All Products
Search
Document Center

Introduction

Last Updated: Mar 18, 2020

Auto Learning is an automatic machine learning platform of Alibaba Cloud Machine Learning Platform for AI. Auto Learning can help you train high-quality models based on small amounts of labeled data. With Auto Learning, you no longer need to learn AI knowledge, compile code, or manually tune parameters.

Overview

Auto Learning supports online data labeling, automatic model training, hyperparameter tuning, and model evaluation. To get started with Auto Learning, you only need to prepare a small amount of labeled data, and specify the maximum model training duration. Auto Learning will generate deeply-optimized models. In addition, Auto Learning is integrated with Elastic Algorithm Service (EAS), allowing you to quickly deploy trained models as online services.

  • Supported regions: China (Beijing) and China (Hangzhou).
  • Supported model services: image classification.

Public preview

Auto Learning is currently in public preview. The model training service is completely free during public preview.

Templates

On the homepage of Auto Learning, two image classification templates are provided: animal classification and apparel classification. Instances created on templates do not require Object Storage Service (OSS) authorization. You can directly use the source data provided by the system to train models. The entire model training procedure includes data labeling, model training and evaluation, and model testing and deployment.

OSS authorization

Auto Learning stores training data on Alibaba Cloud OSS. Before you create an Auto Learning instance, you must complete OSS authorization.Please follow the introduction in the product on this part.

Data labeling

Auto Learning is integrated with OSS, allowing you to use two methods to label images. You can label images online or import an image labeling file. The following sections describe the two data labeling methods in details and their limits.

Label images online

If only a small amount of images is used for model training, such as less than 50 images, you can label these images online. The online labeling function allows you to create labels, delete labels, and add labels to images.

Notes:
  • Make sure that all images are saved in the same folder. You must specify the path of the folder when you create the Auto Learning instance.
  • Supported image formats include JPEG, JPG, PNG, BMP, and TIFF.

    Import an image labeling file

    If a large amount of data is used for model training, you can import an image labeling file. Make sure that the images and the image labeling file are saved in the same OSS path. When you create an Auto Learning instance, you only need to specify the image labeling file. The system will automatically read the associated images. The image and label mappings in the image labeling file must be in the following format:

  1. id,oss data,label
  2. 0,"{""tfspath"":""oss://autodl/yuyi/t4.jpeg""}","{""option"":""Tigers""}"
  3. 1,"{""tfspath"":""oss://autodl/yuyi/pb5.jpeg""}","{""option"":""Polar bears""}"
  4. 2,"{""tfspath"":""oss://autodl/yuyi/cat4.jpeg""}","{""option"":""Cats""}"
Notes:
  • Make sure that the images and image labeling file are saved in the same folder. You must specify the image labeling file when you create the Auto Learning instnace.
  • Supported image formats include JPEG, JPG, PNG, BMP, and TIFF.

Model training settings

After you label all images, you are navigated to the model training settings page.

On this page, you only need to specify the maximum model training duration. Incremental learning is supported.In public preview, you can set the maximum training duration to up to 1 hour.

Maximum Training Duration: The maximum time that Auto Learning takes to train a model. Auto Learning supports Early Stopping. Early Stopping is used to prevent overfitting during the model training process.

Incremental Learning: train an existing model. The input data is used to extend the knowledge of the existing model.

Model evaluation

After the model training settings are configured, you are navigated to the model training and evaluation page.

This page displays the progress of model training. You can click the Model Training drop-down list to select a specific training and view its details, including the accuracy, precision, recall, and F1 score of the model. You can also check the prediction results made on the test dataset and the F1 score evaluation indicators. In the Actions area, you can click the relevant button to view the model training log, or test, deploy, or delete the model.

  • Log: click Log to view the entire model training process and the model training log to verify the training result.
  • Test: after the model is generated, click Test to upload local files to test the model.
  • Deploy: go to EAS to deploy the model as an online model service. For more information, see Deploy and test Auto Learning models.

Model testing

After the model is generated, you are navigated to the model testing page. You can test the model up to five times for free.You can upload local files to test the model. If this is the first time that you have used the model to make prediction, it may take three to five minutes. If the prediction result meets your expectation, click Go to EAS to deploy the model as an online service. For more information about how to deploy a model as an online service in EAS, see Deploy and test Auto Learning models. Fees are charged for online model services deployed by EAS. For more information, see EAS pricing.

Authorize a RAM user account

Auto Learning allows you to use a RAM user account to create instances, train models, and label images. Before you perform these operations, you must authorize your RAM user account. For more information, see Authorize a RAM user account.