Milk

Milk wraps libsvm in Python code.

It also supports k-means clustering with an implementation that is careful not to use too much memory.

Features:

• Random forests
• Self organising maps
• SVMs. Using the libsvm solver with a pythonesque wrapper around it.
• Stepwise Discriminant Analysis for feature selection.
• Non-negative matrix factorisation
• K-means using as little memory as possible.
• Affinity propagation

What is new in this release:

• Added subspace projection kNN.
• Export pdist in milk namespace.
• Added Eigen to source distribution.
• Added measures.curves.roc.
• Added mds_dists function.

What is new in version 0.5:

• Add coordinate-descent based LASSO
• Add unsupervised.center function
• Make zscore work with NaNs (by ignoring them)
• Propagate apply_many calls through transformers

What is new in version 0.4.1:

• Fixed an important bug in gridsearch.

What is new in version 0.4.0:

• Use multiprocessing to take advantage of multi core machines (off by default).
• Add perceptron learner
• Set random seed in random forest learner
• Add warning to milk/__init__.py if import fails
• Add return value to gridminimise
• Set random seed in precluster_learner
• Implemented Error-Correcting Output Codes for reduction of multi-class to binary (including probability estimation)
• Add multi_strategy argument to defaultlearner()
• Make the dot kernel in svm much, much, faster
• Make sigmoidal fitting for SVM probability estimates faster
• Fix bug in randomforest (patch by Wei on milk-users mailing list)

What is new in version 0.3.10:

• Add ext.jugparallel for integration with jug
• Parallel nfold crossvalidation using jug
• Parallel multiple kmeans runs using jug
• cluster_agreement for non-ndarrays
• Add histogram & normali(z|s)e options to milk.kmeans.assign_centroid
• Fix bug in sda when features were constant for a class
• Add select_best_kmeans
• Added defaultlearner as a better name than defaultclassifier
• Add measures.curves.precision_recall
• Add unsupervised.parzen.parzen

What is new in version 0.3.8:

• Fixed compilation on Windows.

What is new in version 0.3.7:

• Logistic regression.
• Source demos included (in source and documentation).
• Add cluster agreement metrics.
• Fix nfoldcrossvalidation bug when using origins.

What is new in version 0.3.5:

• Bugfix for 64 bits.

What is new in version 0.3.4:

• Random forest learners.
• Decision trees sped up 20x.
• Much faster gridsearch (finds optimum without computing all folds).