/
inference.py
280 lines (226 loc) · 8.79 KB
/
inference.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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
# %%
import argparse
import torch
import torch.nn as nn
from rouge import Rouge
import numpy as np
from scipy.stats import pearsonr, spearmanr, kendalltau
from transformers import BartTokenizer, BartForConditionalGeneration
from model.config import Encoder_prefix_prompt_length,Encoder_inter_prompt_length,Decoder_prefix_prompt_length,target_data_set,invalid_sample_id
from utils.wn import predict_with_sync_baseline
from utils.process import *
from utils.dataload import load
from utils.align import convert_bpe2_tokens_space
from utils.align import align_values_frombpe2tokens
from utils.entity_loss import pass_loss_Entity
from utils.wordnet import return_synonyms
from utils.overlap_predictor import NotInSource
from utils.overlap_predictor import returnNone_baseline
from utils.overlap_predictor import synonyms_notin
from utils.process import re_upper
from utils.process import filter_none
from transformers import BartTokenizer
def is_skip(input_text,tokenizer):
encoded_tgt = tokenizer(
[input_text],
max_length=1024,
truncation=True,
padding=True,
return_tensors='pt'
)
tgt_tokens = encoded_tgt['input_ids']
token_list = []
for j in tgt_tokens[0]:
token_list.append(tokenizer._convert_id_to_token(j.cpu().numpy().tolist()))
filtered_token_list = []
for i in token_list:
filtered_token_list.append(tokenizer.convert_tokens_to_string([i]).strip())
filtered_token_list,token2bpe,bpe2tokens = convert_bpe2_tokens_space([input_text],filtered_token_list)
if token2bpe is None:
return True
else:
return False
def create_toy_data():
f = open("/home/shesj/workspace/Data/Data/Sort_XSUM/TConvS2S/TConvS2S.source",'r')
source = f.readlines()
f = open("/home/shesj/workspace/Data/Data/Sort_XSUM/TConvS2S/TConvS2S.target",'r')
target = f.readlines()
max_sample = 10
source = [i.strip() for i in source]
target = [i.strip() for i in target]
f = open("data/toy.json",'w')
data = []
for i,j in zip(source,target):
data.append({'doc':i,'sum':j})
data = data[:max_sample]
import json
json.dump(data,f,indent=2)
sample_infor_dump = []
def load_model(args):
optional_mode = ['zero-shot','full-shot']
if args.mode not in optional_mode:
print("hit unknow mode not in ",optional_mode)
exit()
if args.mode == "zero-shot":
from model.BaselineBARTScorer import BARTScorer
model = BARTScorer(device='cuda:0', checkpoint=args.model_path)
else:
from model.PromptBARTScore import BARTScorer
model = BARTScorer(checkpoint=args.model_path,PromptBART = True)
return model
count = 0
def predict_with_known_num(filtered_token_list,pre_score,after_score,summary):
filtered_token_list,token2bpe,bpe2tokens = convert_bpe2_tokens_space([summary],filtered_token_list)
pre_score[0] = align_values_frombpe2tokens(pre_score[0],token2bpe,bpe2tokens)
pre_score= pre_score[0][1:-1]
pre_score,temp_count = pass_loss_Entity(pre_score,filtered_token_list)
after_score[0] = align_values_frombpe2tokens(after_score[0],token2bpe,bpe2tokens)
after_score= after_score[0][1:-1]
after_score,temp_count = pass_loss_Entity(after_score,filtered_token_list)
diff_score = [j-i for i,j in zip(pre_score,after_score)]
#diff_score,temp_count = pass_loss_Entity(after_score,filtered_token_list)
return diff_score
def zero_shot_predictor(d,s,high_bart_scorer):
filtered_token_list,pre_score,_ = high_bart_scorer.score([d],[s])
filtered_token_list,after_score,_ = high_bart_scorer.score([s + " " + d],[s])
predict_score = predict_with_known_num(filtered_token_list,pre_score,after_score,s)
return predict_score
def few_shot_predictor(d,s,high_bart_scorer):
filtered_token_list,pre_score,_ = high_bart_scorer.score([d],[s])
filtered_token_list,after_score,_ = high_bart_scorer.score([s + " " + d],[s],inserted = s)
predict_score = predict_with_known_num(filtered_token_list,pre_score,after_score,s)
return predict_score
def zero_shot_SummaryPredictor(d,s,high_bart_scorer):
pre_score = high_bart_scorer.score([d],[s],summary_level=True)
after_score = high_bart_scorer.score([s + " " + d],[s],summary_level=True)
assert len(pre_score) == 1
dif_scores = pre_score[0] - after_score[0]
return dif_scores
def predict(args,high_bart_scorer):
tokenizer = BartTokenizer.from_pretrained(args.model_path)
documents = []
summarys = []
result_list = []
import json
f = open(args.data_path,'r')
data = json.load(f)
for i in data:
documents.append(i['doc'])
summarys.append(i['sum'])
sample_id = -1
corpus_predict = []
total_pretent_predict_sample = 0
from tqdm import tqdm
valid_sample = 0
total_not_in_source_predict = []
total_prob_score = []
print(len(documents))
dataset_level_information = []
for i in tqdm(range(len(documents))):
sample_id += 1
document = documents[i]
summary = summarys[i]
if is_skip(summary,tokenizer):
print("HIT SKIP")
continue
pre_sum = summary
if args.Recapital:
summary = summary.capitalize()
summary = re_upper(document,summary)
valid_sample += 1
d = document
s = summary
predict_label1 = NotInSource(d,s)
total_not_in_source_predict = total_not_in_source_predict + predict_label1
sample_info = {}
if args.mode == "zero-shot":
predict_score = zero_shot_predictor(d,s,high_bart_scorer)
summary_level_score = zero_shot_SummaryPredictor(d,s,high_bart_scorer)
if args.mode == 'full-shot':
predict_score = few_shot_predictor(d,s,high_bart_scorer)
total_prob_score += predict_score
sample_info['predict_score'] = predict_score
#sample_info['not_in_score'] = predict_label1
sample_info['document'] = d
sample_info['summary'] = s
sample_info['summary_score'] = summary_level_score
assert len(sample_info['predict_score']) == len(sample_info['summary'].split(" "))
dataset_level_information.append(sample_info)
for no,pro in zip(total_not_in_source_predict,total_prob_score):
sample_info = {}
sample_info['not_in'] = no
sample_info['prob_s'] = pro
corpus_predict.append(sample_info)
not_in_predict = sum(total_not_in_source_predict)
expect_predict_num = int(len(total_not_in_source_predict) * args.predict_raio)
total_pretent_predict_sample += expect_predict_num
idx2score = {}
print(len(total_prob_score))
for i in range(len(total_prob_score)):
idx2score[i] = total_prob_score[i]
idx2score = sorted(idx2score.items(), key = lambda kv:(kv[1], kv[0]),reverse=False)
index = 0
for i in idx2score:
id = i[0]
if total_not_in_source_predict[id] == 0:
total_not_in_source_predict[id] = 1
expect_predict_num -= 1
if expect_predict_num == 0:
break
iter_index = 0
for sample in dataset_level_information:
sample['predicted_label'] = total_not_in_source_predict[iter_index:iter_index+len(sample['predict_score'])]
iter_index += len(sample['predict_score'])
for sample in dataset_level_information:
#print("hit")
tokens = sample['summary'].split(' ')
predic = sample['predicted_label']
assert len(tokens) == len(predic)
tokens_labeled = [i + "[" + str(j) + ']' for i,j in zip(tokens,predic)]
sample['Pre_lab_summary'] = " ".join(tokens_labeled)
del sample['predicted_label']
del sample['predict_score']
f = open(args.output_file_path,'w')
import json
json.dump(dataset_level_information,f,indent=2)
def main(args):
model = load_model(args)
predict(args,model)
if __name__ == "__main__":
create_toy_data()
parser = argparse.ArgumentParser()
parser.add_argument(
"--data_path",
type=str,
default=None,
required=True
)
parser.add_argument(
"--mode",
type=str,
default=None,
required=True
)
parser.add_argument(
"--model_path",
type=str,
default=None,
required=True
)
parser.add_argument(
"--output_file_path",
type=str,
default=None,
required=True
)
parser.add_argument(
"--predict_raio",
type=float,
default=None,
required=True
)
parser.add_argument("--Recapital",
action="store_true",
help="When Summary is lowercase, we Recapital it")
args = parser.parse_args()
main(args)