Platform For AI:Image classification

Prerequisites

Create an OSS bucket. For more information, see Create buckets.

Step 1: Open the model details page

1. Go to the Model Gallery page.

   1. Log in to the PAI console.

   2. In the upper-left corner, select a region.

   3. In the left-side navigation pane, click Workspaces. Click the name of the workspace that you want to open.

   4. In the left-side navigation pane, choose Quick Start > Model Gallery.

2. On the Model Gallery page, in the Computer Vision section, click Image Classification. Then, click the ViT Image Classification-General model card to open the model details page.

   

   On the model list page, multiple models from different open-source communities are available. Choose a model that meets your requirements.

   • Models with more parameters perform better but require more data for fine-tuning and incur higher serving costs.

   • The closer the pretrained dataset is to your use case, the better the model performs. Check the model details page for pretrained dataset information.

Step 2: Deploy the model directly

The cv_vit-base_image-classification_ImageNet-labels model is from the ModelScope community's ViT Image Classification-General model. It is based on the Transformer ViT Base architecture and trained on the ImageNet-1K dataset. It supports classification tasks covered by the ImageNet-1K labels. Model Gallery deploys this model to Elastic Algorithm Service (EAS) to create inference services. Follow these steps:

1. Deploy the service.

   1. On the model details page, click Deploy.

   2. Confirm the Model service info and Deployment info.

      Model Gallery pre-configures computing resources and a service name based on model characteristics. Use the default settings for this example, or customize them. For more information, see Deploy and debug models.

   3. Click Deploy. In the Billing Notification dialog box, click OK.

      The page redirects to the Service details page. In the Basic Information section, wait until the Status changes to Running.

2. Call the model service.

   After deployment, send prediction requests to the model service HTTP API. The service accepts Base64-encoded images and returns results in JSON format.

   1. On the Service details page, in the Resource Information section, click View Call Information.

      

   2. In the Call Information dialog box, on the Public network address call tab, save the Endpoint and Token.

   3. Use the Python Requests library to call the model service. Example:

      import requests
      import base64
      
      image_path = "<PathToLocalImage>"
      url = "<PredictionServiceURL>"
      token = "<PredictionServiceAccessToken>"
      
      
      def encode_file_to_base64(f):
          with open(f, "rb") as file:
              encoded_string = base64.b64encode(file.read())
              base64_str = str(encoded_string, "utf-8")
              return base64_str
      
      base64_string = encode_file_to_base64(image_path)
      request_body = {
          "image": base64_string
      }
      
      headers = {"Authorization": token}
      resp = requests.post(url=url, headers=headers, json=request_body)
      
      print(resp.content.decode())
      print("status code:", resp.status_code)
      

      Replace url with the Endpoint, token with the Token from the previous step, and image_path with the path to a local image file (.png or .jpg).

      A successful call returns the top five most likely labels and their scores. Sample output:

      {
        "scores": [
          0.4078965485095978,
          0.24673610925674438,
          0.1930493414402008,
          0.0026617543771862984,
          0.0009246605914086103
        ],
        "labels": [
          "tiger cat",
          "tabby, tabby cat",
          "Egyptian cat",
          "lynx, catamount",
          "tiger, Panthera tigris"
        ]
      }

Step 3: Fine-tune the model

The cv_vit-base_image-classification_ImageNet-labels model supports classification tasks covered by ImageNet-1K labels. If your target domain is not covered, fine-tune the model with a small set of labeled images for your specific domain.

1. Prepare data.

   1. Prepare your dataset according to the following directory structure:

      The model trains on image data from an OSS bucket. Organize the dataset in a root directory with one subdirectory per category. Each subdirectory name serves as the category label. For example, if the training dataset is at oss://{YourOssBucket}.{OssEndpoint}/{PathToTrainData}/, the structure must be:

      ├── category-1
      │   ├── image1.jpeg
      │   └── image2.jpeg
      ├── category-2
      │   ├── image3.jpeg
      │   └── image4.jpeg
      |...
      |...
      └── category-n
          ├── imagexxx.jpeg
          └── imageyyy.jpeg

      For better performance, prepare a validation dataset with the same directory structure to evaluate the model and tune hyperparameters.

   2. Upload the dataset to your OSS bucket using the directory structure described above. For more information, see Upload files.

2. Submit a training job.

   1. On the Model Details page, click Fine-tune. On the configuration page, specify the Training dataset, Validate dataset, and Output Path.

      Model Gallery pre-configures Computing Resources and Hyperparameters. Use the defaults or customize them. For more information, see Model deployment and training.

      Note

      By default, the model uses the MiniImageNet-100 dataset for fine-tuning, so you can test the training process without preparing your own data.

   2. Click Fine-tune.

      

      The page redirects to the Task details page, where you can monitor progress, logs, and evaluation results. After training completes, the model is saved to the Output Path you specified.

      

3. Deploy the fine-tuned model.

   PAI automatically registers the trained model in AI Assets - Model Management for viewing and deployment. For more information, see Register and manage models.