Image content risk control solution

Prerequisites

• You have activated the PAI (Designer, EAS) Pay-as-you-go service. For more information, see Activate and create a default workspace.

• A general computing resource quota is associated with the workspace. For more information, see Manage workspaces. To create a general computing resource quota, see Add a resource quota.

• You have created an Object Storage Service (OSS) bucket to store raw data, manifest files, and trained model files. For more information about how to create a bucket, see Create buckets.

Limits

OSS does not support cross-region access. Therefore, the OSS bucket that stores your data and the Machine Learning Designer workflow must be in the same region.

Procedure

The following procedure describes how to build an image content risk control solution using the Alibaba Cloud PAI platform:

1. Prepare the data

   Store the raw image data in OSS. Then, use PAI Data Management to scan the raw data and generate a manifest file. Finally, use iTAG to annotate the data and obtain an annotated dataset for model training.

   PAI provides a raw dataset that you can use for data preparation. To download the dataset, see Image classification dataset for content risk control or Object detection dataset for content risk control.

2. Build a model for image content risk control.

   In Machine Learning Designer, you can use a pre-trained model to build either a Resnet50 image classification model or an object detection model for risk detection. Select the appropriate model based on your business scenario:

   • Image Classification Training (torch)

     If your business scenario requires classifying entire images into risk categories, you can build an image classification model.

   • Image Detection Training

     If your business scenario requires detecting and drawing bounding boxes around high-risk entities in images, you can build an object detection model.

3. Deploy and invoke the model service

   You can use EAS to deploy the trained content risk control model as an online service and invoke it for inference in your production environment.

Prepare the data

This section uses a public PAI dataset as an example. You can follow the same steps to prepare your own dataset.

1. Divide the raw images into a training dataset and a test dataset, and upload them to the OSS bucket that you created. For more information about how to upload files to OSS, see Upload files.

2. Use PAI Data Management to scan the raw data and generate a .manifest file. For more information, see Create and manage datasets.

3. Use the iTAG administrator console to create an annotation task. For more information, see Create an annotation task.

4. Use the iTAG annotator console to annotate the data. For more information, see Process an annotation task.

5. On the iTAG page, click the Task Hub tab. Find the completed annotation task and click Get Annotation Result in the Actions column to retrieve the annotated dataset, which is stored in the specified OSS folder.

Deploy and invoke the model service

You can use EAS to deploy the trained image classification or object detection model as an online service and invoke it for inference in your production environment.

1. Go to the Elastic Algorithm Service (EAS) page.

   1. Log on to the PAI console.

   2. In the left-side navigation pane, click Workspaces. On the Workspaces page, click the name of the workspace that you want to manage.

   3. In the left-side navigation pane, choose Model Deployment > Elastic Algorithm Service (EAS).

2. Deploy the model service.

   1. On the Elastic Algorithm Service (EAS) page, click Deploy Service. In the dialog box that appears, select Custom Deployment and click OK.

   2. On the Custom Deployment page, configure the following parameters and click Deploy. This topic describes only the core parameters. For information about other parameters, see Custom Deployment.

      Parameter	Description
      Service Name	Enter a custom name for the model. We recommend that you name it based on your business to distinguish it from other model services.
      Deployment Method	Select Deploy with Processor.
      Model Configuration	The models trained in this example are stored in OSS. Therefore, set the configuration type to Object Storage Service (OSS). In the workflow data storage directory that you configured in the Build a model step, select the OSS path of the epoch_xx_export.pt model file.
      Processor Type	Select EasyCV.
      Model Type	This solution uses Image Classification. If you are deploying an image classification model, select Image Classification. If you are deploying an object detection model, select Object Detection.
      Deployment Resources	This solution uses the ecs.gn7i-c16g1.4xlarge resource specification.

3. View the public endpoint and access token of the model service.

   1. On the Elastic Algorithm Service (EAS) page, click the name of the target service. In the Basic Information area, click View Endpoint Information.

   2. In the Endpoint Information panel, view the endpoint and token for public network access.

4. Use a script for batch invocation.

   1. Create a Python script locally to invoke the model service.

      • Python script for image classification (cv_risk_cls.py):

        import requests
        import base64
        import sys
        import json
        resp = requests.get('<image_url>')
        ENCODING = 'utf-8'
        datas = json.dumps( {
                    "image": base64.b64encode(resp.content).decode(ENCODING),
                    })
        head = {
           "Authorization": "YjdhYWRhYWZhYzNjZTFlMDZlNjAxxxxxxxxxxxxxxxxxxxx" # Replace the service access token with your actual token.
        }
        for x in range(0,1):
            # Replace the public endpoint of the service with your actual endpoint.
            resp = requests.post("<service_url>", data=datas, headers=head)
            print(resp.text)
        print("test endding")

        Replace <image_url> with the URL of the image, such as https://xxxx/1.jpg. Set Authorization to the service token that you obtained. Replace <service_url> with the service endpoint that you obtained.

      • Python script for object detection (cv_risk_det.py):

        import requests
        import base64
        import sys
        import json
        img_file = './xxx.jpg'
        ENCODING = 'utf-8'
        datas = json.dumps( {
                    "image": base64.b64encode(open(img_file, 'rb').read()).decode(ENCODING),
                    })
        head = {
           "Authorization": "NGVkMTVmZjNlNzA3ZGVlNWIzxxxxxxxxxxxxxx" # Replace the service access token with your actual token.
        }
        # Replace the public endpoint of the service with your actual endpoint.
        r = requests.post("<service_url>", data=datas, headers=head)
        print(r.text)

        Set img_file to the path of the image file. Set Authorization to the service token that you obtained. Replace <service_url> with the service endpoint that you obtained.

   2. Upload the Python script for image classification or object detection to an environment. In the directory where you uploaded the script, run the following command.

      $ python3 <cv_risk_xxx.py>

      Replace <cv_risk_xxx.py> with the actual name of your Python script.

5. Monitor service metrics.

   After you invoke the model service, you can view the invocation metrics, such as queries per second (QPS), response time (RT), CPU usage, GPU usage, and memory usage.

   1. On the Elastic Algorithm Service (EAS) page, click the monitoring icon in the Invocation/Logs/Monitoring column of the service that you invoked.

   2. On the Monitoring tab, you can view the model invocation metrics.