Behind-the-Scenes/sample.py at master · Wangpeiyi9979/Behind-the-Scenes · 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
import configs
import json
import os
import utils
import torch
import random
import numpy as np
class Sampler():
    """
    FewEC Dataset
    """

    def __init__(self, opt, case='train'):

        self.type2sentence_index = json.load(open(os.path.join(opt.data_dir, case + 'type2sentence_index.json'), 'r'))
        self.type2trigger2sentence_index = json.load(open(os.path.join(opt.data_dir, case + 'type2trigger2sentence_index.json'), 'r'))
        self.target_classes = list(self.type2sentence_index.keys()) #
        self.label2id = utils.create_id2label_dict('tool_data/{}/{}_classes.txt'.format(opt.dataset, case), reverse=True)
        self.id2label = utils.create_id2label_dict('tool_data/{}/{}_classes.txt'.format(opt.dataset, case), reverse=False)
        self.word2id = json.load(open('tool_data/word2id.json'))
        self.id2word = json.load(open('tool_data/id2word.json'))
        all_data = torch.load(os.path.join(opt.data_dir, "{}_alldata.pth".format(case)))
        try:
            self.synonym_dict = torch.load(os.path.join(opt.data_dir, "{}_synonym_dict.pth".format(case)))  #  self.synonym_dict[event1][event2]: confusion triggers accorss event1 and event2
        except FileNotFoundError:
            print('no synonym dict file')
        self.all_data = all_data
        if case == 'train':
            self.N = opt.N_train
        else:
            self.N = opt.N
        self.K = opt.K
        self.opt = opt

    def next_one(self, method='normal', blurry_p=1, out_sample=False, train=False):
        target_classes = random.sample(self.target_classes, self.N)
        # import ipdb; ipdb.set_trace()
        support_data = [] # N * K
        query_data = [] # N
        support_label_id2rel_label = {i : self.id2label[int(v)] for i, v in enumerate(target_classes)}

        # sampling
        if method == 'normal':
            for i, class_name in enumerate(target_classes):
                sentence_indexes = self.type2sentence_index[class_name]
                indices = np.random.choice(sentence_indexes, self.K + 1, False) # use the last one as query
                for idx in indices[:-1]:
                    support_data.append(self.all_data[idx])
                query_data.append(self.all_data[indices[-1]])

        elif method == 'trigger_uniform':
            for i, class_name in enumerate(target_classes):
                all_trigger = list(self.type2trigger2sentence_index[class_name].keys())
                query_trigger =  np.random.choice(all_trigger, 1, False)
                all_trigger_left = list(set(all_trigger) - set(query_trigger))
                if len(all_trigger_left) >= self.K:
                    trigger_list = np.random.choice(all_trigger_left, self.K, False)
                else:
                    try:
                        trigger_list = np.random.choice(all_trigger_left, self.K, True)
                    except:
                        trigger_list = np.random.choice(all_trigger, self.K, True) # this class has only 1 instance
                all_select_idxs = set()
                for trigger in trigger_list:
                    try:
                        left_idxs = list(set(self.type2trigger2sentence_index[class_name][trigger]) - all_select_idxs)
                        idx = np.random.choice(left_idxs, 1, False)[0]
                    except:
                        idx = np.random.choice(self.type2trigger2sentence_index[class_name][trigger], 1, False)[0]
                    all_select_idxs.add(idx)
                    support_data.append(self.all_data[idx])
                idx = np.random.choice(self.type2trigger2sentence_index[class_name][query_trigger[0]], 1, False)[0]
                query_data.append(self.all_data[idx])
        elif method == 'confusion_uniform':
            for i, class_name in enumerate(target_classes):
                all_trigger = set(self.type2trigger2sentence_index[class_name].keys())
                blurry_trigger = set()
                for j, other_class_name in enumerate(target_classes):
                    if i != j:
                        blurry_trigger = blurry_trigger | set(self.synonym_dict[int(class_name)].get(int(other_class_name), []))
                remain_trigger = all_trigger - blurry_trigger

                query_trigger =  np.random.choice(list(all_trigger), 1, False)
                all_remain_trigger_left = remain_trigger - set(query_trigger)
                all_blurry_trigger_left = blurry_trigger - set(query_trigger)
                all_select_idxs = set()
                for _ in range(self.K):
                    p = np.random.uniform()
                    if len(all_blurry_trigger_left) == 0 and len(all_remain_trigger_left) == 0:
                        trigger = np.random.choice(list(all_trigger-set(query_trigger)), 1, False)[0]
                    elif 0 <= p <= blurry_p and len(all_blurry_trigger_left) !=0 or len(all_remain_trigger_left) == 0:
                        trigger = np.random.choice(list(all_blurry_trigger_left), 1, False)[0]
                        all_blurry_trigger_left = all_blurry_trigger_left - {trigger}
                    else:
                        trigger = np.random.choice(list(all_remain_trigger_left), 1, False)[0]
                        all_remain_trigger_left = all_remain_trigger_left - {trigger}
                    try:
                        left_idxs = list(set(self.type2trigger2sentence_index[class_name][trigger]) - all_select_idxs)
                        idx = np.random.choice(left_idxs, 1, False)[0]
                    except ValueError:
                        idx = np.random.choice(list(set(self.type2trigger2sentence_index[class_name][trigger])), 1, False)[0]
                    support_data.append(self.all_data[idx])
                    all_select_idxs.add(idx)
                idx = np.random.choice(self.type2trigger2sentence_index[class_name][query_trigger[0]], 1, False)[0]
                query_data.append(self.all_data[idx])
        else:
            assert 1 == 2

        # save_data_to_file
        support_tokens = [d['token'] for d in support_data]
        support_trigger_indices = [d['trigger_index'] for d in support_data]

        support_labels = []
        for k, v in support_label_id2rel_label.items():
            support_labels.extend([v] * self.K)
        query_tokens = [d['token'] for d in query_data]
        query_trigger_indices = [d['trigger_index'] for d in query_data]

        query_labels = []
        for k, v in support_label_id2rel_label.items():
            query_labels.append(v)

        return {
            'support_tokens': support_tokens,
            'query_tokens': query_tokens,
            'support_trigger_indices': support_trigger_indices,
            'query_trigger_indices': query_trigger_indices,
            'support_labels': support_labels,
            'query_labels': query_labels,
        }


from tqdm import tqdm
def sample(**keward):
    opt = getattr(configs, 'ProtoConfig')()
    opt.parse(keward)

    train_sample = 20000
    eval_sample = 2000
    test_sample = 10000

    # normal
    Ns=[5, 10]
    Ks=[5, 10]
    test_methods=['normal', 'trigger_uniform', 'confusion_uniform']
    for test_method in test_methods:
        for N in Ns:
            for K in Ks:
                opt.K = K
                opt.N = N
                print(f'method :{test_method} N: {N} K:{K}')
                train_sampler = Sampler(opt, case='train')
                dev_sampler = Sampler(opt, case='val')
                test_sampler = Sampler(opt, case='test')
                all_train_tasks = []
                all_eval_tasks = []
                all_test_tasks = []
                for _ in tqdm(range(train_sample)):
                    task = train_sampler.next_one(method='normal')
                    all_train_tasks.append(task)
                for _ in tqdm(range(eval_sample)):
                    task = dev_sampler.next_one(method=test_method)
                    all_eval_tasks.append(task)
                for _ in tqdm(range(test_sample)):
                    task = test_sampler.next_one(method=test_method)
                    all_test_tasks.append(task)
                out_dir = f'./sampled_tasks/{N}way-{K}shot-{test_method}'
                if not os.path.exists(out_dir):
                    os.mkdir(out_dir)
                with open(out_dir+'/train.jsonl', 'w') as f:
                    for task in all_train_tasks:
                        f.write(json.dumps(task) + '\n')
                with open(out_dir+'/dev.jsonl', 'w') as f:
                    for task in all_eval_tasks:
                        f.write(json.dumps(task) + '\n')
                with open(out_dir+'/test.jsonl', 'w') as f:
                    for task in all_test_tasks:
                        f.write(json.dumps(task) + '\n')


if __name__ == '__main__':
    import fire
    fire.Fire()