Platform For AI:Deploy a Hugging Face model in EAS

Background information

Open source model communities, such as Hugging Face, provide various machine learning models and code implementations. You can access the encapsulated models, frameworks, and related processing logic through interfaces of related libraries. You can perform end-to-end operations such as model training and calling with only a few lines of code without the need to consider environment dependencies, preprocessing and post-processing logic, or framework types. For more information, visit the official site of Hugging Face. This ecology is an upgrade of deep learning frameworks such as TensorFlow and PyTorch.

EAS provides optimized capabilities to facilitate community model deployment.

Deploy a Hugging Face model

PAI allows you to deploy a model in the pipelines library of Hugging Face as an online model service in EAS. Perform the following steps:

1. Select the model that you want to deploy from the pipelines library. In this example, the distilbert-base-uncased-finetuned-sst-2-english model is selected. On the details page of the distilbert-base-uncased-finetuned-sst-2-english model, obtain the model ID, task, and revision information, and save them to your on-premises machine.

   The following table provides the mapping relationship between the TASK that is displayed on the Hugging Face page and the TASK parameter that you need to specify when you deploy services in EAS:

   TASK displayed on the Hugging Face page	TASK to specify in EAS
   Audio Classification	audio-classification
   Automatic Speech Recognition(ASR)	automatic-speech-recognition
   Feature Extraction	feature-extraction
   Fill Mask	fill-mask
   Image Classification	image-classification
   Question Answering	question-answering
   Summarization	summarization
   Text Classification	text-classification
   Sentiment Analysis	sentiment-analysis
   Text Generation	text-generation
   Translation	translation
   Translation (xx-to-yy)	translation_xx_to_yy
   Text-to-Text Generation	text2text-generation
   Zero-Shot Classification	zero-shot-classification
   Document Question Answering	document-question-answering
   Visual Question Answering	visual-question-answering
   Image-to-Text	image-to-text

2. On the EAS page, deploy the Hugging Face model.

   1. On the Create Service page, configure the following parameters. For more information about other parameters, see Model service deployment by using the PAI console.

      

      Parameter	Description
      Service Name	Specify a name for the service.
      Deployment Method	Select Deploy Web App by Using Image.
      Select Image	Click PAI Image, select huggingface-inference from the drop-down list, and then select an image version based on your business requirements.
      Environment Variable	Use the following information that is obtained in Step 1: MODEL_ID: distilbert-base-uncased-finetuned-sst-2-english TASK: text-classification REVISION: main
      Command to Run	The system automatically configures the command to run after you configure the Image Version parameter. You do not need to modify the setting.
      Additional System Disk	Click Extra System Storage in the Resource Deployment Information section and set the Additional System Disk parameter to 100 GB.

   2. Click Deploy. If Service Status changes to Running, the service is deployed.

3. Call the deployed model service.

   Use the console to call the service

   • On the EAS page, click View Web App in the Service Type column to verify the model inference effect on the web UI.

   • Click Online Debugging in the Actions column of the service. On the Body tab, enter the request data, such as, {"data": ["hello"]}, and click Send Request.

     Note

     The format of the input data ({"data": ["XXX"]}) of a text classification model is defined by /api/predict of the Gradio framework. If you use other types of models, such as image classification or speech data processing, you can construct the request data by referring to /api/predict.

   Use APIs to call the service

   1. Click the service name to go to the Service Details tab. On the tab that appears, click View Endpoint Information.

   2. On the Public Endpoint tab of the Invocation Method dialog box, obtain the values of URL and Token.

   3. Use the following code to call the service:

      import requests
      
      resp = requests.post(url="<service_url>",
                    headers={"Authorization": "<token>"},
                    json={"data": ["hello"]})
      
      print(resp.content)
      
      # resp: {"data":[{"label":"POSITIVE","confidences":[{"label":"POSITIVE","confidence":0.9995185136795044}]}],"is_generating":false,"duration":0.280987024307251,"average_duration":0.280987024307251}

      Replace <service_url> and <token> with the values obtained in Step b.