser" section and follows the commands mentioned below to push the Docker image to an AWS Elastic Container Registry (ECR).

1. Build the image with ecr_repository tag.
   sudo docker build -t <ecr_repository> docker-inference/
   
2. Re tag the image with inference_repository_uri name.
   sudo docker tag <ecr_repository> <inference_repository_uri>
   
3. Create an AWS ECR and login into to push the image.
   A. aws ecr get-login-password --region <region> | sudo docker login --username AWS --password-stdin <account_id>.dkr.ecr.<region>.amazonaws.com
   B. aws ecr create-repository --repository-name <ecr_repository>
   C. sudo docker push <inference_repository_uri>

Deploy YOLOv8 Model to Sagemaker

Now deploy the model to a SageMaker endpoint. The instance size to use for deployment is ml.c6i.xlarge. Make sure the image is built and pushed to ECR before trying out this part.

import sagemaker
from sagemaker import Model

ENDPOINT_NAME = "YOLO-V8-DEPLOYMENT"
yolo_model = Model(
                   model_data=model_location,
                   role=role,
                   image_uri=inference_repository_uri,
                   name='YOLO-V8-deployment-1')
                   
yolo_model.deploy(initial_instance_count=1, instance_type='ml.c6i.2xlarge', endpoint_name=ENDPOINT_NAME)

YOLOv8 Inference with Amazon SageMaker

Now we're ready to run our model! In the script below, we demonstrate how to run inference on your YOLOv8 (opens in a new tab) model. First, provide a URL of an image on which you want to run inference on input data. Then, add a comma then specify the device type on which inference should run. Supported image types are 'jpg', 'jpeg', 'png' and device types are 'cpu' & 'CPU_FP32'.

Setting cpu as the device type will perform the ONNXRuntime inference with default CPU EP(MLAS). Setting CPU_FP32 will perform the ONNXRuntime inference with OpenVINO EP.

payload2 = "https://github.com/roboflow/sagemaker-yolov8/blob/main/yolov8-weights/runs/detect/predict/dog.jpeg,CPU_FP32"

# warmup steps
for _ in range(warmup_steps):
    predictor.predict(payload2)

!sleep 10 

total_latency = 0
for _ in range(num_iter_steps):
    response2 = predictor.predict(payload2)
    total_latency += float(response2.decode().split('\n')[0].split(':')[1][:-1].strip())

print("(OpenVINO EP) Avg Latency In Seconds: ", total_latency/num_iter_steps)
print("\nSingle Inference Details:\n",response2.decode())

# Dsiplay the output
# update the labels, points and confidences with your inference output 
labels2 = ['person', 'stop sign', 'person', 'person', 'person', 'bus']
points2 = [[(0, 551), (67, 874)], [(0, 255), (32, 325)], [(221, 406), (345, 857)], [(670, 380), (810, 876)], [(49, 399), (245, 903)], [(17, 231), (802, 768)]]
confidences2 = [0.30, 0.35, 0.82, 0.85, 0.87, 0.87]

display('bus.jpg', labels2, points2, confidences2, class_to_idx, 'ONNXRuntime  Inference With OPENVINO CPU EP')

Here is an example prediction:

Cost Analysis: OpenVINO vs Default CPU

The benchmark results below demonstrate the cost saving comparison of "OpenVINO Execution provider" over "Default CPU(MLAS) Execution provider". The results may be helpful when deciding which ONNXRuntime is best for your applications.

Conclusion

In this guide, we have built a YOLOv8 object detection model and deployed to a AWS SageMaker endpoint, serving inference requests with ONNXRuntime and OpenVino as the execution provider.

A note on licensing: YOLOv8 is released under AGPL-3.0, which affects commercial use:

• On Roboflow cloud: a commercial license is included for every user, including the free plan.
• Self-hosted: you need an Enterprise license add-on, or you must open-source your application under AGPL-3.0.
• Permissive alternative: RF-DETR (opens in a new tab) is Apache 2.0, free for commercial use anywhere.