milk-0.3.4/0000755000175000017500000000000011463416340013252 5ustar luispedroluispedromilk-0.3.4/MANIFEST.in0000644000175000017500000000002411460362604015004 0ustar luispedroluispedroinclude README.rst milk-0.3.4/setup.py0000644000175000017500000000434211462123127014764 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division try: import setuptools except: print ''' setuptools not found. On linux, the package is often called python-setuptools''' from sys import exit exit(1) from numpy.distutils.core import setup, Extension execfile('milk/milk_version.py') long_description = file('README.rst').read() svm_ext = Extension('milk.supervised._svm', sources = ['milk/supervised/_svm.cpp']) som_ext = Extension('milk.unsupervised._som', sources = ['milk/unsupervised/_som.cpp']) tree_ext = Extension('milk.supervised._tree', sources = ['milk/supervised/_tree.cpp']) ext_modules = [svm_ext, som_ext, tree_ext] packages = filter(lambda p: p.startswith('milk'), setuptools.find_packages()) setup(name = 'milk', version = __version__, description = 'Machine Learning Toolkit', long_description = long_description, author = u'Luis Pedro Coelho', author_email = 'lpc@cmu.edu', url = 'http://luispedro.org/software/milk', license = 'MIT', packages = packages, ext_modules = ext_modules, ) milk-0.3.4/PKG-INFO0000644000175000017500000000423111463416340014347 0ustar luispedroluispedroMetadata-Version: 1.0 Name: milk Version: 0.3.4 Summary: Machine Learning Toolkit Home-page: http://luispedro.org/software/milk Author: Luis Pedro Coelho Author-email: lpc@cmu.edu License: MIT Description: ============================== MILK: MACHINE LEARNING TOOLKIT ============================== Machine Learning in Python -------------------------- Milk is a machine learning toolkit in Python. Its focus is on supervised classification and on enabling medium scale learning (defined as data that barely fits in main memory). milk wraps libsvm in a Pythonic way (the models learned have weight arrays that are accessible from Python directly, the models are pickle()able, you can pass any Python function as a kernel,....) It also supports k-means clustering with an implementation that is careful not to use too much memory (if your dataset fits into memory, milk can cluster it). It does not have its own file format or in-memory format, which I consider a feature as it works on numpy arrays directly (or anything that is convertible to a numpy-array) without forcing you to copy memory around. For SVMs, you can even just use any datatype if you have your own kernel function. New in 0.3.4 ------------ - Random forest learners - Decision trees sped up 20x - Much faster gridsearch (finds optimum without computing all folds) 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 License: MIT Author: Luis Pedro Coelho (with code from LibSVM and scikits.learn) Website: http://luispedro.org/software/milk API Documentation: `http://packages.python.org/milk/ `_ Platform: UNKNOWN milk-0.3.4/README.rst0000644000175000017500000000315011463416244014743 0ustar luispedroluispedro============================== MILK: MACHINE LEARNING TOOLKIT ============================== Machine Learning in Python -------------------------- Milk is a machine learning toolkit in Python. Its focus is on supervised classification and on enabling medium scale learning (defined as data that barely fits in main memory). milk wraps libsvm in a Pythonic way (the models learned have weight arrays that are accessible from Python directly, the models are pickle()able, you can pass any Python function as a kernel,....) It also supports k-means clustering with an implementation that is careful not to use too much memory (if your dataset fits into memory, milk can cluster it). It does not have its own file format or in-memory format, which I consider a feature as it works on numpy arrays directly (or anything that is convertible to a numpy-array) without forcing you to copy memory around. For SVMs, you can even just use any datatype if you have your own kernel function. New in 0.3.4 ------------ - Random forest learners - Decision trees sped up 20x - Much faster gridsearch (finds optimum without computing all folds) 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 License: MIT Author: Luis Pedro Coelho (with code from LibSVM and scikits.learn) Website: http://luispedro.org/software/milk API Documentation: `http://packages.python.org/milk/ `_ milk-0.3.4/milk.egg-info/0000755000175000017500000000000011463416340015700 5ustar luispedroluispedromilk-0.3.4/milk.egg-info/dependency_links.txt0000644000175000017500000000000111463416340021746 0ustar luispedroluispedro milk-0.3.4/milk.egg-info/SOURCES.txt0000644000175000017500000000254411463416340017571 0ustar luispedroluispedroMANIFEST.in README.rst setup.py milk/__init__.py milk/milk_version.py milk/nfoldcrossvalidation.py milk.egg-info/PKG-INFO milk.egg-info/SOURCES.txt milk.egg-info/dependency_links.txt milk.egg-info/top_level.txt milk/active/__init__.py milk/active/eimpact.py milk/active/uncertainty.py milk/measures/__init__.py milk/measures/measures.py milk/measures/nfoldcrossvalidation.py milk/supervised/__init__.py milk/supervised/_svm.cpp milk/supervised/_tree.cpp milk/supervised/classifier.py milk/supervised/defaultclassifier.py milk/supervised/featureselection.py milk/supervised/gridsearch.py milk/supervised/grouped.py milk/supervised/knn.py milk/supervised/libsvmclassifier.py milk/supervised/multi.py milk/supervised/multi_view.py milk/supervised/normalise.py milk/supervised/parzen.py milk/supervised/randomforest.py milk/supervised/set2binary_array.py milk/supervised/setup.py milk/supervised/svm.py milk/supervised/tree.py milk/unsupervised/__init__.py milk/unsupervised/_som.cpp milk/unsupervised/affinity.py milk/unsupervised/gaussianmixture.py milk/unsupervised/kmeans.py milk/unsupervised/normalise.py milk/unsupervised/pca.py milk/unsupervised/pdist.py milk/unsupervised/som.py milk/unsupervised/nnmf/__init__.py milk/unsupervised/nnmf/hoyer.py milk/unsupervised/nnmf/lee_seung.py milk/utils/__init__.py milk/utils/utils.py milk/wrapper/__init__.py milk/wrapper/wraplibsvm.pymilk-0.3.4/milk.egg-info/top_level.txt0000644000175000017500000000000511463416340020425 0ustar luispedroluispedromilk milk-0.3.4/milk.egg-info/PKG-INFO0000644000175000017500000000423111463416340016775 0ustar luispedroluispedroMetadata-Version: 1.0 Name: milk Version: 0.3.4 Summary: Machine Learning Toolkit Home-page: http://luispedro.org/software/milk Author: Luis Pedro Coelho Author-email: lpc@cmu.edu License: MIT Description: ============================== MILK: MACHINE LEARNING TOOLKIT ============================== Machine Learning in Python -------------------------- Milk is a machine learning toolkit in Python. Its focus is on supervised classification and on enabling medium scale learning (defined as data that barely fits in main memory). milk wraps libsvm in a Pythonic way (the models learned have weight arrays that are accessible from Python directly, the models are pickle()able, you can pass any Python function as a kernel,....) It also supports k-means clustering with an implementation that is careful not to use too much memory (if your dataset fits into memory, milk can cluster it). It does not have its own file format or in-memory format, which I consider a feature as it works on numpy arrays directly (or anything that is convertible to a numpy-array) without forcing you to copy memory around. For SVMs, you can even just use any datatype if you have your own kernel function. New in 0.3.4 ------------ - Random forest learners - Decision trees sped up 20x - Much faster gridsearch (finds optimum without computing all folds) 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 License: MIT Author: Luis Pedro Coelho (with code from LibSVM and scikits.learn) Website: http://luispedro.org/software/milk API Documentation: `http://packages.python.org/milk/ `_ Platform: UNKNOWN milk-0.3.4/milk/0000755000175000017500000000000011463416340014206 5ustar luispedroluispedromilk-0.3.4/milk/nfoldcrossvalidation.py0000644000175000017500000000013011452411266021001 0ustar luispedroluispedrofrom .measures.nfoldcrossvalidation import foldgenerator, getfold, nfoldcrossvalidation milk-0.3.4/milk/measures/0000755000175000017500000000000011463416340016032 5ustar luispedroluispedromilk-0.3.4/milk/measures/nfoldcrossvalidation.py0000644000175000017500000002022611452411266022635 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division from collections import defaultdict from ..supervised.classifier import normaliselabels from ..supervised.defaultclassifier import defaultclassifier import numpy import numpy as np __all__ = ['foldgenerator', 'getfold', 'nfoldcrossvalidation'] def foldgenerator(labels, nfolds=None, origins=None): ''' for train,test in foldgenerator(labels, nfolds=None, origins=None) ... This generator breaks up the data into `n` folds (default 10). If `origins` is given, then all elements that share the same origin will either be in testing or in training (never in both). This is useful when you have several replicates that shouldn't be mixed together between training&testing but that can be otherwise be treated as independent for learning. Parameters ---------- labels : a sequence the labels nfolds : integer nr of folds (default 10 or minimum label size) origins : if present, must be an array of indices of the same size as labels. Returns ------- iterator over `train, test`, two boolean arrays Bugs ---- Algorithm is very naive w.r.t. unbalanced origins. ''' labels,_ = normaliselabels(labels) nlabels = labels.max()+1 label_locations = [np.where(labels == i)[0] for i in xrange(nlabels)] label_counts = np.array(map(len, label_locations)) if origins is not None: origins = np.asanyarray(origins) if len(origins) != len(labels): raise ValueError( 'milk.nfoldcrossvalidation.foldgenerator: origins must be of same size as labels') label_counts_orig = np.zeros(nlabels, int) label_origins = [[] for i in xrange(nlabels)] counted = set() for lab, orig in zip(labels, origins): if orig in counted: if orig not in label_origins[lab]: for other in xrange(nlabels): if other == lab: continue if orig in label_origins[other]: raise ValueError('milk.nfoldcrossvalidation.foldgenerator: origin %s seems to be present in two labels %s and %s' % (orig, lab, other)) raise AssertionError('This should have been unreachable code') continue label_counts_orig[lab] += 1 label_origins[lab].append(orig) counted.add(orig) label_origins_counts = [np.cumsum([(origins == orig).sum() for orig in label_origins[lab]]) for lab in xrange(nlabels)] else: label_counts_orig = label_counts if nfolds is None: nfolds = min(label_counts_orig.min(), 10) elif nfolds > label_counts_orig.min(): from warnings import warn warn('milk.measures.nfoldcrossvalidation: Reducing the nr. of folds from %s to %s (minimum class size).' % (nfolds, label_counts_orig.min())) nfolds = label_counts_orig.min() if nfolds == 1: note = ('(taking `origins` into account)' if origins is not None else '') raise ValueError(''' milk.nfoldcrossvalidation.foldgenerator: nfolds was reduced to 1 because minimum class size was 1. If you passed in an origins parameter, it might be caused by having a class come from a single origin. The class histogram %s looks like: %s''' % (note, label_counts_orig)) perfolds = (label_counts // nfolds) assert perfolds.min() > 0 testing = np.zeros(len(labels), bool) start = np.zeros(nlabels, int) for fold in xrange(nfolds): testing.fill(False) end = start + perfolds # We need to force the last fold to be past the end to avoid rounding errors # missing some elements if fold == (nfolds - 1): end.fill(len(labels)) for i,locs,s,e in zip(xrange(nlabels), label_locations, start, end): if origins is None: testing[locs[s:e]] = 1 else: s,e = np.searchsorted(label_origins_counts[i], (s,e)) for included in label_origins[i][s:e]: testing[origins == included] = True yield ~testing, testing start = end def getfold(labels, fold, nfolds=None, origins=None): ''' trainingset,testingset = getfold(labels, fold, nfolds=None, origins=None) Get the training and testing set for fold `fold` in `nfolds` Arguments are the same as for `foldgenerator` Parameters ---------- labels : ndarray of labels fold : integer nfolds : integer number of folds (default 10 or size of smallest class) origins : sequence, optional if given, then objects with same origin are *not* scattered across folds ''' if nfolds < fold: raise ValueError('milk.getfold: Attempted to get fold %s out of %s' % (fold, nfolds)) for i,(t,s) in enumerate(foldgenerator(labels, nfolds, origins)): if i == fold: return t,s raise ValueError('milk.getfold: Attempted to get fold %s but the number of actual folds was too small (%s)' % (fold,i)) def nfoldcrossvalidation(features, labels, nfolds=None, classifier=None, origins=None, return_predictions=False): ''' Perform n-fold cross validation cmatrix,labelnames = nfoldcrossvalidation(features, labels, nfolds=10, classifier={defaultclassifier()}, origins=None, return_predictions=False) cmatrix,labelnames,predictions = nfoldcrossvalidation(features, labels, nfolds=10, classifier={defaultclassifier()}, origins=None, return_predictions=True) cmatrix will be a N x N matrix, where N is the number of classes cmatrix[i,j] will be the number of times that an element of class i was classified as class j labelnames[i] will correspond to the label name of class i Parameters ---------- features : a feature matrix or list of feature vectors labels : an array of labels, where label[i] is the label corresponding to features[i] nfolds : Nr of folds. origins : Origin ID (see foldgenerator) return_predictions : whether to return predictions classifier should implement the train() method to return a model (something with an apply() method) ''' assert len(features) == len(labels), 'milk.measures.nfoldcrossvalidation: len(features) should match len(labels)' if classifier is None: classifier = defaultclassifier() labels,labelnames = normaliselabels(labels) predictions = np.zeros_like(labels)-1 features = numpy.asanyarray(features) nclasses = labels.max() + 1 cmatrix = np.zeros((nclasses,nclasses)) for trainingset,testingset in foldgenerator(labels, nfolds, origins=origins): model = classifier.train(features[trainingset], labels[trainingset]) prediction = np.array([model.apply(f) for f in features[testingset]]) predictions[testingset] = prediction.astype(predictions.dtype) for p, r in zip(prediction,labels[testingset]): cmatrix[r,p] += 1 if return_predictions: return cmatrix, labelnames, predictions return cmatrix, labelnames milk-0.3.4/milk/measures/measures.py0000644000175000017500000000307111462113030020216 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # License: MIT from __future__ import division import numpy import numpy as np __all__ = ['accuracy','waccuracy'] def accuracy(cmatrix): ''' acc = accuracy(cmatrix) Accuracy of confusion matrix ''' cmatrix = numpy.asanyarray(cmatrix) return cmatrix.trace()/cmatrix.sum() def zero_one_loss(real, predicted, normalisedlabels=False, names=None): ''' loss = zero_one_loss(real, predicted, normalisedlabels={unused}, names={unused}) Parameters ---------- real : sequence the underlying labels predicted : sequence the predicted labels normalisedlabels : unused names: unused Returns ------- loss : integer the number of instances where `real` differs from `predicted` ''' return np.sum(np.asanyarray(real) != np.asanyarray(predicted)) def waccuracy(cmatrix): ''' wacc = waccuracy(cmatrix) Weighted accuracy of cmatrix ''' cmatrix = numpy.asanyarray(cmatrix) return (cmatrix.diagonal() / cmatrix.sum(1)).mean() def confusion_matrix(real, predicted, normalisedlabels=False, names=None): if not normalisedlabels: real, names = normaliselabels(real) predicted = map(names.index, predicted) n = np.max(real)+1 cmat = np.zeros((n,n), int) for r,p in zip(real, predicted): cmat[r,p] += 1 return cmat ## TODO: Implement http://en.wikipedia.org/wiki/Matthews_Correlation_Coefficient milk-0.3.4/milk/measures/__init__.py0000644000175000017500000000005111452411266020137 0ustar luispedroluispedrofrom measures import accuracy, waccuracy milk-0.3.4/milk/supervised/0000755000175000017500000000000011463416340016377 5ustar luispedroluispedromilk-0.3.4/milk/supervised/parzen.py0000644000175000017500000000333211452411266020251 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np def get_parzen_rbf_loocv(features,labels): xij = np.dot(features,features.T) f2 = np.sum(features**2,1) d = f2-2*xij d = d.T + f2 d_argsorted = d.argsort(1) d_sorted = d.copy() d_sorted.sort(1) e_d = np.exp(-d_sorted) labels_sorted = labels[d_argsorted].astype(np.double) labels_sorted *= 2 labels_sorted -= 1 def f(sigma): k = e_d ** (1./sigma) return (((k[:,1:] * labels_sorted[:,1:]).sum(1) > 0) == labels).mean() return f milk-0.3.4/milk/supervised/svm.py0000644000175000017500000003565711461573003017574 0ustar luispedroluispedro# -*- coding: utf-8 -*- # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division from .classifier import normaliselabels, ctransforms_model from collections import deque import numpy import numpy as np import random from . import _svm def _svm_apply(SVM, q): ''' f_i = _svm_apply(SVM, q) @internal: This is mostly used for testing ''' X,Y,Alphas,b,C,kernel=SVM N = len(X) s = 0.0 for i in xrange(N): s += Alphas[i] * Y[i] * kernel(q, X[i]) return s - b def svm_learn_smo(X,Y,kernel,C,eps=1e-4,tol=1e-2,cache_size=(1<<20)): ''' Learn a svm classifier X: data Y: labels in SVM format (ie Y[i] in (1,-1)) This is a very raw interface. In general, you should use a class like svm_classifier. Implements the Sequential Minimum Optimisation Algorithm from Platt's "Fast training of support vector machines using sequential minimal optimization" in Advances in kernel methods: support vector learning Pages: 185 - 208 Year of Publication: 1999 ISBN:0-262-19416-3 ''' assert numpy.all(numpy.abs(Y) == 1) assert len(X) == len(Y) N = len(Y) Y = Y.astype(numpy.int32) params = numpy.array([0,C,1e-3,1e-5],numpy.double) Alphas0 = numpy.zeros(N, numpy.double) _svm.eval_SMO(X,Y,Alphas0,params,kernel,cache_size) return Alphas0, params[0] def svm_learn_libsvm(X,Y,kernel,C,eps=1e-4,tol=1e-2,cache_size=(1<<20)): ''' Learn a svm classifier using LIBSVM optimiser X: data Y: labels in SVM format (ie Y[i] in (1,-1)) This is a very raw interface. In general, you should use a class like svm_classifier. This uses the LIBSVM optimisation algorithm ''' assert numpy.all(numpy.abs(Y) == 1) assert len(X) == len(Y) N = len(Y) Y = Y.astype(numpy.int32) p = -numpy.ones(N,numpy.double) params = numpy.array([0,C,eps,tol],numpy.double) Alphas0 = numpy.zeros(N, numpy.double) _svm.eval_LIBSVM(X,Y,Alphas0,p,params,kernel,cache_size) return Alphas0, params[0] class preprocessed_rbf_kernel(object): def __init__(self, X, sigma, beta): self.X = X self.Xsum = (X**2).sum(1) self.sigma = sigma self.beta = beta def __call__(self, q): minus_d2_sigma = np.dot(self.X,q) minus_d2_sigma *= 2. minus_d2_sigma -= self.Xsum minus_d2_sigma -= np.dot(q,q) minus_d2_sigma /= self.sigma return self.beta * np.exp(minus_d2_sigma) class rbf_kernel(object): ''' kernel = rbf_kernel(sigma,beta=1) Radial Basis Function kernel Returns a kernel (ie, a function that implements) beta * exp( - ||x1 - x2|| / sigma) ''' def __init__(self, sigma, beta=1): self.sigma = sigma self.beta = beta self.kernel_nr_ = 0 self.kernel_arg_ = float(sigma) def __call__(self, x1, x2): d2 = x1 - x2 d2 **= 2 d2 = d2.sum() res = self.beta*np.exp(-d2/self.sigma) return res def preprocess(self, X): return preprocessed_rbf_kernel(X, self.sigma, self.beta) class polynomial_kernel(object): ''' kernel = polynomial_kernel(d,c=1) returns a kernel (ie, a function) that implements: (+c)**d ''' def __init__(self, d, c=1): self.d = d self.c = c def __call__(self,x1,x2): return (np.dot(x1,x2)+self.c)**self.d class precomputed_kernel(object): ''' kernel = precomputed_kernel(kmatrix) A "fake" kernel which is precomputed. ''' def __init__(self, kmatrix, copy=False): kmatrix = np.ascontiguousarray(kmatrix, np.double, copy=copy) self.kernel_nr_ = 1 self.kernel_arg_ = 0. def __call__(self, x0, x1): return kmatrix[x0,x1] class _call_kernel(object): def __init__(self, k, svs): self.svs = svs self.kernel = k def __call__(self, q): return np.array([self.kernel(s, q) for s in self.svs]) class dot_kernel(object): def __init__(self): self.kernel_nr_ = 2 self.kernel_arg_ = 0. def __call__(self, x0, x1): return np.dot(x0, x1) class svm_raw_model(object): def __init__(self, svs, Yw, b, kernel): self.svs = svs self.Yw = Yw self.b = b self.kernel = kernel try: self.kernelfunction = self.kernel.preprocess(self.svs) except AttributeError: self.kernelfunction = _call_kernel(self.kernel, self.svs) def apply(self, q): Q = self.kernelfunction(q) return np.dot(Q, self.Yw) - self.b class svm_raw(object): ''' svm_raw: classifier classifier = svm_raw(kernel, C, eps=1e-3, tol=1e-8) Parameters ---------- kernel : the kernel to use. This should be a function that takes two data arguments see rbf_kernel and polynomial_kernel. C : the C parameter Other Parameters ---------------- eps : the precision to which to solve the problem (default 1e-3) tol : (|x| < tol) is considered zero ''' def __init__(self, kernel=None, C=1., eps=1e-3, tol=1e-8): self.C = C self.kernel = kernel self.eps = eps self.tol = tol self.algorithm = 'libsvm' def train(self, features, labels, normalisedlabels=False): assert self.kernel is not None, 'milk.supervised.svm_raw.train: kernel not set!' assert self.algorithm in ('libsvm','smo'), 'milk.supervised.svm_raw: unknown algorithm (%s)' % self.algorithm assert not (np.isinf(self.C) or np.isnan(self.C)), 'milk.supervised.svm_raw: setting C to NaN or Inf causes problems.' features = np.asanyarray(features) if normalisedlabels: Y = labels else: Y,_ = normaliselabels(labels) assert Y.max() == 1, 'milk.supervised.svm_raw can only handle binary problems' Y *= 2 Y -= 1 kernel = self.kernel try: kernel = (self.kernel.kernel_nr_, self.kernel.kernel_arg_) features = np.ascontiguousarray(features, np.double) except AttributeError: pass if self.algorithm == 'smo': alphas,b = svm_learn_smo(features,Y,kernel,self.C,self.eps,self.tol) else: alphas,b = svm_learn_libsvm(features,Y,kernel,self.C,self.eps,self.tol) svsi = (alphas != 0) svs = features[svsi] w = alphas[svsi] Y = Y[svsi] Yw = w * Y return svm_raw_model(svs, Yw, b, self.kernel) def get_params(self): return self.C, self.eps,self.tol def set_params(self,params): self.C,self.eps,self.tol = params def set_option(self, optname, value): setattr(self, optname, value) def learn_sigmoid_constants(F,Y, max_iters=None, min_step=1e-10, sigma=1e-12, eps=1e-5): ''' A,B = learn_sigmoid_constants(F,Y) This is a very low-level interface look into the svm_classifier class. Parameters ---------- F : Values of the function F Y : Labels (in boolean format, ie, in (0,1)) Other Parameters ---------------- max_iters : Maximum nr. of iterations min_step : Minimum step sigma : sigma eps : A small number Reference for Implementation ---------------------------- Implements the algorithm from "A Note on Platt's Probabilistic Outputs for Support Vector Machines" by Lin, Lin, and Weng. Machine Learning, Vol. 68, No. 3. (23 October 2007), pp. 267-276 ''' # the deci[i] array is called F[i] in this code F = np.asanyarray(F) Y = np.asanyarray(Y) assert len(F) == len(Y) assert numpy.all( (Y == 1) | (Y == 0) ) from numpy import log, exp N=len(F) if max_iters is None: max_iters = 1000 prior1 = Y.sum() prior0 = N-prior1 small_nr = 1e-4 hi_t = (prior1+1.)/(prior1+2.) lo_t = 1./(prior0+2.) T = Y*hi_t + (1-Y)*lo_t A = 0. B = log( (prior0+1.)/(prior1+1.) ) def target(A,B): fval = 0. for i in xrange(N): fApB = F[i]*A+B if fApB >= 0: fval += T[i]*fApB+log(1+exp(-fApB)) else: fval += (T[i]-1.)*fApB+log(1+exp(fApB)) return fval fval = target(A,B) for iter in xrange(max_iters): h11=sigma h22=sigma h21=0. g1=0. g2=0. for i in xrange(N): fApB = F[i]*A+B if (fApB >= 0): p = exp(-fApB)/(1.+exp(-fApB)) q = 1./(1.+exp(-fApB)) else: p = 1./(1.+exp(fApB)) q = exp(fApB)/(1.+exp(fApB)) d2 = p * q h11 += F[i]*F[i]*d2 h22 += d2 h21 += F[i]*d2 d1 = T[i] - p g1 += F[i]*d1 g2 += d1 if abs(g1) < eps and abs(g2) < eps: # Stopping criteria break det = h11*h22 - h21*h21 dA = - (h22*g1 - h21*g2)/det dB = - (h21*g1 + h11*g2)/det gd = g1*dA + g2*dB stepsize = 1. while stepsize >= min_step: newA = A + stepsize*dA newB = B + stepsize*dB newf = target(newA,newB) if newf < fval+eps*stepsize*gd: A = newA B = newB fval = newf break stepsize /= 2 if stepsize < min_step: print 'Line search fails' break return A,B class svm_binary_model(object): def __init__(self, classes): self.classes = classes def apply(self,f): return self.classes[f >= 0.] class svm_binary(object): ''' classifier = svm_binary() model = classifier.train(features, labels) assert model.apply(f) in labels ''' def train(self, features, labels, normalisedlabels=False): if normalisedlabels: return svm_binary_model( (0,1) ) assert len(labels) >= 2, 'Cannot train from a single example' names = sorted(set(labels)) assert len(names) == 2, 'milk.supervised.svm.svm_binary.train: Can only handle two class problems' return svm_binary_model(names) class svm_to_binary(object): ''' svm_to_binary(base_svm) A simple wrapper so that svm_to_binary(base_svm) is a model that takes the base_svm classifier and then binarises its model output. NOTE: This class does the same job as:: ctransforms(base_svm, svm_binary()) ''' def __init__(self, svm_base): ''' binclassifier = svm_to_binary(svm_base) a classifier that binarises the output of svm_base. ''' self.base = svm_base def train(self, features, labels, normalisedlabels=False): model = self.base.train(features, labels, normalisedlabels=normalisedlabels) binary = svm_binary() binary_model = binary.train(features, labels, normalisedlabels=normalisedlabels) return ctransforms_model([model, binary_model]) def set_option(self, opt, value): self.base.set_option(opt, value) class svm_sigmoidal_correction_model(object): def __init__(self, A, B): self.A = A self.B = B def apply(self,features): return 1./(1.+numpy.exp(features*self.A+self.B)) class svm_sigmoidal_correction(object): ''' svm_sigmoidal_correction : a classifier Sigmoidal approximation for obtaining a probability estimate out of the output of an SVM. ''' def __init__(self): self.max_iters = None def train(self, features, labels, normalisedlabels=False): A,B = learn_sigmoid_constants(features,labels,self.max_iters) return svm_sigmoidal_correction_model(A, B) def get_params(self): return self.max_iters def set_params(self,params): self.max_iters = params def sigma_value_fisher(features,labels): ''' f = sigma_value_fisher(features,labels) value_s = f(s) Computes a function which computes how good the value of sigma is for the features. This function should be *minimised* for a good value of sigma. Parameters ----------- features : features matrix as 2-ndarray. Returns ------- f : a function: float -> float this function should be minimised for a good `sigma` Reference ---------- Implements the measure in "Determination of the spread parameter in the Gaussian kernel for classification and regression" by Wenjian Wanga, Zongben Xua, Weizhen Luc, and Xiaoyun Zhanga ''' features = np.asanyarray(features) xij = np.dot(features,features.T) f2 = np.sum(features**2,1) d = f2-2*xij d = d.T + f2 N1 = (labels==0).sum() N2 = (labels==1).sum() C1 = -d[labels == 0][:,labels == 0] C2 = -d[labels == 1][:,labels == 1] C12 = -d[labels == 0][:,labels == 1] C1 = C1.copy() C2 = C2.copy() C12 = C12.copy() def f(sigma): sigma = float(sigma) N1 = C1.shape[0] N2 = C2.shape[0] if C12.shape != (N1,N2): raise ValueError C1v = np.sum(np.exp(C1/sigma))/N1 C2v = np.sum(np.exp(C2/sigma))/N2 C12v = np.sum(np.exp(C12/sigma))/N1/N2 return (N1 + N2 - C1v - C2v)/(C1v/N1+C2v/N2 - 2.*C12v) return f class fisher_tuned_rbf_svm(object): ''' F = fisher_tuned_rbf_svm(sigmas, base) Returns a wrapper classifier that uses RBF kernels automatically tuned using sigma_value_fisher. ''' def __init__(self, sigmas, base): self.sigmas = sigmas self.base = base def train(self, features, labels, normalisedlabels=False): f = sigma_value_fisher(features, labels) fs = [f(s) for s in self.sigmas] self.sigma = self.sigmas[np.argmin(fs)] self.base.set_option('kernel',rbf_kernel(self.sigma)) return self.base.train(features,labels) milk-0.3.4/milk/supervised/_tree.cpp0000644000175000017500000001021211462300261020166 0ustar luispedroluispedro// Copyright (C) 2010, Luis Pedro Coelho // License: MIT #include #include #include extern "C" { #include #include } namespace { double set_entropy(PyArrayObject* points, double* counts = 0, int clen = -1) { const int* data = reinterpret_cast(PyArray_DATA(points)); const int N = PyArray_DIM(points, 0); for (int i = 0; i != clen; ++i) counts[i] = 0.; for (int i = 0; i != N; ++i) { int value = data[i]; if (value >= clen) { PyErr_SetString(PyExc_RuntimeError, "value too large. aborting"); return 0; } counts[value] += 1.; } // Here is the formula we use: // // H = - \sum px \log(px) // = - \sum (cx/N) \log( cx / N) // = - 1/N \sum { cx [ \log cx - \log N ] } // = - 1/N { (\sum cx \log cx ) - ( \sum cx \log N ) } // = - 1/N { (\sum cx \log cx ) - N \log N } // = ( - 1/N \sum cx \log cx ) + \log N double entropy = 0.; for (int i = 0; i != clen; ++i) { double cx = counts[i]; if (cx) entropy += cx * std::log(cx); } entropy /= -N; entropy += std::log(N); return entropy; } PyObject* py_set_entropy(PyObject* self, PyObject* args) { const char* errmsg = "Arguments were not what was expected for set_entropy.\n" "This is an internal function: Do not call directly unless you know exactly what you're doing.\n"; PyArrayObject* labels; PyArrayObject* counts; if (!PyArg_ParseTuple(args, "OO", &labels, &counts)) { PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } if (!PyArray_Check(labels) || PyArray_TYPE(labels) != NPY_INT32 || !PyArray_ISCONTIGUOUS(labels) || !PyArray_Check(counts) || PyArray_TYPE(counts) != NPY_DOUBLE || !PyArray_ISCONTIGUOUS(counts)) { PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } double* cdata = reinterpret_cast(PyArray_DATA(counts)); const int nlabels = PyArray_DIM(counts, 0); double res = set_entropy(labels, cdata, nlabels); return PyFloat_FromDouble(res); } PyObject* py_information_gain(PyObject* self, PyObject* args) { const char* errmsg = "Arguments were not what was expected for set_entropy.\n" "This is an internal function: Do not call directly unless you know exactly what you're doing.\n"; PyArrayObject* labels0; PyArrayObject* labels1; if (!PyArg_ParseTuple(args, "OO", &labels0, &labels1)) { PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } if (!PyArray_Check(labels0) || PyArray_TYPE(labels0) != NPY_INT32 || !PyArray_ISCONTIGUOUS(labels0) || !PyArray_Check(labels1) || PyArray_TYPE(labels1) != NPY_INT32 || !PyArray_ISCONTIGUOUS(labels1)) { PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } double* counts; double counts_array[8]; int clen = 0; #define GET_MAX(index) \ const int N ## index = PyArray_DIM(labels ## index, 0); \ { \ const int* data = reinterpret_cast(PyArray_DATA(labels ## index)); \ for (int i = 0; i != N ## index; ++i) { \ if (data[i] > clen) clen = data[i]; \ } \ } GET_MAX(0); GET_MAX(1); ++clen; if (clen > 8) { counts = new(std::nothrow) double[clen]; if (!counts) { PyErr_NoMemory(); return 0; } } else { counts = counts_array; } const double N = N0 + N1; double H = - N0/N * set_entropy(labels0, counts, clen) - N1/N * set_entropy(labels1, counts, clen); if (clen > 8) delete [] counts; return PyFloat_FromDouble(H); } PyMethodDef methods[] = { {"set_entropy", py_set_entropy, METH_VARARGS , "Do NOT call directly.\n" }, {"information_gain", py_information_gain, METH_VARARGS , "Do NOT call directly.\n" }, {NULL, NULL,0,NULL}, }; const char * module_doc = "Internal Module.\n" "\n" "Do NOT use directly!\n"; } // namespace extern "C" void init_tree() { import_array(); (void)Py_InitModule3("_tree", methods, module_doc); } milk-0.3.4/milk/supervised/knn.py0000644000175000017500000000446111461573003017542 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division from collections import defaultdict import numpy as np class kNN(object): ''' k-Nearest Neighbour Classifier Naive implementation of a k-nearest neighbour classifier. C = kNN(k) Attributes: ----------- * k: the k to use ''' def __init__(self, k=1): self.k = k def train(self, features, labels, normalisedlabels=False, copy_features=False): features = np.asanyarray(features) labels = np.asanyarray(labels) if copy_features: features = features.copy() labels = labels.copy() return kNN_model(self.k, features, labels) class kNN_model(object): def __init__(self, k, features, labels): self.k = k self.features = features self.labels = labels def apply(self,features): diff2 = ( (self.features - features)**2 ).sum(1) neighbours = diff2.argsort()[:self.k] labels = self.labels[neighbours] votes = defaultdict(int) for L in labels: votes[L] += 1 v,L = max( (v,L) for L,v in votes.items() ) return L # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/milk/supervised/randomforest.py0000644000175000017500000000352611462317331021461 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # License: MIT. See COPYING.MIT file in the milk distribution ''' Random Forest ------------- Main elements ------------- rf_learner : A learner object ''' from __future__ import division import numpy as np import milk.supervised.tree from .normalise import normaliselabels def _sample(features, labels, n, R): ''' features', labels' = _sample(features, labels, n, R) Sample n element from (features,labels) Parameters ---------- features : sequence labels : sequence Same size as labels n : integer R : random object Returns ------- features' : sequence labels' : sequence ''' N = len(features) sfeatures = [] slabels = [] for i in xrange(n): idx = R.randint(N) sfeatures.append(features[idx]) slabels.append(labels[idx]) return np.array(sfeatures), np.array(slabels) class rf_model(object): def __init__(self, forest, names): self.forest = forest self.names = names def apply(self, features): rf = len(self.forest) votes = sum(t.apply(features) for t in self.forest) return (votes > (rf//2)) class rf_learner(object): ''' Random Forest Learner ''' def __init__(self, rf=101): self.rf = rf def train(self, features, labels, normalisedlabels=False, **kwargs): N,M = features.shape m = int(.7*M) n = int(.7*M) R = np.random tree = milk.supervised.tree.tree_learner() forest = [] labels,names = normaliselabels(labels) for i in xrange(self.rf): forest.append(tree.train(*_sample(features, labels, n, R), normalisedlabels=True)) return rf_model(forest, names) milk-0.3.4/milk/supervised/setup.py0000644000175000017500000000273211452411266020115 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from numpy.distutils.core import setup, Extension svm_ext= Extension('_svm', sources = ['_svm.cpp']) if __name__ == '__main__': setup (name = 'svm', version = '0.1', description = "Implement Platt's SMO Algorithm", ext_modules = [svm_ext] ) # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/milk/supervised/defaultclassifier.py0000644000175000017500000001173211462076006022446 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np def defaultclassifier(mode='medium'): ''' learner = defaultclassifier(mode='medium') Return the default classifier learner This is an SVM based classifier using the 1-vs-1 technique for multi-class problems. The features will be first cleaned up (normalised to [-1, +1]) and go through SDA feature selection. Parameters ----------- mode : string, optional One of ('fast','medium','slow', 'really-slow'). This defines the speed accuracy trade-off. It essentially defines how large the SVM parameter range is. Returns ------- learner : classifier learner object See Also -------- feature_selection_simple : Just perform the feature selection svm_simple : Perform classification ''' # These cannot be imported at module scope! # The reason is that they introduce a dependency loop: # gridsearch depends on nfoldcrossvalidation # nfoldcrossvalidation depends on defaultclassifier # which cannot depend on gridsearch # # Importing at function level keeps all these issues at bay # from .classifier import ctransforms from .gridsearch import gridsearch from . import svm from .normalise import chkfinite, interval_normalise from .featureselection import sda_filter, featureselector, linear_independent_features from .multi import one_against_one assert mode in ('really-slow', 'slow', 'medium', 'fast'), \ "milk.supervised.defaultclassifier: mode must be one of 'fast','slow','medium'." if mode == 'fast': c_range = np.arange(-2,4) sigma_range = np.arange(-2,3) elif mode == 'medium': c_range = np.arange(-2,4) sigma_range = np.arange(-4,4) elif mode == 'really-slow': c_range = np.arange(-4,10) sigma_range = np.arange(-7,7) else: # mode == 'slow' c_range = np.arange(-9,5) sigma_range = np.arange(-7,4) return ctransforms( chkfinite(), interval_normalise(), featureselector(linear_independent_features), sda_filter(), gridsearch(one_against_one(svm.svm_to_binary(svm.svm_raw())), params={ 'C': 2.**c_range, 'kernel': [svm.rbf_kernel(2.**i) for i in sigma_range], } )) def feature_selection_simple(): ''' selector = feature_selection_simple() Standard feature normalisation and selection This fills in NaNs and Infs (to 0 and large numbers), normalises features to [-1, +1] and uses SDA for feature selection. Returns ------- selector : supervised learner See Also -------- defaultclassifier : perform feature selection *and* classification ''' from .classifier import ctransforms from .normalise import chkfinite, interval_normalise from .featureselection import sda_filter, featureselector, linear_independent_features return ctransforms( chkfinite(), interval_normalise(), featureselector(linear_independent_features), sda_filter(), ) def svm_simple(C, kernel): ''' learner = svm_simple(C, kernel) Returns a one-against-one SVM based classifier with `C` and `kernel` Parameters ---------- C : double C parameter kernel : kernel Kernel to use Returns ------- learner : supervised learner See Also -------- feature_selection_simple : Perform feature selection defaultclassifier : feature selection and gridsearch for SVM parameters ''' from . import svm from .multi import one_against_one return one_against_one(svm.svm_to_binary(svm.svm_raw(C=C, kernel=kernel))) milk-0.3.4/milk/supervised/gridsearch.py0000644000175000017500000001344111462050203021056 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from .classifier import normaliselabels def _allassignments(options): try: from itertools import product except ImportError: def product(*args, **kwds): # from http://docs.python.org/library/itertools.html#itertools.product pools = map(tuple, args) * kwds.get('repeat', 1) result = [[]] for pool in pools: result = [x+[y] for x in result for y in pool] for prod in result: yield tuple(prod) from itertools import repeat, izip for ks,vs in izip(repeat(options.keys()), product(*options.values())): yield zip(ks,vs) def _set_assignment(obj,assignments): for k,v in assignments: obj.set_option(k,v) def gridminimise(learner, features, labels, params, measure=None, nfolds=10): ''' best = gridminimise(learner, features, labels, params, measure={0/1 loss}) Grid search for the settings of parameters that maximises a given measure This function is equivalent to searching the grid, but does not actually search the whole grid. Parameters ---------- learner : a classifier object features : sequence of features labels : sequence of labels params : dictionary of sequences keys are the options to change, values are sequences of corresponding elements to try measure : function, optional a function that takes labels and outputs and returns the loss. Default: 0/1 loss. This must be an *additive* function. nfolds : integer, optional nr of folds to run, default: 10 Returns ------- best : a sequence of assignments ''' # The algorithm is as follows: # # for all assignments: error = 0, next_iteration = 0 # # at each iteration: # look for assignment with smallest error # if that is done: return it # else: perform one more iteration # # When the function returns, that assignment has the lowest error of all # assignments and all the iterations are done. Therefore, other assignments # could only be worse even if we never computed the whole error! from ..measures.nfoldcrossvalidation import foldgenerator if measure is None: def measure(real, preds): return np.sum(np.asarray(real) != np.asarray(preds)) labels,_ = normaliselabels(labels) allassignments = list(_allassignments(params)) N = len(allassignments) iteration = np.zeros(N, int) error = np.zeros(N, float) folds = [(Tr.copy(), Te.copy()) for Tr,Te in foldgenerator(labels, nfolds)] while True: next_pos = (error == error.min()) iter = iteration[next_pos].max() if iter == nfolds: (besti,) = np.where(next_pos & (iteration == iter)) besti = besti[0] return allassignments[besti] (ps,) = np.where(next_pos & (iteration == iter)) p = ps[0] _set_assignment(learner, allassignments[p]) train, test = folds[iter] model = learner.train(features[train], labels[train], normalisedlabels=True) preds = [model.apply(f) for f in features[test]] error[p] += measure(labels[test], preds) iteration[p] += 1 class gridsearch(object): ''' G = gridsearch(base, measure=accuracy, nfolds=10, params={ 'param1 : [...], param2 : [...]}) Perform a grid search for the best parameter values. When G.train() is called, then for each combination of p1 in param1, p2 in param2, ... it performs:: base_classifier.param1 = p1 base_classifier.param2 = p2 ... value[p1, p2,...] = measure(crossvaliation(base_classifier) it then picks the highest set of parameters and re-learns a model on the whole data. Parameters ----------- base_classifier : classifier to use measure : function, optional a function that takes labels and outputs and returns the loss. Default: 0/1 loss. This must be an *additive* function. nfolds : integer, optional Nr of folds params : dictionary ''' def __init__(self, base, measure=None, nfolds=10, params={}): self.params = params self.base = base self.nfolds = 10 self.measure = measure def is_multi_class(self): return self.base.is_multi_class() def train(self, features, labels, normalisedlabels=False): self.best = gridminimise(self.base, features, labels, self.params, self.measure, self.nfolds) _set_assignment(self.base, self.best) return self.base.train(features, labels, normalisedlabels=normalisedlabels) milk-0.3.4/milk/supervised/libsvmclassifier.py0000644000175000017500000000552411461573003022316 0ustar luispedroluispedro#s -*- coding: utf-8 -*- # Copyright (C) 2008 Murphy Lab # Carnegie Mellon University # # Written by Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. # from __future__ import division from numpy import array, asanyarray _svm = None try: from libsvm import svm _svm = svm except: pass __all__=['libsvmClassifier'] class libsvmClassifier(object): def __init__(self,probability = False): classifier.__init__(self) if _svm is None: raise RuntimeError('SVM Library not found. Cannot use this classifier.') self.param = _svm.svm_parameter(kernel_type = svm.RBF, probability = probability) self.output_probability = False def set_option(self,optname,value): setattr(self.param,optname,value) def train(self, features, labels, normalisedlabels=False): problem = _svm.svm_problem(labels,features) return libsvmClassifier_model(_svm.svm_model(problem,self.param)) class libsvmClassifier_model(object): def __init__(self, model): self.model = model def apply(self,feats): if self.output_probability: return self.model.predict_probability(feats) return self.model.predict(feats) def __getstate__(self): # This is really really really hacky, but it works N=NamedTemporaryFile() self.model.save(N.name) S=N.read() return S,self.output_probability,self._trained,self._labelnames def __setstate__(self,state): if _svm is None: raise RuntimeError('LibSVM Library not found. Cannot use this classifier.') S,self.output_probability,self._trained,self._labelnames = state N=NamedTemporaryFile() N.write(S) N.flush() self.model = _svm.svm_model(N.name) # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/milk/supervised/__init__.py0000644000175000017500000000530211457212370020510 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. ''' milk.supervised This hold the supervised classification modules: Submodules ---------- - defaultclassifier: contains a default "good enough" classifier - svm: related to SVMs - grouped: contains objects to transform single object classifiers into group classifiers by voting - multi: transforms binary classifiers into multi-class classifiers (1-vs-1 or 1-vs-rest) - featureselection: feature selection - knn: k-nearest neighbours - tree: decision tree classifiers Classifiers ----------- All classifiers have a `train` function which takes 2 arguments: - features : sequence of features - labels : sequence of labels They return a `model` object, which has an `apply` function which takes a single input and returns its label. Note that there are always two objects: the learned and the model and they are independent. Every time you call learner.train() you get a new model. Both classifiers and models are pickle()able. Example ------- :: features = np.random.randn(100,20) features[:50] *= 2 labels = np.repeat((0,1), 50) classifier = milk.defaultclassifier() model = classifier.train(features, labels) new_label = model.apply(np.random.randn(100)) new_label2 = model.apply(np.random.randn(100)*2) ''' from .defaultclassifier import defaultclassifier, svm_simple from .classifier import normaliselabels from .gridsearch import gridsearch from .tree import tree_learner __all__ = [ 'normaliselabels', 'defaultclassifier', 'svm_simple', 'gridsearch', 'tree_learner', ] milk-0.3.4/milk/supervised/multi.py0000644000175000017500000001367011461573003020110 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division from .classifier import normaliselabels import numpy as np class one_against_rest(object): ''' Implements one vs. rest classification strategy to transform a binary classifier into a multi-class classifier. classifier = one_against_rest(base) base must obey the classifier interface Example ------- :: multi = one_against_rest(milk.supervised.simple_svm()) model = multi.train(training_features,labels) print model.apply(testing_features) See Also -------- one_against_one ''' def __init__(self,base): self.base = base self.is_multi_class = True self.options = {} def set_option(self, k, v): self.options[k] = v def train(self, features, labels, normalisedlabels=False): labels, names = normaliselabels(labels) nclasses = labels.max() + 1 models = [] for i in xrange(nclasses): for k,v in self.options.iteritems(): self.base.set_option(k, v) model = self.base.train(features, (labels == i).astype(int), normalisedlabels=True) models.append(model) return one_against_rest_model(models, names) class one_against_rest_model(object): def __init__(self, models, names): self.models = models self.nclasses = len(self.models) self.names = names def apply(self, feats): vals = np.array([c.apply(feats) for c in self.models]) (idxs,) = np.where(vals) if len(idxs) == 1: (label,) = idxs elif len(idxs) == 0: label = 0 else: label = idxs[0] return self.names[label] class one_against_one(object): ''' Implements one vs. one classification strategy to transform a binary classifier into a multi-class classifier. classifier = one_against_one(base) base must obey the classifier interface Example ------- :: multi = one_against_one(milk.supervised.simple_svm()) multi.train(training_features,labels) print multi.apply(testing_features) See Also -------- one_against_rest ''' def __init__(self, base): self.base = base self.is_multi_class = True self.options = {} def set_option(self, k, v): self.options[k] = v def train(self, features, labels, **kwargs): ''' one_against_one.train(objs,labels) ''' labels, names = normaliselabels(labels) features = np.asanyarray(features) nclasses = labels.max() + 1 models = [ [None for i in xrange(nclasses)] for j in xrange(nclasses)] for i in xrange(nclasses): for j in xrange(i+1, nclasses): for k,v in self.options.iteritems(): self.base.set_option(k, v) idxs = (labels == i) | (labels == j) assert idxs.sum() > 0, 'milk.multi.one_against_one: Pair-wise classifier has no data' # Fixme: here I could add a Null model or something model = self.base.train(features[idxs], (labels[idxs]==i).astype(int), normalisedlabels=True) models[i][j] = model return one_against_one_model(models, names) class one_against_one_model(object): def __init__(self, models, names): self.models = models self.names = names self.nclasses = len(models) def apply(self,feats): ''' one_against_one.apply(objs) Classify one single object. ''' nc = self.nclasses votes = np.zeros(nc) for i in xrange(nc): for j in xrange(i+1,nc): c = self.models[i][j].apply(feats) if c: votes[i] += 1 else: votes[j] += 1 return self.names[votes.argmax(0)] class one_against_rest_multi_model(object): def __init__(self, models): self.models = models def apply(self, feats): return [lab for lab,model in self.models.iteritems() if model.apply(feats)] class one_against_rest_multi(object): ''' learner = one_against_rest_multi() model = learner.train(features, labels) classes = model.apply(f_test) ''' def __init__(self, base): self.base = base if hasattr(base, 'set_option'): self.set_option = base.set_option def train(self, features, labels, normalisedlabels=False): ''' ''' import operator all_labels = set() for ls in labels: all_labels.update(ls) models = {} for label in all_labels: models[label] = self.base.train(features, [(label in ls) for ls in labels]) return one_against_rest_multi_model(models) milk-0.3.4/milk/supervised/set2binary_array.py0000644000175000017500000000347311461573003022236 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np class set2binary_array_model(object): def __init__(self, universe): self.universe = list(universe) def apply(self, elems): res = np.zeros(len(self.universe) + 1, bool) for e in elems: try: res[self.universe.index(e)] = True except : res[-1] = True return res class set2binary_array(object): def train(self, features, labels, normalisedlabels=False): allfeatures = set() for f in features: allfeatures.update(f) return set2binary_array_model(allfeatures) milk-0.3.4/milk/supervised/multi_view.py0000644000175000017500000000315111461573003021133 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # License: MIT. See COPYING.MIT file in the milk distribution import numpy as np class multi_view_model(object): def __init__(self, models): self.models = models def apply(self, features): if len(features) != len(self.models): raise ValueError('milk.supervised.two_view: Nr of features does not match training data (got %s, expected %s)' % (len(features) ,len(self.models))) Ps = np.array([model.apply(f) for model,f in zip(self.models, features)]) if np.any(Ps <= 0.): return False if np.any(Ps >= 1.): return True # This is binary only: # if \prod Pi > \prod (1-Pi) return 1 # is equivalent to # if \prod Pi/(1-Pi) > 1. return 1 # if \sum \log( Pi/(1-Pi) ) > 0. return 1 return np.sum( np.log(Ps/(1-Ps)) ) > 0 class multi_view_classifier(object): def __init__(self, bases): self.bases = bases def train(self, features, labels, normalisedlabels=False): features = zip(*features) if len(features) != len(self.bases): raise ValueError('milk.supervised.two_view: Nr of features does not match classifiser construction (got %s, expected %s)' % (len(features) ,len(self.bases))) models = [] for basis,f in zip(self.bases, features): try: f = np.array(f) except: f = np.array(f, dtype=object) models.append(basis.train(f, labels)) return multi_view_model(models) milk-0.3.4/milk/supervised/grouped.py0000644000175000017500000000731311461573003020420 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # License: MIT. See COPYING.MIT file in the milk distribution from __future__ import division import numpy as np from collections import defaultdict from .classifier import normaliselabels def _concatenate_features_labels(gfeatures, glabels): if type(gfeatures) == np.ndarray and gfeatures.dtype == object: gfeatures = list(gfeatures) features = np.concatenate(gfeatures) labels = [] for feats,label in zip(gfeatures, glabels): labels.extend( [label] * len(feats) ) return features, labels class voting_classifier(object): ''' Implements a voting scheme for multiple sub-examples per example. classifier = voting_classifier(base) base should be a binary classifier Example ------- :: voterlearn = voting_classifier(milk.supervised.simple_svm()) voter = voterlearn.train(training_groups, labeled_groups) res = voter.apply([ [f0, f1, f3] ]) ''' def __init__(self, base): self.base = base def train(self, gfeatures, glabels, normalisedlabels=False): features, labels = _concatenate_features_labels(gfeatures, glabels) return voting_model(self.base.train(features, labels)) class voting_model(object): def __init__(self, base): self.base = base def apply(self, gfeatures): votes = defaultdict(int) for feats in gfeatures: votes[self.base.apply(feats)] += 1 best = None most_votes = 0 for k,v in votes.iteritems(): if v > most_votes: best = k most_votes = v return best class mean_classifier(object): ''' Implements a mean scheme for multiple sub-examples per example. classifier = mean_classifier(base) `base` should be a classifier that returns a numeric confidence value `classifier` will return the **mean** Example ------- :: meanlearner = mean_classifier(milk.supervised.raw_svm()) model = meanlearner.train(training_groups, labeled_groups) res = model.apply([ [f0, f1, f3] ]) ''' def __init__(self, base): self.base = base if hasattr(base, 'set_option'): self.set_option = base.set_option def train(self, gfeatures, glabels, normalisedlabels=False): features, labels = _concatenate_features_labels(gfeatures, glabels) return mean_model(self.base.train(features, labels)) class mean_model(object): def __init__(self, base): self.base = base def apply(self, gfeatures): return np.mean([self.base.apply(feats) for feats in gfeatures]) def remove_outliers(features, limit, min_size): ''' features = remove_outliers(features, limit, min_size) ''' nsize = int(limit * len(features)) if nsize < min_size: return features normed = features - features.mean(0) std = normed.std(0) std[std == 0] = 1 normed /= std f2_sum1 = (normed**2).mean(1) values = f2_sum1.copy() values.sort() top = values[nsize] selected = f2_sum1 < top return features[selected] class filter_outliers_model(object): def __init__(self, limit, min_size): self.limit = limit self.min_size = min_size def apply(self, features): return remove_outliers(features, self.limit, self.min_size) class filter_outliers(object): def __init__(self, limit=.9, min_size=3): self.limit = limit self.min_size = min_size def train(self, features, labels, normalisedlabels=False): return filter_outliers_model(self.limit, self.min_size) milk-0.3.4/milk/supervised/tree.py0000644000175000017500000001477311462146402017722 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. ''' ================ Tree Classifier ================ Decision tree based classifier ------------------------------ ''' from __future__ import division import numpy as np from .classifier import normaliselabels class Leaf(object): ''' v : value w : weight ''' def __init__(self, v, w): self.v = v self.w = w def __repr__(self): return 'Leaf(%s,%s)' % (self.v, self.w) class Node(object): # This could be replaced by a namedtuple def __init__(self, featid, featval, left, right): self.featid = featid self.featval = featval self.left = left self.right = right def _split(features, labels, criterion, subsample, R): N,f = features.shape if subsample is not None: samples = np.array(R.sample(xrange(features.shape[1]), subsample)) features = features[:, samples] f = subsample best = None best_val = -1. for i in xrange(f): domain_i = sorted(set(features[:,i])) for d in domain_i[1:]: cur_split = (features[:,i] < d) value = criterion(labels[cur_split],labels[~cur_split]) if value > best_val: best_val = value if subsample is not None: ti = samples[i] else: ti = i best = ti,d return best from ._tree import set_entropy from ._tree import information_gain as _information_gain def information_gain(labels0, labels1, include_entropy=False): ''' ig = information_gain(labels0, labels1, include_entropy=False) Information Gain See http://en.wikipedia.org/wiki/Information_gain_in_decision_trees The function calculated here does not include the original entropy unless you explicitly ask for it (by passing include_entropy=True) ''' if include_entropy: return set_entropy(np.concatenate( (labels0, labels1) )) + \ _information_gain(labels0, labels1) return _information_gain(labels0, labels1) def build_tree(features, labels, criterion, min_split=4, subsample=None, R=None): ''' tree = build_tree(features, labels, criterion, min_split=4, subsample=None, R=None) Parameters ---------- features : sequence features to use labels : sequence labels criterion : function {labels} x {labels} -> float function to measure goodness of split min_split : integer minimum size to split on subsample : integer, optional if given, then, at each step, choose Returns ------- tree : Tree ''' assert len(features) == len(labels), 'build_tree: Nr of labels does not match nr of features' features = np.asanyarray(features) labels = np.asanyarray(labels) if subsample is not None: if subsample <= 0: raise ValueError('milk.supervised.tree.build_tree: `subsample` must be > 0.\nDid you mean to use None to signal no subsample?') from ..utils import get_pyrandom R = get_pyrandom(R) def recursive(features, labels): N = float(len(labels)) if N < min_split: return Leaf(labels.sum()/N, N) S = _split(features, labels, criterion, subsample, R) if S is None: return Leaf(labels.sum()/N, N) i,thresh = S split = features[:,i] < thresh return Node(featid=i, featval=thresh, left =recursive(features[ split], labels[ split]), right=recursive(features[~split], labels[~split])) return recursive(features, labels) def apply_tree(tree, features): ''' conf = apply_tree(tree, features) Applies the decision tree to a set of features. ''' if type(tree) is Leaf: return tree.v if features[tree.featid] < tree.featval: return apply_tree(tree.left, features) return apply_tree(tree.right, features) class tree_learner(object): ''' tree = tree_learner() model = tree.train(features,labels) predicted = model.apply(testfeatures) A decision tree classifier (currently, implements the greedy ID3 algorithm without any pruning. Attributes ---------- criterion : function, optional criterion to use for tree construction, this should be a function that receives a set of labels (default: information_gain). min_split : integer, optional minimum size to split on (default: 4). ''' def __init__(self, criterion=information_gain, min_split=4, return_label=True, subsample=None, R=None): self.criterion = criterion self.min_split = min_split self.return_label = return_label self.subsample = subsample self.R = R def train(self, features, labels, normalisedlabels=False): if not normalisedlabels: labels,names = normaliselabels(labels) tree = build_tree(features, labels, self.criterion, self.min_split, self.subsample, self.R) return tree_model(tree, self.return_label) tree_classifier = tree_learner class tree_model(object): ''' tree model ''' def __init__(self, tree, return_label): self.tree = tree self.return_label = return_label def apply(self,feats): value = apply_tree(self.tree, feats) if self.return_label: return value > .5 return value milk-0.3.4/milk/supervised/featureselection.py0000644000175000017500000001674411461573003022324 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from numpy.linalg import det from . classifier import normaliselabels _TOLERANCE = 0 _SIGNIFICANCE_IN = .15 _SIGNIFICANCE_OUT = .15 def _sweep(A, k, flag): Akk = A[k,k] # cross[i,j] = A[i,k] * A[k,j] cross = (A[:,k][:, np.newaxis] * A[k]) B = A - cross/Akk # currently: B[i,j] = A[i,j] - A[i,k]*A[k,j]/Akk # Now fix row k and col k, followed by Bkk B[k] = flag * A[k]/A[k,k] B[:,k] = flag * A[:,k]/A[k,k] B[k,k] = -1./Akk return B def sda(features, labels, tolerance=None, significance_in=None, significance_out=None): ''' features_idx = sda(features,labels) Perform Stepwise Discriminant Analysis for feature selection Pre filter the feature matrix to remove linearly dependent features before calling this function. Behaviour is undefined otherwise. This implements the algorithm described in Jennrich, R.I. (1977), "Stepwise Regression" & "Stepwise Discriminant Analysis," both in Statistical Methods for Digital Computers, eds. K. Enslein, A. Ralston, and H. Wilf, New York; John Wiley & Sons, Inc. ''' import scipy.stats if tolerance is None: tolerance = _TOLERANCE if significance_in is None: significance_in = _SIGNIFICANCE_IN if significance_out is None: significance_out = _SIGNIFICANCE_OUT assert len(features) == len(labels), 'milk.supervised.featureselection.sda: length of features not the same as length of labels' N, m = features.shape labels,labelsu = normaliselabels(labels) q = len(labelsu) df = features - features.mean(0) T = np.dot(df.T, df) dfs = [(features[labels == i] - features[labels == i].mean(0)) for i in xrange(q)] W = np.sum(np.dot(d.T, d) for d in dfs) ignoreidx = ( W.diagonal() == 0 ) if ignoreidx.any(): idxs, = np.where(~ignoreidx) F = sda(features[:,~ignoreidx],labels) return idxs[F] output = [] D = W.diagonal() df1 = q-1 last_enter_k = -1 while True: V = W.diagonal()/T.diagonal() W_d = W.diagonal() V_neg = (W_d < 0) p = V_neg.sum() if V_neg.any(): V_m = V[V_neg].min() k, = np.where(V == V_m) k = k[0] Fremove = (N-p-q+1)/(q-1)*(V_m-1) df2 = N-p-q+1 PrF = 1 - scipy.stats.f.cdf(Fremove,df1,df2) if PrF > significance_out: #print 'removing ',k, 'V(k)', 1./V_m, 'Fremove', Fremove, 'df1', df1, 'df2', df2, 'PrF', PrF if k == last_enter_k: # We are going into an infinite loop. import warnings warnings.warn('pyslic.featureselection.sda: infinite loop detected (maybe bug?).') break W = _sweep(W,k,1) T = _sweep(T,k,1) continue ks = ( (W_d / D) > tolerance) if ks.any(): V_m = V[ks].min() k, = np.where(V==V_m) k = k[0] Fenter = (N-p-q)/(q-1) * (1-V_m)/V_m df2 = N-p-q PrF = 1 - scipy.stats.f.cdf(Fenter,df1,df2) if PrF < significance_in: #print 'adding ',k, 'V(k)', 1./V_m, 'Fenter', Fenter, 'df1', df1, 'df2', df2, 'PrF', PrF W = _sweep(W,k,-1) T = _sweep(T,k,-1) if PrF < .0001: output.append((Fenter,k)) last_enter_k = k continue break output.sort(reverse=True) return np.array([x[1] for x in output]) def linearly_independent_subset(V, threshold=1.e-5, return_orthogonal_basis=False): ''' subset = linearly_independent_subset(V, threshold=1.e-5) subset,U = linearly_independent_subset(V, threshold=1.e-5, return_orthogonal_basis=True) Discover a linearly independent subset of `V` Parameters ---------- V : sequence of input vectors threshold : float, optional vectors with 2-norm smaller or equal to this are considered zero (default: 1e.-5) return_orthogonal_basis : Boolean, optional whether to return orthogonal basis set Returns ------- subset : ndarray of integers indices used for basis U : 2-array orthogonal basis into span{V} Implementation Reference ------------------------ Use Gram-Schmidt with a check for when the v_k is close enough to zero to ignore See http://en.wikipedia.org/wiki/Gram-Schmidt_process ''' V = np.array(V, copy=True) orthogonal = [] used = [] for i,u in enumerate(V): for v in orthogonal: u -= np.dot(u,v)/np.dot(v,v) * v if np.dot(u,u) > threshold: orthogonal.append(u) used.append(i) if return_orthogonal_basis: return np.array(used),np.array(orthogonal) return np.array(used) def linear_independent_features(features, labels=None): ''' indices = linear_independent_features(features, labels=None) Returns the indices of a set of linearly independent features (columns). Parameters ---------- features : ndarray labels : ignored This argument is only here to conform to the learner interface. Returns ------- indices : ndarray of integers indices of features to keep See Also -------- `linearly_independent_subset` : this function is equivalent to `linearly_independent_subset(features.T)` ''' return linearly_independent_subset(features.T) class filterfeatures(object): ''' selector = filterfeatures(idxs) Returns a transformer which selects the features given by idxs ''' def __init__(self, idxs): self.idxs = idxs def apply(self, features): return features[self.idxs] def __repr__(self): return 'filterfeatures(%s)' % self.idxs class featureselector(object): ''' selector = featureselector(function) Returns a transformer which selects features according to selected_idxs = function(features,labels) ''' def __init__(self, selector): self.selector = selector def train(self, features, labels, normalisedlabels=False): idxs = self.selector(features, labels) return filterfeatures(idxs) def __repr__(self): return 'featureselector(%s)' % self.selector def sda_filter(): return featureselector(sda) milk-0.3.4/milk/supervised/classifier.py0000644000175000017500000000556311461573003021104 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy import numpy as np __all__ = ['normaliselabels', 'ctransforms'] def normaliselabels(labels): ''' normalised, names = normaliselabels(labels) Normalises the labels to be integers from 0 through N-1 `normalised` is a np.array, while `names` is a list mapping the indices to the old names. Parameters ---------- labels : any iterable of labels Returns ------ normalised : a numpy ndarray of integers 0 .. N-1 names : list of label names ''' names = sorted(set(labels)) normalised = map(names.index, labels) return np.array(normalised), names class ctransforms_model(object): ''' model = ctransforms_model(models) A model that consists of a series of transformations. See Also -------- ctransforms ''' def __init__(self, models): self.models = models def apply(self,features): for T in self.models: features = T.apply(features) return features class ctransforms(object): ''' ctransf = ctransforms(c0, c1, c2, ...) Concatenate transforms. ''' def __init__(self,*args): self.transforms = args def train(self, features, labels, normalisedlabels=False): models = [] model = None for T in self.transforms: if model is not None: features = numpy.array([model.apply(f) for f in features]) model = T.train(features, labels, normalisedlabels=normalisedlabels) models.append(model) return ctransforms_model(models) def set_option(self, opt, val): idx, opt = opt self.transforms[idx].set_option(opt,val) milk-0.3.4/milk/supervised/_svm.cpp0000644000175000017500000010234111453701533020050 0ustar luispedroluispedro// Copyright (C) 2008, Luís Pedro Coelho // Copyright (c) 2000-2008 Chih-Chung Chang and Chih-Jen Lin (LIBSVM Code) // // Permission is hereby granted, free of charge, to any person obtaining a copy // of this software and associated documentation files (the "Software"), to deal // in the Software without restriction, including without limitation the rights // to use, copy, modify, merge, publish, distribute, sublicense, and/or sell // copies of the Software, and to permit persons to whom the Software is // furnished to do so, subject to the following conditions: // // The above copyright notice and this permission notice shall be included in // all copies or substantial portions of the Software. // // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR // IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, // FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE // AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER // LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, // OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN // THE SOFTWARE. #include #include #include #include #include #include #include #include //#include //#include //using __gnu_debug::vector; //using __gnu_debug::list; using std::vector; using std::list; extern "C" { #include #include } namespace { const double INF = HUGE_VAL; // This is a boost function // Copied here for convenience. template inline Iter prior(Iter it) { --it; return it; } using std::max; using std::min; template T median(T bot, T mid, T hi) { if (mid < bot) return bot; if (mid > hi) return hi; return mid; } struct SMO_Exception { SMO_Exception(const char* msg): msg(msg) { } const char* msg; }; struct Python_Exception { }; void check_for_interrupts() { PyErr_CheckSignals(); if (PyErr_Occurred()) { throw Python_Exception(); } } class KernelComputation { public: virtual ~KernelComputation() { } virtual double do_kernel(int i1, int i2) const = 0; }; class KernelCache { public: KernelCache(std::auto_ptr computation_, int N, int cache_nr_doubles); virtual ~KernelCache(); double* get_kline(int idx, int size = -1); double* get_diag(); double kernel_apply(int i1, int i2) const; protected: virtual double do_kernel(int i1, int i2) const { return computation_->do_kernel(i1,i2); } const int N_; private: std::auto_ptr computation_; double ** cache_; double * dcache_; int cache_free_; list cache_lru; vector::iterator> cache_iter; }; KernelCache::KernelCache(std::auto_ptr computation, int N, int cache_nr_floats): N_(N), computation_(computation), dcache_(0) { cache_ = new double*[N_]; for (int i = 0; i != N_; ++i) cache_[i] = 0; cache_free_ = (cache_nr_floats/N_); cache_iter.resize(N_,cache_lru.end()); } KernelCache::~KernelCache() { for (int i = 0; i != N_; ++i) delete [] cache_[i]; delete [] cache_; delete [] dcache_; } double* KernelCache::get_kline(int idx, int s) { if (s == -1) s = N_; assert (s <= N_); if (!cache_[idx]) { if (!cache_free_) { int to_remove = cache_lru.front(); cache_lru.pop_front(); cache_[idx] = cache_[to_remove]; cache_[to_remove] = 0; } else { cache_[idx] = new double[N_]; --cache_free_; } for (int i = 0; i != N_; ++i) { if (i == idx && dcache_) cache_[i][i] = dcache_[i]; else if(i != idx && cache_[i]) cache_[idx][i] = cache_[i][idx]; else cache_[idx][i] = do_kernel(idx,i); } } else { cache_lru.erase(cache_iter[idx]); } cache_lru.push_back(idx); cache_iter[idx]=prior(cache_lru.end()); return cache_[idx]; } double* KernelCache::get_diag() { if (!dcache_) { dcache_ = new double[N_]; for (int i = 0; i != N_; ++i) { if (cache_[i]) dcache_[i] = cache_[i][i]; else dcache_[i] = do_kernel(i,i); } } return dcache_; } /*** * Returns the value of Kernel(X_i1, X_i2). * Uses the cache if possible, but does not update it. */ double KernelCache::kernel_apply(int i1, int i2) const { if (cache_[i1]) { assert(do_kernel(i1,i2) == cache_[i1][i2]); return cache_[i1][i2]; } if (cache_[i2]) { assert(do_kernel(i1,i2) == cache_[i2][i1]); return cache_[i2][i1]; } if (i1 == i2 && dcache_) { assert(do_kernel(i1,i2) == dcache_[i1]); return dcache_[i1]; } return do_kernel(i1,i2); } class PyKernel : public KernelComputation { PyKernel(const PyKernel&); public: PyKernel(PyObject* X, PyObject* kernel, int N); ~PyKernel(); protected: virtual double do_kernel(int i1, int i2) const; private: PyObject* const X_; PyObject* const pykernel_; const int N_; }; PyKernel::PyKernel(PyObject* X, PyObject* kernel, int N): X_(X), pykernel_(kernel), N_(N) { Py_INCREF(X); Py_INCREF(kernel); } PyKernel::~PyKernel() { Py_DECREF(X_); Py_DECREF(pykernel_); } double PyKernel::do_kernel(int i1, int i2) const { assert(i1 < N_); assert(i2 < N_); PyObject* obj1 = PySequence_GetItem(X_,i1); PyObject* obj2 = PySequence_GetItem(X_,i2); if (!obj1 || !obj2) { Py_XDECREF(obj1); Py_XDECREF(obj2); throw SMO_Exception("svm.eval_SMO: Unable to access element in X array"); } PyObject* arglist = Py_BuildValue("(OO)",obj1,obj2); PyObject* result = PyEval_CallObject(pykernel_,arglist); Py_XDECREF(obj1); Py_XDECREF(obj2); Py_DECREF(arglist); if (!result) { check_for_interrupts(); throw SMO_Exception("svm.eval_SMO: Unable to call kernel"); } double val = PyFloat_AsDouble(result); Py_DECREF(result); return val; } class RBFKernel : public KernelComputation { public: RBFKernel(PyArrayObject* X, double gamma); ~RBFKernel(); protected: virtual double do_kernel(int i1, int i2) const; private: PyArrayObject* X_; double ngamma_; const int N_; const int N1_; }; RBFKernel::RBFKernel(PyArrayObject* X, double gamma): X_(reinterpret_cast(X)), ngamma_(-1./gamma), N_(PyArray_DIM(X,0)), N1_(PyArray_DIM(X,1)) { if (!PyArray_Check(X)) { throw SMO_Exception("RBF Kernel used, but not with numpy array."); } if (!PyArray_ISCARRAY(X)) { throw SMO_Exception("RBF Kernel used but not with CARRAY."); } Py_INCREF(X); } RBFKernel::~RBFKernel() { Py_DECREF(X_); } double RBFKernel::do_kernel(int i1, int i2) const { assert(i1 < N_); assert(i2 < N_); const double* data1 = static_cast(PyArray_GETPTR1(X_,i1)); const double* data2 = static_cast(PyArray_GETPTR1(X_,i2)); double sumdiff = 0.; for (int i = 0; i != N1_; ++i) { double diff = data1[i]-data2[i]; sumdiff += diff * diff; } sumdiff *= ngamma_; double res = std::exp(sumdiff); return res; } class DotKernel : public KernelComputation { public: explicit DotKernel(PyArrayObject* X); ~DotKernel(); protected: virtual double do_kernel(int i1, int i2) const; private: PyArrayObject* X_; const int N_; const int N1_; }; DotKernel::DotKernel(PyArrayObject* X): X_(reinterpret_cast(X)), N_(PyArray_DIM(X,0)), N1_(PyArray_DIM(X,1)) { if (!PyArray_Check(X)) { throw SMO_Exception("Dot Kernel used, but not with numpy array."); } if (!PyArray_ISCARRAY(X)) { throw SMO_Exception("Dot Kernel used but not with CARRAY."); } Py_INCREF(X); } DotKernel::~DotKernel() { Py_DECREF(X_); } double DotKernel::do_kernel(int i1, int i2) const { assert(i1 < N_); assert(i2 < N_); const double* data1 = static_cast(PyArray_GETPTR1(X_,i1)); const double* data2 = static_cast(PyArray_GETPTR1(X_,i2)); double dotsum = 0.; for (int i = 0; i != N1_; ++i) { dotsum += data1[i] * data2[i]; } return dotsum; } class PrecomputedKernel : public KernelComputation { public: PrecomputedKernel(PyArrayObject* X); ~PrecomputedKernel(); protected: virtual double do_kernel(int i1, int i2) const; private: PyArrayObject* X_; }; PrecomputedKernel::PrecomputedKernel(PyArrayObject* X): X_(X) { if (!PyArray_Check(X)) { throw SMO_Exception("PrecomputedKernel used, but not with numpy array."); } if (!PyArray_ISCARRAY(X)) { throw SMO_Exception("Precomputed used but not with CARRAY."); } Py_INCREF(X); } PrecomputedKernel::~PrecomputedKernel() { Py_DECREF(X_); } double PrecomputedKernel::do_kernel(int i1, int i2) const { const double* data = static_cast(PyArray_GETPTR2(X_,i1,i2)); return *data; } std::auto_ptr get_kernel(PyObject* X, PyObject* kernel) { typedef std::auto_ptr res_type; if (PyCallable_Check(kernel)) return res_type(new PyKernel(X, kernel, PySequence_Length(X))); if (!PyTuple_Check(kernel) || PyTuple_Size(kernel) != 2) throw SMO_Exception("Cannot parse kernel."); PyObject* type = PyTuple_GET_ITEM(kernel,0); PyObject* arg = PyTuple_GET_ITEM(kernel,1); if (!PyInt_Check(type) || !PyFloat_Check(arg)) throw SMO_Exception("Cannot parse kernel (wrong types)"); long type_nr = PyInt_AsLong(type); double arg_value = PyFloat_AsDouble(arg); switch (type_nr) { case 0: return res_type(new RBFKernel(reinterpret_cast(X), arg_value)); case 1: return res_type(new PrecomputedKernel(reinterpret_cast(X))); case 2: return res_type(new DotKernel(reinterpret_cast(X))); default: throw SMO_Exception("Unknown kernel type!"); } } class LIBSVM_KernelCache : public KernelCache { public: LIBSVM_KernelCache(const int* Y, std::auto_ptr kernel, int N, int cache_nr_doubles) :KernelCache(kernel,N,cache_nr_doubles), Y_(Y) { } private: double do_kernel(int i, int j) const { double res = KernelCache::do_kernel(i,j); return res * Y_[i] * Y_[j]; } const int* Y_; }; class SMO { public: SMO(PyObject* X, int* Y, double* Alphas, double& b, double C, int N, PyObject* kernel, double eps, double tol, int cache_size): Alphas(Alphas), Y(Y), b(b), C(C), N(N), cache_(get_kernel(X,kernel),N,cache_size), eps(eps), tol(tol) { } void optimise(); double apply(int) const; bool take_step(int,int); bool examine_example(int); double get_error(int) const; private: double* Alphas; int* Y; double& b; double C; int N; mutable KernelCache cache_; const double eps; const double tol; }; double SMO::get_error(int j) const { return apply(j) - Y[j]; } double SMO::apply(int j) const { double sum = -b; double* Kernel_Line = cache_.get_kline(j); for (int i = 0; i != N; ++i) { if (Alphas[i] != C) { sum += Y[i] * Alphas[i] * Kernel_Line[i]; } } //std::cout << "SMO::apply( " << j << " ): " << sum << '\n'; return sum; } bool SMO::take_step(int i1, int i2) { //std::cout << "take_step( " << i1 << ", " << i2 << " );\n"; if (i1 == i2) return false; const double alpha1 = Alphas[i1]; const double alpha2 = Alphas[i2]; const int y1 = Y[i1]; const int y2 = Y[i2]; double L, H; if (y1 != y2) { L = max(0.,alpha2-alpha1); H = min(C,C+alpha2-alpha1); } else { L = max(0.,alpha1+alpha2-C); H = min(C,alpha1+alpha2); } if (L == H) return false; const int s = y1*y2; const double E1 = get_error(i1); const double E2 = get_error(i2); const double k11 = cache_.kernel_apply(i1,i1); const double k12 = cache_.kernel_apply(i1,i2); const double k22 = cache_.kernel_apply(i2,i2); const double eta = 2*k12-k11-k22; double a1,a2; if (eta < 0) { a2 = alpha2-y2*(E1-E2)/eta; a2 = median(L,a2,H); } else { double gamma = alpha1+s*alpha2; // Eq. (12.22) double v1=E1+y1+b-y1*alpha1*k11-y2*alpha2*k12; // Eq. (12.21) // Note that f(x1) = E1 + y1 double v2=E2+y2+b-y1*alpha1*k12-y2*alpha2*k22; // Eq. (12.21) double L_obj = gamma-s*L+L-.5*k11*(gamma-s*L)*(gamma-s*L)-.5*k22*L*L-s*k12*(gamma-s*L)*L-y1*(gamma-s*L)*v1-y2*L*v2; // # + W_const # Eq. (12.23) double H_obj = gamma-s*H+H-.5*k11*(gamma-s*H)*(gamma-s*L)-.5*k22*H*H-s*k12*(gamma-s*H)*H-y1*(gamma-s*H)*v1-y2*H*v2; // # + W_const # Eq. (12.23) if (L_obj > H_obj + eps) { a2 = L; } else if (L_obj < H_obj - eps) { a2 = H; } else { a2 = alpha2; } } if (a2 < tol) a2 = 0; else if (a2 > C-tol) a2 = C; if (std::abs(a2-alpha2) < eps*(a2+alpha2+eps)) return false; a1 = alpha1+s*(alpha2-a2); if (a1 < tol) a1 = 0; if (a1 > C-tol) a1 = C; // update everything Alphas[i1]=a1; Alphas[i2]=a2; double b1 = E1 + Y[i1]*(a1-alpha1)*k11+Y[i2]*(a2-alpha2)*k12+b; // Eq. (12.9) double b2 = E2 + Y[i1]*(a1-alpha1)*k12+Y[i2]*(a2-alpha2)*k22+b; // Eq. (12.10) const double new_b = (b1+b2)/2.; /* #for i in xrange(N): # if Alphas[i] in (0,C): # continue # elif i == i1 or i == i2: # E[i] = 0 # else: # E[i] += y1*(a1-alpha1)*kernel_apply(i1,i)+y2*(a2-alpha2)*kernel_apply(i2,i) + (b-new_b) # Eq. (12.11) #E[i1]=f_at(i1)-y1 #E[i2]=f_at(i2)-y2 */ b = new_b; return true; } bool SMO::examine_example(int i2) { //std::cout << "examine_example( " << i2 << " ) " << std::endl; const int y2 = Y[i2]; const double alpha2 = Alphas[i2]; const double E2 = get_error(i2); const double r2 = E2 * y2; //#print 'alpha2', alpha2, 'E2', E2, 'r2', r2 if ( ( (r2 < -tol) && (alpha2 < C) ) || ( (r2 > tol) && (alpha2 > 0) )){ int best_i1 = -1; double bestE = -1; for (int i = 0; i != N; ++i) { if (Alphas[i] != 0 && Alphas[i] != C) { double dE = E2-get_error(i); if (dE < 0.) dE = -dE; if (dE > bestE) { bestE=dE; best_i1 = i; } } } if (best_i1 != -1 && take_step(best_i1,i2)) return true; for (int i1 = 0; i1 != N; ++i1) { if (Alphas[i1] && Alphas[i1] != C && take_step(i1,i2)) return true; } for (int i1 = 0; i1 != N; ++i1) { if (take_step(i1,i2)) return true; } } return false; } void SMO::optimise() { b = 0; for (int i = 0; i != N; ++i) Alphas[i] = 0; int changed = 0; bool examineAll = true; //int iter = 0; while (changed || examineAll) { //std::cout << "SMO::optimize loop: " << iter++ << "\n"; check_for_interrupts(); changed = 0; for (int i = 0; i != N; ++i) { if (examineAll || (Alphas[i] != 0 && Alphas[i] != C)) { changed += examine_example(i); } } if (examineAll) examineAll = false; else if (!changed) examineAll = true; } } void assert_type_contiguous(PyArrayObject* array,int type) { if (!PyArray_Check(array) || PyArray_TYPE(array) != type || !PyArray_ISCONTIGUOUS(array)) { throw SMO_Exception("Arguments to eval_(SMO|LIBSVM) don't conform to expectation. Are you calling this directly? This is an internal function!"); } } PyObject* eval_SMO(PyObject* self, PyObject* args) { try { PyObject* X; PyArrayObject* Y; PyArrayObject* Alphas0; PyArrayObject* params; PyObject* kernel; int cache_size; if (!PyArg_ParseTuple(args, "OOOOOi", &X, &Y, &Alphas0, ¶ms,&kernel,&cache_size)) { const char* errmsg = "Arguments were not what was expected for eval_SMO.\n" "This is an internal function: Do not call directly unless you know what you're doing.\n"; PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } assert_type_contiguous(Y,NPY_INT32); assert_type_contiguous(Alphas0,NPY_DOUBLE); assert_type_contiguous(params,NPY_DOUBLE); if (PyArray_DIM(params,0) < 4) throw SMO_Exception("eval_SMO: Too few parameters"); int * Yv = static_cast(PyArray_DATA(Y)); double* Alphas = static_cast(PyArray_DATA(Alphas0)); unsigned N = PyArray_DIM(Y,0); double b = *static_cast(PyArray_GETPTR1(params,0)); double C = *static_cast(PyArray_GETPTR1(params,1)); double eps = *static_cast(PyArray_GETPTR1(params,2)); double tol = *static_cast(PyArray_GETPTR1(params,3)); SMO optimiser(X,Yv,Alphas,b,C,N,kernel,eps,tol,cache_size); optimiser.optimise(); *static_cast(PyArray_GETPTR1(params,0)) = b; // Write back b Py_RETURN_NONE; } catch (const Python_Exception&) { // if Python_Exception was thrown, then PyErr is already set. return 0; } catch (const SMO_Exception& exc) { PyErr_SetString(PyExc_RuntimeError,exc.msg); return 0; } catch (...) { PyErr_SetString(PyExc_RuntimeError,"Some sort of exception in eval_SMO."); return 0; } } // The code for LIBSVM_Solver is taken from LIBSVM and adapted to work well in milk // Copyright (c) 2000-2008 Chih-Chung Chang and Chih-Jen Lin // Changes were made to make it more similar to our formulation. // An SMO algorithm in Fan et al., JMLR 6(2005), p. 1889--1918 // Solves: // // min 0.5(\alpha^T Q \alpha) + p^T \alpha // // y^T \alpha = \delta // y_i = +1 or -1 // 0 <= alpha_i <= Cp for y_i = 1 // 0 <= alpha_i <= Cn for y_i = -1 // // Given: // // Q, p, y, Cp, Cn, and an initial feasible point \alpha // l is the size of vectors and matrices // eps is the stopping tolerance // // solution will be put in \alpha, objective value will be put in obj // #define info printf void info_flush() { } class LIBSVM_Solver { public: LIBSVM_Solver(PyObject* X, int* Y, double* Alphas, double* p, double& b, double C, int N, PyObject* kernel, double eps, double tol, int cache_size, bool shrinking): Alphas(Alphas), Y(Y), b(b), C(C), N(N), cache_(Y,get_kernel(X,kernel),N,cache_size), eps(eps), tol(tol), alpha_status(N), active_size(N), active_set(N), p(p), shrinking(shrinking), tau(eps) { } virtual ~LIBSVM_Solver() { } struct SolutionInfo { double obj; double rho; double upper_bound_p; double upper_bound_n; double r; // for LIBSVM_Solver_NU }; void optimise(); protected: double* Alphas; int* Y; double& b; double C; const int N; LIBSVM_KernelCache cache_; const double eps; const double tol; vector G; // gradient of objective function enum alpha_status_e { lower_bound, upper_bound, free}; vector alpha_status; int active_size; vector active_set; double *p; vector G_bar; // gradient, if we treat free variables as 0 bool unshrinked; // XXX bool shrinking; double tau; double get_C(int i) const { return C; //return (Y[i] > 0)? Cp : Cn; } void update_alpha_status(int i) { if(Alphas[i] >= get_C(i)) alpha_status[i] = upper_bound; else if(Alphas[i] <= 0) alpha_status[i] = lower_bound; else alpha_status[i] = free; } bool is_upper_bound(int i) const { return alpha_status[i] == upper_bound; } bool is_lower_bound(int i) const { return alpha_status[i] == lower_bound; } bool is_free(int i) const { return alpha_status[i] == free; } void swap_index(int i, int j); void reconstruct_gradient(); virtual bool select_working_set(int &i, int &j); virtual double calculate_rho(); virtual void do_shrinking(); private: bool be_shrunken(int i, double Gmax1, double Gmax2); void print_status() const; }; void LIBSVM_Solver::swap_index(int i, int j) { // We *do not* swap in the cache or kernel // Therefore *all acesses* to the cache or kernel // must be of the form cache_.get_kernel(active_set[i]) // instead of cache_.get_kernel(i)! std::swap(Y[i],Y[j]); std::swap(G[i],G[j]); std::swap(alpha_status[i],alpha_status[j]); std::swap(Alphas[i],Alphas[j]); std::swap(p[i],p[j]); std::swap(active_set[i],active_set[j]); std::swap(G_bar[i],G_bar[j]); } void LIBSVM_Solver::reconstruct_gradient() { // reconstruct inactive elements of G from G_bar and free variables if(active_size == N) return; for(int i=active_size;i= 0) && (i < active_size)); assert((j >= 0) && (j < active_size)); // update alpha[i] and alpha[j], handle bounds carefully const double *Q_i = cache_.get_kline(active_set[i], active_size); const double *Q_j = cache_.get_kline(active_set[j], active_size); const double C_i = get_C(i); const double C_j = get_C(j); const double old_alpha_i = Alphas[i]; const double old_alpha_j = Alphas[j]; if(Y[i]!=Y[j]) { double quad_coef = Q_i[active_set[i]]+Q_j[active_set[j]]+2.0*Q_i[active_set[j]]; if (quad_coef <= 0) quad_coef = tau; const double delta = (-G[i]-G[j])/quad_coef; const double diff = Alphas[i] - Alphas[j]; Alphas[i] += delta; Alphas[j] += delta; if(diff > 0) { if(Alphas[j] < 0) { Alphas[j] = 0; Alphas[i] = diff; } } else { if(Alphas[i] < 0) { Alphas[i] = 0; Alphas[j] = -diff; } } if(diff > C_i - C_j) { if(Alphas[i] > C_i) { Alphas[i] = C_i; Alphas[j] = C_i - diff; } } else { if(Alphas[j] > C_j) { Alphas[j] = C_j; Alphas[i] = C_j + diff; } } //print_status(); } else { double quad_coef = Q_i[active_set[i]]+Q_j[active_set[j]]-2*Q_i[active_set[j]]; if (quad_coef <= 0) quad_coef = tau; const double delta = (G[i]-G[j])/quad_coef; double sum = Alphas[i] + Alphas[j]; Alphas[i] -= delta; Alphas[j] += delta; if(sum > C_i) { if(Alphas[i] > C_i) { Alphas[i] = C_i; Alphas[j] = sum - C_i; } } else { if(Alphas[j] < 0) { Alphas[j] = 0; Alphas[i] = sum; } } if(sum > C_j) { if(Alphas[j] > C_j) { Alphas[j] = C_j; Alphas[i] = sum - C_j; } } else { if(Alphas[i] < 0) { Alphas[i] = 0; Alphas[j] = sum; } } } // update G const double delta_Alphas_i = Alphas[i] - old_alpha_i; const double delta_Alphas_j = Alphas[j] - old_alpha_j; // print_status(); for(int k=0;krho = calculate_rho(); b = calculate_rho(); // calculate objective value double v = 0; for(int i=0;iobj = v/2; // put back the solution for (int i = 0; i != N; ++i) { while (active_set[i] != i) { int j = active_set[i]; std::swap(Y[i],Y[j]); // It's not polite to clobber Y, so put it back std::swap(Alphas[i],Alphas[j]); std::swap(active_set[i],active_set[j]); } } //si->upper_bound_p = Cp; //si->upper_bound_n = Cn; } void LIBSVM_Solver::print_status() const { std::cout << " active_set_size: " << active_size << "\n"; for (int i = 0; i != N; ++i) { std::cout << " active_set[i]: " << active_set[i] << "\n"; std::cout << " p[i]: " << p[i] << "\n"; std::cout << " G[i]: " << G[i] << "\n"; std::cout << " G_bar[i]: " << G_bar[i] << "\n"; std::cout << " alpha_status[i]: " << alpha_status[i] << "\n"; std::cout << " Y[i]: " << Y[i] << "\n"; std::cout << " Alphas[i]: " << Alphas[i] << "\n"; std::cout << "\n"; } } // return true if already optimal, return false otherwise bool LIBSVM_Solver::select_working_set(int &out_i, int &out_j) { // return i,j such that // i: maximizes -y_i * grad(f)_i, i in I_up(\alpha) // j: minimizes the decrease of obj value // (if quadratic coefficeint <= 0, replace it with tau) // -y_j*grad(f)_j < -y_i*grad(f)_i, j in I_low(\alpha) //print_status(); double Gmax = -INF; double Gmax2 = -INF; int Gmax_idx = -1; int Gmin_idx = -1; double obj_diff_min = INF; for(int t=0;t= Gmax) { Gmax = -Y[t]*G[t]; Gmax_idx = t; } } } if (Gmax_idx == -1) return true; const int i = Gmax_idx; const double* Q_i = cache_.get_kline(active_set[i], N); const double* QDiag = cache_.get_diag(); for(int j=0;j= Gmax2) Gmax2 = YGj; if (grad_diff > 0) { const double quad_coef = Q_i[active_set[i]] + QDiag[active_set[j]] - 2*Y[j]*Y[i]*Q_i[active_set[j]]; const double obj_diff_factor = (quad_coef > 0) ? quad_coef : tau; const double obj_diff = -(grad_diff*grad_diff)/obj_diff_factor; if (obj_diff <= obj_diff_min) { Gmin_idx = j; obj_diff_min = obj_diff; } } } } if (Gmin_idx == -1) return true; if(Gmax+Gmax2 < eps) return true; out_i = Gmax_idx; out_j = Gmin_idx; return false; } bool LIBSVM_Solver::be_shrunken(int i, const double Gmax1, const double Gmax2) { if(is_upper_bound(i)) { if(Y[i]==+1) return -G[i] > Gmax1; return -G[i] > Gmax2; } else if(is_lower_bound(i)) { if (Y[i]==+1) return G[i] > Gmax2; return G[i] > Gmax1; } return false; } void LIBSVM_Solver::do_shrinking() { double Gmax1 = -INF; // max { -y_i * grad(f)_i | i in I_up(\alpha) } double Gmax2 = -INF; // max { y_i * grad(f)_i | i in I_low(\alpha) } // find maximal violating pair first for(int i=0;i i) { if (!be_shrunken(active_size, Gmax1, Gmax2)) { swap_index(i,active_size); break; } --active_size; } } } // unshrink, check all variables again before final iterations if(unshrinked || Gmax1 + Gmax2 > eps*10) return; unshrinked = true; reconstruct_gradient(); for(int i= N-1; i >= active_size; --i) { if (!be_shrunken(i, Gmax1, Gmax2)) { while (active_size < i) { if (be_shrunken(active_size, Gmax1, Gmax2)) { swap_index(i,active_size); break; } ++active_size; } ++active_size; } } } double LIBSVM_Solver::calculate_rho() { int nr_free = 0; double ub = INF, lb = -INF, sum_free = 0; for(int i=0;i0) return sum_free/nr_free; return (ub+lb)/2; } PyObject* eval_LIBSVM(PyObject* self, PyObject* args) { try { PyObject* X; PyArrayObject* Y; PyArrayObject* Alphas0; PyArrayObject* p; PyArrayObject* params; PyObject* kernel; int cache_size; if (!PyArg_ParseTuple(args, "OOOOOOi", &X, &Y, &Alphas0, &p, ¶ms,&kernel,&cache_size)) { const char* errmsg = "Arguments were not what was expected for eval_LIBSVM.\n" "This is an internal function: Do not call directly unless you know exactly what you're doing.\n"; PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } assert_type_contiguous(Y,NPY_INT32); assert_type_contiguous(Alphas0,NPY_DOUBLE); assert_type_contiguous(p,NPY_DOUBLE); assert_type_contiguous(params,NPY_DOUBLE); if (PyArray_DIM(params,0) < 4) throw SMO_Exception("eval_LIBSVM: Too few parameters"); int * Yv = static_cast(PyArray_DATA(Y)); double* Alphas = static_cast(PyArray_DATA(Alphas0)); double* pv = static_cast(PyArray_DATA(p)); unsigned N = PyArray_DIM(Y,0); double& b = *static_cast(PyArray_GETPTR1(params,0)); double C = *static_cast(PyArray_GETPTR1(params,1)); double eps = *static_cast(PyArray_GETPTR1(params,2)); double tol = *static_cast(PyArray_GETPTR1(params,3)); LIBSVM_Solver optimiser(X,Yv,Alphas,pv,b,C,N,kernel,eps,tol,cache_size, true); optimiser.optimise(); Py_RETURN_NONE; } catch (const Python_Exception&) { // if Python_Exception was thrown, then PyErr is already set. return 0; } catch (const SMO_Exception& exc) { PyErr_SetString(PyExc_RuntimeError,exc.msg); return 0; } catch (...) { PyErr_SetString(PyExc_RuntimeError,"Some sort of exception in eval_LIBSVM."); return 0; } } PyMethodDef methods[] = { {"eval_SMO",eval_SMO, METH_VARARGS , "Do NOT call directly.\n" }, {"eval_LIBSVM",eval_LIBSVM, METH_VARARGS , "Do NOT call directly.\n" }, {NULL, NULL,0,NULL}, }; const char * module_doc = "Internal SVM Module.\n" "\n" "Do NOT use directly!\n"; } // namespace extern "C" void init_svm() { import_array(); (void)Py_InitModule3("_svm", methods, module_doc); } milk-0.3.4/milk/supervised/normalise.py0000644000175000017500000001011411462317314020737 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division from collections import defaultdict import numpy as np from .classifier import normaliselabels __all__ = [ 'zscore', 'zscore_normalise', 'interval_normalise', 'chkfinite', 'sample_to_2min', 'normaliselabels' ] def zscore(features): """ features = zscore(features) Returns a copy of features which has been normalised to zscores """ mu = features.mean(0) sigma = np.std(features,0) sigma[sigma == 0] = 1 return (features - mu) / sigma class subtract_divide_model(object): def __init__(self, shift, factor): factor[factor == 0] = 1 # This makes the division a null op. self.shift = shift self.factor = factor def apply(self, features): return (features - self.shift)/self.factor def __repr__(self): return 'subtract_divide_model(%s, %s)' % (self.shift, self.factor) class zscore_normalise(object): ''' Normalise to z-scores A preprocessor that normalises features to z scores. ''' def train(self, features, labels, normalisedlabels=False): shift = features.mean(0) factor = np.std(features,0) return subtract_divide_model(shift, factor) class interval_normalise(object): ''' Linearly scale to the interval [-1,1] (per libsvm recommendation) ''' def train(self, features, labels, normalisedlabels=False): ptp = features.ptp(0) shift = features.min(0) + ptp/2. factor = ptp/2. return subtract_divide_model(shift, factor) def __repr__(self): return 'interval_normalise()' def sample_to_2min(labels): ''' selected = sample_to_2min(labels) Select examples so that the ratio of size of the largest class to the smallest class is at most two (i.e., min_label_count = min { (labels == L).sum() | for L in set(labels) } for L' in set(labels): assert (labels == L').sum() <= 2 * min_label_count ) Parameters ---------- * labels: sequence of labels Output ------ * selected: a Boolean numpy.ndarray ''' counts = defaultdict(int) for n in labels: counts[n] += 1 labels = np.asanyarray(labels) max_entries = np.min(counts.values())*2 selected = np.zeros(len(labels), bool) for c in counts.iterkeys(): p, = np.where(labels == c) p = p[:max_entries] selected[p] = 1 return selected class chkfinite(object): ''' Fill NaN & Inf values Replaces NaN & Inf values with zeros. ''' def __init__(self): pass def train(self, features, labels, normalisedlabels=False): return self def apply(self, features): nans = np.isnan(features) + np.isinf(features) if nans.any(): features = features.copy() features[nans] = 0 return features def __repr__(self): return 'chkfinite()' milk-0.3.4/milk/milk_version.py0000644000175000017500000000002611462635265017267 0ustar luispedroluispedro__version__ = '0.3.4' milk-0.3.4/milk/__init__.py0000644000175000017500000000403711457212370016323 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. ''' Milk Machine learning in Python Toplevel functions ------------------ - nfoldcrossvalidation: n-fold crossvalidation - defaultclassifier: get a general purpose classifier - kmeans: kmeans clustering Modules ------- - supervised - unsupervised - measures Example ------- :: features = np.random.randn(100,20) features[:50] *= 2 labels = np.repeat((0,1), 50) classifier = milk.defaultclassifier() model = classifier.train(features, labels) new_label = model.apply(np.random.randn(100)) new_label2 = model.apply(np.random.randn(100)*2) ''' from .nfoldcrossvalidation import nfoldcrossvalidation from .supervised.defaultclassifier import defaultclassifier from .unsupervised.kmeans import kmeans from milk_version import __version__ __all__ = [ '__version__', 'kmeans', 'defaultclassifier', 'nfoldcrossvalidation', ] milk-0.3.4/milk/utils/0000755000175000017500000000000011463416340015346 5ustar luispedroluispedromilk-0.3.4/milk/utils/utils.py0000644000175000017500000000534511452411266017067 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np import random __all__ = [ 'get_nprandom', 'get_pyrandom', ] def get_nprandom(R): ''' R' = get_nprandom(R) Returns a numpy.RandomState from R Parameters ---------- R : can be one of: None : Returns the default numpy global state integer : Uses it as a seed for constructing a new random generator RandomState : returns R Returns ------- R' : np.RandomState ''' if R is None: return np.random.mtrand._rand if type(R) == int: return np.random.RandomState(R) if type(R) is random.Random: return np.random.RandomState(R.randint(0, 2**30)) if type(R) is np.random.RandomState: return R raise TypeError,"get_nprandom() does not know how to handle type %s." % type(R) def get_pyrandom(R): ''' R = get_pyrandom(R) Returns a random.Random object based on R Parameters ---------- R : can be one of: None : Returns the default numpy global state integer : Uses it as a seed for constructing a new random generator RandomState : returns R Returns ------- R' : random.Random ''' if R is None: return random.seed.im_self if type(R) is int: return random.Random(R) if type(R) is np.random.RandomState: return random.Random(R.randint(2**30)) if type(R) is random.Random: return R raise TypeError,"get_pyrandom() does not know how to handle type %s." % type(R) milk-0.3.4/milk/utils/__init__.py0000644000175000017500000000002411231643006017445 0ustar luispedroluispedrofrom utils import * milk-0.3.4/milk/unsupervised/0000755000175000017500000000000011463416340016742 5ustar luispedroluispedromilk-0.3.4/milk/unsupervised/kmeans.py0000644000175000017500000001567311453357506020615 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from numpy import linalg from ..utils import get_pyrandom from .normalise import zscore __all__ = ['kmeans','repeated_kmeans'] def _mahalanobis2(fmatrix, x, icov): diff = fmatrix-x # The expression below seems to be faster than looping over the elements and summing return np.dot(diff, np.dot(icov, diff.T)).diagonal() def centroid_errors(fmatrix, assignments, centroids): ''' errors = centroid_errors(fmatrix, assignments, centroids) Computes the following for all i, j: ci = assignments[i] errors[i,j] = fmatrix[ci, j] - centroids[ci, j] Parameters ---------- fmatrix : feature matrix assignments : Assignments array centroids : centroids Returns ------- errors : Difference between fmatrix and corresponding centroid ''' errors = [] for k in xrange(len(centroids)): errors.append(fmatrix[assignments == k] - centroids[k]) return np.concatenate(errors) def residual_sum_squares(fmatrix,assignments,centroids,distance='euclidean',**kwargs): ''' rss = residual_sum_squares(fmatrix, assignments, centroids, distance='euclidean', **kwargs) Computes residual sum squares Parameters ---------- fmatrix : feature matrix assignments : Assignments array centroids : centroids Returns ------- rss : residual sum squares ''' if distance != 'euclidean': raise NotImplemented, "residual_sum_squares only implemented for 'euclidean' distance" rss = 0.0 for k, c in enumerate(centroids): diff = fmatrix[assignments == k] - c diff = diff.ravel() rss += np.dot(diff, diff) return rss def kmeans(fmatrix,K,distance='euclidean',max_iter=1000,R=None,**kwargs): ''' assignments, centroids = kmean(fmatrix, K, distance='euclidean', max_iter=1000, R=None, icov=None, covmat=None) K-Means Clustering Parameters ---------- distance: one of: - 'euclidean' : euclidean distance (default) - 'seuclidean' : standartised euclidean distance. This is equivalent to first normalising the features. - 'mahalanobis' : mahalanobis distance. This can make use of the following keyword arguments: + 'icov' (the inverse of the covariance matrix), + 'covmat' (the covariance matrix) If neither is passed, then the function computes the covariance from the feature matrix max_iter : Maximum number of iteration (default: 1000) Returns ------- assignments : An 1-D array of size `len(fmatrix)` centroids : An array of `k'` centroids ''' fmatrix = np.asanyarray(fmatrix) if distance == 'seuclidean': fmatrix = zscore(fmatrix) distance = 'euclidean' if distance == 'euclidean': def distfunction(fmatrix, cs): dists = np.dot(fmatrix, -2*cs.T) dists += np.array([np.dot(c,c) for c in cs]) # For a distance, we'd need to add the fmatrix**2 components, but # it doesn't matter because we are going to perform an argmin() on # the result. return dists elif distance == 'mahalanobis': icov = kwargs.get('icov', None) if icov is None: covmat = kwargs.get('covmat', None) if covmat is None: covmat = np.cov(fmatrix.T) icov = linalg.inv(covmat) def distfunction(fmatrix, cs): return np.array([_mahalanobis2(fmatrix, c, icov) for c in cs]).T else: raise ValueError('Distance argument unknown (%s)' % distance) R = get_pyrandom(R) centroids = np.array(R.sample(fmatrix,K), fmatrix.dtype) prev = np.zeros(len(fmatrix), np.int32) bins = np.arange(K+1) for i in xrange(max_iter): dists = distfunction(fmatrix, centroids) assignments = dists.argmin(1) if np.all(assignments == prev): break empty = [] counts,_ = np.histogram(assignments, bins) for ci,count in enumerate(counts): if count: where = (assignments == ci) mean = np.dot(fmatrix.T, where) mean /= count centroids[ci] = mean else: empty.append(ci) if empty: centroids = np.delete(centroids, empty, axis=0) K = len(centroids) bins = np.arange(K+1) prev[:] = assignments return assignments, centroids def repeated_kmeans(fmatrix,k,iterations,distance='euclidean',max_iter=1000,R=None,**kwargs): ''' assignments,centroids = repeated_kmeans(fmatrix, k, repeats, distance='euclidean',max_iter=1000,**kwargs) Runs kmeans repeats times and returns the best result as evaluated according to distance See Also -------- `kmeans` Parameters ---------- fmatrix : feature matrix k : nr of centroids iterations : Nr of repetitions distance : 'euclidean' (default) or 'seuclidean' max_iter : Max nr of iterations per kmeans run R : random source Returns ------- assignments : 1-D array of assignments centroids : centroids These are the same returns as `kmeans` ''' if distance == 'seuclidean': fmatrix = zscore(fmatrix) distance = 'euclidean' if distance != 'euclidean': raise NotImplementedError, "repeated_kmeans is only implemented for 'euclidean' or 'seuclidean' distance" best = np.inf R = get_pyrandom(R) for i in xrange(iterations): A,C = kmeans(fmatrix, k, distance, max_iter=max_iter, R=R,**kwargs) rss = residual_sum_squares(fmatrix,A,C,distance,**kwargs) if rss < best: Ab,Cb = A,C best = rss return Ab,Cb milk-0.3.4/milk/unsupervised/_som.cpp0000644000175000017500000000750211453677127020422 0ustar luispedroluispedro#include #include #include extern "C" { #include #include } namespace { struct SOM_Exception { SOM_Exception(const char* msg): msg(msg) { } const char* msg; }; void assert_type_contiguous(PyArrayObject* array,int type) { if (!PyArray_Check(array) || PyArray_TYPE(array) != type || !PyArray_ISCONTIGUOUS(array)) { throw SOM_Exception("Arguments to putpoints don't conform to expectation. Are you calling this directly? This is an internal function!"); } } void putpoints(PyArrayObject* grid, PyArrayObject* points, float L, int radius) { if (PyArray_NDIM(grid) != 3) throw SOM_Exception("grid should be three dimensional"); if (PyArray_NDIM(points) != 2) throw SOM_Exception("points should be two dimensional"); const int rows = PyArray_DIM(grid, 0); const int cols = PyArray_DIM(grid, 1); const int d = PyArray_DIM(grid, 2); const int n = PyArray_DIM(points, 0); if (PyArray_DIM(points, 1) != d) throw SOM_Exception("second dimension of points is not third dimension of grid"); Py_BEGIN_ALLOW_THREADS for (int i = 0; i != n; i++){ const float* p = static_cast(PyArray_GETPTR1(points, i)); int min_y = 0; int min_x = 0; float best = std::numeric_limits::max(); for (int y = 0; y != rows; ++y) { for (int x = 0; x != cols; ++x) { float dist = 0.; const float* gpoint = static_cast(PyArray_GETPTR2(grid, y, x)); for (int j = 0; j != d; ++j) { dist += (p[j] - gpoint[j])*(p[j] - gpoint[j]); } if (dist < best) { best = dist; min_y = y; min_x = x; } } } const int start_y = std::max(0, min_y - radius); const int start_x = std::max(0, min_x - radius); const int end_y = std::min(rows, min_y + radius); const int end_x = std::min(rows, min_x + radius); for (int y = start_y; y != end_y; ++y) { for (int x = start_x; x != end_x; ++x) { const float L2 = L /(1 + std::abs(min_y - y) + std::abs(min_x - x)); float* gpoint = static_cast(PyArray_GETPTR2(grid, y, x)); for (int j = 0; j != d; ++j) { gpoint[j] *= (1.-L2); gpoint[j] += L2 * p[j]; } } } } Py_END_ALLOW_THREADS } PyObject* py_putpoints(PyObject* self, PyObject* args) { try { PyArrayObject* grid; PyArrayObject* points; float L; int radius; if (!PyArg_ParseTuple(args, "OOfi", &grid, &points, &L, &radius)) { const char* errmsg = "Arguments were not what was expected for putpoints.\n" "This is an internal function: Do not call directly unless you know exactly what you're doing.\n"; PyErr_SetString(PyExc_RuntimeError,errmsg); return 0; } assert_type_contiguous(grid, NPY_FLOAT); assert_type_contiguous(points, NPY_FLOAT); putpoints(grid, points, L, radius); Py_RETURN_NONE; } catch (const SOM_Exception& exc) { PyErr_SetString(PyExc_RuntimeError,exc.msg); return 0; } catch (...) { PyErr_SetString(PyExc_RuntimeError,"Some sort of exception in putpoints."); return 0; } } PyMethodDef methods[] = { {"putpoints", py_putpoints, METH_VARARGS , "Do NOT call directly.\n" }, {NULL, NULL,0,NULL}, }; const char * module_doc = "Internal SOM Module.\n" "\n" "Do NOT use directly!\n"; } // namespace extern "C" void init_som() { import_array(); (void)Py_InitModule3("_som", methods, module_doc); } milk-0.3.4/milk/unsupervised/gaussianmixture.py0000644000175000017500000001325211452411266022547 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from numpy import log, pi, array from numpy.linalg import det, inv from kmeans import residual_sum_squares, centroid_errors __all__ = ['BIC','AIC','log_likelihood','nr_parameters'] def log_likelihood(fmatrix,assignments,centroids,model='one_variance',covs=None): ''' log_like = log_likelihood(feature_matrix, assignments, centroids, model='one_variance', covs=None) Compute the log likelihood of feature_matrix[i] being generated from centroid[i] ''' N,q = fmatrix.shape k = len(centroids) if model == 'one_variance': Rss = residual_sum_squares(fmatrix,assignments,centroids) #sigma2=Rss/N return -N/2.*log(2*pi*Rss/N)-N/2 elif model == 'diagonal_covariance': errors = centroid_errors(fmatrix,assignments,centroids) errors *= errors errors = errors.sum(1) Rss = np.zeros(k) counts = np.zeros(k) for i in xrange(fmatrix.shape[0]): c = assignments[i] Rss[c] += errors[i] counts[c] += 1 sigma2s = Rss/(counts+(counts==0)) return -N/2.*log(2*pi) -N/2. -1/2.*np.sum(counts*np.log(sigma2s+(counts==0))) elif model == 'full_covariance': res = -N*q/2.*log(2*pi) for k in xrange(len(centroids)): diff = (fmatrix[assignments == k] - centroids[k]) if covs is None: covm = np.cov(diff.T) else: covm = covs[k] if covm.shape == (): covm = np.matrix([[covm]]) icov = np.matrix(inv(covm)) diff = np.matrix(diff) Nk = diff.shape[0] res += -Nk/2.*log(det(covm)) + \ -.5 * (diff * icov * diff.T).diagonal().sum() return res raise ValueError, "log_likelihood: cannot handle model '%s'" % model def nr_parameters(fmatrix,k,model='one_variance'): ''' nr_p = nr_parameters(fmatrix, k, model='one_variance') Compute the number of parameters for a model of k clusters on Parameters ---------- fmatrix : feature matrix k : nr of clusters model : one of 'one_variance' (default), 'diagonal_covariance', or 'full_covariance' Returns ------- nr_p : Number of parameters ''' N,q = fmatrix.shape if model == 'one_variance': return k*q+1 elif model == 'diagonal_covariance': return k*(q+1) elif model == 'full_covariance': return k*+q*q raise ValueError, "milk.unsupervised.gaussianmixture.nr_parameters: cannot handle model '%s'" % model def _compute(type, fmatrix, assignments, centroids, model='one_variance', covs=None): N,q = fmatrix.shape k = len(centroids) log_like = log_likelihood(fmatrix, assignments, centroids, model, covs) n_param = nr_parameters(fmatrix,k,model) if type == 'BIC': return -2*log_like + n_param * log(N) elif type == 'AIC': return -2*log_like + 2 * n_param else: assert False def BIC(fmatrix, assignments, centroids, model='one_variance', covs=None): ''' B = BIC(fmatrix, assignments, centroids, model='one_variance', covs={From Data}) Compute Bayesian Information Criterion Parameters ---------- fmatrix : feature matrix assignments : Centroid assignments centroids : Centroids model : one of: 'one_variance': All features share the same variance parameter sigma^2. Default 'full_covariance': Estimate a full covariance matrix or use covs[i] for centroid[i] covs : Covariance matrices Returns ------- B = BIC value See Also -------- `AIC` ''' return _compute('BIC', fmatrix, assignments, centroids, model, covs) def AIC(fmatrix,assignments,centroids,model='one_variance',covs=None): ''' A = AIC(fmatrix,assignments,centroids,model) Compute Akaike Information Criterion Parameters ---------- fmatrix : feature matrix assignments : Centroid assignments centroids : Centroids model : one of: 'one_variance': All features share the same variance parameter sigma^2. Default 'full_covariance': Estimate a full covariance matrix or use covs[i] for centroid[i] covs : Covariance matrices Returns ------- A = AIC value See Also -------- `BIC` ''' return _compute('AIC', fmatrix, assignments, centroids, model, covs) milk-0.3.4/milk/unsupervised/som.py0000644000175000017500000000627511454604522020125 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # License: MIT. See COPYING.MIT file in the milk distribution from __future__ import division import numpy as np from ..utils import get_pyrandom from . import _som def putpoints(grid, points, L=.2, radius=4, iterations=1, shuffle=True, R=None): ''' putpoints(grid, points, L=.2, radius=4, iterations=1, shuffle=True, R=None) Feeds elements of `points` into the SOM `grid` Parameters ---------- grid : ndarray Self organising map points : ndarray data to feed to array L : float, optional How much to influence neighbouring points (default: .2) radius : integer, optional Maximum radius of influence (in L_1 distance, default: 4) iterations : integer, optional Number of iterations shuffle : boolean, optional Whether to shuffle the points before each iterations R : source of randomness ''' if radius is None: radius = 4 if type(L) != float: raise TypeError("milk.unsupervised.som: L should be floating point") if type(radius) != int: raise TypeError("milk.unsupervised.som: radius should be an integer") if grid.dtype != np.float32: raise TypeError('milk.unsupervised.som: only float32 arrays are accepted') if points.dtype != np.float32: raise TypeError('milk.unsupervised.som: only float32 arrays are accepted') if len(grid.shape) == 2: grid = grid.reshape(grid.shape+(1,)) if shuffle: random = get_pyrandom(R) for i in xrange(iterations): if shuffle: random.shuffle(points) _som.putpoints(grid, points, L, radius) def closest(grid, f): ''' y,x = closest(grid, f) Finds the coordinates of the closest point in the `grid` to `f` :: y,x = \\argmin_{y,x} { || grid[y,x] - f ||^2 } Parameters ---------- grid : ndarray of shape Y,X,J self-organised map f : ndarray of shape J point Returns ------- y,x : integers coordinates into `grid` ''' delta = grid - f delta **= 2 delta = delta.sum(2) return np.unravel_index(delta.argmin(), delta.shape) def som(data, shape, iterations=1000, L=.2, radius=4, R=None): ''' grid = som(data, shape, iterations=1000, L=.2, radius=4, R=None): Self-organising maps Parameters ---------- points : ndarray data to feed to array shape : tuple Desired shape of output. Must be 2-dimensional. L : float, optional How much to influence neighbouring points (default: .2) radius : integer, optional Maximum radius of influence (in L_1 distance, default: 4) iterations : integer, optional Number of iterations R : source of randomness Returns ------- grid : ndarray Map ''' R = get_pyrandom(R) d = data.shape[1] if data.dtype != np.float32: data = data.astype(np.float32) grid = np.array(R.sample(data, np.product(shape))).reshape(shape + (d,)) putpoints(grid, data, L=L, radius=radius, iterations=iterations, shuffle=True, R=R) return grid milk-0.3.4/milk/unsupervised/__init__.py0000644000175000017500000000303211452411266021051 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. ''' milk.unsupervised Unsupervised Learning --------------------- - kmeans: This is a highly optimised implementation of kmeans - PCA: Simple implementation - Non-negative matrix factorisation: both direct and with sparsity constraints ''' from kmeans import kmeans,repeated_kmeans from gaussianmixture import * from pca import pca from nnmf import * from pdist import pdist, plike from .som import som milk-0.3.4/milk/unsupervised/pdist.py0000644000175000017500000000302011452411266020432 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # License: MIT. See COPYING.MIT file in the milk distribution from __future__ import division import numpy as np def pdist(X, Y=None, distance='euclidean2'): ''' D = pdist(X, Y={X}, distance='euclidean2') Compute distance matrix:: D[i,j] == np.sum( (X[i] - Y[j])**2 ) Parameters ---------- X : feature matrix Y : feature matrix (default: use `X`) distance : one of 'euclidean' or 'euclidean2' (default) Returns ------- D : matrix of doubles ''' # Use Dij = np.dot(Xi, Xi) + np.dot(Xj,Xj) - 2.*np.dot(Xi,Xj) if Y is None: D = np.dot(X, X.T) x2 = D.diagonal() y2 = x2 else: D = np.dot(X, Y.T) x2 = np.array([np.dot(x,x) for x in X]) y2 = np.array([np.dot(y,y) for y in Y]) D *= -2. D += x2[:,np.newaxis] D += y2 # Because of numerical imprecision, we might get negative numbers # (which cause problems down the road, .e.g., when doing the sqrt): np.maximum(D, 0, D) if distance == 'euclidean': np.sqrt(D, D) return D def plike(X, sigma2=None): ''' L = plike(X, sigma2={guess based on X}) Parameters ---------- X : feature matrix sigma2 : bandwidth Returns ------- L : likelihood matrix ''' L = pdist(X) if sigma2 is None: sigma2 = np.median(L) L /= -sigma2 np.exp(L, L) return L milk-0.3.4/milk/unsupervised/nnmf/0000755000175000017500000000000011463416340017700 5ustar luispedroluispedromilk-0.3.4/milk/unsupervised/nnmf/lee_seung.py0000644000175000017500000000573111452411266022226 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from numpy import dot from ...utils import get_nprandom __all__ = ['nnmf'] def nnmf(V, r, cost='norm2', max_iter=int(1e4), tol=1e-8, R=None): ''' A,S = nnmf(X, r, cost='norm2', tol=1e-8, R=None) Implement Lee & Seung's algorithm Parameters ---------- V : 2-ndarray input matrix r : integer nr of latent features cost : one of: 'norm2' : minimise || X - AS ||_2 (default) 'i-div' : minimise D(X||AS), where D is I-divergence (generalisation of K-L divergence) max_iter : integer, optional maximum number of iterations (default: 10000) tol : double tolerance threshold for early exit (when the update factor is with tol of 1., the function exits) R : integer, optional random seed Returns ------- A : 2-ndarray S : 2-ndarray Reference --------- "Algorithms for Non-negative Matrix Factorization" by Daniel D Lee, Sebastian H Seung (available at http://citeseer.ist.psu.edu/lee01algorithms.html) ''' # Nomenclature in the function follows lee & seung, while outside nomenclature follows eps = 1e-8 n,m = V.shape R = get_nprandom(R) W = R.standard_normal((n,r))**2 H = R.standard_normal((r,m))**2 for i in xrange(max_iter): if cost == 'norm2': updateH = dot(W.T,V)/(dot(dot(W.T,W),H)+eps) H *= updateH updateW = dot(V,H.T)/(dot(W,dot(H,H.T))+eps) W *= updateW elif cost == 'i-div': raise NotImplementedError,'I-Div not implemented in lee_seung.nnmf' if True or (i % 10) == 0: max_update = max(updateW.max(),updateH.max()) if abs(1.-max_update) < tol: break return W,H milk-0.3.4/milk/unsupervised/nnmf/__init__.py0000644000175000017500000000246411452411266022017 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. import lee_seung as ls from hoyer import sparse_nnmf import hoyer lee_seung = ls.nnmf __all__ = ['lee_seung','sparse_nnmf'] # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/milk/unsupervised/nnmf/hoyer.py0000644000175000017500000001023111452411266021375 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from ...utils import get_nprandom __all__ = ['hoyer_sparse_nnmf'] def sp(s): L2 = np.sqrt(np.dot(s,s)) L1 = np.abs(s).sum() sn = np.sqrt(len(s)) return (sn-L1/L2)/(sn-1) def _solve_alpha(s,m,L2): sm = s-m s2 = np.dot(s,s) sm2 = np.dot(sm, sm) m2 = np.dot(m, m) dot = np.dot(m, sm) alpha = (-dot + np.sqrt(dot**2 - sm2*(m2-L2**2)))/sm2 return alpha def _project(x,L1,L2): ''' Implement projection onto sparse space ''' x = np.asanyarray(x) n = len(x) s = x + (L1 - x.sum())/n Z = np.zeros(n,bool) while True: m = (~Z) * L1/(n-Z.sum()) alpha = _solve_alpha(s,m,L2) s = m + alpha * (s - m) negs = (s < 0) if not negs.any(): return s Z |= negs s[Z] = 0 c = (s.sum() - L1)/(~Z).sum() s -= c*(~Z) def _L1for(s,x,L2): ''' Solve for L1 in s = [ sqrt(n) - L1/L2] / [sqrt(n) - 1] ''' sn = np.sqrt(len(x)) return L2*(s+sn-s*sn) def sparse_nnmf(V, r, sparsenessW=None, sparsenessH=None, max_iter=10000, R=None): ''' W,H = hoyer.sparse_nnmf(V, r, sparsenessW = None, sparsenessH = None, max_iter=10000, R=None) Implement sparse nonnegative matrix factorisation. Parameters ---------- V : 2-D matrix input feature matrix r : integer number of latent features sparsenessW : double, optional sparseness contraint on W (default: no sparsity contraint) sparsenessH : double, optional sparseness contraint on H (default: no sparsity contraint) max_iter : integer, optional maximum nr of iterations (default: 10000) R : integer, optional source of randomness Returns ------- W : 2-ndarray H : 2-ndarray Reference --------- "Non-negative Matrix Factorisation with Sparseness Constraints" by Patrik Hoyer in Journal of Machine Learning Research 5 (2004) 1457--1469 ''' n,m = V.shape R = get_nprandom(R) mu_W = .15 mu_H = .15 eps = 1e-8 W = R.standard_normal((n,r))**2 H = R.standard_normal((r,m))**2 def fix(X, sparseness): for i in xrange(r): row = X[i] L2 = np.sqrt(np.dot(row, row)) row /= L2 X[i] = _project(row, _L1for(sparseness, row, 1.), 1.) def fixW(): fix(W.T, sparsenessW) def fixH(): fix(H, sparsenessH) if sparsenessW is not None: fixW() if sparsenessH is not None: fixH() for i in xrange(max_iter): if sparsenessW is not None: W -= mu_W * np.dot(np.dot(W,H)-V,H.T) fixW() else: updateW = np.dot(V,H.T)/(np.dot(W,np.dot(H,H.T))+eps) W *= updateW if sparsenessH is not None: H -= mu_H * np.dot(W.T,np.dot(W,H)-V) fixH() else: updateH = np.dot(W.T,V)/(np.dot(np.dot(W.T,W),H)+eps) H *= updateH return W,H milk-0.3.4/milk/unsupervised/pca.py0000644000175000017500000000335211452411266020062 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np from numpy import linalg from . import normalise def pca(X, zscore=True): ''' Y,V = pca(X, zscore=True) Principal Component Analysis Performs principal component analysis. Returns transformed matrix and principal components Parameters ---------- * X: data matrix * zscore: whether to normalise to zscores (default: True) ''' if zscore: X = normalise.zscore(X) C = np.cov(X.T) w,v = linalg.eig(C) Y = np.dot(v,X.T).T return Y,v # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/milk/unsupervised/affinity.py0000644000175000017500000000763311452411266021136 0ustar luispedroluispedro"""Algorithms for clustering : Meanshift and Affinity propagation Original Authors (for scikits.learn): Alexandre Gramfort alexandre.gramfort@inria.fr Gael Varoquaux gael.varoquaux@normalesup.org """ import numpy as np def affinity_propagation(S, p=None, convit=30, maxit=200, damping=0.5, copy=True, R=0): """Perform Affinity Propagation Clustering of data Parameters ---------- S: array [n_points, n_points] Matrix of similarities between points p: array [n_points,] or float, optional Preferences for each point damping : float, optional Damping factor copy: boolean, optional If copy is False, the affinity matrix is modified inplace by the algorithm, for memory efficiency R : source of randomness Returns ------- cluster_centers_indices: array [n_clusters] index of clusters centers labels : array [n_points] cluster labels for each point Notes ----- See examples/plot_affinity_propagation.py for an example. Reference: Brendan J. Frey and Delbert Dueck, "Clustering by Passing Messages Between Data Points", Science Feb. 2007 """ if copy: # Copy the affinity matrix to avoid modifying it inplace S = S.copy() n_points = S.shape[0] assert S.shape[0] == S.shape[1] if p is None: p = np.median(S) if damping < 0.5 or damping >= 1: raise ValueError('damping must be >= 0.5 and < 1') random_state = np.random.RandomState(R) # Place preferences on the diagonal of S S.flat[::(n_points+1)] = p A = np.zeros((n_points, n_points)) R = np.zeros((n_points, n_points)) # Initialize messages # Remove degeneracies noise = random_state.randn(n_points, n_points) typeinfo = np.finfo(S.dtype) noise *= typeinfo.tiny*100 S += noise del noise # Execute parallel affinity propagation updates e = np.zeros((n_points, convit)) ind = np.arange(n_points) for it in range(maxit): Aold = A.copy() Rold = R.copy() A += S I = np.argmax(A, axis=1) Y = A[ind, I]#np.max(A, axis=1) A[ind, I] = typeinfo.min Y2 = np.max(A, axis=1) R = S - Y[:, np.newaxis] R[ind, I[ind]] = S[ind, I] - Y2 Rold *= damping R *= (1-damping) R += Rold # Compute availabilities Rd = R.diagonal().copy() np.maximum(R, 0, R) R.flat[::n_points+1] = Rd A = np.sum(R, axis=0)[np.newaxis, :] - R dA = np.diag(A) A = np.minimum(A, 0) A.flat[::n_points+1] = dA Aold *= damping A *= (1-damping) A += Aold # Check for convergence E = (np.diag(A) + np.diag(R)) > 0 e[:, it % convit] = E K = np.sum(E, axis=0) if it >= convit: se = np.sum(e, axis=1); unconverged = np.sum((se == convit) + (se == 0)) != n_points if (not unconverged and (K>0)) or (it==maxit): print "Converged after %d iterations." % it break else: print "Did not converge" I = np.where(np.diag(A+R) > 0)[0] K = I.size # Identify exemplars if K > 0: c = np.argmax(S[:, I], axis=1) c[I] = np.arange(K) # Identify clusters # Refine the final set of exemplars and clusters and return results for k in range(K): ii = np.where(c==k)[0] j = np.argmax(np.sum(S[ii, ii], axis=0)) I[k] = ii[j] c = np.argmax(S[:, I], axis=1) c[I] = np.arange(K) labels = I[c] # Reduce labels to a sorted, gapless, list cluster_centers_indices = np.unique(labels) labels = np.searchsorted(cluster_centers_indices, labels) else: labels = np.empty((n_points, 1)) cluster_centers_indices = None labels.fill(np.nan) return cluster_centers_indices, labels milk-0.3.4/milk/unsupervised/normalise.py0000644000175000017500000000300211452411266021300 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy as np __all__ = ['zscore'] def zscore(features): """ features = zscore(features) Returns a copy of features which has been normalised to zscores """ mu = features.mean(0) sigma = np.std(features,0) sigma[sigma == 0] = 1 return (features - mu) / sigma # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/milk/wrapper/0000755000175000017500000000000011463416340015666 5ustar luispedroluispedromilk-0.3.4/milk/wrapper/__init__.py0000644000175000017500000000000011452411266017765 0ustar luispedroluispedromilk-0.3.4/milk/wrapper/wraplibsvm.py0000644000175000017500000000456411452411266020437 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # License: MIT. See COPYING.MIT file in the milk distribution from __future__ import division from milk.supervised.classifier import normaliselabels try: from libsvm import svm as libsvm except ImportError: try: import svm as libsvm except ImportError: libsvm = None from tempfile import NamedTemporaryFile class libsvmModel(object): def __init__(self, model, names, output_probability): self.model = model self.names = names self.output_probability = output_probability def apply(self,feats): if self.output_probability: return self.model.predict_probability(feats) res = self.model.predict(feats) return self.names[int(res)] def __getstate__(self): # This is really really really hacky, but it works N = NamedTemporaryFile() self.model.save(N.name) S = N.read() return S,self.output_probability,self.names def __setstate__(self,state): if libsvm is None: raise RuntimeError('LibSVM Library not found. Cannot use this classifier.') S,self.output_probability,self.names = state N = NamedTemporaryFile() N.write(S) N.flush() self.model = libsvm.svm_model(N.name) class libsvmClassifier(object): def __init__(self,probability = False, auto_weighting = True): if libsvm is None: raise RuntimeError('LibSVM Library not found. Cannot use this classifier.') self.param = libsvm.svm_parameter(kernel_type = libsvm.RBF, probability = probability) self.output_probability = probability self.auto_weighting = auto_weighting def set_option(self,optname,value): setattr(self.param, optname, value) def train(self, features, labels): labels,names = normaliselabels(labels) if self.auto_weighting: nlabels = labels.max() + 1 self.param.nr_weight = int(nlabels) self.param.weight_label = range(nlabels) self.param.weight = [(labels != i).mean() for i in xrange(nlabels)] problem = libsvm.svm_problem(labels.astype(float), features) model = libsvm.svm_model(problem, self.param) return libsvmModel(model, names, self.output_probability) milk-0.3.4/milk/active/0000755000175000017500000000000011463416340015461 5ustar luispedroluispedromilk-0.3.4/milk/active/__init__.py0000644000175000017500000000000011231643005017551 0ustar luispedroluispedromilk-0.3.4/milk/active/uncertainty.py0000644000175000017500000000420711452411266020403 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # vim: set ts=4 sts=4 sw=4 expandtab smartindent: # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. ''' Uncertainty ============ Implements uncertainty-based active learning strategies. These are strategies that are based on querying those elements in the pool which we are most uncertain about Functions ---------- * entropy * one_minus_max ''' from __future__ import division import numpy def entropy(model, pool): ''' entropies = entropy(model, pool) Returns the entropy of each classification output for members in the pool. ''' def _entropy(labels, ps): H = 0. for p in ps: if p > 1e-9: H += p * np.log(p) return H return [_entropy(model.apply(u)) for u in pool] def one_minus_max(model,pool): ''' oneminus = one_minus_max(model,pool) oneminus[i] = 1 - max_L { P(pool_i == L) } Returns one minus the probability for the best label guess. ''' def _minus1(labels, ps): return 1. - np.max(ps) return [_minus1(model.apply(u)) for u in pool] milk-0.3.4/milk/active/eimpact.py0000644000175000017500000000457511452411266017470 0ustar luispedroluispedro# -*- coding: utf-8 -*- # Copyright (C) 2008-2010, Luis Pedro Coelho # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN # THE SOFTWARE. from __future__ import division import numpy from ..supervised import svm from ..supervised.classifier import ctransforms def expected_impacts(D,labels,U): ''' EIs = expected_impacts(D,labels,U) Compute Expected impact for each element of U Eis[i]: P(label[i] == 1) * IMPACT(label[i] == 1) + P(label[i] == 0) * IMPACT(label[i] == 0) ''' assert len(D) == len(labels), 'Nr of labeled examples should match lenght of labels vector' K = svm.rbf_kernel(20000) prob_classifier = ctransforms(svm.svm_raw(kernel=K,C=4),svm.svm_sigmoidal_correction()) label_classifier = ctransforms(svm.svm_raw(kernel=K,C=4),svm.svm_binary()) prob_classifier.train(D,labels) u_probs = prob_classifier(U) u_labels = (u_probs > .5) impacts = [] for u,p in zip(U,u_probs): print len(impacts) label_classifier.train(numpy.vstack((D,u)),numpy.hstack((labels,[0]))) u_labels_0 = label_classifier(U) label_classifier.train(numpy.vstack((D,u)),numpy.hstack((labels,[1]))) u_labels_1 = label_classifier(U) e_impact = (1.-p)*(u_labels != u_labels_0).sum() + p*(u_labels != u_labels_1).sum() impacts.append(e_impact) return impacts # vim: set ts=4 sts=4 sw=4 expandtab smartindent: milk-0.3.4/setup.cfg0000644000175000017500000000007311463416340015073 0ustar luispedroluispedro[egg_info] tag_build = tag_date = 0 tag_svn_revision = 0