dummy-link

SemanticModels

A julia package for representing and manipulating model semantics

Readme

SemanticModels.jl

A julia package for representing and manipulating models at the semantic level.

Traditional scientific computing happens by translating conceptual models of natural phenomena into mathematical models on a chalkboard and then implementing those models in code that is then compiled into executable instructions and run on a machine. However, changes to these models traditionally require modelers to go back to the drawing board and change the conceptual and mathematical model before implementing new software to analyze the new model. The new software is always built by changing the old software until you build up enough cruft to declare it legacy code and start over. SemanticModels changes this by representing models at a semantic level and allowing programs to be expressed as transformations on these models.

SemanticModels Diagram

The domains of software security and programming language theory (PLT) have spent a lot of time developing software and theory for the analysis of computer programs, but these tools have not been adopted by the scientific community. This is because the tools understand the programs as software, without consideration of the conceptual and mathematical structure above them. SemanticModels.jl addresses this problem.

General purpose solvers such as Jump and Stan introduce domain specific languages to describe the problems that they can solve. This is a great step in the right direction because the DSL often contains the semantic structures of the modeling domain embedded in the language. If all scientific software was written in these DSLs we would be able to apply program analysis to the models and enable powerful program transformations to build better systems for scientists and enable AI algorithms to write scientific codes. Packages like ModelingToolkit.jl, which builds a tools to design these DSLs will help achieve that vision.

SemanticModels takes an alternative approach, which is to learn the DSL from actual usage of the libraries. Every software library defines an implicit embedded DSL for its users. We aim to leverage that fact, along with large collections of open source software to learn the modeling frameworks from the corpus of code.

Getting Started

Install this package with

Pkg.add("SemanticModels")
Pkg.test("SemanticModels")

Note that running the tests for the first time can take a while because DifferentialEquations is a large library that requires a long precompilation step. Various functions in the SemanticModels.Dubstep module can also have long precompile times, due to heavy use of generated functions.

Then you can load it at the julia REPL with using SemanticModels

You should start exploring the notebooks in the examples folder. These notebooks are represented in jupytext format, and are stored as julia programs you can run at the repl or in the notebook interface after installing the jupytext plugin for jupyter.

  1. Model augmentation: an example script examples/agentgraft.jl shows how to augment an agent based simulation to add new modeling components using an API for changing models at the semantic level.

  2. Model Representations: SemanticModels supports extracting knowledge graph representations of scripts. See the examples/agenttypes2.jl notebook for a demonstration.

There are scripts in the folder SemanticModels/bin which provide command line access to some functionality of the package. For example julia --project bin/extract.jl examples/epicookbook/notebooks/SimpleDeterministicModels/SEIRmodel.jl will extract code based knowledge elements from the julia source code file examples/epicookbook/notebooks/SimpleDeterministicModels/SEIRmodel.jl.

See the tests and documentation for more example usage.

Docker

You can easily spin up a SemanticModels.jl Jupyterlab instance with docker.

  1. docker run -it --rm -p 88889:8888 jpfa/semanticmodels:stretch
  2. Navigate to the link it returns: localhost:8888/?token=...
  3. From there you can run the examples included in this repository, or write your own code to explore the functionality of SemanticModels.jl

Note: to open a .jl file as a notebook in the jupyterlab interface right click and select "Open in > Notebook".

Documentation

There is a docs folder which contains the documentation, including reports sent to our sponsor, DARPA.

Documentation is currently published https://aske.gtri.gatech.edu/docs/latest

Our documentation and examples are built with Jupyter notebooks. We use jupytext to support diff friendly outputs in the repo. Please follow the jupytext readme to install this jupyter plugin. If you use the docker container, jupytext is already installed.

Examples

In addition to the examples in the documentation, there are fully worked out examples in the folder https://github.com/jpfairbanks/SemanticModels.jl/tree/master/examples/. Each subdirectory represents one self contained example, starting with epicookbook.

Concepts

Here is a preview of the concepts used in SemanticModels, please see the full documentation for a more thorough description.

Model Augmentation

The primary usecase for SemanticModels.jl is to assist scientists in what we call model augmentation. This is the process of taking a known model developed by another researcher (potentially a past version of yourself) and transforming the model to create a novel model. This process can help fit an existing theory to new data, explore alternative hypotheses about the mechanisms of a natural phenomena, or conduct counterfactual thought experiments.

SemanticModels.ModelTool is the current home for this capability. You can call m = ModelTool.model(ExpAgentProblem, expr) to lift an agent based model up to the semantic level, then apply transformations on that m and then call eval(m.expr) to generate code for that new model. This allows you to compare different variations on a theme to conduct your research.

If you are working with ODEProblem or agent based odels, there are prebuilt types for representing those models, but if you want to add a new class of models you will just have to write:

  1. a struct T that holds a structured representation of instances of the model class
  2. extend ModelTool.model(::DataType{T}, expr::Expr) to extract that information from a code representation of your model
  3. a set of valid transforms that can be done to your model.

SemanticModels.jl provides library functions to help with steps 2 and 3 and functions for executing and comparing then outputs of different variations of the model.

We think of SemanticModels as a post hoc modeling framework the enters the scene after scientific code has been written. As opposed to a standard modeling framework that you develop before you write the scientific code.

Overdubbing

SemanticModels.Dubstep provides functionality for manipulating models at execution time. Where ModelTool allows you to manipulate models at the syntactic level, Dubstep allows you to manipulate their execution. This falls along similar lines of static vs dynamic analysis.

You can modify a program's execution using Cassette.overdub and replace function calls with your own functions. For an example, see test/transform/ode.jl. Or you can use a new compiler pass if you need more control over the values that you want to manipulate.

Knowledge Graphs

MetaGraphs.jl is used to model the relationships between models and concepts in a knowledge graph.

There are a few different forms of knowledge graphs that can be extracted from codes.

  1. The type graph: Vertices are types, edges are functions between types see examples/agenttypes2.jl.

  2. Vertices are functions and variables, edges represent dataflow, function references variable or function calls function.

  3. Conceptual knowledge graph from text, vertices are concepts edges are relations between concepts.

Acknowledgements

This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00111990008.

First Commit

11/01/2018

Last Touched

22 days ago

Commits

542 commits

Used By: