-
Notifications
You must be signed in to change notification settings - Fork 28
/
npm.py
165 lines (138 loc) · 6.94 KB
/
npm.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
import numpy as np
import torch
import time
from collections import defaultdict
from transformers import AutoModelForMaskedLM, AutoTokenizer
from factscore.lm import LM
from factscore.retrieval import Retrieval
def softmax(x):
return(np.exp(x - np.max(x)) / np.exp(x - np.max(x)).sum())
class NPM(LM):
def __init__(self, bm25, model_name, cache_file):
assert model_name.startswith("npm")
self.bm25 = bm25
self.model_name = model_name
self.model = None
self.tokenizer = AutoTokenizer.from_pretrained("facebook/" + self.model_name)
self.mask_id = self.tokenizer.mask_token_id
with open("roberta_stopwords.txt", "r") as f:
self.stopwords = set()
for line in f:
self.stopwords.add(int(line.strip()))
super().__init__(cache_file=cache_file)
def load_model(self):
self.model = AutoModelForMaskedLM.from_pretrained("facebook/" + self.model_name)
self.model.cuda()
self.model.eval()
def save_cache(self):
super().save_cache()
self.bm25.save_cache()
def tokenize(self, texts, skip_special_tokens=False, padding=True):
assert type(texts)==list
all_input_ids = self.tokenizer(texts)["input_ids"]
if skip_special_tokens:
for i, input_ids in enumerate(all_input_ids):
assert input_ids[0]==0 and input_ids[-1]==2
all_input_ids[i] = input_ids[1:-1]
if not padding:
return all_input_ids
max_length = np.max([len(_ids) for _ids in all_input_ids])
_all_input_ids = []
_all_attention_mask = []
for i, input_ids in enumerate(all_input_ids):
n_valid = len(input_ids)
n_masks = max_length - n_valid
_all_input_ids.append(input_ids + [0 for _ in range(n_masks)])
_all_attention_mask.append([1 for _ in range(n_valid)] + [0 for _ in range(n_masks)])
return torch.LongTensor(_all_input_ids), torch.LongTensor(_all_attention_mask)
def decode(self, input_ids):
return self.tokenizer.decode(input_ids)
def encode(self, texts, skip_special_tokens=False, gt_input_ids=None):
assert type(texts)==list
if self.model is None:
self.load_model()
if gt_input_ids is not None:
assert len(texts)==len(gt_input_ids)
all_input_ids, all_attention_mask = self.tokenize(texts, skip_special_tokens=skip_special_tokens)
with torch.no_grad():
outputs = self.model(all_input_ids.cuda(),
all_attention_mask.cuda(),
output_hidden_states=True,
return_dict=True)
all_logits = outputs["logits"].detach().cpu().numpy()
all_hidden_states = outputs["hidden_states"][-1].detach().cpu().numpy()
results = []
for i, (text, input_ids, logits, hidden_states) in enumerate(zip(texts, all_input_ids, all_logits, all_hidden_states)):
input_ids = input_ids.numpy().tolist()
if self.mask_id in input_ids:
idx = input_ids.index(self.mask_id)
assert gt_input_ids is not None
prob = softmax(logits[idx])[gt_input_ids[i]]
results.append((prob, hidden_states[idx]))
else:
_input_ids = [_id for _id in input_ids if _id not in [0, 2]]
_hidden_states = [h for _id, h in zip(input_ids, hidden_states) if _id not in [0, 2]]
results.append((_input_ids, _hidden_states))
return results
def get_probabilty(self, topic, question):
passages = self.bm25.get_passages(topic, question, k=3)
passages = [p["text"].strip() for p in passages]
cache_key = question + "#" + "#".join(passages)
if cache_key not in self.cache_dict:
encoded = self.encode(passages, skip_special_tokens=True)
stacked_passage_tokens, stacked_passage_vectors = [], []
for input_ids, vectors in encoded:
stacked_passage_tokens += input_ids
if len(vectors)>0:
stacked_passage_vectors.append(vectors)
stacked_passage_vectors = np.concatenate(stacked_passage_vectors, 0)
question_input_ids = self.tokenize(["Fact: " + question], skip_special_tokens=False, padding=False)[0]
if 2 in question_input_ids:
question_input_ids = question_input_ids[:question_input_ids.index(2)]
question_input_ids = question_input_ids[1:]
'''
triples = []
prefix = True
for i, input_id in enumerate(question_input_ids):
if prefix:
if input_id==35: # the end of prefix
prefix = False
continue
if input_id in [0, 2] or input_id in self.stopwords:
continue
new_question = self.decode(question_input_ids[:i] + [self.mask_id] + question_input_ids[i+1:])
prob, vector = self.encode(new_question, gt_input_id=input_id)
triples.append((prob, vector, input_id))
'''
triples = []
batch = []
gt_input_ids = []
prefix = True
for i, input_id in enumerate(question_input_ids):
if prefix:
if input_id==35: # the end of prefix
prefix = False
continue
if input_id in [0, 2] or input_id in self.stopwords:
continue
batch.append(self.decode(question_input_ids[:i] + [self.mask_id] + question_input_ids[i+1:]))
gt_input_ids.append(input_id)
for (prob, vector), gt_input_id in zip(self.encode(batch, gt_input_ids=gt_input_ids), gt_input_ids):
triples.append((prob, vector, gt_input_id))
stacked_question_vectors = np.stack([v for _, v, _ in triples], 0)
all_scores = np.exp(np.inner(stacked_question_vectors, stacked_passage_vectors) / np.sqrt(stacked_passage_vectors.shape[-1]))
probs = []
for (softmax_prob, vector, input_id), scores in zip(triples, all_scores):
assert len(stacked_passage_tokens)==len(scores)
if input_id not in stacked_passage_tokens:
probs.append(0)
else:
aggregated_scores = defaultdict(list)
for token, score in zip(stacked_passage_tokens, scores):
aggregated_scores[token].append(score)
tot = np.sum([np.sum(v) for v in aggregated_scores.values()])
prob = np.sum(aggregated_scores[input_id]) / tot
probs.append(prob)
self.cache_dict[cache_key] = np.mean(probs)
self.add_n += 1
return self.cache_dict[cache_key]