Update models/qa_model.py
Browse files- models/qa_model.py +55 -19
models/qa_model.py
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
import numpy as np
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
-
from transformers import AutoModelForQuestionAnswering, pipeline
|
5 |
from features.text_utils import post_process_answer
|
6 |
from features.graph_utils import find_best_cluster
|
7 |
from optimum.onnxruntime import ORTModelForQuestionAnswering
|
@@ -11,13 +11,18 @@ class QAEnsembleModel(nn.Module):
|
|
11 |
def __init__(self, model_name, model_checkpoints, entity_dict,
|
12 |
thr=0.1, device="cpu"):
|
13 |
super(QAEnsembleModel, self).__init__()
|
14 |
-
self.nlps = []
|
|
|
|
|
15 |
for model_checkpoint in model_checkpoints:
|
16 |
model = ORTModelForQuestionAnswering.from_pretrained(model_name, from_transformers=True)#.half()
|
17 |
model.load_state_dict(torch.load(model_checkpoint, map_location=torch.device('cpu')), strict=False)
|
18 |
-
nlp = pipeline('question-answering', model=model,
|
19 |
-
|
20 |
-
self.nlps.append(nlp)
|
|
|
|
|
|
|
21 |
self.entity_dict = entity_dict
|
22 |
self.thr = thr
|
23 |
|
@@ -28,22 +33,53 @@ class QAEnsembleModel(nn.Module):
|
|
28 |
curr_answers = []
|
29 |
curr_scores = []
|
30 |
best_score = 0
|
31 |
-
for i, nlp in enumerate(self.nlps):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
for text, score in zip(texts, ranking_scores):
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
if i == 0:
|
44 |
-
if
|
45 |
-
answer =
|
46 |
-
best_score =
|
47 |
if len(curr_answers) == 0:
|
48 |
return None
|
49 |
curr_answers = [post_process_answer(x, self.entity_dict) for x in curr_answers]
|
|
|
1 |
import numpy as np
|
2 |
import torch
|
3 |
import torch.nn as nn
|
4 |
+
# from transformers import AutoModelForQuestionAnswering, pipeline
|
5 |
from features.text_utils import post_process_answer
|
6 |
from features.graph_utils import find_best_cluster
|
7 |
from optimum.onnxruntime import ORTModelForQuestionAnswering
|
|
|
11 |
def __init__(self, model_name, model_checkpoints, entity_dict,
|
12 |
thr=0.1, device="cpu"):
|
13 |
super(QAEnsembleModel, self).__init__()
|
14 |
+
# self.nlps = []
|
15 |
+
self.models = []
|
16 |
+
self.tokenizers = []
|
17 |
for model_checkpoint in model_checkpoints:
|
18 |
model = ORTModelForQuestionAnswering.from_pretrained(model_name, from_transformers=True)#.half()
|
19 |
model.load_state_dict(torch.load(model_checkpoint, map_location=torch.device('cpu')), strict=False)
|
20 |
+
# nlp = pipeline('question-answering', model=model,
|
21 |
+
# tokenizer=model_name, device=device)
|
22 |
+
# self.nlps.append(nlp)
|
23 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
24 |
+
self.models.append(model)
|
25 |
+
self.tokenizers.append(tokenizer)
|
26 |
self.entity_dict = entity_dict
|
27 |
self.thr = thr
|
28 |
|
|
|
33 |
curr_answers = []
|
34 |
curr_scores = []
|
35 |
best_score = 0
|
36 |
+
# for i, nlp in enumerate(self.nlps):
|
37 |
+
# for text, score in zip(texts, ranking_scores):
|
38 |
+
# QA_input = {
|
39 |
+
# 'question': question,
|
40 |
+
# 'context': text
|
41 |
+
# }
|
42 |
+
# res = nlp(QA_input)
|
43 |
+
# # print(res)
|
44 |
+
# if res["score"] > self.thr:
|
45 |
+
# curr_answers.append(res["answer"])
|
46 |
+
# curr_scores.append(res["score"])
|
47 |
+
# res["score"] = res["score"] * score
|
48 |
+
# if i == 0:
|
49 |
+
# if res["score"] > best_score:
|
50 |
+
# answer = res["answer"]
|
51 |
+
# best_score = res["score"]
|
52 |
+
|
53 |
+
for i, (model, tokenizer) in enumerate(zip(self.models, self.tokenizers)):
|
54 |
for text, score in zip(texts, ranking_scores):
|
55 |
+
# Encode the question and context as input ids and attention mask
|
56 |
+
inputs = tokenizer(question, text, return_tensors="pt")
|
57 |
+
input_ids = inputs["input_ids"]
|
58 |
+
attention_mask = inputs["attention_mask"]
|
59 |
+
# Get the start and end logits from the model
|
60 |
+
outputs = model(input_ids, attention_mask=attention_mask)
|
61 |
+
start_logits = outputs.start_logits
|
62 |
+
end_logits = outputs.end_logits
|
63 |
+
# Get the most likely start and end indices
|
64 |
+
start_idx = torch.argmax(start_logits)
|
65 |
+
end_idx = torch.argmax(end_logits)
|
66 |
+
# Get the answer span from the input ids
|
67 |
+
answer_ids = input_ids[0][start_idx:end_idx+1]
|
68 |
+
# Decode the answer ids to get the answer text
|
69 |
+
answer_text = tokenizer.decode(answer_ids)
|
70 |
+
# Get the answer score from the start and end logits
|
71 |
+
answer_score = torch.max(start_logits) + torch.max(end_logits)
|
72 |
+
# Convert to numpy values
|
73 |
+
answer_text = answer_text.numpy()
|
74 |
+
answer_score = answer_score.numpy()
|
75 |
+
if answer_score > self.thr:
|
76 |
+
curr_answers.append(answer_text)
|
77 |
+
curr_scores.append(answer_score)
|
78 |
+
answer_score = answer_score * score
|
79 |
if i == 0:
|
80 |
+
if answer_score > best_score:
|
81 |
+
answer = answer_text
|
82 |
+
best_score = answer_score
|
83 |
if len(curr_answers) == 0:
|
84 |
return None
|
85 |
curr_answers = [post_process_answer(x, self.entity_dict) for x in curr_answers]
|