Electron microscopy
Example of Vertex AutoML Vision
- Python Automation and Machine Learning for ICs -
- An Online Book: Python Automation and Machine Learning for ICs by Yougui Liao -
Python Automation and Machine Learning for ICs                                                           http://www.globalsino.com/ICs/        

Chapter/Index: Introduction | A | B | C | D | E | F | G | H | I | J | K | L | M | N | O | P | Q | R | S | T | U | V | W | X | Y | Z | Appendix


Vertex AI, from Google Cloud, is a platform designed to simplify the deployment and maintenance of machine learning models. Within Vertex AI, there's a tool called AutoML Vision which allows users to train custom machine learning models to perform tasks related to image recognition without requiring deep machine learning expertise. That is, AutoML Vision enables individuals with minimal ML expertise to develop high-quality image classification models.

Here’s an example scenario of how Vertex AutoML Vision, which is Retail Store Shelf Analysis, might be used:

  • Objective: A retail company wants to optimize shelf stocking and ensure promotional compliance in their stores. They plan to use image recognition to analyze the configuration of products on shelves.
  • Steps Using Vertex AutoML Vision:
    • Data Collection: The company collects thousands of photographs from various store locations, showing different shelf setups.
    • Labeling: Each image is labeled with information about the products displayed, their arrangement, presence of promotional materials, and compliance with planned layouts.
    • Uploading (see details below) and Training:
      • The labeled images are uploaded to Vertex AI.
      • The user selects AutoML Vision as the model type.
      • AutoML Vision automatically splits the data into training, validation, and test sets.
      • The platform uses the training data to train a model to recognize and classify the different elements in the images.
    • Model Evaluation:
      • After the training process, Vertex AI provides an evaluation of the model’s accuracy using the test set.
      • The user reviews metrics like precision, recall, and overall accuracy to determine if the model meets the business requirements.
    • Deployment:
      • Once satisfied with the model’s performance, the user deploys the model on Vertex AI.
      • The model can then be accessed via an API to analyze new images from store cameras in real-time.
    • Application:
      • Store managers receive reports on their mobile devices about shelf status, discrepancies in product placement, and promotional compliance.
      • They can quickly address any issues, optimize shelf arrangements, and ensure promotions are visible as planned.

Upload training images to cloud storage:

  • Storage Capacity: Training deep learning models requires potentially vast datasets, and local storage on personal or even corporate devices might not be sufficient. Cloud storage provides scalable and flexible storage options that can accommodate large volumes of data.
  • Accessibility: Cloud storage allows your data to be accessed from anywhere, which is particularly useful for teams that are distributed geographically. This means that anyone working on the project can access the training images from any location without needing to transfer large files back and forth.
  • Data Security and Backup: Cloud providers typically offer robust security measures that can be more effective than local solutions, including automated backups and disaster recovery protocols. This helps protect your data against loss due to hardware failures or other issues.
  • Integration with ML Platforms: Many cloud platforms offer integrated machine learning services that can directly interact with data stored on their servers. This integration can simplify the workflow of data preprocessing, model training, and deployment.
  • Scalability of Computational Resources: Training modern image processing models can be resource-intensive. Cloud environments allow you to easily scale up the computing power as needed, which can be more cost-effective and less cumbersome than maintaining your own high-performance computing infrastructure.
  • Collaboration: Cloud storage solutions often come with tools that facilitate collaboration, like simultaneous editing and version control, which are beneficial for teams working on machine learning projects.
To obtain predictions from the model, you need to submit prediction requests to an endpoint within your project. This endpoint will relay your request to the hosted model and deliver the results back to you. The preparation for utilizing a trained machine learning model to make predictions in Vertex AutoML Vision typically involves:
  • Training the Model: First, you train your model using Vertex AI AutoML Vision, which involves feeding it a dataset containing images labeled with classes. AutoML Vision automates the process of model selection and training.
  • Creating an Endpoint: Once the model is trained, you need to deploy it to an endpoint. An endpoint in Vertex AI is a machine learning model that has been deployed for online or batch predictions. It is a URL or a service that listens for prediction requests.
  • Using the Endpoint for Predictions: After deploying the model to an endpoint, you can send new data (images, in the case of AutoML Vision) to this endpoint to get predictions. The statement indicates that in the next section of the guide or documentation you are following, they will show you how to send these prediction requests.

In essence, the endpoint serves as the interface between your application and the trained machine learning model hosted in the cloud. By sending images to the endpoint, you can leverage the model's capabilities to analyze the images and return predictions based on what it has learned during training. This process is crucial for applications that require real-time or batch processing of image data to classify, detect, or segment visual information.

One practical example is that a semiconductor company wants to build a machine learning model that can detect defects on wafers and microchips using their own custom images. Then, the company should use Google AutoML Vision for this use case since this model is trained on the company's dataset with the company's labels. That is, AutoML Vision allows organizations to train custom machine learning models tailored to their specific requirements using their own labeled datasets. This makes it well-suited for specialized tasks such as detecting defects in semiconductors where proprietary and highly specific data is used.