In this article we will create an ARM template that will deploy a Custom Vision Service and output the Training and Prediction API Keys.

Recently Custom Vision Service has been made available on the Azure Portal in public preview. Previously, it was only available via www.customvision.ai.

Today we will focus on creating an Azure Resource Manager template to easily setup and deploy a Custom Vision service to a resource group. Our ARM template will be created in a new Azure Resource Group deployment project in Visual Studio.

Creation

Let's declare the parameters of the ARM template:

{
  "$schema": "http://schema.management.azure.com/schemas/2014-04-01-preview/deploymentTemplate.json#",
  "contentVersion": "1.0.0.0",
  "parameters": {
    "nameTraining": {
      "type": "string",
      "defaultValue": "myCustomVisionService"
    },
    "location": {
      "type": "string",
      "defaultValue": "southcentralus",
      "allowedValues": [
        "southcentralus"
      ]
    },
    "skuTraining": {
      "type": "string",
      "defaultValue": "F0",
      "allowedValues": [
        "F0",
        "S0"
      ],
      "metadata": {
        "description": "Describes the pricing tier for Training"
      }
    },
    "skuPrediction": {
      "type": "string",
      "defaultValue": "F0",
      "allowedValues": [
        "F0",
        "S0"
      ],
      "metadata": {
        "description": "Describes the pricing tier for Prediction"
      }
    }
  }
  ...
}

• nameTraining: name of the training service. If no parameter is provided the default name myCustomVisionService is used.
• location: the location where the service will be deployed. Currently in the preview only southcentralus is available.
• skuTraining: the pricing tier of the training service. If no parameter is provided the pricing tier will be F0.
• skuPrediction: the pricing tier of the prediction service. If no parameter is provided the pricing tier will be F0.

We will also need one variable in this ARM template:

{
  ...
  "variables": {
    "namePrediction": "[concat(take(replace(parameters('nameTraining'),'-',''), 18), '_Prediction')]"
  }
  ...
}

• namePrediction: here we generate the name of the prediction service based on the name of the training service.

Now we will declare the resources necessary to deploy a Custom Vision Service:

{
  ...
  "resources": [
    {
      "apiVersion": "2016-02-01-preview",
      "name": "[parameters('nameTraining')]",
      "type": "Microsoft.CognitiveServices/accounts",
      "location": "[parameters('location')]",
      "kind": "CustomVision.Training",
      "sku": {
        "name": "[parameters('skuTraining')]"
      },
      "properties": {}
    },
    {
      "apiVersion": "2016-02-01-preview",
      "name": "[variables('namePrediction')]",
      "type": "Microsoft.CognitiveServices/accounts",
      "location": "[parameters('location')]",
      "kind": "CustomVision.Prediction",
      "sku": {
        "name": "[parameters('skuPrediction')]"
      },
      "properties": {}
    }
  ]
  ...
}

We can pay attention to two things here:

• The training and prediction services are declared with the following type: Microsoft.CognitiveServices/accounts.
• The two services are differentiated with the kind attribute: CustomVision.Training and CustomVision.Prediction.

And to finish, we will output the API keys for the training service and prediction service:

{
  ...
  "outputs": {
    "Training API Key": {
      "type": "string",
      "value": "[listKeys(resourceId('Microsoft.CognitiveServices/accounts', parameters('nameTraining')), providers('Microsoft.CognitiveServices', 'accounts').apiVersions[0]).key1]"
    },
    "Prediction API Key": {
      "type": "string",
      "value": "[listKeys(resourceId('Microsoft.CognitiveServices/accounts', variables('namePrediction')), providers('Microsoft.CognitiveServices', 'accounts').apiVersions[0]).key1]"
    }
  }
}

Example of use

The ARM template is now ready, let's open a Windows PowerShell and try it:

.\Deploy-AzureResourceGroup.ps1 -ResourceGroupName 'MyResourceGroupName' -ResourceGroupLocation 'southcentralus' -TemplateFile '.\azuredeploy.json'

...

OutputsString      :
                     Name             Type                       Value
                     ===============  =========================  ==========
                     training API Key  String                     bcxx2e598139477e975d71d688549c7c
                     prediction API Key  String                     90b1f3b84xxx4dfca342fd56d42df1dc

If everything goes well, you should see the same kind of output as above.

To go further

If you try to deploy the service again you will get the following error message:

Resource Microsoft.CognitiveServices/accounts 'myCustomVisionService' failed with message '{
   "error": {
     "code": "CanNotCreateMultipleFreeAccounts",
     "message": "Operation failed. Only one free account is allowed for account type 'CustomVision.Training'."
   }
}'

As stated in the message, you can't deploy multiple free Custom Vision service.

Summary

Today we have learned how to create an ARM template that will deploy a Custom Vision service and output the API keys for the training service and prediction service.

You can download the example solution here:

Browse the GitHub repository

Please feel free to comment or contact me if you have any question about this article.

• English

1. Microsoft
2. Azure
3. Cognitive Services
4. Vision

Back to articles