SPECTRUM/dataset.py at main · NiceRingNode/SPECTRUM · 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
202
# -*- coding:utf-8 -*-

import numpy as np
from torch.utils.data import Dataset
from utils import extract_temporal_features

class Numbers(Dataset):
    def __init__(self,handwriting_info,transform=None,train=True):
        super().__init__()
        self.users = handwriting_info.keys()
        self.users_cnt = len(self.users)
        self.train = train
        self.features = []
        self.genuine_cnt = np.zeros(self.users_cnt,dtype=np.int32)
        self.forgery_cnt = np.zeros(self.users_cnt,dtype=np.int32)
        self.session_labels = []
        for i,k in enumerate(self.users):
            extract_temporal_features(handwriting_info[k]['genuine'],self.features)
            extract_temporal_features(handwriting_info[k]['forgery'],self.features)
            self.genuine_cnt[i] = len(handwriting_info[k]['genuine'])
            self.forgery_cnt[i] = len(handwriting_info[k]['forgery'])
        self.features_cnt = len(self.features)
        accu_indices = np.cumsum(self.genuine_cnt + self.forgery_cnt)
        self.accu_indices = np.roll(accu_indices,1)
        self.accu_indices[0] = 0
        self.feature_dims = np.shape(self.features[0])[1]
        user_interval = self.genuine_cnt[0] + self.forgery_cnt[0]
        self.user_labels = []
        for i in range(self.users_cnt):
            self.user_labels.extend([i] * user_interval)
        self.binary_labels = ([1] * self.genuine_cnt[0] + [0] * self.forgery_cnt[0]) * self.users_cnt
        self.transform = transform

    @property
    def config(self):
        config_info = {
            'users_cnt':self.users_cnt,
            'accu_indices':self.accu_indices,
            'forgery_cnt':self.forgery_cnt,
            'genuine_cnt':self.genuine_cnt,
            'features_cnt':self.features_cnt,
        }
        return config_info

    def __len__(self):
        return self.features_cnt

    def __getitem__(self,idx):
        if self.train:
            feature = self.features[idx]
            if self.transform is not None:
                feature = self.transform(feature)
        else:
            feature = self.features[idx]
        return feature,len(feature),self.user_labels[idx],self.binary_labels[idx]

class DeepSignDB(Dataset):
    def __init__(self,handwriting_info,transform=None,train=True,finger_mode=False):
        super().__init__()
        self.users = handwriting_info.keys()
        self.users_cnt = len(self.users)
        self.train = train
        self.features = []
        self.genuine_cnt = np.zeros(self.users_cnt,dtype=np.int32)
        self.forgery_cnt = np.zeros(self.users_cnt,dtype=np.int32)
        self.binary_labels = []
        for i,k in enumerate(self.users):
            extract_temporal_features(handwriting_info[k][True],self.features,finger_mode=finger_mode)
            extract_temporal_features(handwriting_info[k][False],self.features,finger_mode=finger_mode)
            self.genuine_cnt[i] = len(handwriting_info[k][True])
            self.forgery_cnt[i] = len(handwriting_info[k][False])
            self.binary_labels.extend([1] * self.genuine_cnt[i] + [0] * self.forgery_cnt[i])
        self.features_cnt = len(self.features)
        accu_indices = np.cumsum(self.genuine_cnt + self.forgery_cnt)
        self.accu_indices = np.roll(accu_indices,1)
        self.accu_indices[0] = 0
        self.feature_dims = np.shape(self.features[0])[1]
        self.transform = transform
        self.user_labels = []
        user_interval = self.genuine_cnt[0] + self.forgery_cnt[0]
        for i in range(self.users_cnt):
            self.user_labels.extend([i] * user_interval)

    @property
    def config(self):
        config_info = {
            'users_cnt':self.users_cnt,
            'accu_indices':self.accu_indices,
            'forgery_cnt':self.forgery_cnt,
            'genuine_cnt':self.genuine_cnt,
            'features_cnt':self.features_cnt,
        }
        return config_info

    def __len__(self):
        return self.features_cnt

    def __getitem__(self,idx):
        feature = self.features[idx]
        if self.train:
            if self.transform is not None:
                feature = self.transform(feature)
        return feature,len(feature),self.user_labels[idx],self.binary_labels[idx]

    def add_data(self,handwriting_info,finger_mode=False):
        cur_users = handwriting_info.keys()
        cur_users_cnt = len(handwriting_info)
        genuine_cnt = np.zeros(cur_users_cnt,dtype=np.int32)
        forgery_cnt = np.zeros(cur_users_cnt,dtype=np.int32)
        for i,k in enumerate(cur_users):
            extract_temporal_features(handwriting_info[k][True],self.features,finger_mode=finger_mode)
            extract_temporal_features(handwriting_info[k][False],self.features,finger_mode=finger_mode)
            genuine_cnt[i] = len(handwriting_info[k][True])
            forgery_cnt[i] = len(handwriting_info[k][False])
            self.binary_labels.extend([1] * genuine_cnt[i] + [0] * forgery_cnt[i])
        self.features_cnt = len(self.features)
        self.forgery_cnt = np.concatenate((self.forgery_cnt,forgery_cnt))
        self.genuine_cnt = np.concatenate((self.genuine_cnt,genuine_cnt))
        accu_indices = np.cumsum(self.genuine_cnt + self.forgery_cnt)
        self.accu_indices = np.roll(accu_indices,1)
        self.accu_indices[0] = 0
        last_users_cnt = self.users_cnt
        self.users_cnt += cur_users_cnt
        user_interval = genuine_cnt[0] + forgery_cnt[0]
        for i in range(last_users_cnt,self.users_cnt):
            self.user_labels.extend([i] * user_interval)

class TrainSampler:
    def __init__(self,dataset_config:dict,user_sample,genuine_sample=5,forgery_sample=5):
        self.__dict__.update(dataset_config)
        self.users_indices = np.arange(0,self.users_cnt,dtype=np.int32)
        self.user_sample = user_sample
        self.genuine_sample = genuine_sample
        self.forgery_sample = forgery_sample

    def __len__(self):
        return self.users_cnt

    def __iter__(self):
        np.random.shuffle(self.users_indices)
        for _ in range(self.users_cnt):
            batch_indices = []
            indices = np.random.choice(self.users_indices,size=self.user_sample,replace=False)
            for idx in indices:
                genuine_indices = np.random.choice(np.arange(0,self.genuine_cnt[idx]),
                    size=self.genuine_sample,replace=False)
                forgery_indices = np.random.choice(np.arange(0,self.forgery_cnt[idx]),
                    size=self.forgery_sample // 2,replace=False) + self.genuine_cnt[idx]
                batch_indices.append(genuine_indices + self.accu_indices[idx])
                batch_indices.append(forgery_indices + self.accu_indices[idx])
                random_forgery_users = (idx + np.random.randint(1,self.users_cnt - 1,size=self.forgery_sample // 2)) % self.users_cnt
                for rf_user in random_forgery_users:
                    rf_idx = np.random.choice(self.genuine_cnt[rf_user],size=1,replace=False)
                    batch_indices.append(rf_idx + self.accu_indices[rf_user])
            batch_indices = np.concatenate(batch_indices,axis=0).astype(np.int32)
            yield batch_indices

class TestSampler:
    def __init__(self,users_cnt,genuine_sample=5,forgery_sample=5):
        self.users_cnt = users_cnt
        self.users_indices = np.arange(0,self.users_cnt,dtype=np.int32)
        self.user_sample = genuine_sample + forgery_sample
        self.genuine_sample = genuine_sample
        self.forgery_sample = forgery_sample

    def __len__(self):
        return self.users_cnt

    def __iter__(self):
        for i in range(self.users_cnt):
            batch_indices = np.arange(i * self.user_sample,(i + 1) * self.user_sample,dtype=np.int32)
            yield batch_indices

class TestSamplerDeepSign:
    def __init__(self,dataset_config):
        self.__dict__.update(dataset_config)
        self.users_indices = np.arange(0,self.users_cnt,dtype=np.int32)
        self.user_sample = self.genuine_cnt[0] + self.forgery_cnt[0] # 每个batch的用户个数

    def __len__(self):
        return self.users_cnt

    def __iter__(self):
        # np.random.shuffle(self.users_indices)
        for i in range(self.users_cnt): # 每次都取一个用户，每个都拿全部20个，然后55个就取完了
            batch_indices = np.arange(i * self.user_sample,(i + 1) * self.user_sample,dtype=np.int32)
            # print(batch_indices)
            yield batch_indices

def collate_fn(batch:list):
    # batch: 列表，第一个是笔迹信息，第二个是长度
    batch_size = len(batch)
    handwriting = [i[0] for i in batch]
    hw_len = np.array([i[1] for i in batch],dtype=np.float32)
    user_labels = np.array([i[2] for i in batch])
    binary_labels = np.array([i[3] for i in batch])
    max_len = int(np.max(hw_len))
    time_function_cnts = np.shape(handwriting[0])[1]
    handwriting_padded = np.zeros((batch_size,max_len,time_function_cnts),dtype=np.float32)
    for i,hw in enumerate(handwriting):
        handwriting_padded[i,:hw.shape[0]] = hw
    return handwriting_padded,hw_len,user_labels,binary_labels