forked from studio-ousia/luke
-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluate.py
152 lines (139 loc) · 6.3 KB
/
evaluate.py
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
import click
import logging
import os
from typing import List, Optional
import torch
from tqdm import tqdm
from transformers import AutoTokenizer
from luke.utils.entity_vocab import EntityVocab, MASK_TOKEN, PAD_TOKEN
from dataloader import create_dataloader
from dataset import load_dataset
from model import LukeForEntityDisambiguation
logger = logging.getLogger(__name__)
@click.command()
@click.option("--model-dir", type=click.Path(exists=True), required=True)
@click.option("--dataset-dir", type=click.Path(exists=True), required=True)
@click.option("--titles-file", type=click.Path(exists=True), required=True)
@click.option("--redirects-file", type=click.Path(exists=True), required=True)
@click.option("--ppr-for-ned-dir", type=click.Path(exists=True))
@click.option(
"-t",
"--test-set",
default=["test_b", "ace2004", "aquaint", "msnbc", "wikipedia", "clueweb"],
type=click.Choice(["test_b", "test_b_ppr", "ace2004", "aquaint", "msnbc", "wikipedia", "clueweb"]),
multiple=True,
)
@click.option("--device", type=str, default="cuda")
@click.option("--max-seq-length", type=int, default=512)
@click.option("--max-entity-length", type=int, default=128)
@click.option("--max-candidate-length", type=int, default=30)
@click.option("--max-mention-length", type=int, default=30)
@click.option(
"--inference-mode", type=click.Choice(["global", "local"]), default="global",
)
@click.option(
"--document-split-mode", type=click.Choice(["simple", "per_mention"]), default="simple",
)
def evaluate(
model_dir: str,
dataset_dir: str,
titles_file: str,
redirects_file: str,
ppr_for_ned_dir: Optional[str],
test_set: List[str],
device: str,
max_seq_length: int,
max_entity_length: int,
max_candidate_length: int,
max_mention_length: int,
inference_mode: str,
document_split_mode: str,
):
model = LukeForEntityDisambiguation.from_pretrained(model_dir).eval()
model.to(device)
tokenizer = AutoTokenizer.from_pretrained(model_dir)
entity_vocab_path = os.path.join(model_dir, "entity_vocab.jsonl")
entity_vocab = EntityVocab(entity_vocab_path)
pad_entity_id = entity_vocab[PAD_TOKEN]
mask_entity_id = entity_vocab[MASK_TOKEN]
dataset = load_dataset(
dataset_dir=dataset_dir,
titles_file=titles_file,
redirects_file=redirects_file,
ppr_for_ned_dir=ppr_for_ned_dir,
)
for dataset_name in test_set:
print(f"========== Dataset: {dataset_name} ==========")
documents = dataset.get_dataset(dataset_name)
dataloader = create_dataloader(
documents=documents,
tokenizer=tokenizer,
entity_vocab=entity_vocab,
batch_size=1,
fold="eval",
document_split_mode=document_split_mode,
max_seq_length=max_seq_length,
max_entity_length=max_entity_length,
max_candidate_length=max_candidate_length,
max_mention_length=max_mention_length,
)
candidate_indices_list = []
eval_entity_mask_list = []
for input_dict in tqdm(dataloader, leave=False):
inputs = {k: v.to(device) for k, v in input_dict.items()}
entity_ids = inputs.pop("entity_ids")
entity_length = inputs["entity_attention_mask"].sum()
input_entity_ids = entity_ids.new_full(entity_ids.size(), pad_entity_id)
input_entity_ids[0, :entity_length] = mask_entity_id
eval_entity_mask = inputs.pop("eval_entity_mask")
eval_entity_mask_list.append(eval_entity_mask[0, :entity_length])
with torch.no_grad():
candidate_indices = torch.zeros(entity_length, dtype=torch.long, device=device)
if inference_mode == "local":
logits = model(entity_ids=input_entity_ids, **inputs)[0]
for n, entity_id in enumerate(torch.argmax(logits, dim=2)[0, :entity_length]):
if inputs["entity_candidate_ids"][0, n].sum() != 0:
candidate_indices[n] = (inputs["entity_candidate_ids"][0, n] == entity_id).nonzero(
as_tuple=True
)[0][0]
else:
for _ in range(entity_length):
logits = model(entity_ids=input_entity_ids, **inputs)[0]
probs = torch.nn.functional.softmax(logits, dim=2) * (
input_entity_ids == mask_entity_id
).unsqueeze(-1).type_as(logits)
max_probs, max_indices = torch.max(probs.squeeze(0), dim=1)
target_index = torch.argmax(max_probs, dim=0)
input_entity_ids[0, target_index] = max_indices[target_index]
if inputs["entity_candidate_ids"][0, target_index].sum() != 0:
candidate_indices[target_index] = (
inputs["entity_candidate_ids"][0, target_index] == max_indices[target_index]
).nonzero(as_tuple=True)[0][0]
candidate_indices_list.append(candidate_indices)
all_candidate_indices = torch.cat(candidate_indices_list)
all_eval_entity_mask = torch.cat(eval_entity_mask_list)
last_index = -1
num_correct = 0
num_mentions = 0
num_mentions_with_candidates = 0
for document in documents:
for mention in document.mentions:
num_mentions += 1
index = last_index + 1
while True:
if all_eval_entity_mask[index] == 1:
break
index += 1
last_index = index
if mention.candidates:
num_mentions_with_candidates += 1
predicted_candidate_index = all_candidate_indices[index]
predicted_title = mention.candidates[predicted_candidate_index].title
if predicted_title == mention.title:
num_correct += 1
precision = num_correct / num_mentions_with_candidates
recall = num_correct / num_mentions
f1 = 2.0 * precision * recall / (precision + recall)
print(f"F1: {f1:.3f} Precision: {precision:.3f} Recall: {recall:.3f}")
if __name__ == "__main__":
evaluate()