Kiwi Model Analyzer

Kiwi is a port of the excellent LIME algorithm. It is a Python library for use in scripts/notebooks but the hard numerical processing is implemented in Rust for performance.

It is particularly designed for processing large numbers of records in a timely and efficient manner. The current benchmark is the adult dataset with XDGBoost, currently standing at ~10_000 records per minute.

Currently only classification of tabular data is supported.

Building

Currently there is no packaged solution for building Kiwi, here is a list of instructions which should work cross platform.

1. Obtaining tools

You will need:

- The Rust compiler
- Python development headers
- A BLAS implementation (possibly with a fortran compiler)

You will first need a the Rust compiler, see here for details.

Documentation on obtaining Python headers is here.

Avaliable BLAS implementations are documented here. By default this library uses openblas-static for building. (And thus, from the documentation requires the GNU C compiler and the GNU fortran compiler as well).

2. Building

Having obtained all these tools, this should work.

  $ git clone https://github.com/jameswelchman/kiwi.git
  $ cd kiwi
  $ cargo build --release

Assuming the build was succesful, one can now "install" the library to the main repo path.

Linux

  $ cp target/release/libkiwi.so kiwi.so

OS/X

  $ cp target/release/libkiwi.dylib kiwi.so

Win32

  $ cp target/release/kiwi.dll kiwi.pyd

Getting started

There are two notebooks iris and the more in-depth adult. These two notebooks descibe most of the tabular API for Kiwi.

API Docs

Currently API notes are maintained in the file lib.rs.

The author is optimistic about the pyo3/sphinx ecosystem providing a better solution in the future.

Troubleshooting

Kernel Crashed / Memory allocation error with explain_instance_iter

Try reducing the max_memory keyword parameter from the default.

    explainer = KiwiTabularExplainer(training_data)
    explainer.explain_instance_iter(data, predict_fn, max_memory=16_000_000)

explain_instance_iter running very slowly

Try using the sample_background_thread=True when creating the tabular explainer. Usually this decreases performance, however it never falls below a certain minimum.

    explainer = KiwiTabularExplainer(training_data, sample_background_thread=True)
    explainer.explain_instance_iter(data, predict_fn)