/omesa

Experiment, Storage and Visualization Framework for Machine Learning research.

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Omesa

Travis-CI Docs Landscape

Python 3.4

A small framework for easily deployable, reproducible machine learning research that largely builds on top the scikit-learn API. Its goal is to make common research procedures templated, optimized, and well recorded. To this end it features:

  • Flexible wrappers to plug in your tools and features of choice.
  • Sparse and multi-threaded pipeline through hashing and pooling - from data to feature space.
  • Storage of all settings and fitted parts of the entire experiment, promoting reproducibility.
  • Dump an easily deployable version of the final model for plug-and-play demos.
  • Visual overview in an web-environment for comparison and intepretation of experiments.

Read the documentation at readthedocs.

Important Note

This repository is currently in alpha development, and is therefore not stable.

Why use Omesa?

Omesa can be used in various scenarios:

  • As an end-to-end template-to-results framework.
  • As several fast classes to aid your machine learning workflow.
  • As a storage and overview for your experiments.

As such, you can vary between restricted options and little code, and many options (thus more code) that require some restricted storage format. Either way, Omesa aids in structuring and comparing experiments.

Front-end Preview

The front-end provides visualization and comparison of model performance. Currently only the 'Results' section works, preview below:

Front

Front Prop

To test, do the following:

$ cd /dir/to/omesa/examples
$ python3 n_gram.py
$ cd ../front
$ python3 ./app.wsgi

And follow the localhost link that is shown to access the web app. Please note that this part can be quite unstable. Bug reports are welcome.

Dependencies

Omesa currently heavily relies on numpy, scipy and sklearn. To use Frog as a Dutch back-end, we strongly recommend using LaMachine. For English, there is a spaCy wrapper available.

Omesa Only - End-To-End In 2 Minutes

With the end-to-end Experiment pipeline and a configuration dictionary, several experiments or set-ups can be ran and evaluated with a very minimal piece of code. One of the test examples provided is that of n-gram classification of Wikipedia documents. In this experiment, we are provided with a toy set n_gram.csv that features 10 articles about Machine Learning, and 10 random other articles. To run the experiment, the following configuration is used:

Example

With the end-to-end Experiment pipeline and a configuration dictionary, several experiments or set-ups can be ran and evaluated with a very minimal piece of code. One of the test examples provided is that of n-gram classification of Wikipedia documents. In this experiment, we are provided with a toy set n_gram.csv that features 10 articles about Machine Learning, and 10 random other articles. To run the experiment, the following configuration is used:

    from omesa.experiment import Experiment
    from omesa.featurizer import Ngrams
    from omesa.containers import Pipe
    from omesa.components import Vectorizer, Evaluator
    from sklearn.naive_bayes import MultinomialNB

Experiment(
    project="unit_tests",
    name="20_news_grams",
    data=[CSV("n_gram.csv", data="intro", label="label")],
    pipeline=[
        Vectorizer(
            features=[
                Ngrams(level='char', n_list=[3])
            ]),
        Pipe('clf', MultinomialNB()),
        Evaluator(scoring='f1', average='micro',
                  lime_docs=CSV("n_gram.csv", data="intro", label="label")),
    ],
    save=("model", "db")
)

This will cross validate performance on the .csv, selecting text and label columns and indicating a header is present in the .csv document. We provide the Ngrams function and parameters to be used as features, and store the log.

Output

The log file will be printed during run time, as well as stored in the script's directory. A sample from the output of the current experiment is as follows:

---- Omesa ----

 Config:

        feature:   char_ngram
        n_list:    [3]

    name: gram_experiment
    seed: 42

 Sparse train shape: (20, 1301)

 Performance on test set:

             precision    recall  f1-score   support

         DF       0.83      0.50      0.62        10
         ML       0.64      0.90      0.75        10

avg / total       0.74      0.70      0.69        20


 Experiment took 0.2 seconds

----------

Adding own Features

Here's an example of the most minimum word frequency feature class:

class SomeFeaturizer(object):

    def __init__(self, some_params):
        """Set parameters for SomeFeaturizer."""
        self.name = 'hookname'
        self.some_params = some_params

    def transform(self, raw, parse):
        """Return a dictionary of feature values."""
        return Counter([x for x in raw])

This returns a {word: frequency} dict per instance that can easily be transformed into a sparse matrix.

Acknowledgements

Part of the work on Omesa was carried out in the context of the AMiCA (IWT SBO-project 120007) project, funded by the government agency for Innovation by Science and Technology (IWT).