Reproducible Experiment Platform (REP)
REP is ipython-based environment for conducting data-driven research in a consistent and reproducible way.
- unified python wrapper for different ML libraries (wrappers follow extended scikit-learn interface)
- parallel training of classifiers on cluster
- classification/regression reports with plots
- interactive plots supported
- smart grid-search algorithms with parallelized execution
- versioning of research using git
- pluggable quality metrics for classification
- meta-algorithm design (aka 'rep-lego')
REP is not trying to substitute scikit-learn, but extends it and provides better user experience.
To get started, look at the notebooks in /howto/
Notebooks can be viewed (not executed) online at nbviewer <br />
There are basic introductory notebooks (about python, IPython) and more advanced ones (about the REP itself)
Installation with Docker
We provide the docker image with
REP and all it's dependencies
Installation with bare hands
- gitter chat, troubleshooting
- API, contributing new estimator
- API, contributing new metric
- If you use REP in research, please consider citing
Apache 2.0, library is open-source.
REP wrappers are sklearn compatible:
from rep.estimators import XGBoostClassifier, SklearnClassifier, TheanetsClassifier clf = XGBoostClassifier(n_estimators=300, eta=0.1).fit(trainX, trainY) probabilities = clf.predict_proba(testX)
Beloved trick of kagglers is to run bagging over complex algorithms. This is how it is done in REP:
from sklearn.ensemble import BaggingClassifier clf = BaggingClassifier(base_estimator=XGBoostClassifier(), n_estimators=10) # wrapping sklearn to REP wrapper clf = SklearnClassifier(clf)
Another useful trick is to use folding instead of splitting data into train/test.
This is specially useful when you're using some kind of complex stacking
from rep.metaml import FoldingClassifier clf = FoldingClassifier(TheanetsClassifier(), n_folds=3) probabilities = clf.fit(X, y).predict_proba(X)
In example above all data are splitted into 3 folds,
and each fold is predicted by classifier which was trained on other 2 folds.
Also REP classifiers provide report:
report = clf.test_on(testX, testY) report.roc().plot() # plot ROC curve from rep.report.metrics import RocAuc # learning curves are useful when training GBDT! report.learning_curve(RocAuc(), steps=10)
You can read about other REP tools (like smart distributed grid search, folding and factory)
in documentation and howto examples.