Yolov26 Segmentation in Azure Machine Learning Compute Instance

Prerequisites

• Azure Subscription
• Azure Machine Learning Workspace
• Quota for Standard_D13_v2 cores

Introduction

• This notebook demonstrates how to use the yolov26 model for image segmentation using Azure Machine Learning.
• Download from Hugging face and run the inferencing in a notebook.
• Using image and text prompts to generate segmentation masks.
• Vision based object detection and segmentation.

Create Azure Machine Learning Workspace

• Create a brand new Azure Machine Learning Workspace in Azure Portal.
• Create a new resource group or use an existing one.
• Now for storage account provide these RBAC roles to the AML workspace managed identity.
• Storage Blob Data Contributor
• Storage File Privileged Data Contributor
• Create a new Azure Machine Learning Compute Instance in the AML workspace.
• Choose the appropriate size for the compute instance based on your requirements.
• SKU i am using is Standard_D13_v2

Code Example

Setup Compute Instance

• First create a new compute instance in Azure Machine Learning workspace.
• Now go to terminal and run the following commands to create a sam3env python virtual environment with required dependencies.

conda create --name yolo26env -y python=3.12
conda activate yolo26env
conda install pip -y
conda install ipykernel -y
python -m ipykernel install --user --name yolo26env --display-name "yolo26env"pip install ultralytics
pip install openai-clip
pip install azureml-inference-server-http
pip install azure-monitor-opentelemetry-exporter
pip install pip install azure-ai-ml
pip install pip install azure-identity
pip install matplotlib

Get Balamurugan Balakreshnan’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

Notebook Code

• make sure install openai-clip package to avoid errors while loading the model.

%pip install openai-clip

• now lets download the model

!mkdir -p ./models
!cd ./models
!curl -L -o ./models/yoloe-26n-seg.pt https://github.com/ultralytics/assets/releases/download/v8.4.0/yoloe-26n-seg.pt

• To avoid insufficient memory issues we will use yolov26seg model which is optimized for segmentation tasks.
• in each cell let’s run command to create ultralytics home directory.

!mkdir -p /home/azureuser/.ultralytics
!export ULTRALYTICS_HOME=/home/azureuser/.ultralytics

• Now download the mobile clip

!cd /home/azureuser/.ultralytics
!curl -L -o mobileclip2_b.ts https://github.com/ultralytics/assets/releases/download/v8.4.0/mobileclip2_b.ts

• now set the home directory path for ultralytics

import os
os.environ["ULTRALYTICS_HOME"] = "/home/azureuser/.ultralytics"

• Check for token

import clip
print(clip)
print(hasattr(clip, "tokenize"))

• now write the code to load the model

import os
os.environ["ULTRALYTICS_HOME"] = "/home/azureuser/.ultralytics"
from ultralytics import YOLO
import cv2
import matplotlib.pyplot as plt
segmodel = YOLO("./models/yoloe-26n-seg.pt")
names = ["person", "bus"]
segmodel.set_classes(names, segmodel.get_text_pe(names))
results = segmodel.predict(
    "https://ultralytics.com/images/bus.jpg",
    conf=0.1
)
r = results[0]
img = r.plot()
plt.figure(figsize=(12,8))
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.axis("off")
plt.show()

• Run the cell and wait for output

• You should see the segmentation masks and bounding boxes for the detected objects in the image.
• You can change the image URL and the class names to test with different images and objects.
• open ai clip, mobileclip2_b.ts and yolov26seg model are required for this to work.
• Make sure the curl URL is current versions are used from ultralytics official github repo.

original article at — Samples2026/AML/yolov26seg.md at main · balakreshnan/Samples2026