Mastering Model Export from Azure Custom Vision for Offline Use

Introduction:

In the realm of Artificial Intelligence (AI), Azure Custom Vision provides powerful tools for object detection and image classification. However, one critical aspect that developers often encounter is the need to export these models for use in environments where internet connectivity is unavailable. This blog will guide you through the essential concepts, steps, and best practices required to effectively export models from Azure Custom Vision for offline use, ensuring that you capture the most critical 20% of knowledge that will provide you with 80% of the understanding you need.

Table of Contents:

1. Understanding Domains in Custom Vision
2. Steps to Export a Model for Offline Use
   • Change the Domain
   • Retrain the Model
   • Export the Model
3. Export Formats and Their Usage
   • TensorFlow
   • ONNX
   • CoreML
4. Application Scenarios
   • Edge Devices
   • Disconnected Networks
5. Best Practices
   • Test Your Exported Model
   • Optimize Performance
6. Memory Techniques
   • Mnemonics
   • Story-Based Memory Technique
7. Conclusion

---

Understanding Domains in Custom Vision

Azure Custom Vision operates using various domains, each tailored for specific tasks like General, Food, Landmark, and Retail. Compact Domains are particularly crucial for exporting models for offline use because they are optimized for lightweight deployment on edge devices, where internet connectivity may not be available.

Steps to Export a Model for Offline Use

Exporting a model from Azure Custom Vision for offline use involves three primary steps:

Change the Domain

If your model is not in a compact domain, switch it to a compact domain, such as General (compact). This step is necessary because only compact domains support export.

Retrain the Model

Once you change the domain, retrain the model to ensure it adapts to the new domain’s parameters and performs effectively within the new constraints.

Export the Model

Finally, export the model in a format compatible with offline or edge environments, such as TensorFlow, ONNX, or CoreML.

Export Formats and Their Usage

TensorFlow

TensorFlow is ideal for Android devices or any environment that supports TensorFlow.

ONNX

ONNX is a versatile format, compatible across different platforms, including Windows, Linux, and various machine learning frameworks.

CoreML

CoreML is specifically designed for Apple devices, allowing seamless integration into iOS applications.

Application Scenarios

Edge Devices

Edge devices, such as IoT gadgets, mobile phones, or on-premises servers, require models that can run without an internet connection.

Disconnected Networks

Some scenarios, such as those involving data security or regulatory concerns, necessitate models that operate independently of cloud services.

Best Practices

Test Your Exported Model

Before deploying, ensure that the model performs as expected in the target environment.

Optimize Performance

While compact models are already optimized, you may need to fine-tune parameters like batch size or input dimensions to achieve optimal performance.

Memory Techniques

Mnemonics

• DREAM: Domain change, Retrain, Export, Application scenarios, Model testing.
• FLOOR: Formats (TensorFlow, ONNX, CoreML), Lightweight deployment, Optimize performance, Offline use, Retest.

Story-Based Memory Technique

Imagine you’re preparing for a long journey to a remote island (representing a disconnected network). Before leaving, you need to pack only the essentials (Compact Domain). You then test each item to ensure it works under the island’s conditions (Retrain the Model). Finally, you carefully place everything in your suitcase, ensuring it’s ready for the journey (Export the Model). On the island, you efficiently use each tool for survival (Edge Devices, Disconnected Networks) and fine-tune your strategies as you go (Optimize Performance).

Conclusion

Exporting models from Azure Custom Vision for offline use is a critical skill for developers working with AI in environments where internet connectivity is unavailable. By understanding compact domains, following the correct steps to export, and choosing the right formats, you can ensure your models perform optimally in disconnected settings. Remember to apply best practices, test thoroughly, and optimize where needed to achieve the best results.