Important notes

Invocation

This model is expected to be run as real-time inference only and does not support batch transform. Unlike classical ML model, this product supports both making inference and training on the same deployed endpoint.

General usage steps

1. Add texts to the system.
2. Facilitate learning (training) or prediction (inference).
   • For training, assign inter-phrase relationships for the newly added phrase and existing (previously imported) phrases.
   • For prediction, call a dedicated method to get list of the top phrases ordered by semantic similarity with input phrase. To retrieve predicted category from the returned data you can use a wrapper call as shown in the example below.

Limitations and Processing times

To achieve optimal performance the following limits should be met:

• Each message should be up to 100 words and 1000 characters long.
• Total number of messages in the system should not exceed 10 000.

For ml.m5.4xlarge instance benchmarks are as follows:

• Importing of 100-message categorized corpus takes 1hr.
• Prediction for a single message takes 3 to 30 seconds.

Invoking endpoint

Deploying the model endpoint (more detailed example can be found in sample notebook):

import dynamic_ai as dynai
model_arn = dynai.get_model_package_arn()
dynai.deploy_model(model_arn=model_arn, instance_type='ml.m5.4xlarge')

Real-time training snippet using categories:

msg_id = dynai.add_message('Very simple test message')
dynai.set_category(msg_id, 'Test')

Real-time category prediction:

msg_id = dynai.add_message('Hey, here I am - a new test message')
dynai.predict_category(id)

Resources

• Python helper library
• Sample notebook
• Demo video