Instruction

It’s time to begin the exciting journey of building your custom image classification model using Create ML! To get started, you’ll download and set up the Create ML app.

Here are the steps to download and install the Create ML app:

1. Open Xcode: If Xcode isn’t installed, download it from the Mac App Store.
2. Access Create ML: Launch Xcode and, from the top menu, navigate to “Xcode” > “Open Developer Tool” > “Create ML”.
3. Install Create ML: Follow the prompts to install Create ML if it isn’t already installed.

With Create ML installed and ready, you’ll begin building your custom image classification model. This lesson will clarify how an image classifier fits into the broader scope of machine-learning domains and how you’ll apply it in the MoodTracker app.

What Is an Image Classifier?

An image classifier is a machine-learning model designed to recognize and categorize images. By training the model with numerous labeled examples, such as images describing emotions such as angry, fear, and happy, it learns to identify different emotional states accurately. The process involves showing the model many examples, allowing it to associate specific features with particular emotional labels.

Navigating the Create ML App

Before you cover the step-by-step process of creating the image classifier in the demo, you’ll focus here on the app itself and each part of it.

• Left Menu: Navigate between sections of the project. It has the main project details section, the model sources, and the data sources.
• Main Screen: The central area where most interactions take place, such as training the model and viewing results. This screen is divided into several tabs for the stages of the machine-learning process.
• Activity: Displays additional details and settings related to the current project, such as accuracy of different steps, ongoing processes, and project history.

You’ll notice a bunch of tabs in the center the user interface. Here’s the breakdown:

The Settings tab allows you to import the dataset to be used for both training and testing. Here, you can also configure various settings such as the feature extractor, iterations, and data augmentation options to fine-tune the training process.

In the Training tab, you can view the current status of the training process, including the accuracy achieved for both the training and validation data. The tab also provides an accuracy graph, which offers visual representations of the model’s performance during training.

The Evaluation tab displays detailed metrics such as accuracy, precision, recall for training, validation, and testing data. Additionally, the “Evaluation Summary” provides a comprehensive overview of the evaluation results, highlighting areas of strength and potential improvement.

Use the Preview tab to upload new images and test the model’s predictions. This tab provides immediate feedback on how the model classifies the uploaded images, including confidence scores. Real-time testing allows you to interactively test the model with different images to gauge its real-world performance.

The Output tab offers detailed information about the final model, including its size, structure, and class labels. In this tab, you can export your model and preview it in Xcode.

Train Button

The Train button is located to the left of the five main tabs. Clicking this button starts the training process using the settings configured in the Settings tab. Monitor the training progress and make adjustments as needed to optimize the model’s performance.

With these sections and tabs, the Create ML app has all the tools needed to build, train, and evaluate the custom image classification model.

General Recommendations for Preparing Your Image Classification Model

Data Preparation

Follow best practices for data preparation to ensure a well-trained model. These guidelines are based on Apple’s documentation for image classification and help achieve accurate and reliable results.

• Use at least 10 images per category but aim for a more varied set of images.
• Incorporate images from different angles and lighting conditions.
• Ensure balance in the number of images across categories. For example, avoid having an uneven distribution like 15 images in one category and 800 in another.
• Acceptable image formats include JPEG and PNG, viewable in QuickTime Player. For optimal results, use images with a minimum size of 299 x 299 pixels.
• Gather images that closely represent the expected input for the model in the app. If the app will classify images from a camera in various lighting conditions, collect images from similar scenarios.

Structuring Training and Testing Data

Organize the training dataset by placing images into subfolders named after their categories, such as “Happy” for images describing happiness. Create a testing dataset from the original training dataset. Duplicate the folder structure of the training dataset and move approximately 20 percent of the images from each category into the corresponding folders of the testing dataset.

Additional Tips for Success

• Use data augmentation techniques to increase the variety of your training dataset. Augmentation can include rotating, flipping, and cropping images to simulate different conditions.
• Start with a smaller dataset and gradually increase its size as you refine your model. Iteratively improving the model helps identify issues early and make necessary adjustments.

With these general recommendations in mind, you are now ready to build your custom image classification model. The next section will provide a detailed demo that walks you through the process step-by-step, helping you apply these principles effectively.