MaaS: the Models as a Service API

The goal of the Models as a Service (MaaS) project is to provide an easy to use, descriptive middleware layer API to facilitate model search and discovery, exploration, configuration, and execution. MaaS support an array of agricultural, economic, hydrological, weather, and health models. MaaS provides two key innovations:

1. MaaS allows users to to work with a single lightweight and well-documented API to execute a variety of models
2. MaaS extracts model output from various proprietary or complex formats and provides it to users in a normalized, tabular format

MaaS allows users to interact with a variety of expert models across a wide ranging parameter space, enabling the creation of sophisticated sensitivity analyses such as the one above, which was made using a 3rd party tool, but powered by MaaS.

Contents

• Design
• Available Models
• Architecture
• Documentation
• Collaborators

Design

Model discovery

Model exploration is managed through the Exploration Controller. The Exploration Controller allows a user to obtain a model’s description, understand its parameters, obtain an example/default configuration, and understand its outputs.

This is further enabled through the Concepts Controller which maps concepts to models, model outputs, and model parameters. A client may request a list of all concepts that are mapped to models in MaaS and may submit a single concept to learn which related models, outputs and parameters are tied to it in MaaS. This is powered by Eidos, an open-domain machine reading system designed by the Computational Language Understanding (CLU) Lab at University of Arizona.

Model execution

Model execution is managed by the Execution Controller. Models consist of pre-built Docker images that are hosted on an arbitrary server. Running a model requires the creation of a specific model controller, such as this one for Kimetrica’s malnutrition model. The model controller is responsible for obtaining a model configuration and tasking Docker to run the model image inside a container with the given configuration. The model controller specifies a Docker container entrypoint, such as this one. The model controller is then responsible for storing the output results file(s) to S3 and ingesting the normalized results into the MaaS database.

Models may be run using the /run_model endpoint. For more information on model execution, refer to model-execution.

Available Models

Currently, MaaS supports the following models:

Architecture

Collaborators

MaaS collaborators include:

• Atlas AI
• UCSB
• University of Florida
• USC Information Sciences Institute
• Penn State
• Kimetrica
• Columbia University
• CSIRO
• Potsdam Institute for Climate Impact Research