Create Azure AI Document Intelligence Machine Learning Model

September 9, 2024 ยท View on GitHub

Follow the steps below to train a Azure AI Document Intelligence custom extraction machine learning (ML) model.

Step 1: Upload Pre-labeled Form data to Azure Data Lake Storage

In the repository folder data/samples/train, there are pre-labeled training forms:

FR Deployment Data Samples

These need to be uploaded to the Azure Storage account.

  1. Go to the Azure Portal and select the Azure Data Lake Storage Account that was created by deployment script. The name of this account should be storage.
  2. On the left side of the menu pane, select Data storage, then Containers and go to the samples container.
  3. Create a new folder and name it train.
  4. In the train folder, create two folders. One named contoso_set_1 and the other named contoso_set_2.
  5. Upload the sample labeling files in data/samples/train/contoso_set_1 and data/samples/train/contoso_set_2 into the corresponding folders. You now have two full sets of pre-labeled data to create the machine learning models.

Step 2: Train and Merge Document Intelligence ML Models

In this step, you will train custom Azure AI Document intelligence custom extraction models and merge them into a composite model. For more information, please refer to Azure online document Compose Custom Models.

  1. Go to Document Intelligence Studio, scroll down to Custom Extraction Model and select Create new, as illustrated below.

    FR Create Custom Model

  2. Select +Create a project to create a project.

  3. Enter a project name. For example SafetyFormProject-Set-1 or any other project name of your choice.

  4. Enter a project description. For example Custom document intelligence model with samples contoso_set_1 and click Continue.

  5. Select your Subscription, Resource Group and the Document Intelligence resource.

  6. Select the latest, non-preview API Version.

  7. Now you will be prompted to enter the training data source, as illustrated below. Select your subscription. Select Resource Group, and Azure storage created by the deployment scripts. Enter samples in the Blob container field. Enter train/contoso_set_1 in the Folder path field. Click Continue.

    FR Training Data Set

  8. Review Information and click Create Project. This step connects the form recognizer studio to Azure data lake storage/container resource in your subscription to access the training data.

  9. After the project is created, forms with OCR, field key and value pair will appear as illustrated below. Click 'Train' on upper right corner.

    FR Label Data and Train

  10. Fill in information as below, and select the dropdown "Build Mode" to Template, and then click Train.

    FR Train New Model

  11. Once the training for contoso_set_1 samples is done, the model will be located in Models section with confidence score of each field, as illustrated below.

    FR Field Confidence

  12. Train a second model with files stored in train/contoso_set_2, using above steps to create a new project and model. Name your second model as consoto-set-2 or choice of your own.

  13. Click 'Models' from your project. You will see a list of models already created. You can now merge individual models into a composite model. Select contoso-set-1 and contoso-set-2, then click Compose. The system will prompt you for a new model name and description. Name it contoso-safety-forms and provide a description. Click Compose.

    FR Merge Models

    FR Compose Model

  14. Now your model id contoso-safety-forms will appear in the Model ID list, as illustrated below.

    FR List of Models

  15. If you called your composite model contoso-safety-forms, you can go on to to Run the Solution

  16. If you did NOT call your composite model contoso-safety-forms, follow the instructions below:

    1. From the the Azure Portal, open the resource group you deployed this solution to.
    2. Find the Azure Functions App, click the resource and get to its overview page.
    3. On left panel, under section Settings, click Environment variables. Under the App, locate CUSTOM_BUILT_MODEL_ID click it and replace the default value with your composite model id.
    4. click OK and then Save. After this, your Azure Functions app will work with this document intelligence extraction model.ModelID

Now go back to Run the Solution