Azure AI Custom Vision

Quick Answer: Custom Vision lets you train image classification and object detection models with your own labeled data. Choose from domain-specific base models, train with as few as 5 images per class, and export models as TensorFlow, CoreML, ONNX, or Docker containers for edge deployment.

Classification vs. Object Detection

Classification Types

• Multiclass: Each image gets exactly one label (cat OR dog)
• Multilabel: Each image can have multiple labels (outdoor AND sunny AND beach)

Project Setup and Training

Step 1: Create a Custom Vision Project

from azure.cognitiveservices.vision.customvision.training import (
    CustomVisionTrainingClient
)
from azure.cognitiveservices.vision.customvision.training.models import (
    Domain
)
from msrest.authentication import ApiKeyCredentials

credentials = ApiKeyCredentials(
    in_headers={"Training-key": "<training-key>"}
)
trainer = CustomVisionTrainingClient(
    endpoint="https://my-customvision.cognitiveservices.azure.com/",
    credentials=credentials
)

# Create a classification project
project = trainer.create_project(
    name="Product Quality Inspection",
    domain_id="general",  # or "food", "landmarks", "retail"
    classification_type="Multiclass"
)

Step 2: Add Tags and Upload Images

# Create tags
good_tag = trainer.create_tag(project.id, "Good")
defective_tag = trainer.create_tag(project.id, "Defective")

# Upload and tag images
import os
good_images_dir = "./training_data/good/"
for filename in os.listdir(good_images_dir):
    with open(os.path.join(good_images_dir, filename), "rb") as f:
        trainer.create_images_from_data(
            project.id,
            f.read(),
            tag_ids=[good_tag.id]
        )

Step 3: Train the Model

# Start training (quick training)
iteration = trainer.train_project(project.id)

# Wait for training to complete
while iteration.status != "Completed":
    iteration = trainer.get_iteration(project.id, iteration.id)
    print(f"Training status: {iteration.status}")
    time.sleep(10)

Step 4: Evaluate Performance

Key metrics for model evaluation:

Step 5: Publish and Use the Model

# Publish the trained iteration
publish_iteration_name = "v1"
prediction_resource_id = "/subscriptions/.../Microsoft.CognitiveServices/accounts/my-prediction"

trainer.publish_iteration(
    project.id,
    iteration.id,
    publish_iteration_name,
    prediction_resource_id
)

# Make predictions
from azure.cognitiveservices.vision.customvision.prediction import (
    CustomVisionPredictionClient
)

predictor = CustomVisionPredictionClient(
    endpoint="https://my-customvision.cognitiveservices.azure.com/",
    credentials=ApiKeyCredentials(
        in_headers={"Prediction-key": "<prediction-key>"}
    )
)

with open("test_image.jpg", "rb") as f:
    results = predictor.classify_image(
        project.id,
        publish_iteration_name,
        f.read()
    )

for prediction in results.predictions:
    print(f"{prediction.tag_name}: {prediction.probability:.2%}")

Domain-Specific Base Models

On the Exam: Compact domains are specifically designed for export to edge devices. If a question mentions offline or edge deployment, choose a compact domain. Standard (non-compact) domains provide higher accuracy but can only be used via the cloud API.

Model Export for Edge Deployment

Custom Vision models trained with compact domains can be exported:

Export Workflow

# Export the model in ONNX format
export = trainer.export_iteration(
    project.id,
    iteration.id,
    platform="ONNX"
)
# Download the exported model from export.download_uri

Training Data Best Practices