CED/knn_eval.py at main · MLAI-Yonsei/CED · 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
import os
import torch
import argparse
import numpy as np
import sklearn.covariance
from loguru import logger
from lib.metrics import get_metrics
from torch.nn import CosineSimilarity
from sklearn.neighbors import NearestNeighbors
import torch.nn.functional as F
from sklearn.covariance import EmpiricalCovariance, MinCovDet
import pandas as pd

# inter-layer pooling
def pooling_features(features, pooling='last'):
    num_layers = features.shape[0]
    if pooling == 'last':
        return features[-1, :, :]
    elif pooling == 'avg':
        return np.mean(features[1:], axis=0)
    else:
        raise NotImplementedError


def load_and_pool_features(input_dir, dataset_patterns, token_pooling, layer_pooling):
    pooled_features_dict = {}

    for pattern in dataset_patterns:
        pooled_features_list = []
        i = 1
        while True:
            filepath = f"{input_dir}/{token_pooling}_{pattern}{i}.npy"
            if not os.path.exists(filepath):
                break
            features = np.load(filepath, allow_pickle=True)
            pooled_features = pooling_features(features, layer_pooling)
            pooled_features_list.append(pooled_features)
            i += 1

        pooled_features_dict[pattern] = pooled_features_list
    return pooled_features_dict


def average_scores(feature_sets, ind_train_features, knn):
    num_feature_sets = len(feature_sets)
    num_samples = feature_sets[0].shape[0]
    knn = NearestNeighbors(n_neighbors=10, algorithm='brute')
    knn.fit(ind_train_features)
    scores_sum = np.zeros(num_samples)

    for features in feature_sets:
        normalized_features = features / (np.linalg.norm(features, axis=-1, keepdims=True) + 1e-10)
        scores = 1 - np.max(knn.kneighbors(normalized_features)[0], axis=1)
        scores_sum += scores
    average_scores = scores_sum / num_feature_sets
    return average_scores


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--dataset', type=str)
    parser.add_argument('--ood_datasets',
                        type=str, required=False)
    parser.add_argument('--layer_pooling', type=str, default='last')
    parser.add_argument('--token_pooling', type=str, default='avg',
                        help='token pooling way', choices=['cls', 'avg', 'max'])
    parser.add_argument('--distance_metric', type=str, default='maha',
                        help='distance metric')
    parser.add_argument('--log_file', type=str, default='./log/default.log')
    parser.add_argument('--base_dir',required=False,type=str)
    parser.add_argument('--seed',required=False,type=str)
    parser.add_argument('--input_dir', default='./log/embeddings/roberta-base/sst-2/seed13',
        type=str, required=False, help='save directory')
    args = parser.parse_args()

    log_file_name = args.log_file
    logger.add(log_file_name)
    logger.info('args:\n' + args.__repr__())

    input_dir = args.input_dir
    base_dir = args.base_dir
    token_pooling = args.token_pooling
    layer_pooling = args.layer_pooling
    ood_dataset = args.ood_datasets
    dataset_patterns = ['ood_combined', 'ood_oracles', 'oracle_combined', 'test_combined', 'oracle_combined_test','test_oracles']

    ind_train_features = np.load(
        '{}/{}_ind_train_features.npy'.format(base_dir, token_pooling))
    ind_train_labels = np.load(
        '{}/{}_ind_train_labels.npy'.format(base_dir, token_pooling))

    ind_train_features = pooling_features(ind_train_features, layer_pooling)
    ind_train_features = ind_train_features / \
            np.linalg.norm(ind_train_features, axis=-1, keepdims=True) + 1e-10

    ind_test_features = np.load(
        '{}/{}_ind_test_features.npy'.format(base_dir, token_pooling))
    ind_test_labels = np.load( '{}/{}_ind_test_labels.npy'.format(base_dir, token_pooling))
    ind_test_features = pooling_features(ind_test_features, layer_pooling)
    ind_test_features = ind_test_features / \
        np.linalg.norm(ind_test_features, axis=-1, keepdims=True) + 1e-10

    knn = NearestNeighbors(n_neighbors=10, algorithm='brute')
    knn.fit(ind_train_features)

    ind_scores = 1 - np.mean(knn.kneighbors(ind_test_features)[0], axis=1)

    ood_features = np.load(
        '{}/{}_ood_features_{}.npy'.format(base_dir, token_pooling, ood_dataset))
    ood_pairs = np.load(
            '{}/{}_ood_pairs_{}.npy'.format(base_dir, token_pooling, ood_dataset),allow_pickle=True)
    ood_features = pooling_features(ood_features, layer_pooling)
    ood_features = ood_features / \
        np.linalg.norm(ood_features, axis=-1, keepdims=True) + 1e-10
    ood_scores = 1 - np.max(knn.kneighbors(ood_features)[0], axis=1)
    metrics = get_metrics(ind_scores, ood_scores)
    print("KNN score")
    logger.info('ood dataset: {}'.format(ood_dataset))
    logger.info('metrics: {}'.format(metrics))


    dataset_patterns = ['ood_combined', 'ood_oracles', 'oracle_combined','test_combined', 'oracle_combined_test','test_oracles']
    pooled_features_dict = load_and_pool_features(input_dir, dataset_patterns, token_pooling, layer_pooling)

    ood_features = np.load(
        '{}/{}_ood_features_{}.npy'.format(base_dir, token_pooling, ood_dataset))
    ood_features = pooling_features(ood_features, layer_pooling)
    ood_features = ood_features / \
        np.linalg.norm(ood_features, axis=-1, keepdims=True) + 1e-10
    ood_scores = 1 - np.max(knn.kneighbors(ood_features)[0], axis=1)

    ind_test_features = np.load(
        '{}/{}_ind_test_features.npy'.format(base_dir, token_pooling))
    ind_test_features = pooling_features(ind_test_features, layer_pooling)
    ind_test_features = ind_test_features / \
            np.linalg.norm(ind_test_features, axis=-1, keepdims=True) + 1e-10
    test_scores = 1 - np.mean(knn.kneighbors(ind_test_features)[0], axis=1)


    ood_oracles_scores = average_scores(pooled_features_dict['ood_oracles'], ind_train_features, knn)
    test_oracles_scores = average_scores(pooled_features_dict['test_oracles'], ind_train_features, knn)
    ood_combined_scores = average_scores(pooled_features_dict['ood_combined'], ind_train_features, knn)
    test_combined_scores = average_scores(pooled_features_dict['test_combined'], ind_train_features, knn)
    oracle_combined_scores = average_scores(pooled_features_dict['oracle_combined'], ind_train_features, knn)
    oracle_combined_test_scores = average_scores(pooled_features_dict['oracle_combined_test'], ind_train_features, knn)

    ind_scores = - 0.7*test_oracles_scores + test_scores + 0.2*(test_combined_scores - oracle_combined_test_scores)
    ood_scores = - 0.7*ood_oracles_scores + ood_scores + 0.2*(ood_combined_scores -  oracle_combined_scores)
    metrics = get_metrics(ind_scores, ood_scores)
    print("KNN+CED score")
    logger.info('metrics: {}'.format(metrics))

if __name__ == '__main__':
    main()