1+ import numpy as np
2+ from sklearn .ensemble import RandomForestClassifier , VotingClassifier
3+ from sklearn .metrics .classification import accuracy_score , recall_score , f1_score
4+ import scipy .stats as st
5+ import sys
6+
7+ def A (sample ):
8+ feat = []
9+ for col in range (0 ,sample .shape [1 ]):
10+ average = np .average (sample [:,col ])
11+ feat .append (average )
12+
13+ return feat
14+
15+ def SD (sample ):
16+ feat = []
17+ for col in range (0 , sample .shape [1 ]):
18+ std = np .std (sample [:, col ])
19+ feat .append (std )
20+
21+ return feat
22+
23+ def AAD (sample ):
24+ feat = []
25+ for col in range (0 , sample .shape [1 ]):
26+ data = sample [:, col ]
27+ add = np .mean (np .absolute (data - np .mean (data )))
28+ feat .append (add )
29+
30+ return feat
31+
32+ def ARA (sample ):
33+ #Average Resultant Acceleration[1]:
34+ # Average of the square roots of the sum of the values of each axis squared √(xi^2 + yi^2+ zi^2) over the ED
35+ feat = []
36+ sum_square = 0
37+ sample = np .power (sample , 2 )
38+ for col in range (0 , sample .shape [1 ]):
39+ sum_square = sum_square + sample [:, col ]
40+
41+ sample = np .sqrt (sum_square )
42+ average = np .average (sample )
43+ feat .append (average )
44+ return feat
45+
46+ def TBP (sample ):
47+ from scipy import signal
48+ feat = []
49+ sum_of_time = 0
50+ for col in range (0 , sample .shape [1 ]):
51+ data = sample [:, col ]
52+ peaks = signal .find_peaks_cwt (data , np .arange (1 ,4 ))
53+
54+ feat .append (peaks )
55+
56+ return feat
57+
58+ def feature_extraction (X ):
59+ #Extracts the features, as mentioned by Catal et al. 2015
60+ # Average - A,
61+ # Standard Deviation - SD,
62+ # Average Absolute Difference - AAD,
63+ # Average Resultant Acceleration - ARA(1),
64+ # Time Between Peaks - TBP
65+ X_tmp = []
66+ for sample in X :
67+ features = A (sample )
68+ features = np .hstack ((features , A (sample )))
69+ features = np .hstack ((features , SD (sample )))
70+ features = np .hstack ((features , AAD (sample )))
71+ features = np .hstack ((features , ARA (sample )))
72+ #features = np.hstack((features, TBP(sample)))
73+ X_tmp .append (features )
74+
75+ X = np .array (X_tmp )
76+ return X
77+
78+ def train_j48 (X , y ):
79+ from sklearn import tree
80+ clf = tree .DecisionTreeClassifier ()
81+ #clf = clf.fit(X, y)
82+ return clf
83+
84+ def train_mlp (X , y ):
85+ from sklearn .neural_network import MLPClassifier
86+ a = int ((X .shape [1 ] + np .amax (y )) / 2 )#Default param of weka, amax(y) gets the number of classes
87+ clf = MLPClassifier (solver = 'lbfgs' , alpha = 1e-5 , hidden_layer_sizes = (a ,),
88+ learning_rate_init = 0.3 , momentum = 0.2 , max_iter = 500 , #Default param of weka
89+ )
90+ #clf.fit(X, y)
91+ return clf
92+
93+ def train_logistic_regression (X , y ):
94+ from sklearn .linear_model import LogisticRegression
95+ clf = LogisticRegression (multi_class = 'ovr' )
96+ #clf.fit(X, y)
97+ return clf
98+
99+ if __name__ == '__main__' :
100+ #Paper: On the use of ensemble of classifiers for accelerometer-based activity recognition
101+ np .random .seed (12227 )
102+
103+ if (len (sys .argv ) > 1 ):
104+ data_input_file = sys .argv [1 ]
105+ else :
106+ data_input_file = 'E:/datasets/sensors/TemporalWindow/LOSO/UTD-MHAD2_1s.npz'
107+
108+ tmp = np .load (data_input_file )
109+ X = tmp ['X' ]
110+ X = X [:, 0 , :, :]
111+ y = tmp ['y' ]
112+ folds = tmp ['folds' ]
113+
114+ n_class = y .shape [1 ]
115+
116+ avg_acc = []
117+ avg_recall = []
118+ avg_f1 = []
119+ y = np .argmax (y , axis = 1 )
120+
121+ print ('Catal et al. 2015 {}' .format (data_input_file ))
122+
123+ for i in range (0 , len (folds )):
124+ train_idx = folds [i ][0 ]
125+ test_idx = folds [i ][1 ]
126+
127+ X_train = X [train_idx ]
128+ X_test = X [test_idx ]
129+
130+ X_train = feature_extraction (X_train )
131+ X_test = feature_extraction (X_test )
132+
133+ j_48 = train_j48 (X_train ,y [train_idx ])
134+ mlp = train_mlp (X_train , y [train_idx ])
135+ logistic_regression = train_logistic_regression (X_train , y [train_idx ])
136+
137+ majority_voting = VotingClassifier (estimators = [('dt' , j_48 ), ('mlp' , mlp ), ('lr' , logistic_regression )], voting = 'soft' )
138+ majority_voting .fit (X_train , y [train_idx ])
139+ tmp = majority_voting .predict (X_test )
140+
141+ acc_fold = accuracy_score (y [test_idx ], tmp )
142+ avg_acc .append (acc_fold )
143+
144+ recall_fold = recall_score (y [test_idx ], tmp , average = 'macro' )
145+ avg_recall .append (recall_fold )
146+
147+ f1_fold = f1_score (y [test_idx ], tmp , average = 'macro' )
148+ avg_f1 .append (f1_fold )
149+
150+ print ('Accuracy[{:.4f}] Recall[{:.4f}] F1[{:.4f}] at fold[{}]' .format (acc_fold , recall_fold , f1_fold ,i ))
151+ print ('______________________________________________________' )
152+
153+ ic_acc = st .t .interval (0.9 , len (avg_acc ) - 1 , loc = np .mean (avg_acc ), scale = st .sem (avg_acc ))
154+ ic_recall = st .t .interval (0.9 , len (avg_recall ) - 1 , loc = np .mean (avg_recall ), scale = st .sem (avg_recall ))
155+ ic_f1 = st .t .interval (0.9 , len (avg_f1 ) - 1 , loc = np .mean (avg_f1 ), scale = st .sem (avg_f1 ))
156+ print ('Mean Accuracy[{:.4f}] IC [{:.4f}, {:.4f}]' .format (np .mean (avg_acc ), ic_acc [0 ], ic_acc [1 ]))
157+ print ('Mean Recall[{:.4f}] IC [{:.4f}, {:.4f}]' .format (np .mean (avg_recall ), ic_recall [0 ], ic_recall [1 ]))
158+ print ('Mean F1[{:.4f}] IC [{:.4f}, {:.4f}]' .format (np .mean (avg_f1 ), ic_f1 [0 ], ic_f1 [1 ]))
0 commit comments