PyDecode is a dynamic programming toolkit developed for research in natural langauge processing. Its aim is to be simple enough for fast prototyping, but efficient enough for research use.

.. _documentation: http://pydecode.readthedocs.org/

.. image:: _images/parsing_9_0.png :width: 500 px :align: center

| |

Features

• Simple specifications. Dynamic programming algorithms specified through pseudo-code. ::

  Viterbi algorithm.

  ... c.init(items[0, :]) for i in range(1, n): for t in range(len(tags)): c.set(items[i, t], items[i-1, :], labels=labels[i, t, :]) graph = c.finish()

• Efficient implementation. Core code in C++, python interfaces through numpy. ::

  Compute path.

  label_weights = numpy.random.random(graph.label_size) weights = pydecode.transform_label_array(graph, label_weights) path = pydecode.best_path(graph, weights)

• High-level algorithms. Includes a set of widely-used algorithms. ::

  Inside probabilities.

  inside = pydecode.inside(graph, weights, kind=pydecode.LogProb)

  (Max)-marginals.

  marginals = pydecode.marginals(graph, weights)

  Pruning

  mask = marginals > threshold pruned_graph = pydecode.filter(graph, mask)

• Integration with machine learning toolkits. Train structured models. ::

  Train a discriminative tagger.

  perceptron_tagger = StructuredPerceptron(tagger) perceptron_tagger.fit(X, Y) Y_test = perceptron_tagger.predict(X_test)

• Visualization tools. IPython integrated tools for debugging and teaching. ::

  pydecode.draw(graph, paths=paths)

.. image:: _images/hmm.png :width: 500 px :align: center

.. Documentation, Tutorial and Gallery .. ----------------------

.. .. hlist:: .. :columns: 2

.. * documentation_ .. * tutorial_ .. * gallery_ .. * api_

.. Features .. -------------

.. Currently the toolkit is in development. It includes the following features:

.. * Simple construction of dynamic programs. .. * Customizable GraphViz output for debugging. .. * Algorithms for best path, inside scores, outside scores, and oracle scores. .. * Several types of pruning. .. * Integration with an (I)LP solver for constrained problems. .. * Lagrangian Relaxation optimization tools. .. * Semiring operations over hypergraph structures. .. * Hooks into PyStruct for structured training. .. * Fast k-best algorithms.

.. .. image:: https://travis-ci.org/srush/PyDecode.png?branch=master .. :target: https://travis-ci.org/srush/PyDecode

.. _gallery: http://pydecode.readthedocs.org/en/latest/notebooks/index.html .. _tutorial: http://pydecode.readthedocs.org/en/latest/notebooks/index.html .. _api: http://pydecode.readthedocs.org/en/latest/api.html