FaceRecognition/adaboost.py at master · Diooooo/FaceRecognition · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
import numpy as np
import time


class AdaBoost:
    def __init__(self, pos_features, neg_features, weak_cls):
        self.pos_f = pos_features
        self.neg_f = neg_features
        self.p_count = self.pos_f.shape[0]
        self.n_count = self.neg_f.shape[0]
        self.weak_cls = weak_cls

    def train(self, iteration, indices=None, pos=None, neg=None):
        strong_cls = []
        if indices is None:
            indices = np.array([i for i in range(self.weak_cls.shape[0])])
        if pos is None:
            pos = self.pos_f
        if neg is None:
            neg = self.neg_f
        p_count = pos.shape[0]
        n_count = neg.shape[0]
        w_p = np.ones(p_count) * 1 / (2 * p_count)
        w_n = np.ones(n_count) * 1 / (2 * n_count)
        cur = time.time()
        for i in range(iteration):
            norm_sum = np.sum(w_p) + np.sum(w_n)
            w_p = w_p / norm_sum
            w_n = w_n / norm_sum
            selected_index = -1
            selected_threshold = 0
            selected_parity = 1
            selected_p_res = None
            selected_n_res = None
            error = 1
            for j, index in enumerate(indices):
                threshold = self.weak_cls[index][0]
                parity = self.weak_cls[index][1]
                err, p_res, n_res = self.get_cls_err(pos[:, index], neg[:, index], w_p, w_n, threshold,
                                                     parity)
                if err < error:
                    # print('index: {}, errpr: {}'.format(index, err))
                    error = err
                    selected_index = index
                    selected_threshold = threshold
                    selected_parity = parity
                    selected_p_res = p_res
                    selected_n_res = n_res
            beta = error / (1 - error)
            w_p = w_p * np.float_power(beta, selected_p_res)
            w_n = w_n * np.float_power(beta, 1 - selected_n_res)
            alpha = np.log(1 / beta)
            strong_cls.append([selected_index, selected_threshold, selected_parity, alpha])
            print('Classifier {}--index: {}, threshold: {}, parity: {}, error: {}, alpha: {}'.format(i + 1,
                                                                                                     selected_index,
                                                                                                     selected_threshold,
                                                                                                     selected_parity,
                                                                                                     error, alpha))
        save_file = './cascade_v3/strong_cls-{}k-{}.npy'.format(int(len(indices) / 1000), iteration)
        np.save(save_file, strong_cls)
        print('*' * 10, 'Saved [{}], training using: {}s'.format(save_file, time.time() - cur), '*' * 10)
        return np.asarray(strong_cls)

    def predict(self, strong_cls, features, label=None):
        if len(features.shape) > 2:
            n_sample = 1
            features = features.reshape(1, -1)
        else:
            n_sample = features.shape[0]
        pred = np.zeros(n_sample)
        tmp = np.zeros_like(pred)
        for i in range(strong_cls.shape[0]):
            alpha = strong_cls[i, 3]
            index = strong_cls[i, 0]
            threshold = strong_cls[i, 1]
            parity = strong_cls[i, 2]
            weak_pred = (((features[:, int(index)] - threshold) * parity) > 0)
            weak_pred = weak_pred.astype(np.int)
            tmp += alpha * weak_pred
        strong_threshold = 0.5 * np.sum(strong_cls[:, 3])
        pred[tmp >= strong_threshold] = 1
        if not label:
            return pred
        else:
            return pred

    def get_cls_err(self, p_feature, n_feature, w_p, w_n, threshold, parity):
        p_res = np.zeros(p_features.shape[0])
        n_res = np.zeros(n_features.shape[0])
        p_res[(p_feature - threshold) * parity > 0] = 1
        n_res[(n_feature - threshold) * parity > 0] = 1
        err = np.sum(w_p * np.abs(p_res - 1)) + np.sum(w_n * n_res)
        return err, p_res, n_res


def generate_cls(p_features, n_features, kernels):
    weak_classifiers = []
    m = kernels.shape[0]
    for f_num in range(m):
        pf = p_features[:3000, f_num]
        nf = n_features[:3000, f_num]
        elems = np.unique(np.concatenate((pf, nf), axis=0))
        print('min: {}, max: {}'.format(elems.min(), elems.max()))
        err = 0
        threshold = 0
        parity = 1
        cur = time.time()
        for e in elems:
            parity = 1
            p_res = np.zeros(pf.shape[0])
            n_res = np.zeros(nf.shape[0])
            p_res[pf > e] = 1
            n_res[nf > e] = 1
            num_pf1 = len(p_res[p_res == 1])
            num_nf0 = len(n_res[n_res == 0])
            num_pf = len(pf)
            num_nf = len(nf)
            err_rate = (num_pf1 + num_nf0) / (num_pf + num_nf)
            if 1. - err_rate > err_rate:
                parity = -1
                err_rate = 1. - err_rate
            if err_rate > err:
                err = err_rate
                threshold = e
        weak_classifiers.append([threshold, parity])
        if (f_num + 1) % 10000 == 0:
            print('*' * 10, 'saved weak_classifiers-{}.npy'.format(f_num + 1), '*' * 10)
            np.save('./weak_classifiers-{}.npy'.format(f_num + 1), np.asarray(weak_classifiers))
            weak_classifiers = []
        print(
            'feature: {}/{}, threshold: {}, error rate: {}, parity: {}, using: {}s'.format(f_num + 1, m, threshold, err,
                                                                                           parity, time.time() - cur))
    weak_classifiers = np.asarray(weak_classifiers)
    np.save('./weak_classifiers-last.npy', weak_classifiers)
    cls1 = np.load('./weak_classifiers-10000.npy')
    cls2 = np.load('./weak_classifiers-20000.npy')
    cls3 = np.load('./weak_classifiers-30000.npy')
    cls4 = np.load('./weak_classifiers-40000.npy')
    cls5 = np.load('./weak_classifiers-50000.npy')
    cls6 = np.load('./weak_classifiers-60000.npy')
    cls7 = np.load('./weak_classifiers-70000.npy')
    cls8 = np.load('./weak_classifiers-80000.npy')
    cls9 = np.load('./weak_classifiers-90000.npy')
    cls10 = np.load('./weak_classifiers-100000.npy')
    cls11 = np.load('./weak_classifiers-110000.npy')
    cls12 = np.load('./weak_classifiers-120000.npy')
    cls13 = np.load('./weak_classifiers-130000.npy')
    cls14 = np.load('./weak_classifiers-140000.npy')
    cls15 = np.load('./weak_classifiers-150000.npy')
    cls16 = np.load('./weak_classifiers-160000.npy')
    cls17 = np.load('./weak_classifiers-last.npy')
    cls = np.concatenate(
        (cls1, cls2, cls3, cls4, cls5, cls6, cls7, cls8, cls9, cls10, cls11, cls12, cls13, cls14, cls15, cls16, cls17),
        axis=0)
    np.save('./weak_classifiers.npy', cls)
    return cls


if __name__ == '__main__':
    # p = np.load('./dataset/positive.npy')
    # n = np.load('./dataset/negative-pre.npy')
    n_features = np.load('./n_features-pre.npy')
    # np.random.shuffle(n_features)
    # n_features = n_features[:3000]
    print('load negative features')
    p_features = np.load('./p_features.npy')
    np.random.shuffle(p_features)
    p_features = p_features[:3000]
    print('load positive features')

    # kernels = np.load('./haar-(24,24).npy')
    cls = np.load('./weak_classifiers.npy')
    print('load weak classifiers')
    adaboost = AdaBoost(p_features, n_features, cls)
    # rand_idx = np.random.permutation(cls.shape[0])
    layers = [1, 5, 10, 15, 20]
    for i in layers:
        # np.random.shuffle(rand_idx)
        # indices = rand_idx
        strong_cls = adaboost.train(i, indices=None, neg=n_features)
        # strong_cls = np.load('./strong_cls-160k-10.npy')
        cur = time.time()
        pred_p = adaboost.predict(strong_cls, p_features[:1000])
        pred_n = adaboost.predict(strong_cls, n_features)
        n_count = len(pred_n)
        total = n_count + 1000
        tp = len(pred_p[pred_p == 1])
        fp = 1000 - tp
        tn = len(pred_n[pred_n == 0])
        fn = n_count - tn
        print(
            'test using: {}s, tp: {}, fp: {}, fn: {}, tn: {}, total: {}'.format(time.time() - cur,
                                                                                tp / total,
                                                                                fp / total,
                                                                                fn / total,
                                                                                tn / total,
                                                                                (tp + tn) / total))
        # n_features_tn = n_features[pred_n == 0]
        n_features = n_features[pred_n == 1]
        # n_features = np.concatenate((n_features, n_features_tn[:int(0.5 * n_features_tn.shape[0])]), axis=0)
        print('negative: {}'.format(n_features.shape[0]))