In [15]:
# code to regenerate input files for results for paper
#
# http://www.fabiangieseke.de/pdfs/neucom2013_draft.pdf
#  see Table 2 results real-sim data set
#
# see also: 
#  http://www.fabiangieseke.de/pdfs/icpram2012.pdf
# 
import sys
from time import time
from pprint import pprint

import numpy as np
import scipy
import scipy.sparse as sp
import joblib

import io
import os.path

import sklearn
import sklearn.svm
import sklearn.datasets
import sklearn.metrics
import sklearn.cross_validation


from sklearn.externals.six import u, b

import warnings
warnings.filterwarnings('ignore')

%pylab inline


Populating the interactive namespace from numpy and matplotlib

In [16]:
X, y = sklearn.datasets.load_svmlight_file('real-sim')

In [17]:
X.shape


Out[17]:
(72309, 20958)

In [18]:
from collections import Counter
Counter(y)


Out[18]:
Counter({-1.0: 50071, 1.0: 22238})

In [19]:
splits = sklearn.cross_validation.StratifiedShuffleSplit(y, n_iter=1, test_size=0.50)
train_indices, test_indices = splits.__iter__().next()

In [20]:
instance_ids = np.arange(y.size)
X_train = X[train_indices]
train_ids = instance_ids[train_indices]

X_test = X[test_indices]
test_ids = instance_ids[test_indices]

train_labels = y[train_indices]
test_labels = y[test_indices]

In [21]:
svm = sklearn.svm.LinearSVC(penalty='l2', C=1, dual=False)
svm.fit(X_train, train_labels)


Out[21]:
LinearSVC(C=1, class_weight=None, dual=False, fit_intercept=True,
     intercept_scaling=1, loss='l2', multi_class='ovr', penalty='l2',
     random_state=None, tol=0.0001, verbose=0)

Ground Truth
Best possible HO accuracy


In [22]:
accuracy = sklearn.metrics.accuracy_score(test_labels, svm.predict(X_test))
print 100.0*(accuracy)
#ncv = 10
#print sklearn.cross_validation.cross_val_score(svm, X_train, train_labels, cv=10).sum()/ncv
#print sklearn.cross_validation.cross_val_score(svm, X_test, test_labels, cv=10).sum()/ncv


97.1262619278

In [23]:
print X_train.shape
print X_test.shape


(36154, 20958)
(36155, 20958)

In [24]:
def print_svmlight_infiles(L, y_l, U, y_u, HO, y_ho):

    A = scipy.sparse.vstack((L,U))
    unk_l = y_u*0   
    a_l = np.hstack((y_l, unk_l))
    print A.shape, a_l.shape 
    numL = L.shape[0] 
    
    training_file = 'svmlight.train.%d'%numL
    sklearn.datasets.dump_svmlight_file(A,a_l,training_file,zero_based=False)
    
    testL_file = 'svmlight.testL.%d'%numL
    sklearn.datasets.dump_svmlight_file(L,y_l,testL_file,zero_based=False)
    
    testU_file = 'svmlight.testU.%d'%numL
    sklearn.datasets.dump_svmlight_file(U,y_u,testU_file,zero_based=False)
    
    testHO_file = 'svmlight.testHO.%d'%numL
    sklearn.datasets.dump_svmlight_file(HO,y_ho,testHO_file,zero_based=False)

In [29]:
def dump_svmlin(X, y, fX, fy):
    X_value_pattern = u("%d:%.16g")
    
    is_sp = int(hasattr(X, "tocsr"))
    
    fy.write(''.join(["%d\n" % label for label in y]))
    

    for i in range(X.shape[0]):
        if is_sp:
            span = slice(X.indptr[i], X.indptr[i + 1])
            row = zip(X.indices[span], X.data[span])
        else:
            nz = X[i] != 0
            row = zip(np.where(nz)[0], X[i, nz])

        s = " ".join(X_value_pattern % (j + 1, x) for j, x in row)
        fX.write(("%s\n" % s).encode('ascii'))

In [26]:
def print_svmlin_infiles(L, y_l, U, y_u, HO, y_ho):
    print "print_svmlin_infiles..."
    A = scipy.sparse.vstack((L,U))
    unk_l = y_u*0
    a_l = np.hstack((y_l, unk_l))
    print A.shape, a_l.shape 
    
    numL = L.shape[0]
    
    examples_file = open('svmlin.train.examples.%d'%numL, "wb")
    labels_file = open('svmlin.train.labels.%d'%numL, "wb")
    dump_svmlin(A, a_l, examples_file, labels_file)
    
    examples_file = open('svmlin.testL.examples.%d'%numL, "wb")
    labels_file = open('svmlin.testL.labels.%d'%numL, "wb")
    dump_svmlin(L, y_l, examples_file, labels_file)
    
    examples_file = open('svmlin.testU.examples.%d'%numL, "wb")
    labels_file = open('svmlin.testU.labels.%d'%numL, "wb")
    dump_svmlin(U, y_u, examples_file, labels_file)
    
    examples_file = open('svmlin.testHO.examples.%d'%numL, "wb")
    labels_file = open('svmlin.testHO.labels.%d'%numL, "wb")
    dump_svmlin(HO, y_ho, examples_file, labels_file)

In [27]:
def print_universvm_infiles(L, y_l, U, y_u, HO, y_ho):
    A = scipy.sparse.vstack((L,U))
    unk_l = y_u*0 - 3
    a_l = np.hstack((y_l, unk_l))
        
    print A.shape, a_l.shape 
    numL = L.shape[0]
    
    training_file = 'universvm.train.%d'%numL
    sklearn.datasets.dump_svmlight_file(A,a_l,training_file,zero_based=False)
    
    testL_file = 'universvm.testL.%d'%numL
    sklearn.datasets.dump_svmlight_file(L,y_l,testL_file,zero_based=False)
    
    testU_file = 'universvm.testU.%d'%numL
    sklearn.datasets.dump_svmlight_file(U,y_u,testU_file,zero_based=False)
    
    testHO_file = 'universvm.testHO.%d'%numL
    sklearn.datasets.dump_svmlight_file(HO,y_ho,testHO_file,zero_based=False)

In [30]:
#split_sizes = [0.0025, 0.005, 0.01, 0.04, 0.08]
split_sizes = [0.08]

for U_size in split_sizes:
    splits = sklearn.cross_validation.StratifiedShuffleSplit(train_labels, n_iter=1, test_size=1-U_size)
    labeled_indices, unlabeled_indices = splits.__iter__().next()
    
    L = X_train[labeled_indices]
    L_ids = instance_ids[labeled_indices]

    U = X_train[unlabeled_indices]
    U_ids = instance_ids[unlabeled_indices]

    y_l = train_labels[labeled_indices]
    y_u = train_labels[unlabeled_indices]

    HO = X_test
    y_ho = test_labels
    print "making..."
   
    #print_universvm_infiles(L, y_l, U, y_u, X_test, test_labels)
    print_svmlin_infiles(L, y_l, U, y_u, X_test, test_labels)
    print_svmlight_infiles(L, y_l, U, y_u, HO, y_ho)
    
    svm_small = sklearn.svm.LinearSVC(C=10,fit_intercept=False)
    svm_small.fit(L, y_l)
    y_p = svm_small.predict(HO)
    score = svm_small.score(HO,y_ho)
#    print np.mean((y_ho+1)/2),np.mean((y_p+1)/2)
    yy = y_p*y_ho
    acc =  float(np.where(yy>0)[0].shape[0])/y_ho.shape[0]
   # print acc
    
    print  L.shape, y_l.shape, X_test.shape, test_labels.shape

    print "baseline HO accuracy for ",U_size," l=",L.shape[0] ,"  ",(score)*100.0
    print "\n"


making...
print_svmlin_infiles...
(36154, 20958) (36154,)
(36154, 20958) (36154,)
(2892, 20958) (2892,) (36155, 20958) (36155,)
baseline HO accuracy for  0.08  l= 2892    94.5152814272



In [ ]: