Update models/handler.py
Browse files- models/handler.py +35 -32
models/handler.py
CHANGED
@@ -6,16 +6,14 @@ from models.qa_model import *
|
|
6 |
from tqdm.auto import tqdm
|
7 |
tqdm.pandas()
|
8 |
from datasets import load_dataset
|
9 |
-
|
10 |
-
# from transformers import pipeline, AutoTokenizer
|
11 |
-
|
12 |
|
13 |
class EndpointHandler():
|
14 |
def __init__(self, path=""):
|
15 |
df_wiki_windows = load_dataset("foxxy-hm/e2eqa-wiki", data_files="processed/wikipedia_20220620_cleaned_v2.csv")["train"].to_pandas()
|
16 |
df_wiki = load_dataset("foxxy-hm/e2eqa-wiki", data_files="wikipedia_20220620_short.csv")["train"].to_pandas()
|
17 |
df_wiki.title = df_wiki.title.apply(str)
|
18 |
-
|
19 |
entity_dict = load_dataset("foxxy-hm/e2eqa-wiki", data_files="processed/entities.json")["train"].to_dict()
|
20 |
new_dict = dict()
|
21 |
for key, val in entity_dict.items():
|
@@ -25,68 +23,73 @@ class EndpointHandler():
|
|
25 |
new_dict[key.lower()] = val
|
26 |
entity_dict.update(new_dict)
|
27 |
title2idx = dict([(x.strip(), y) for x, y in zip(df_wiki.title, df_wiki.index.values)])
|
28 |
-
|
29 |
qa_model = QAEnsembleModel("nguyenvulebinh/vi-mrc-large", ["qa_model_robust.bin"], entity_dict)
|
30 |
-
pairwise_model_stage1 = PairwiseModel("nguyenvulebinh/vi-mrc-base")
|
31 |
pairwise_model_stage1.load_state_dict(torch.load("pairwise_v2.bin", map_location=torch.device('cpu')))
|
32 |
pairwise_model_stage1.eval()
|
33 |
-
|
34 |
-
pairwise_model_stage2 = PairwiseModel("nguyenvulebinh/vi-mrc-base")
|
35 |
pairwise_model_stage2.load_state_dict(torch.load("pairwise_stage2_seed0.bin", map_location=torch.device('cpu')))
|
36 |
-
|
37 |
bm25_model_stage1 = BM25Gensim("bm25_stage1/", entity_dict, title2idx)
|
38 |
bm25_model_stage2_full = BM25Gensim("bm25_stage2/full_text/", entity_dict, title2idx)
|
39 |
bm25_model_stage2_title = BM25Gensim("bm25_stage2/title/", entity_dict, title2idx)
|
40 |
-
# # create inference pipeline
|
41 |
-
# self.pipeline = pipeline("text-classification", model=model, tokenizer=tokenizer)
|
42 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
def get_answer_e2e(self, question):
|
44 |
-
#Bm25 retrieval for top200 candidates
|
45 |
query = preprocess(question).lower()
|
46 |
-
top_n, bm25_scores = bm25_model_stage1.get_topk_stage1(query, topk=200)
|
47 |
titles = [preprocess(df_wiki_windows.title.values[i]) for i in top_n]
|
48 |
texts = [preprocess(df_wiki_windows.text.values[i]) for i in top_n]
|
49 |
-
|
50 |
-
#Reranking with pairwise model for top10
|
51 |
question = preprocess(question)
|
52 |
-
ranking_preds = pairwise_model_stage1.stage1_ranking(question, texts)
|
53 |
ranking_scores = ranking_preds * bm25_scores
|
54 |
-
|
55 |
-
#Question answering
|
56 |
best_idxs = np.argsort(ranking_scores)[-10:]
|
57 |
ranking_scores = np.array(ranking_scores)[best_idxs]
|
58 |
texts = np.array(texts)[best_idxs]
|
59 |
-
best_answer = qa_model(question, texts, ranking_scores)
|
60 |
if best_answer is None:
|
61 |
return "Chịu"
|
62 |
bm25_answer = preprocess(str(best_answer).lower(), max_length=128, remove_puncts=True)
|
63 |
-
|
64 |
-
#Entity mapping
|
65 |
if not check_number(bm25_answer):
|
66 |
bm25_question = preprocess(str(question).lower(), max_length=128, remove_puncts=True)
|
67 |
bm25_question_answer = bm25_question + " " + bm25_answer
|
68 |
-
candidates, scores = bm25_model_stage2_title.get_topk_stage2(bm25_answer, raw_answer=best_answer)
|
69 |
titles = [df_wiki.title.values[i] for i in candidates]
|
70 |
texts = [df_wiki.text.values[i] for i in candidates]
|
71 |
-
ranking_preds = pairwise_model_stage2.stage2_ranking(question, best_answer, titles, texts)
|
72 |
if ranking_preds.max() >= 0.1:
|
73 |
final_answer = titles[ranking_preds.argmax()]
|
74 |
else:
|
75 |
-
candidates, scores = bm25_model_stage2_full.get_topk_stage2(bm25_question_answer)
|
76 |
titles = [df_wiki.title.values[i] for i in candidates] + titles
|
77 |
texts = [df_wiki.text.values[i] for i in candidates] + texts
|
78 |
ranking_preds = np.concatenate(
|
79 |
-
[pairwise_model_stage2.stage2_ranking(question, best_answer, titles, texts), ranking_preds])
|
80 |
final_answer = "wiki/"+titles[ranking_preds.argmax()].replace(" ","_")
|
81 |
else:
|
82 |
final_answer = bm25_answer.lower()
|
83 |
-
|
84 |
|
85 |
|
|
|
|
|
|
|
|
|
86 |
def __call__(self, question):
|
87 |
-
|
88 |
-
""
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
|
|
6 |
from tqdm.auto import tqdm
|
7 |
tqdm.pandas()
|
8 |
from datasets import load_dataset
|
9 |
+
from transformers import pipeline
|
|
|
|
|
10 |
|
11 |
class EndpointHandler():
|
12 |
def __init__(self, path=""):
|
13 |
df_wiki_windows = load_dataset("foxxy-hm/e2eqa-wiki", data_files="processed/wikipedia_20220620_cleaned_v2.csv")["train"].to_pandas()
|
14 |
df_wiki = load_dataset("foxxy-hm/e2eqa-wiki", data_files="wikipedia_20220620_short.csv")["train"].to_pandas()
|
15 |
df_wiki.title = df_wiki.title.apply(str)
|
16 |
+
|
17 |
entity_dict = load_dataset("foxxy-hm/e2eqa-wiki", data_files="processed/entities.json")["train"].to_dict()
|
18 |
new_dict = dict()
|
19 |
for key, val in entity_dict.items():
|
|
|
23 |
new_dict[key.lower()] = val
|
24 |
entity_dict.update(new_dict)
|
25 |
title2idx = dict([(x.strip(), y) for x, y in zip(df_wiki.title, df_wiki.index.values)])
|
26 |
+
|
27 |
qa_model = QAEnsembleModel("nguyenvulebinh/vi-mrc-large", ["qa_model_robust.bin"], entity_dict)
|
28 |
+
pairwise_model_stage1 = PairwiseModel("nguyenvulebinh/vi-mrc-base")
|
29 |
pairwise_model_stage1.load_state_dict(torch.load("pairwise_v2.bin", map_location=torch.device('cpu')))
|
30 |
pairwise_model_stage1.eval()
|
31 |
+
|
32 |
+
pairwise_model_stage2 = PairwiseModel("nguyenvulebinh/vi-mrc-base")
|
33 |
pairwise_model_stage2.load_state_dict(torch.load("pairwise_stage2_seed0.bin", map_location=torch.device('cpu')))
|
34 |
+
|
35 |
bm25_model_stage1 = BM25Gensim("bm25_stage1/", entity_dict, title2idx)
|
36 |
bm25_model_stage2_full = BM25Gensim("bm25_stage2/full_text/", entity_dict, title2idx)
|
37 |
bm25_model_stage2_title = BM25Gensim("bm25_stage2/title/", entity_dict, title2idx)
|
|
|
|
|
38 |
|
39 |
+
self.qa_model = qa_model
|
40 |
+
self.pairwise_model_stage1 = pairwise_model_stage1
|
41 |
+
self.pairwise_model_stage2 = pairwise_model_stage2
|
42 |
+
self.bm25_model_stage1 = bm25_model_stage1
|
43 |
+
self.bm25_model_stage2_full = bm25_model_stage2_full
|
44 |
+
self.bm25_model_stage2_title = bm25_model_stage2_title
|
45 |
+
|
46 |
def get_answer_e2e(self, question):
|
|
|
47 |
query = preprocess(question).lower()
|
48 |
+
top_n, bm25_scores = self.bm25_model_stage1.get_topk_stage1(query, topk=200)
|
49 |
titles = [preprocess(df_wiki_windows.title.values[i]) for i in top_n]
|
50 |
texts = [preprocess(df_wiki_windows.text.values[i]) for i in top_n]
|
51 |
+
|
|
|
52 |
question = preprocess(question)
|
53 |
+
ranking_preds = self.pairwise_model_stage1.stage1_ranking(question, texts)
|
54 |
ranking_scores = ranking_preds * bm25_scores
|
55 |
+
|
|
|
56 |
best_idxs = np.argsort(ranking_scores)[-10:]
|
57 |
ranking_scores = np.array(ranking_scores)[best_idxs]
|
58 |
texts = np.array(texts)[best_idxs]
|
59 |
+
best_answer = self.qa_model(question, texts, ranking_scores)
|
60 |
if best_answer is None:
|
61 |
return "Chịu"
|
62 |
bm25_answer = preprocess(str(best_answer).lower(), max_length=128, remove_puncts=True)
|
63 |
+
|
|
|
64 |
if not check_number(bm25_answer):
|
65 |
bm25_question = preprocess(str(question).lower(), max_length=128, remove_puncts=True)
|
66 |
bm25_question_answer = bm25_question + " " + bm25_answer
|
67 |
+
candidates, scores = self.bm25_model_stage2_title.get_topk_stage2(bm25_answer, raw_answer=best_answer)
|
68 |
titles = [df_wiki.title.values[i] for i in candidates]
|
69 |
texts = [df_wiki.text.values[i] for i in candidates]
|
70 |
+
ranking_preds = self.pairwise_model_stage2.stage2_ranking(question, best_answer, titles, texts)
|
71 |
if ranking_preds.max() >= 0.1:
|
72 |
final_answer = titles[ranking_preds.argmax()]
|
73 |
else:
|
74 |
+
candidates, scores = self.bm25_model_stage2_full.get_topk_stage2(bm25_question_answer)
|
75 |
titles = [df_wiki.title.values[i] for i in candidates] + titles
|
76 |
texts = [df_wiki.text.values[i] for i in candidates] + texts
|
77 |
ranking_preds = np.concatenate(
|
78 |
+
[self.pairwise_model_stage2.stage2_ranking(question, best_answer, titles, texts), ranking_preds])
|
79 |
final_answer = "wiki/"+titles[ranking_preds.argmax()].replace(" ","_")
|
80 |
else:
|
81 |
final_answer = bm25_answer.lower()
|
82 |
+
return final_answer
|
83 |
|
84 |
|
85 |
+
class InferencePipeline:
|
86 |
+
def __init__(self):
|
87 |
+
self.endpoint_handler = EndpointHandler() # Instantiate the EndpointHandler class
|
88 |
+
|
89 |
def __call__(self, question):
|
90 |
+
answer = self.endpoint_handler.get_answer_e2e(question) # Call the get_answer_e2e method from EndpointHandler
|
91 |
+
return {"answer": answer} # Return the answer as a dictionary
|
92 |
+
|
93 |
+
|
94 |
+
inference_pipeline = InferencePipeline()
|
95 |
+
pipeline = pipeline("qa-model", model=inference_pipeline, tokenizer=None)
|