HaMI/utils.py at main · mala-lab/HaMI · 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
import pickle
import numpy as np
import random
import torch
from torch.utils.data import Dataset

from sklearn import metrics
from collections import defaultdict


def load_embedding(args, mode='train'):
    datapath = f"{args.data_dir}/{mode}_{args.data_name}_{args.model_name}.pkl"
    acc_datapath = f"{args.data_dir}/acc_double_{mode}_{args.data_name}_{args.model_name}.pkl"
    se_datapath = f"{args.data_dir}/se_{mode}_{args.data_name}_{args.model_name}.pkl"
    loaded_data = pickle.load(open(datapath, 'rb'))
    acc_loaded_data = pickle.load(open(acc_datapath, 'rb'))
    se_loaded_data = pickle.load(open(se_datapath, 'rb'))
    assert list(loaded_data.keys()) == list(acc_loaded_data.keys()) == se_loaded_data['id'], "ID mismatch"

    # Processing train data
    se_prob_list = idx2prob(se_loaded_data['semantic_ids'], used_sample=args.used_sample) # turn class identity to probability - list of lists
    if mode == 'train':
        nor_hidden_list = []
        abn_hidden_list = []
    else:
        emb_hidden_list = defaultdict(list)
        label_hidden_list = defaultdict(list)
        random.seed(args.seed)
        val_indices = random.sample(range(len(loaded_data)), args.val_sample)
        print(f"val_indices: {len(val_indices)}")

    used_acc = 1 if mode == 'train' else args.used_sample
    for id, tid in enumerate(loaded_data):
        example = loaded_data[tid]
        high_t_answer = example['high_t_answer']
        for m in range(used_acc):
            acc = acc_loaded_data[tid][m]
            if acc == 2.0 and args.is_refined:
                continue

            hiddens = high_t_answer[m][-1][:, args.start_layer::args.interval_layer,:] if args.end_layer is None \
                else high_t_answer[m][-1][:,args.start_layer:args.end_layer:args.interval_layer,:]
            logits = high_t_answer[m][2]
            log_mean = np.mean(np.array(logits))
            se_prob = se_prob_list[id][m]
            # print(se_prob_list[id], se_prob)

            if args.mode == 'logits':
                logit_prob = [np.exp(logits[i]) for i in range(len(logits))]
                hiddens = hiddens*((args.a+args.b*torch.tensor(logit_prob)).reshape(-1,1,1))
            elif args.mode == 'log_mean':
                hiddens = hiddens*(args.a-args.b*log_mean)
            elif args.mode == 'se':
                hiddens = hiddens*(args.a+args.b*se_prob)
            elif args.mode == 'token_log_mean':
                token_log_mean = [np.mean(np.array(logits[:(i+1)])) for i in range(len(logits))]
                hiddens = hiddens*((args.a-args.b*torch.tensor(token_log_mean)).reshape(-1,1,1))

            if mode == 'train':
                if acc == 1.0 or acc == 2.0:
                    nor_hidden_list.append(hiddens)
                elif acc == 0.0:
                    abn_hidden_list.append(hiddens)

            else:
                if id in val_indices:
                    m='val'

                emb_hidden_list[m].append(hiddens)
                if acc == 2.0:
                    label_hidden_list[m].append(0.0)
                else:
                    label_hidden_list[m].append(1.0-acc)
    if mode == 'train':
        n_label = [0.0]*len(nor_hidden_list)
        a_label = [1.0]*len(abn_hidden_list)
        n_idx = list(range(len(nor_hidden_list)))
        a_idx = list(range(len(abn_hidden_list)))
        return nor_hidden_list, abn_hidden_list, n_label, a_label, n_idx, a_idx
    else:
        for m in range(used_acc):
            assert len(emb_hidden_list[m]) == len(label_hidden_list[m]), f"{m} Length mismatch"
        return emb_hidden_list, label_hidden_list


def idx2prob(semantic_ids, used_sample=6):
    """
    Converts semantic cluster results to probabilities.
    Args:
        semantic_ids (list): List of semantic_id lists.
        used_sample (int): Number of samples to use.
    Returns:
        list: List of probabilitity list for each question.
    """
    se_uncertainty = []
    for count in semantic_ids:
        if len(count) > used_sample:
            count = count[:used_sample]
        elif len(count) < used_sample:
            count = count
        values, nums = np.unique(count, return_counts=True)
        probabilities = nums / len(count)
        pro_dict = dict(zip(values, probabilities))
        prob_list = [pro_dict[i] for i in count]
        se_uncertainty.append(prob_list)

    return se_uncertainty


# ref: DeepOD: https://github.com/xuhongzuo/DeepOD/tree/main/deepod
class CustomDataset(Dataset):
    def __init__(self, x_data, y_data):

        self.x_data = torch.tensor(x_data, dtype=torch.long)  # features
        self.y_data = torch.tensor(y_data, dtype=torch.long)  # labels

    def __len__(self):
        return len(self.x_data)

    def __getitem__(self, index):
        # return a tuple of (x, y)
        return self.x_data[index], self.y_data[index]


def tabular_metrics(y_true, y_score):


    # F1@k, using real percentage to calculate F1-score
    ratio = 100.0 * len(np.where(y_true == 0)[0]) / len(y_true)
    thresh = np.percentile(y_score, ratio)
    y_pred = (y_score >= thresh).astype(int)
    y_true = y_true.astype(int)
    p, r, f1, support = metrics.precision_recall_fscore_support(y_true, y_pred, average='binary')

    return metrics.roc_auc_score(y_true, y_score), metrics.average_precision_score(y_true, y_score), f1