Rep Versions Save

• python2 and python3 dockers
• updated libraries
• added CacheClassifier
• minimized size of docker image, simplified building process
• some fixes for ML libraries
• some documentation updates
• deleted plot.ly
• solved theanets reproducibility

Fixes:

• TMVA process correct termination
• TMVA fix for MAX OS El Capitan (problems with dynamic libraries paths)
• fix travis (show not passed tests, create docker on dockerhub)
• fix wget in notebooks
• fix errors calculation in efficiencies (for flatness property)
• added Makefile
• fix normalization in the multi dimentional metric

• Add continuous integration
• Python 3 support
• Conda installation in docker and travis
• Kitematic-friendly docker
• Update all libraries versions
• added Folding Regressor, added feature importances for folding
• added minimization to gridsearch, added random gridsearch from distributions
• added folding scorer for regressor to gridsearch
• faster tests
• updated notebooks
• Fixes:
  • tmva termination
  • documentation for grid search
  • Gridsearch bugs with metrics (metric fit)
  • learning curve with mask for folding

• Fixed bug in IPython profile creation inside Dockerfile.
• Nothing else was changed.

• Support of neural networks in common interface:

  • theanets
  • neurolab
  • pybrain

  Now all the REP stuff is available for classifiers and regressors from these libraries:

  • usage inside sklearn pipeline
  • grid_search for hyper parameter optimization
  • reports, parallel training on cluster

• New lovely documentation, check it out!

• Fixes in metaclassifiers connected with usage of expressions-as-features

• Rewritten FeatureSplitter

• Switched to sklearn 0.16

• New method train_test_split_group - splitting into train and test by the value of special column. Samples with same values are either both in train or both in test.

• Update howto/notebooks with new open physical datasets

• Tmva implementation enhancement with root_numpy https://github.com/yandex/rep/issues/2.
• Add FPRatTPR (return fpr value at fixed tpr) and TPRatFPR (return tpr value at fixed fpr) metrics, which are required, e.g. for tuning online triggering system. Moreover learning curves are available for these metrics now.
• Many improvements in documentation.

• unified classifiers wrapper for variety of implementations: TMVA, Sklearn, XGBoost, uBoost
• parallel training of classifiers on cluster
• classification/regression reports with plots
• support of interactive plots (bokeh, plotly)
• grid-search with parallelized execution on a cluster
• git, versioning of research
• computation of different classification metrics
• partial support of python 3.