Advanced Features

The AI Model Share library can help deploy interactive web dashboards for making predictions with your models in minutes.

However, sometimes there is a need for higher levels of customization than the base features include.

This page describes additional pathways to allow for the greatest programming flexibility, while also leveraging the power of the AI Model Share library.

Custom Deployments

The base ModelPlayground.deploy method deploys a pre-written lambda handler optimized for efficiency with specific types of prediction domains.

For projects requiring more flexibility in the lambda handler, the AI ModelShare library allows for “Custom Deployments”. Custom Deployments allow users to leverage the AI ModelShare infrastructure through AWS, while also creating opportunities for additional customization.

Tutorial

Guide to Custom Deployments

Using PySpark

AI Model Share supports PySpark. Note that the current prediction API runtime can only accept Pandas Dataframes, which may require additional steps for PySpark preprocessors.

Tutorial

Quick Start Tutorial with PySpark

Connecting Custom Web Apps

AI Model Share allows users to leverage the AI Model Share deployment infrastructure to power their own custom web apps. Web apps can be displayed through the AI Model Share website and be highlighted as part of a developer’s portfolio.

Users can connect their web apps in 3 easy steps:

1. Deploy a Model Playground (See tutorials HERE).

2. In the code for your web app, set the path for your Model Playground’s prediction API (built for you by AI Model Share in the “deploy” step) as the endpoint for the API request.

   Note

   Owners can find their Model Playground’s API URL on the “Predict” tab of their Model Playground.

   AI Model Share API URLs follow this format: “https://example.execute-api.us-east-1.amazonaws.com” (Remove “/prod/m” from the end of the url string.)

   

3. Authorization tokens are generated for users when they log in to the AI Model Share website. Unpack the token parameter within your Streamlit code, then format the headers in your API call to expect a token as a query parameter.

   auth_token = st.experimental_get_query_params()['token'][0]
   
   headers = {
       "Content-Type": "application/json",
       "authorizationToken": auth_token,
   }
   

Done! Design your web-app to customize the AI Model Share infrastructure to your own needs. See examples below:

Examples

• Text Classification: Webapp

• Text Classification: Github

• Tabular Classification: Webapp

• Tabular Classification: Github

• Image Classification: Webapp

• Image Classification: Github

Model Reproducibility

AI Model Share encourages users to share, replicate, and build on each other’s work by offering full model reproducibility functionality. Users can leverage Competitions & Experiments as a way to exchange trained & untrained models.

To share a reproducible model, take the following steps:

1. Export the model’s reproducibility environment into a json file with the export_reproducibility_env function. This function captures all of the necessary information to exactly reproduce a machine learning model.

   Example:

   from aimodelshare import export_reproducibility_env
   
   mode = "gpu" # or "cpu", depending on model type
   seed = 2021
   
   export_reproducibility_env(
    seed=seed,
     directory="", #use "" for current directory
     mode=mode,
   )
   

Note

The captured reproducibility environment only applies for one training iteration (data prep, preprocessing, fit model, submit model), so it is recommended to only train one model per training iteration.

2. Build your model with your preferred ML library.

3. Submit model with the reproducibility_env_filepath argument set.

   Example:

   #Submit Reproducible Model:
   reproducibility_env_filepath="reproducibility.json"
   
   # generate predicted y values (for keras models)
   y_pred = model.predict(X_test).argmax(axis=1)
   prediction_labels = [y_test.columns[i] for i in y_pred]
   
   # Submit Model to Competition Leaderboard
   mycompetition.submit_model(model_filepath = "model.onnx",
              preprocessor_filepath="preprocessor.zip",
              prediction_submission=prediction_labels,
              reproducibility_env_filepath=reproducibility_env_filepath)
   

To instantiate a previously submitted reproducible model, use the Competition.instantiate_model method.

Example:

reproducible_model = mycompetition.instantiate_model(version=1, reproduce=True)