Spaces:
Paused
Paused
create models.py with AlbertQA model
Browse files- app/models.py +66 -0
app/models.py
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AlbertForQuestionAnswering
|
2 |
+
import torch
|
3 |
+
|
4 |
+
class QAModel:
|
5 |
+
def __init__(self, model_name: str = "twmkn9/albert-base-v2-squad2"):
|
6 |
+
"""
|
7 |
+
Initialize the QA model and tokenizer.
|
8 |
+
"""
|
9 |
+
self.model_name = model_name
|
10 |
+
self.tokenizer = None
|
11 |
+
self.model = None
|
12 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
13 |
+
self.load_model()
|
14 |
+
|
15 |
+
def load_model(self):
|
16 |
+
"""
|
17 |
+
Load the tokenizer and model.
|
18 |
+
"""
|
19 |
+
self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
|
20 |
+
self.model = AlbertForQuestionAnswering.from_pretrained(self.model_name).to(self.device)
|
21 |
+
logger.info(f"Loaded QA model: {self.model_name}")
|
22 |
+
|
23 |
+
def inference_qa(self, context: str, question: str):
|
24 |
+
"""
|
25 |
+
Perform question-answering inference.
|
26 |
+
Args:
|
27 |
+
context (str): The text passage or document.
|
28 |
+
question (str): The question to be answered.
|
29 |
+
Returns:
|
30 |
+
str: The predicted answer.
|
31 |
+
"""
|
32 |
+
if self.tokenizer is None or self.model is None:
|
33 |
+
raise ValueError("Model or tokenizer is not loaded.")
|
34 |
+
|
35 |
+
# Tokenize inputs
|
36 |
+
inputs = self.tokenizer(question, context, return_tensors="pt", truncation=True, padding=True)
|
37 |
+
inputs = {key: value.to(self.device) for key, value in inputs.items()}
|
38 |
+
|
39 |
+
# Perform inference
|
40 |
+
with torch.no_grad():
|
41 |
+
outputs = self.model(**inputs)
|
42 |
+
|
43 |
+
# Extract answer
|
44 |
+
answer_start_index = outputs.start_logits.argmax()
|
45 |
+
answer_end_index = outputs.end_logits.argmax()
|
46 |
+
predict_answer_tokens = inputs["input_ids"][0, answer_start_index : answer_end_index + 1]
|
47 |
+
answer = self.tokenizer.decode(predict_answer_tokens, skip_special_tokens=True)
|
48 |
+
|
49 |
+
return answer
|
50 |
+
|
51 |
+
# Global instance of the QA model
|
52 |
+
qa_model = QAModel()
|
53 |
+
|
54 |
+
def load_qa_pipeline(model_name: str = "twmkn9/albert-base-v2-squad2"):
|
55 |
+
"""
|
56 |
+
Load the QA model and tokenizer.
|
57 |
+
"""
|
58 |
+
global qa_model
|
59 |
+
qa_model = QAModel(model_name)
|
60 |
+
return qa_model
|
61 |
+
|
62 |
+
def inference_qa(qa_pipeline, context: str, question: str):
|
63 |
+
"""
|
64 |
+
Perform QA inference using the loaded pipeline.
|
65 |
+
"""
|
66 |
+
return qa_pipeline.inference_qa(context, question)
|