Distributed Python Active Learning library
DPyACL Distributed Python Framework for Active Learning May 2020 Alfredo Lorie Bernardo version 0.3.3
DPyACL is a flexible Distributed Active Learning library written in Python, aimed to make active learning experiments
simpler and faster. Its leverage Dask distributed features to execute active learning experiments computations among a
cluster of computers, allowing to speed up computation and tackle scenarios where data doesn't fit in a single computer.
It also has been developed with a modular object-oriented design to provide an intuitive, ease of use interface and
to allow reuse, modification, and extensibility. It also offers full compatibility with libraries like NumPy, SciPy,
Pandas, Scikit-learn and Keras. This library is available in PyPI and distributed under the GNU license.4
Up to date, DPyACL heavily uses Dask library to implement in a distributed and parallel fashion the the most significant strategies strategies that have appeared on the single_label-label. For future releases, we hope to include strategies strategies related with multi-label learning paradigms.
The fastest way to use
DPyACL is from a Jupyter Notebook.
Preparing an experiment
When defining an Active Learning experiment
DPyACL offers set pre-defined components that can be configured and
combined by the user to better fit its needs. The required components to setup and experiment are listed below
- The Dataset
- Labelled and unlabelled sets: Optional - The experiment might be configured to randomly choose an initial labeled and unlabeled sets
- An Experiment: HoldOut and KFold experiments are provided
- The AL scenario: The current release provides a Pool Based Scenario
- The Machine Learning Technique: It can be a machine learning technique from any library that provides an API compatible with the fit, predict and predict_proba definitions. Sklearn, Dask-ML, Keras are compatible
- The Evaluation Method(s)
- The Query Strategy
- The Stopping Criteria
- The Oracle: The current release provides a Simulated Oracle
Configuring the experiment
ml_technique = LogisticRegression(solver='liblinear') stopping_criteria = MaxIteration(50) query_strategy = QueryMarginSampling() performance_metrics = [ Accuracy(), F1(average='macro'), Precision(average='macro'), Recall(average='macro')] experiment = HoldOutExperiment( client=None, X=_X, Y=_y, scenario_type=PoolBasedSamplingScenario, train_idx=train_idx, test_idx=test_idx, label_idx=label_idx, unlabel_idx=unlabel_idx, ml_technique=ml_technique, performance_metrics=performance_metrics, query_strategy=query_strategy, oracle=SimulatedOracle(labels=_y), stopping_criteria=stopping_criteria, self_partition=False )
Execute the experiment
result = experiment.evaluate(verbose=True)
Analyze the experiment results
query_analyser = ExperimentAnalyserFactory.experiment_analyser( performance_metrics= [metric.metric_name for metric in performance_metrics], method_name=query_strategy.query_function_name, method_results=result, type="queries" ) # get a brief description of the experiment query_analyser.plot_learning_curves(title='Active Learning experiment results')
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size dpyacl-0.3.3.tar.gz (52.2 kB)||File type Source||Python version None||Upload date||Hashes View|