Liyan06
commited on
Commit
·
3201a95
1
Parent(s):
ca16988
add customized handler
Browse files- handler.py +13 -0
- minicheck/inference.py +210 -0
- minicheck/minicheck.py +51 -0
handler.py
ADDED
@@ -0,0 +1,13 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from minicheck.minicheck import MiniCheck
|
2 |
+
|
3 |
+
class EndpointHandler():
|
4 |
+
def __init__(self, path="./"):
|
5 |
+
self.scorer = MiniCheck(path=path)
|
6 |
+
|
7 |
+
def __call__(self, data):
|
8 |
+
docs = data.pop("docs",data)
|
9 |
+
claims = data.pop("claims", None)
|
10 |
+
|
11 |
+
_, raw_prob, _, _ = self.scorer.score(docs=docs, claims=claims)
|
12 |
+
|
13 |
+
return raw_prob
|
minicheck/inference.py
ADDED
@@ -0,0 +1,210 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Adapt code from https://github.com/yuh-zha/AlignScore/tree/main
|
2 |
+
|
3 |
+
import sys
|
4 |
+
sys.path.append("..")
|
5 |
+
|
6 |
+
from nltk.tokenize import sent_tokenize
|
7 |
+
import torch
|
8 |
+
from transformers import AutoConfig, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
|
9 |
+
import torch.nn as nn
|
10 |
+
from tqdm import tqdm
|
11 |
+
import torch.nn.functional as F
|
12 |
+
import os
|
13 |
+
|
14 |
+
|
15 |
+
def sent_tokenize_with_newlines(text):
|
16 |
+
blocks = text.split('\n')
|
17 |
+
|
18 |
+
tokenized_blocks = [sent_tokenize(block) for block in blocks]
|
19 |
+
tokenized_text = []
|
20 |
+
for block in tokenized_blocks:
|
21 |
+
tokenized_text.extend(block)
|
22 |
+
tokenized_text.append('\n')
|
23 |
+
|
24 |
+
return tokenized_text[:-1]
|
25 |
+
|
26 |
+
|
27 |
+
class Inferencer():
|
28 |
+
def __init__(self, path, chunk_size, max_input_length, batch_size) -> None:
|
29 |
+
|
30 |
+
self.path = path
|
31 |
+
self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
32 |
+
|
33 |
+
self.model = AutoModelForSeq2SeqLM.from_pretrained(path).to(self.device)
|
34 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
35 |
+
|
36 |
+
self.chunk_size=500 if chunk_size is None else chunk_size
|
37 |
+
self.max_input_length=2048 if max_input_length is None else max_input_length
|
38 |
+
self.max_output_length = 256
|
39 |
+
|
40 |
+
self.model.eval()
|
41 |
+
self.batch_size = batch_size
|
42 |
+
self.softmax = nn.Softmax(dim=-1)
|
43 |
+
|
44 |
+
def inference_example_batch(self, doc: list, claim: list):
|
45 |
+
"""
|
46 |
+
inference a example,
|
47 |
+
doc: list
|
48 |
+
claim: list
|
49 |
+
using self.inference to batch the process
|
50 |
+
"""
|
51 |
+
|
52 |
+
assert len(doc) == len(claim), "doc must has the same length with claimthesis!"
|
53 |
+
|
54 |
+
max_support_probs = []
|
55 |
+
used_chunks = []
|
56 |
+
support_prob_per_chunk = []
|
57 |
+
|
58 |
+
for one_doc, one_claim in tqdm(zip(doc, claim), desc="Evaluating", total=len(doc)):
|
59 |
+
output = self.inference_per_example(one_doc, one_claim)
|
60 |
+
max_support_probs.append(output['max_support_prob'])
|
61 |
+
used_chunks.append(output['used_chunks'])
|
62 |
+
support_prob_per_chunk.append(output['support_prob_per_chunk'])
|
63 |
+
|
64 |
+
return {
|
65 |
+
'max_support_probs': max_support_probs,
|
66 |
+
'used_chunks': used_chunks,
|
67 |
+
'support_prob_per_chunk': support_prob_per_chunk
|
68 |
+
}
|
69 |
+
|
70 |
+
def inference_per_example(self, doc:str, claim: str):
|
71 |
+
"""
|
72 |
+
inference a example,
|
73 |
+
doc: string
|
74 |
+
claim: string
|
75 |
+
using self.inference to batch the process
|
76 |
+
"""
|
77 |
+
def chunks(lst, n):
|
78 |
+
"""Yield successive chunks from lst with each having approximately n tokens.
|
79 |
+
|
80 |
+
For flan-t5, we split using the white space;
|
81 |
+
For roberta and deberta, we split using the tokenization.
|
82 |
+
"""
|
83 |
+
current_chunk = []
|
84 |
+
current_word_count = 0
|
85 |
+
for sentence in lst:
|
86 |
+
sentence_word_count = len(sentence.split())
|
87 |
+
if current_word_count + sentence_word_count > n:
|
88 |
+
yield ' '.join(current_chunk)
|
89 |
+
current_chunk = [sentence]
|
90 |
+
current_word_count = sentence_word_count
|
91 |
+
else:
|
92 |
+
current_chunk.append(sentence)
|
93 |
+
current_word_count += sentence_word_count
|
94 |
+
if current_chunk:
|
95 |
+
yield ' '.join(current_chunk)
|
96 |
+
|
97 |
+
doc_sents = sent_tokenize_with_newlines(doc)
|
98 |
+
doc_sents = doc_sents or ['']
|
99 |
+
|
100 |
+
doc_chunks = [chunk.replace(" \n ", '\n').strip() for chunk in chunks(doc_sents, self.chunk_size)]
|
101 |
+
|
102 |
+
'''
|
103 |
+
[chunk_1, chunk_2, chunk_3, chunk_4, ...]
|
104 |
+
[claim]
|
105 |
+
'''
|
106 |
+
claim_repeat = [claim] * len(doc_chunks)
|
107 |
+
|
108 |
+
output = self.inference(doc_chunks, claim_repeat)
|
109 |
+
|
110 |
+
return output
|
111 |
+
|
112 |
+
def inference(self, doc, claim):
|
113 |
+
"""
|
114 |
+
inference a list of doc and claim
|
115 |
+
|
116 |
+
Standard aggregation (max) over chunks of doc
|
117 |
+
|
118 |
+
Note: We do not have any post-processing steps for 'claim'
|
119 |
+
and directly check 'doc' against 'claim'. If there are multiple
|
120 |
+
sentences in 'claim'. Sentences are not splitted and are checked
|
121 |
+
as a single piece of text.
|
122 |
+
|
123 |
+
If there are multiple sentences in 'claim', we suggest users to
|
124 |
+
split 'claim' into sentences beforehand and prepares data like
|
125 |
+
(doc, claim_1), (doc, claim_2), ... for a multi-sentence 'claim'.
|
126 |
+
|
127 |
+
**We leave the user to decide how to aggregate the results from multiple sentences.**
|
128 |
+
|
129 |
+
Note: AggreFact-CNN is the only dataset that contains three-sentence
|
130 |
+
summaries and have annotations on the whole summaries, so we do not
|
131 |
+
split the sentences in each 'claim' during prediciotn for simplicity.
|
132 |
+
Therefore, for this dataset, our result is based on treating the whole
|
133 |
+
summary as a single piece of text (one 'claim').
|
134 |
+
|
135 |
+
In general, sentence-level prediciton performance is better than that on
|
136 |
+
the full-response-level.
|
137 |
+
"""
|
138 |
+
|
139 |
+
if isinstance(doc, str) and isinstance(claim, str):
|
140 |
+
doc = [doc]
|
141 |
+
claim = [claim]
|
142 |
+
|
143 |
+
batch_input, _, batch_org_chunks = self.batch_tokenize(doc, claim)
|
144 |
+
|
145 |
+
label_probs_list = []
|
146 |
+
used_chunks = []
|
147 |
+
|
148 |
+
for mini_batch_input, batch_org_chunk in zip(batch_input, batch_org_chunks):
|
149 |
+
|
150 |
+
mini_batch_input = {k: v.to(self.device) for k, v in mini_batch_input.items()}
|
151 |
+
|
152 |
+
with torch.no_grad():
|
153 |
+
|
154 |
+
decoder_input_ids = torch.zeros((mini_batch_input['input_ids'].size(0), 1), dtype=torch.long).to(self.device)
|
155 |
+
outputs = self.model(input_ids=mini_batch_input['input_ids'], attention_mask=mini_batch_input['attention_mask'], decoder_input_ids=decoder_input_ids)
|
156 |
+
logits = outputs.logits.squeeze(1)
|
157 |
+
|
158 |
+
# 3 for no support and 209 for support
|
159 |
+
label_logits = logits[:, torch.tensor([3, 209])].cpu()
|
160 |
+
label_probs = torch.nn.functional.softmax(label_logits, dim=-1)
|
161 |
+
|
162 |
+
label_probs_list.append(label_probs)
|
163 |
+
used_chunks.extend(batch_org_chunk)
|
164 |
+
|
165 |
+
label_probs = torch.cat(label_probs_list)
|
166 |
+
support_prob_per_chunk = label_probs[:, 1].cpu().numpy()
|
167 |
+
max_support_prob = label_probs[:, 1].max().item()
|
168 |
+
|
169 |
+
return {
|
170 |
+
'max_support_prob': max_support_prob,
|
171 |
+
'used_chunks': used_chunks,
|
172 |
+
'support_prob_per_chunk': support_prob_per_chunk
|
173 |
+
}
|
174 |
+
|
175 |
+
def batch_tokenize(self, doc, claim):
|
176 |
+
"""
|
177 |
+
input doc and claims are lists
|
178 |
+
"""
|
179 |
+
assert isinstance(doc, list) and isinstance(claim, list)
|
180 |
+
assert len(doc) == len(claim), "doc and claim should be in the same length."
|
181 |
+
|
182 |
+
original_text = [self.tokenizer.eos_token.join([one_doc, one_claim]) for one_doc, one_claim in zip(doc, claim)]
|
183 |
+
|
184 |
+
batch_input = []
|
185 |
+
batch_concat_text = []
|
186 |
+
batch_org_chunks = []
|
187 |
+
for mini_batch in self.chunks(original_text, self.batch_size):
|
188 |
+
model_inputs = self.tokenizer(
|
189 |
+
['predict: ' + text for text in mini_batch],
|
190 |
+
max_length=self.max_input_length,
|
191 |
+
truncation=True,
|
192 |
+
padding=True,
|
193 |
+
return_tensors="pt"
|
194 |
+
)
|
195 |
+
|
196 |
+
batch_input.append(model_inputs)
|
197 |
+
batch_concat_text.append(mini_batch)
|
198 |
+
batch_org_chunks.append([item[:item.find('</s>')] for item in mini_batch])
|
199 |
+
|
200 |
+
return batch_input, batch_concat_text, batch_org_chunks
|
201 |
+
|
202 |
+
def chunks(self, lst, n):
|
203 |
+
"""Yield successive n-sized chunks from lst."""
|
204 |
+
for i in range(0, len(lst), n):
|
205 |
+
yield lst[i:i + n]
|
206 |
+
|
207 |
+
def fact_check(self, doc, claim):
|
208 |
+
|
209 |
+
outputs = self.inference_example_batch(doc, claim)
|
210 |
+
return outputs['max_support_probs'], outputs['used_chunks'], outputs['support_prob_per_chunk']
|
minicheck/minicheck.py
ADDED
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Adapt code from https://github.com/yuh-zha/AlignScore/tree/main
|
2 |
+
|
3 |
+
import sys
|
4 |
+
sys.path.append("..")
|
5 |
+
|
6 |
+
from minicheck.inference import Inferencer
|
7 |
+
from typing import List
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
|
11 |
+
class MiniCheck:
|
12 |
+
def __init__(self, path, chunk_size=None, max_input_length=None, batch_size=16) -> None:
|
13 |
+
|
14 |
+
self.model = Inferencer(
|
15 |
+
path=path,
|
16 |
+
batch_size=batch_size,
|
17 |
+
chunk_size=chunk_size,
|
18 |
+
max_input_length=max_input_length,
|
19 |
+
)
|
20 |
+
|
21 |
+
def score(self, docs: List[str], claims: List[str]) -> List[float]:
|
22 |
+
'''
|
23 |
+
pred_labels: 0 / 1 (0: unsupported, 1: supported)
|
24 |
+
max_support_probs: the probability of "supported" for the chunk that determin the final pred_label
|
25 |
+
used_chunks: divided chunks of the input document
|
26 |
+
support_prob_per_chunk: the probability of "supported" for each chunk
|
27 |
+
'''
|
28 |
+
|
29 |
+
assert isinstance(docs, list) or isinstance(docs, np.ndarray), "docs must be a list or np.ndarray"
|
30 |
+
assert isinstance(claims, list) or isinstance(claims, np.ndarray), "claims must be a list or np.ndarray"
|
31 |
+
|
32 |
+
max_support_prob, used_chunk, support_prob_per_chunk = self.model.fact_check(docs, claims)
|
33 |
+
pred_label = [1 if prob > 0.5 else 0 for prob in max_support_prob]
|
34 |
+
|
35 |
+
return pred_label, max_support_prob, used_chunk, support_prob_per_chunk
|
36 |
+
|
37 |
+
|
38 |
+
if __name__ == '__main__':
|
39 |
+
|
40 |
+
path = "./"
|
41 |
+
|
42 |
+
doc = "A group of students gather in the school library to study for their upcoming final exams."
|
43 |
+
claim_1 = "The students are preparing for an examination."
|
44 |
+
claim_2 = "The students are on vacation."
|
45 |
+
|
46 |
+
# flan-t5-large
|
47 |
+
scorer = MiniCheck(path)
|
48 |
+
pred_label, raw_prob, _, _ = scorer.score(docs=[doc, doc], claims=[claim_1, claim_2])
|
49 |
+
|
50 |
+
print(pred_label) # [1, 0]
|
51 |
+
print(raw_prob) # [0.9805923700332642, 0.007121307775378227]
|