ChapTER/evaluate.py at main · GKNL/ChapTER · 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
import os
import json
import tqdm
import torch

from time import time
from typing import List, Tuple
from dataclasses import dataclass, asdict

from config import args
from input_temporal.doc_temporal import load_data, Example
from predict import BertPredictor
from input_temporal.dict_hub_temporal import get_entity_dict, get_all_filter_quadruple_dict
from input_temporal.quadruple import EntityDict
from rerank import rerank_by_graph
from logger_config import logger


def _setup_entity_dict() -> EntityDict:
    return get_entity_dict()


entity_dict = _setup_entity_dict()
all_filter_quadruple_dict = get_all_filter_quadruple_dict()


@dataclass
class PredInfo:
    head: str
    relation: str
    tail: str
    pred_tail: str
    pred_score: float
    topk_score_info: str
    rank: int
    correct: bool


@torch.no_grad()
def compute_metrics(hr_tensor: torch.tensor,
                    entities_tensor: torch.tensor,
                    target: List[int],
                    examples: List[Example],
                    k=3, batch_size=256) -> Tuple:
    assert hr_tensor.size(1) == entities_tensor.size(1)
    total = hr_tensor.size(0)
    entity_cnt = len(entity_dict)
    assert entity_cnt == entities_tensor.size(0)
    target = torch.LongTensor(target).unsqueeze(-1).to(hr_tensor.device)
    topk_scores, topk_indices = [], []
    ranks = []

    mean_rank, mrr, hit1, hit3, hit10 = 0, 0, 0, 0, 0

    for start in tqdm.tqdm(range(0, total, batch_size)):
        end = start + batch_size
        # batch_size * entity_cnt
        batch_score = torch.mm(hr_tensor[start:end, :], entities_tensor.t())
        assert entity_cnt == batch_score.size(1)
        batch_target = target[start:end]

        # re-ranking based on topological structure
        rerank_by_graph(batch_score, examples[start:end], entity_dict=entity_dict)

        # filter
        # for idx in range(batch_score.size(0)):
        #     mask_indices = []
        #     cur_ex = examples[start + idx]
        #     gold_neighbor_ids = all_filter_quadruple_dict.get_neighbors(cur_ex.head_id, cur_ex.relation, cur_ex.time)  # time filter
        #     if len(gold_neighbor_ids) > 10000:
        #         logger.debug('{} - {} has {} neighbors'.format(cur_ex.head_id, cur_ex.relation, len(gold_neighbor_ids)))
        #     for e_id in gold_neighbor_ids:
        #         if e_id == cur_ex.tail_id:
        #             continue
        #         mask_indices.append(entity_dict.entity_to_idx(e_id))
        #     mask_indices = torch.LongTensor(mask_indices).to(batch_score.device)
        #     batch_score[idx].index_fill_(0, mask_indices, -1)

        batch_sorted_score, batch_sorted_indices = torch.sort(batch_score, dim=-1, descending=True)
        target_rank = torch.nonzero(batch_sorted_indices.eq(batch_target).long(), as_tuple=False)
        assert target_rank.size(0) == batch_score.size(0)
        for idx in range(batch_score.size(0)):
            idx_rank = target_rank[idx].tolist()
            assert idx_rank[0] == idx
            cur_rank = idx_rank[1]

            # 0-based -> 1-based
            cur_rank += 1
            mean_rank += cur_rank
            mrr += 1.0 / cur_rank
            hit1 += 1 if cur_rank <= 1 else 0
            hit3 += 1 if cur_rank <= 3 else 0
            hit10 += 1 if cur_rank <= 10 else 0
            ranks.append(cur_rank)

        topk_scores.extend(batch_sorted_score[:, :k].tolist())
        topk_indices.extend(batch_sorted_indices[:, :k].tolist())

    metrics = {'mean_rank': mean_rank, 'mrr': mrr, 'hit@1': hit1, 'hit@3': hit3, 'hit@10': hit10}
    metrics = {k: round(v / total, 4) for k, v in metrics.items()}
    assert len(topk_scores) == total
    return topk_scores, topk_indices, metrics, ranks


def predict_by_split():
    assert os.path.exists(args.valid_path)
    assert os.path.exists(args.train_path)

    predictor = BertPredictor()
    predictor.load(ckt_path=args.eval_model_path, use_data_parallel=True)
    entity_tensor = predictor.predict_by_entities(entity_dict.entity_exs)

    forward_metrics = eval_single_direction(predictor,
                                            entity_tensor=entity_tensor,
                                            eval_forward=True)
    backward_metrics = eval_single_direction(predictor,
                                             entity_tensor=entity_tensor,
                                             eval_forward=False)
    metrics = {k: round((forward_metrics[k] + backward_metrics[k]) / 2, 4) for k in forward_metrics}
    logger.info('Averaged metrics: {}'.format(metrics))

    prefix, basename = os.path.dirname(args.eval_model_path), os.path.basename(args.eval_model_path)
    split = os.path.basename(args.valid_path)
    with open('{}/metrics_{}_{}.json'.format(prefix, split, basename), 'w', encoding='utf-8') as writer:
        writer.write('forward metrics: {}\n'.format(json.dumps(forward_metrics)))
        writer.write('backward metrics: {}\n'.format(json.dumps(backward_metrics)))
        writer.write('average metrics: {}\n'.format(json.dumps(metrics)))


def eval_single_direction(predictor: BertPredictor,
                          entity_tensor: torch.tensor,
                          eval_forward=True,
                          batch_size=256) -> dict:
    start_time = time()
    examples = load_data(args.valid_path, add_forward_quadruple=eval_forward, add_backward_quadruple=not eval_forward)

    hr_tensor, _ = predictor.predict_by_examples(examples)
    hr_tensor = hr_tensor.to(entity_tensor.device)
    target = [entity_dict.entity_to_idx(ex.tail_id) for ex in examples]
    logger.info('predict tensor done, compute metrics...')

    topk_scores, topk_indices, metrics, ranks = compute_metrics(hr_tensor=hr_tensor, entities_tensor=entity_tensor,
                                                                target=target, examples=examples,
                                                                batch_size=batch_size)
    eval_dir = 'forward' if eval_forward else 'backward'
    logger.info('{} metrics: {}'.format(eval_dir, json.dumps(metrics)))

    pred_infos = []
    for idx, ex in enumerate(examples):
        cur_topk_scores = topk_scores[idx]
        cur_topk_indices = topk_indices[idx]
        pred_idx = cur_topk_indices[0]
        cur_score_info = {entity_dict.get_entity_by_idx(topk_idx).entity: round(topk_score, 3)
                          for topk_score, topk_idx in zip(cur_topk_scores, cur_topk_indices)}

        pred_info = PredInfo(head=ex.head, relation=ex.relation,
                             tail=ex.tail, pred_tail=entity_dict.get_entity_by_idx(pred_idx).entity,
                             pred_score=round(cur_topk_scores[0], 4),
                             topk_score_info=json.dumps(cur_score_info),
                             rank=ranks[idx],
                             correct=pred_idx == target[idx])
        pred_infos.append(pred_info)

    prefix, basename = os.path.dirname(args.eval_model_path), os.path.basename(args.eval_model_path)
    split = os.path.basename(args.valid_path)
    with open('{}/eval_{}_{}_{}.json'.format(prefix, split, eval_dir, basename), 'w', encoding='utf-8') as writer:
        writer.write(json.dumps([asdict(info) for info in pred_infos], ensure_ascii=False, indent=4))

    logger.info('Evaluation takes {} seconds'.format(round(time() - start_time, 3)))
    return metrics


if __name__ == '__main__':
    predict_by_split()