add Classifier.py

1 files changed, 133 insertions, 0 deletions

diff --git a/Classifier.py b/Classifier.py
new file mode 100644
index 0000000..67e84fb
--- /dev/null
+++ b/Classifier.py
@@ -0,0 +1,133 @@
+import numpy as np
+from itertools import combinations
+from sklearn.preprocessing import MinMaxScaler
+from sklearn.grid_search import GridSearchCV
+from sklearn.neighbors import KNeighborsClassifier
+import sklearn.svm.libsvm
+from sklearn.svm import SVC
+
+import random
+import threading
+
+NUM_OF_LABELS = 4
+WINDOW_SIZE = 500
+WINDOW_SHIFT_TRAIN = 50
+WINDOW_SHIFT_TEST = 50
+
+def data_of_two_labels(data, ans, y1, y2):
+    tmp = []
+    tmpa = []
+    for i in range(len(data)):
+        if ans[i] != y1 and ans[i] != y2:
+            continue
+        tmp.append(data[i])
+        tmpa.append(ans[i])
+    return tmp, tmpa
+
+class SVM_train_thread(threading.Thread):
+    def __init__(self, data, ans, clf, labels):
+        super(SVM_train_thread, self).__init__()
+        self.data = data
+        self.ans = ans
+        self.clf = clf
+        self.label = labels
+    def run(self):
+        self.clf.fit( self.data, self.ans )
+
+def gen_model(train_data, train_ans, verbose=True):
+    scalers = []
+    classifiers = []
+    scores = []
+    threads = []
+    tuned_parameters = [
+            {'kernel': ['rbf'], 'gamma': [], 'C': []},
+        ]
+    now = 4.0
+    while now > 0.00001:
+        tuned_parameters[0]['gamma'].append(now)
+        if now >= 1:
+            tuned_parameters[0]['C'].append(now) 
+        now /= 2;
+
+    for i in combinations(range(NUM_OF_LABELS), 2):
+        data, ans = data_of_two_labels(train_data, train_ans, i[0], i[1])
+        data = np.array(data)
+
+        tmp = zip(data, ans)
+        random.shuffle(tmp)
+        data, ans = zip(*tmp)
+        
+        scalers.append( MinMaxScaler(feature_range=(-1,1)) )
+        #scalers.append( StandardScaler() )
+        data = scalers[-1].fit_transform( data )
+        clf = GridSearchCV( SVC(), tuned_parameters, cv=5, refit=True )
+        threads.append( SVM_train_thread(data, ans, clf, i) )
+        threads[-1].start()
+
+    for t in threads:
+        t.join()
+        classifiers.append( t.clf.best_estimator_ )
+        scores.append(t.clf.best_score_)
+        if verbose:
+            print(t.label, t.clf.best_params_, t.clf.best_score_)
+
+    scores = [1] * NUM_OF_LABELS
+    return scalers, classifiers, scores
+
+
+def multi_classification(datas, scalers, clfs, scores):
+    predictions = []
+    ret = []
+    j = 0
+    for i in combinations(range(NUM_OF_LABELS), 2):
+        predictions.append(clfs[j].predict(scalers[j].transform(np.array(datas))))
+        j += 1
+    predictions = np.array(predictions).T
+    #print predictions
+    for i in predictions:
+        label_score = [0.0] * NUM_OF_LABELS
+        for j in i:
+            label_score[j] += scores[j]
+        #print label_score
+        ret.append(label_score.index(max(label_score)))
+    return ret
+
+def main():
+    import json
+    with open("data/rawdata_0_2_4_serial", "r") as f:
+        y, X1, X2 = zip(*json.loads(f.readline()))
+
+    testX1 = np.array(X1[-3 * NUM_OF_LABELS:]).reshape(1, -1)[0]
+    testX2 = np.array(X2[-3 * NUM_OF_LABELS:]).reshape(1, -1)[0]
+    testy = y[-3 * NUM_OF_LABELS:]
+    X1 = X1[:-3 * NUM_OF_LABELS]
+    X2 = X2[:-3 * NUM_OF_LABELS]
+    y = y[:-3 * NUM_OF_LABELS]
+    X = []
+    y_2 = []
+
+    print len(testX1)
+    for yi, x1, x2 in zip(y, X1, X2):
+        for i in range(700, 1300, WINDOW_SHIFT_TRAIN):
+            X.append( np.concatenate(( 
+                    np.absolute(np.fft.fft(x1[i: i+WINDOW_SIZE])) , 
+                    np.absolute(np.fft.fft(x2[i: i+WINDOW_SIZE])) ) ).tolist())
+            y_2.append( yi )
+
+    y = y_2
+    scalers, classifiers, scores = gen_model(X, y)
+
+    print testy
+    ans = []
+    for i in range(0, len(testX1)-WINDOW_SIZE, WINDOW_SHIFT_TEST):
+        x = np.concatenate(( 
+                np.absolute(np.fft.fft(testX1[i: i+WINDOW_SIZE])) , 
+                np.absolute(np.fft.fft(testX2[i: i+WINDOW_SIZE])) ) ).tolist()
+        ans.append( multi_classification(x, scalers, classifiers, scores)[0] )
+
+    ans += [0] * (WINDOW_SIZE / WINDOW_SHIFT_TEST)
+    for i in np.array(ans).reshape(-1,40).tolist():
+        print i
+
+if __name__ == '__main__':
+    main()