-
Notifications
You must be signed in to change notification settings - Fork 22
/
predictor.py
173 lines (150 loc) · 6.84 KB
/
predictor.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
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
import torch
from fastNLP import cache_results
from data.pipe import BartNERPipe
from fastNLP import SequentialSampler, SortedSampler
from fastNLP import DataSetIter
from fastNLP.core.utils import _move_dict_value_to_device
from tqdm import tqdm
import json
dataset_name = 'conll2003'
model_path = 'save_models/best_SequenceGeneratorModel_f_2021-06-09-01-47-26-903275' # you can set args.save_model=1 in train.py
bart_name = 'facebook/bart-large'
target_type = 'word'
device = 'cuda:0'
cache_fn = f"caches/data_{bart_name}_{dataset_name}_{target_type}.pt"
@cache_results(cache_fn, _refresh=False)
def get_data():
pipe = BartNERPipe(tokenizer=bart_name, dataset_name=dataset_name, target_type=target_type)
if dataset_name == 'conll2003':
paths = {'test': "../data/conll2003/test.txt",
'train': "../data/conll2003/train.txt",
'dev': "../data/conll2003/dev.txt"}
data_bundle = pipe.process_from_file(paths, demo=False)
elif dataset_name == 'en-ontonotes':
paths = '../data/en-ontonotes/english'
data_bundle = pipe.process_from_file(paths)
else:
data_bundle = pipe.process_from_file(f'../data/{dataset_name}', demo=False)
return data_bundle, pipe.tokenizer, pipe.mapping2id
data_bundle, tokenizer, mapping2id = get_data()
model = torch.load(model_path)
device = torch.device(device)
model.to(device)
model.eval()
eos_token_id = 0
word_start_index = len(mapping2id) + 2
not_bpe_start = 0
if dataset_name == 'conll2003': # if you use other dataset, please change this mapping
mapping = {
'<<location>>': 'LOC',
'<<person>>': 'PER',
'<<organization>>': 'ORG',
'<<others>>': 'MISC'
}
elif dataset_name == 'en_ace04':
mapping = {v: k for k, v in {
'loc': '<<location>>', "gpe": "<<government>>", "wea": "<<weapon>>", 'veh': "<<vehicle>>",
'per': '<<person>>',
'org': '<<organization>>',
'fac': '<<buildings>>',
}.items()}
id2label = {k: mapping[v] for k, v in enumerate(mapping2id.keys())}
def get_pairs(ps, word_start_index, target_type):
pairs = []
cur_pair = []
for j in ps:
if j < word_start_index:
if target_type == 'span':
if len(cur_pair) > 0 and len(cur_pair) % 2 == 0:
if all([cur_pair[i] <= cur_pair[i + 1] for i in range(len(cur_pair) - 1)]):
pairs.append(tuple(cur_pair + [j]))
else:
if len(cur_pair) > 0:
if all([cur_pair[i] < cur_pair[i + 1] for i in range(len(cur_pair) - 1)]):
pairs.append(tuple(cur_pair + [j]))
cur_pair = []
else:
cur_pair.append(j)
return pairs
def get_spans(pairs, cum_lens, mapping2id, dataset_name, id2label):
spans = []
pred_y = ['O' for _ in range(len(raw_words_i))]
for pair in pairs:
label = pair[-1]
try:
idxes = [cum_lens.index(p - len(mapping2id) - 2) for p in pair[:-1]]
start_idx = idxes[0]
end_idx = idxes[-1]
if dataset_name in ('en_ace04', 'en_ace05'):
spans.append((start_idx, end_idx, id2label[label - 2]))
else:
pred_y[start_idx] = f'B-{id2label[label - 2]}'
for _ in range(start_idx + 1, end_idx + 1):
pred_y[_] = f'I-{id2label[label - 2]}'
except Exception as e:
pass
return pairs, pred_y
for name in ['test']:
ds = data_bundle.get_dataset(name)
ds.set_ignore_type('raw_words', 'raw_target')
ds.set_target('raw_words', 'raw_target')
with open(f'preds/{name}.conll', 'w', encoding='utf-8') as f:
data_iterator = DataSetIter(ds, batch_size=32, sampler=SequentialSampler())
for batch_x, batch_y in tqdm(data_iterator, total=len(data_iterator)):
_move_dict_value_to_device(batch_x, batch_y, device=device)
src_tokens = batch_x['src_tokens']
first = batch_x['first']
src_seq_len = batch_x['src_seq_len']
tgt_seq_len = batch_x['tgt_seq_len']
raw_words = batch_y['raw_words']
raw_targets = batch_y['raw_target']
pred_y = model.predict(src_tokens=src_tokens, src_seq_len=src_seq_len, first=first)
pred = pred_y['pred']
tgt_tokens = batch_y['tgt_tokens']
pred_eos_index = pred.flip(dims=[1]).eq(eos_token_id).cumsum(dim=1).long()
pred = pred[:, 1:] # 去掉</s>
tgt_tokens = tgt_tokens[:, 1:]
pred_seq_len = pred_eos_index.flip(dims=[1]).eq(pred_eos_index[:, -1:]).sum(dim=1) # bsz
pred_seq_len = (pred_seq_len - 2).tolist()
tgt_seq_len = (tgt_seq_len - 2).tolist()
for i, ps in enumerate(pred.tolist()):
em = 0
ps = ps[:pred_seq_len[i]]
ts = tgt_tokens[i, :tgt_seq_len[i]]
pairs, t_pairs = [], []
if len(ps):
pairs = get_pairs(ps, word_start_index, target_type)
if len(ts):
t_pairs = get_pairs(ts, word_start_index, target_type)
raw_words_i = raw_words[i]
src_tokens_i = src_tokens[i, :src_seq_len[i]].tolist()
src_tokens_i = tokenizer.convert_ids_to_tokens(src_tokens_i)
cum_lens = [1]
start_idx = 1
for token in raw_words_i:
start_idx += len(tokenizer.tokenize(token, add_prefix_space=True))
cum_lens.append(start_idx)
cum_lens.append(start_idx + 1)
target_y = raw_targets[i]
pred_spans, pred_y = get_spans(pairs, cum_lens, mapping2id, dataset_name, id2label)
target_spans, _ = get_spans(t_pairs, cum_lens, mapping2id, dataset_name, id2label)
if dataset_name in ('en_ace04', 'en_ace05'):
f.write(json.dumps({'sentence': ' '.join(raw_words_i), 'spans': target_spans, 'pred_spans': pred_spans}))
else:
assert len(pred_y) == len(raw_words_i) == len(target_y)
for raw_word, t, p in zip(raw_words_i, target_y, pred_y):
f.write(f'{raw_word} {t} {p}\n')
f.write('\n')
print(f"In total, has {not_bpe_start} predictions on the non-word start.")
# the output file for flat NER will be similar to the following(empty line separate two sentences)
# SOCCER O O
# - O O
# JAPAN B-LOC B-LOC
# GET O O
# LUCKY O O
# WIN O O
# , O O
# CHINA B-PER B-LOC
# the output file for nested NER will be similar to the following(empty line separate two sentences)
# {'sentence': "xxx xxx", 'pred_spans': [(start, end, label), (start, end, label)], 'spans': [(start, end, label)...]}
# {'sentence': "xxx xxx", 'pred_spans': [(start, end, label), (start, end, label)], 'spans': [(start, end, label)...]}