PAI Visualized Modeling: Enroute from Development to Deployment

The model deployment section below has Elastic Algorithm Service (EAS) which houses the deployment models already available and acts as a platform to deploy the customized models created by the users.

Visualized Modeling is a drag and drop service available with Platform for AI (PAI) for developing and deploying machine learning models. This helps the users to conceptualize the flowchart they developed into a fully functional machine learning model and applications. The development functionality is by connecting the functional blocks available with the library and it is similar to connecting Lego blocks. Once the blocks are connected, the parameters are fine tuned in order to get the pipeline of functionalities get aligned in the way it is designated to work. This blog walks through the procedures to develop a machine learning model, deploy the same using Elastic Algorithm Service (EAS) and using the deployed model with API connectivity.
Step 1: Get familiarized with the Visualized Modeling tool
Get into the PAI console and in the left pane, the Visualized Modeling (Designer) will be available under model development and training.
blogpic
The model deployment section below has Elastic Algorithm Service (EAS) which houses the deployment models already available and acts as a platform to deploy the customized models created by the users. When you scroll down a bit, the AI Computing Asset Management will be found. This section manages the datasets, models, images, jobs etc., as resources. We can add our custom datasets and model files as we store in the OSS in this asset management.
blogpic
Step 2: Development of Visualized Modeling
Click on the Visualized Modeling (Designer) in the left pane under the Model Development and Training.
blogpic
The Preset Templates tab has some preloaded pipelines developed and ready to use. While clicking on create pipeline under Pipelines, you can create a new pipeline or load pipeline from files saved. While the pipeline is created, click on the pipeline file. Then browse through the left pane to look for the components available. Drag and drop the components into the pipeline area and establish connections between them as needed. Here in this blog, I have loaded the heart disease prediction pipeline.
blogpic
Click on each component and set the parameters corresponding to the nodes created and connected. After all the blocks connected are confirmed to work good enough, we need to create PMML file.
Step 3: Development of Machine Learning model
Most of the machine learning models available with visualized modeling support PMML format. Click on to the machine learning algorithm block in the pipeline. Here it is logistic regression.
blogpic
Check on the “Whether to Generate PMML” in the right pane. Now it is set to develop the machine learning model. Now click on somewhere outside of the blocks and observe the right pane.
blogpic
Configure the right pane to store the corresponding files to an OSS bucket you have designated for this purpose. Now click on the green play button on the left top corner of the pipeline area. This will run the processes designated in the pipeline. This execution process will give a look and feel experience. The entire flow of the pipeline can be monitored by the flow it visualizes. The green tick on the blocks represents the blocks which has completed execution and the green dotted lines on the connections show the ongoing execution.
blogpic
Click on the task button on the right top corner of the pipeline to watch over the ongoing processes.
blogpic
Until the outputs are available, the visualization button on the center top of the pipeline remains disabled. We got to wait until the blocks completed execution. Upon completion of execution, the visualization button becomes active.
blogpic
When the visualization button is clicked, it displays all the graphs and metric statistics generated.
blogpic
Then we can save the pipeline. As we enabled the PMML generation earlier, the model file is generated already. This will be seen active by clicking on the ‘Models’ menu available above the pipeline area.
blogpic
Step 4: Registering the model
Depending on the number of machine learning algorithms available, model files in PMML gets generated and displayed in models tab. Select the algorithm from which you need the model to be saved and click “Register Model”.
blogpic
Enter the name to save the model and scroll down. Check on all the parameters available and click on “Determine”. Now the machine learning model generated will get stored under the “Models” section.
blogpic
We can deploy this model by clicking on “Deploy to EAS” option and we can deploy by getting into the EAS section as well. On the console of PAI’s left pane, click on Elastic Algorithm Service (EAS) under Model Deployment.
blogpic
Click on Deploy Service and proceed further.
blogpic
Enter a name for the service. Under deployment method, choose “Deploy Service by Using Model and Processor”. In Model File, click “Select Model” Then specify the name of the model registered earlier and its version. Then for the Processor Type, select “PMML”. Now scroll down further and enter the details as needed.
blogpic
Choose the CPU/GPU required as per the application scope and choose the VPC and security groups.
blogpic
The configuration is auto generated and it can be saved as JSON format. The JSON file can be loaded to create new deployment too. Then Click on Deploy.
blogpic
To confirm the deployment of model service, click “OK”.
blogpic
The deployment is done. We need to wait until the service is created. Upon creation when we click on Invocation method, it shows Public Endpoint and a Token. These two information are needed to call the model from application as API call.
Step 5: API call on the model invocation using public endpoint : Using Jupyter Notebook in local PC/Laptop or PAI DSW
Open Jupyter Notebook on your PC or Laptop. Create a new notebook.
blogpic
Install the library function of eas-prediction using pip to ensure the endpoints can be used.
blogpic