Skip to content

Latest commit

 

History

History
64 lines (54 loc) · 3.75 KB

README.md

File metadata and controls

64 lines (54 loc) · 3.75 KB
Raw

RembrandtML     

RembrandtML is an intuitive machine learning API.

It can be used as an instructional model of robust coding practices applied to ML tasks.

The project intends to make the ML workflow easier to understand and implement by abstracting common ML concepts and tasks into entities which can be easily manipulated in code.

Key features include:

  1. Test-first development is modeled.
  2. The flexible, modular design makes modifying and adding implementations easy.
  3. The code is fully instrumented for comprehensive logging and time measurement.
  4. Methods to tune model hyperparameters.
  5. Plotting accuracy scores of many configurations for comparison.
  6. Feature management makes it simple to add and remove features to tune the model.

Getting Started:

  1. Work through the Quickstart
  2. Find a test that covers an aspect of ML and a framework that you want to learn about. There are lots of examples that demonstrate
    • How to load scikit-learn data
    • How to load data from a csv file using Pandas
    • How to use Linear Regression using both scikit-learn and TensorFlow
  3. Call the test from a test runner or test_runner.py
  4. Step through the code in a debugger

#More advanced software engineer techniques:

Dependency Injection

A logger and time are used by all custom types in the project. These services are provided to each object through the Instrumentation singleton.

Custom Errors

While it is a trivial savings a keystrokes, the custom FunctionNotImplementedError demonstrates how to extend Errors for customized functionality.

Design Patterns

The DataProvider classes give an example of the Template Patterns.

The abstract base class defines the algorithm of retrieving data from a dataset.

Each concrete subclass overrides methods when customized functionality is required.

For example, training data and label data is accessed very differently with a scikit-learn Bunch compared to a Pandas DataFrame. The scikit-learn Bunch object stores the label data(y) in ndarray accessible through the 'target' key in the Bunch. If the data was loaded from a csv into a Pandas DataFrame, the label data needs to be accessed by feature name and removed from the training data explicitly.

Initialization Steps

  1. Create DataConfig
  2. Create ModelConfig
  3. Create ContextConfig
  4. Create Context using ContextFactory.create(context_config)

Initialization Implementation

  1. ContextFactor.create() instantiates
    1. Logger
    2. Instrumentation
    3. DataContainer
  2. ContextFactory.create() calls
  3. ModelFactory.create(), which instantiates
    1. Model
  4. The model constructor instantiates
    1. ModelImpl for MLSingleModelBase
    2. ModelImpl collection for MLEnsembleModelBase
      • Since an ensemble model may need a collection of estimator models before it is initialized, such as Scikit Learn VotingClassifier, different subclasses handle ModelImpl initialization.

Test-First Development

The implementation of ensemble models is a good example of test-first development.

  1. The first step is create a series of test to ensure proper initialization.
    1. See TestEnsembleModels as an example
    2. Testing error conditions is important to ensure bothe that proper validation is happening and that errors are being properly reported.
      • This is a good time to think about what your assumptions are about the state of the data and the code. Add a few tests to verify those assumption. This will make troubleshooting much easier in the future.
      • See TestEnsembleModels.test_voting_sklearn_estimators_error()
  2. Next, implement each piece of functionality to that the tests pass.