Kitware Tech-Tip 04-03-2013 Part 3: Computing a BoW representation from a feature matrix and training / testing a SVM classifier in a N-fold cross-validation setup.

kw-techtip-04-03-2013-p3.py

"""kw-techtip-04-03-2013-p3.py

Compute BoW representations and train / test a SVM classifier in a N-fold 
cross-validation setup. 

Usage
-----

$ python kw-techtip-04-03-2013-p3.py <DataFile> <IndexFile> <ClassInfoFile>
"""

import sys
import numpy as np

# scikit-learn imports 
from sklearn.cluster import KMeans
from sklearn.metrics import accuracy_score
from sklearn.cross_validation import KFold
from sklearn.cross_validation import ShuffleSplit
from sklearn.grid_search import GridSearchCV
from sklearn import preprocessing
from sklearn import svm


dat = np.genfromtxt(sys.argv[1]) # Data
idx = np.genfromtxt(sys.argv[2]) # Index
lab = np.genfromtxt(sys.argv[3]) # Class labels

# Create cross-validation folds
cv = ShuffleSplit(len(lab), n_iter=5, test_size=0.2) 

for n, (trn, tst) in enumerate(cv):
    p = []
    for i in trn:
        p.extend(list(np.where(idx==i)[0]))
    pos = np.asarray(p)
    
    # Learn a codebook (i.e., vector-quantization of features)
    print "[Fold: %d (%d/%d)] Learning the codebook ..." % (n, len(trn),len(tst))
    cb = KMeans(32, init="k-means++", n_init=10, max_iter=500)
    cb.fit(dat[pos,:])

    X = np.zeros((len(trn), 32))
    L = np.zeros((len(trn),))
    for cnt, i in enumerate(trn):
        pos = np.where(idx==i)[0]
        X[cnt,:] = np.asarray(np.histogram(cb.predict(dat[pos,:]),
                              bins=range(0,33),
                              density=True)[0])

    # Cross-validate SVM classifier parameters (gamma, C) on the 
    # training portion of the data in a 5-fold cross-validation 
    # setup.
    parameters = [{'kernel': ['rbf'],   
                   'gamma': np.logspace(-6,2,10), 
                   'C': [1, 10, 100, 1000]}]
    clf = GridSearchCV(svm.SVC(C=1), parameters)
    clf.fit(X, np.asarray(lab)[trn], cv=5)
    
    # And now, test ...
    Y = np.zeros((len(tst), 32))
    for cnt, i in enumerate(tst):
        pos =  np.where(idx==i)[0]
        Y[cnt,:] = np.asarray(np.histogram(cb.predict(dat[pos,:]),
                              bins=range(0,33), density=True)[0])
    score = clf.score(Y,np.asarray(lab)[tst])
    print score*100