In [1]:
import numpy as np
import matplotlib.pyplot as plt
from sklearn.datasets.samples_generator import make_blobs

%matplotlib inline

Order of models:
I. a OneVsAllClassifier for the three class clusters below
-> Ia. normalizing the boundary vectors
-> Ib. tuning parameter C
II. a multiclass SVM for the same data
III. a multiclass SVM for NYC urban sound files (not trained yet)


In [2]:
# Create the  training data
np.random.seed(2)
X, y = make_blobs(n_samples=300,cluster_std=.25, centers=np.array([(-3,1),(0,2),(3,1)]))
plt.scatter(X[:, 0], X[:, 1], c=y, s=50)


Out[2]:
<matplotlib.collections.PathCollection at 0x67d36a8ac8>

In [3]:
from sklearn.base import BaseEstimator, ClassifierMixin, clone
from numpy import linalg as L
   
class OneVsAllClassifier(BaseEstimator, ClassifierMixin):  
    """
    One-vs-all classifier
    We assume that the classes will be the integers 0,..,(n_classes-1).
    We assume that the estimator provided to the class, after fitting, has a "decision_function" that 
    returns the score for the positive class.
    """
    def __init__(self, estimator, n_classes):      
        """
        Constructed with the number of classes and an estimator (e.g. an
        SVM estimator from sklearn)
        @param estimator : binary base classifier used
        @param n_classes : number of classes
        """
        self.n_classes = n_classes 
        self.estimators = [clone(estimator) for _ in range(n_classes)]
        self.fitted = False
      
    def fit(self, X, y=None):
        """
        This should fit one classifier for each class.
        self.estimators[i] should be fit on class i vs rest
        @param X: array-like, shape = [n_samples,n_features], input data
        @param y: array-like, shape = [n_samples,] class labels
        @return returns self
        """
        for i in range(self.n_classes):
            self.estimators[i].fit(X, y==i)
        self.fitted = True  
        
        return self   

    def decision_function(self, X):
        """
        Returns the score of each input for each class. Assumes
        that the given estimator also implements the decision_function method (which sklearn SVMs do), 
        and that fit has been called.
        @param X : array-like, shape = [n_samples, n_features] input data
        @return array-like, shape = [n_samples, n_classes]
        """
        if not self.fitted:
            raise RuntimeError("You must train classifer before predicting data.")

        if not hasattr(self.estimators[0], "decision_function"):
            raise AttributeError(
                "Base estimator doesn't have a decision_function attribute.")
        
        classes_score = [ est.decision_function(X) for est in self.estimators ]
        score_ = classes_score[0]
        
        for thing in classes_score[1:]:
            score_ = np.column_stack( (score_, thing ) )
            
        return score_
    
    def predict(self, X):
        """
        Predict the class with the highest score.
        @param X: array-like, shape = [n_samples,n_features] input data
        @returns array-like, shape = [n_samples,] the predicted classes for each input
        """
        predictions = np.zeros(X.shape[0])
        decisions = self.decision_function(X)
        for i in range(predictions.shape[0]):
            predictions[i] = np.argmax( decisions[i] )
        
        return predictions

In [4]:
#Here we test the OneVsAllClassifier
from sklearn import svm
svm_estimator = svm.LinearSVC(loss='hinge', fit_intercept=False, C=5)
clf_onevsall = OneVsAllClassifier(svm_estimator, n_classes=3)
clf_onevsall.fit(X,y)

for i in range(3) :
    print("Coeffs %d"%i)
    print(clf_onevsall.estimators[i].coef_) #Will fail if you haven't implemented fit yet
    print( clf_onevsall.estimators[i].coef_/L.norm(clf_onevsall.estimators[i].coef_))
    
# create a mesh to plot in
h = .02  # step size in the mesh
x_min, x_max = min(X[:,0])-3,max(X[:,0])+3
y_min, y_max = min(X[:,1])-3,max(X[:,1])+3
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
                     np.arange(y_min, y_max, h))
mesh_input = np.c_[xx.ravel(), yy.ravel()]

Z = clf_onevsall.predict(mesh_input)
Z = Z.reshape(xx.shape)
plt.contourf(xx, yy, Z, cmap=plt.cm.coolwarm, alpha=0.8)
# Plot also the training points
plt.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.coolwarm)

from sklearn import metrics
metrics.confusion_matrix(y, clf_onevsall.predict(X))


Coeffs 0
[[-1.05854193 -0.90296009]]
[[-0.7608034  -0.64898242]]
Coeffs 1
[[ 0.34572347  0.35643156]]
[[ 0.69624222  0.71780692]]
Coeffs 2
[[ 0.89050156 -0.82570517]]
[[ 0.73328156 -0.67992511]]
Out[4]:
array([[100,   0,   0],
       [  0, 100,   0],
       [  0,   9,  91]])

Notice the margin vs classification accuracy trade-off by tuning parameter C


In [5]:
# create a mesh to plot in
h = .02  # step size in the mesh
x_min, x_max = min(X[:,0])-3,max(X[:,0])+3
y_min, y_max = min(X[:,1])-3,max(X[:,1])+3
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
                     np.arange(y_min, y_max, h))
mesh_input = np.c_[xx.ravel(), yy.ravel()]

def f(parameter):
    svm_estimator = svm.LinearSVC(loss='hinge', fit_intercept=False, C=parameter)
    clf_onevsall = OneVsAllClassifier(svm_estimator, n_classes=3)
    clf_onevsall.fit(X,y)

    Z = clf_onevsall.predict(mesh_input)
    Z = Z.reshape(xx.shape)
    plt.contourf(xx, yy, Z, cmap=plt.cm.coolwarm, alpha=0.2)
    # Plot also the training points
    plt.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.coolwarm)

for c in [1, 50, 100, 200, 500, 1000]:
    f(c)


By normalizing the three boundary vectors, discrepancies between intra-boundary angles have lowered.


In [6]:
soa = np.array([[0, 0, -0.76080386, -0.64898189], [0, 0,0.68477689,  0.72875278], [0, 0,  0.73358915, -0.67959323]])
A, B, C, D = zip(*(-1*soa) )
plt.figure()
ax = plt.gca()
ax.quiver(A, B, C, D, scale=1)
ax.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.coolwarm)

ax.set_xlim([-6, 6])
ax.set_ylim([-2, 4])
plt.draw()
plt.show()


Multiclass SVM


In [17]:
def zeroOne(y,a) :
    '''
    Computes the zero-one loss.
    @param y: output class
    @param a: predicted class
    @return 1 if different, 0 if same
    '''
    return int(y != a)

def featureMap(X,y,num_classes) :
    '''
    Computes the class-sensitive features.
    @param X: array-like, shape = [n_samples,n_inFeatures] or [n_inFeatures,], input features for input data
    @param y: a target class (in range 0,..,num_classes-1)
    @return array-like, shape = [n_samples,n_outFeatures], the class sensitive features for class y
    '''
    #The following line handles X being a 1d-array or a 2d-array
    num_samples, num_inFeatures = (1,X.shape[0]) if len(X.shape) == 1 else (X.shape[0],X.shape[1])
    num_outFeatures = num_inFeatures*num_classes
    output = np.zeros((num_samples, num_outFeatures))
    
    if X.ndim == 1:
        output[0, y*num_inFeatures: y*num_inFeatures+2] = X
    
    else:
        for i in range(num_samples):
            output[i, y*num_inFeatures: y*num_inFeatures +2  ] = X[i]  
    
    return output

def sgd(X, y, num_outFeatures, subgd, lam = 0.01, eta = 0.1, T = 10000):
    '''
    Runs subgradient descent, and outputs resulting parameter vector.
    @param X: array-like, shape = [n_samples,n_features], input training data 
    @param y: array-like, shape = [n_samples,], class labels
    @param num_outFeatures: number of class-sensitive features
    @param subgd: function taking x,y and giving subgradient of objective
    @param weight w
    @param eta: learning rate for SGD
    @param T: maximum number of iterations
    @return: vector of weights
    '''
    num_samples = X.shape[0]
    w_store = np.zeros(num_outFeatures)
    w = np.zeros(num_outFeatures)
    #w = np.array([0,0.1,0,0.1,0,0.1])
    for t in range(T):
        #eta /= (t+1)
        #sample_ind = np.random.choice(num_samples, 1)
        for sample_ind in range(num_samples):
            w -= eta*subgd(X[sample_ind], y[sample_ind], w)
            #if sample_ind == t:
             #   print(subgd(X[sample_ind], y[sample_ind], w))
            w -= eta*lam*w
    return w#_store/T

class MulticlassSVM(BaseEstimator, ClassifierMixin):
    '''
    Implements a Multiclass SVM estimator.
    '''
    def __init__(self, num_outFeatures, lam=0.01, num_classes=3, Delta=zeroOne, Psi=featureMap):       
        '''
        Creates a MulticlassSVM estimator.
        @param num_outFeatures: number of class-sensitive features produced by Psi
        @param lam: l2 regularization parameter
        @param num_classes: number of classes (assumed numbered 0,..,num_classes-1)
        @param Delta: class-sensitive loss function taking two arguments (i.e., target margin)
        @param Psi: class-sensitive feature map taking two arguments
        '''
        self.num_outFeatures = num_outFeatures
        self.lam = lam
        self.num_classes = num_classes
        self.Delta = Delta
        self.Psi = lambda X,y : Psi(X,y,num_classes)
        self.fitted = False
    
    def subgradient(self,x,y,w):
        '''
        Computes the subgradient at a given data point x,y
        @param x: sample input
        @param y: sample class
        @param w: parameter vector
        @return returns subgradient vector at given x,y,w
        '''
        scores_ = []
        for i in [0,1,2]:
            scores_.append( zeroOne(i,y) + np.dot(w, (self.Psi(x,i) - self.Psi(x,y) ).reshape(6) ) )
        y_hat = (np.argmax( scores_ )    )
        
        #print(x, x.shape, (self.Psi(x,y_hat) - self.Psi(x,y)).reshape(6))
        return (self.lam*w + self.Psi(x, y_hat) - self.Psi(x, y)).reshape(6) 
        
    def fit(self,X,y,eta=0.1,T=10000):
        '''
        Fits multiclass SVM
        @param X: array-like, shape = [num_samples,num_inFeatures], input data
        @param y: array-like, shape = [num_samples,], input classes
        @param eta: learning rate for SGD
        @param T: maximum number of iterations
        @return returns self
        '''
        self.coef_ = sgd(
            X,y, self.num_outFeatures,self.subgradient,self.lam,eta,T) #when to use num_outFeatures?
        self.fitted = True
        return self
    
    def decision_function(self, data):
        '''
        Returns the score on each input for each class. Assumes
        that fit has been called.
        @param X : array-like, shape = [n_samples, n_inFeatures]
        @return array-like, shape = [n_samples, n_classes]
        giving scores for each sample,class pairing
        '''
        if not self.fitted:
            raise RuntimeError("You must train classifer before predicting data.")
        
        samples = data.shape[0]
        scores = np.zeros((samples, 3))
        
        for s in range(samples):
            for c in [0,1,2]:
                scores[s,c] = np.dot(self.coef_, self.Psi(data[s], c).T)
            
        return scores
            
    def predict(self, X):
        '''
        Predict the class with the highest score.
        @param X: array-like, shape = [n_samples, n_inFeatures], input data to predict
        @return array-like, shape = [n_samples,], class labels predicted for each data point
        '''
        labels = np.zeros(X.shape[0])
        table = self.decision_function(X)
        #print(self.decision_function(X))
        for i in range(X.shape[0]):
            labels[i] = np.argmax( table[i,:] )
        return labels

In the plot below we see this more complex model has also classified all data correctly.


In [18]:
#the following code tests the MulticlassSVM and sgd
#will fail if MulticlassSVM is not implemented yet

est = MulticlassSVM(6,lam=0.01)
est.fit(X,y)
print("w:")
print(est.coef_)
#print(mesh_input.shape)
Z = est.predict(mesh_input)
Z = Z.reshape(xx.shape)
plt.contourf(xx, yy, Z, cmap=plt.cm.coolwarm, alpha=0.8)
# Plot also the training points
plt.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.coolwarm)

from sklearn import metrics
metrics.confusion_matrix(y, est.predict(X))


w:
[-1.10680193 -0.3459119  -0.0149213   0.55996935  1.12172323 -0.21405745]
Out[18]:
array([[100,   0,   0],
       [  0, 100,   0],
       [  0,   0, 100]])

Now for multiclass classification of NYC urban sound files!


In [35]:
import librosa as l
path ='../UrbanSound8K/audio/fold1/'
y, sr = l.load(path + '7061-6-0-0.wav')#+'7061-6-0-0.wav')

In [47]:
import IPython
IPython.display.Audio(path + '7061-6-0-0.wav')


Out[47]:

In [45]:
from librosa.display import waveplot, specshow

plt.figure()
plt.subplot(3, 1, 1)
waveplot(y, sr=sr)
plt.title('Waves')


Out[45]:
<matplotlib.text.Text at 0x67d5557ba8>

In [44]:
D = l.amplitude_to_db(l.stft(y), ref=np.max)
plt.subplot(4, 2, 1)
specshow(D, y_axis='linear')
plt.colorbar(format='%+2.0f dB')
plt.title('Linear-frequency power spectrogram')


Out[44]:
<matplotlib.text.Text at 0x67d549a518>

In [122]:
def features(file_path):
    try:
        y, sr = l.load(file_path, duration=5.0, res_type='scipy')
    except:
        print(file_path)
    #    return 'r'
    mfcc = l.feature.mfcc(y).mean(axis = 1).reshape(20)
    #print(l.feature.delta(y).mean(), mfcc.shape)
    features = np.append(mfcc,(l.feature.delta(y).mean()) ) #np.concatenate((l.feature.delta(y).mean(), mfcc ))
    if features.shape[0] != 21:
        #print(l.feature.delta(y).shape , l.feature.mfcc(y).shape)
        raise Exception('sample w/ bad feature')
    return features

#features(path + '7061-6-0-0.wav')

In [57]:
import pandas as pd
meta = pd.read_csv('../UrbanSound8K/metadata/UrbanSound8k.csv')
data_set = meta.sample(n=4000)
data_set.head()


Out[57]:
slice_file_name fsID start end salience fold classID class
2447 157695-3-0-8.wav 157695 4.000000 8.000000 1 4 3 dog_bark
2200 155044-9-0-37.wav 155044 18.500000 22.500000 1 9 9 street_music
7379 7064-6-0-0.wav 7064 0.948524 2.358256 1 4 6 gun_shot
3249 167464-0-0-14.wav 167464 7.000000 11.000000 2 10 0 air_conditioner
557 108041-9-0-7.wav 108041 3.500000 7.500000 1 1 9 street_music

In [ ]:
array = np.zeros((4000, 21))
labels = np.zeros((4000,))

from os import listdir 

i = 0
for folder in range(1,11):
    path ='../UrbanSound8K/audio/fold%d/' %(folder)
    for file in listdir(path):
        if file in data_set['slice_file_name'].values:
            labels[i] = data_set[data_set['slice_file_name'] == file]['classID'].values[0] 
            array[i] = features(path+file)
            i += 1

mean = array.mean(axis=1)
array -= mean[:, np.newaxis]
array_normed = array/(array.std(axis=1))

In [60]:
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(
    data_set['slice_file_name'], data_set['classID'], test_size=0.5, random_state=42)
X_train.shape, X_test.shape, y_train.shape


Out[60]:
((2000,), (2000,), (2000,))