Spaces:
Sleeping
Sleeping
Update train_model.py
Browse files- train_model.py +74 -24
train_model.py
CHANGED
@@ -7,25 +7,35 @@ from transformers import BertTokenizer, BertModel
|
|
7 |
import numpy as np
|
8 |
|
9 |
# MongoDB Atlas 연결 설정
|
10 |
-
client = MongoClient(
|
|
|
|
|
11 |
db = client["two_tower_model"]
|
12 |
train_dataset = db["train_dataset"]
|
13 |
|
14 |
-
#
|
15 |
-
tokenizer = BertTokenizer.from_pretrained(
|
16 |
-
|
|
|
17 |
|
18 |
# 상품 임베딩 함수
|
19 |
def embed_product_data(product):
|
20 |
"""
|
21 |
-
상품 데이터를 임베딩하는 함수.
|
22 |
"""
|
23 |
-
text =
|
24 |
-
|
25 |
-
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
return embedding
|
28 |
|
|
|
29 |
# PyTorch Dataset 정의
|
30 |
class TripletDataset(Dataset):
|
31 |
def __init__(self, dataset):
|
@@ -41,31 +51,64 @@ class TripletDataset(Dataset):
|
|
41 |
negative = torch.tensor(data["negative_embedding"], dtype=torch.float32)
|
42 |
return anchor, positive, negative
|
43 |
|
|
|
44 |
# MongoDB에서 데이터셋 로드 및 임베딩 변환
|
45 |
-
def prepare_training_data():
|
46 |
-
dataset = list(train_dataset.find())
|
47 |
if not dataset:
|
48 |
raise ValueError("No training data found in MongoDB.")
|
49 |
|
50 |
# Anchor, Positive, Negative 임베딩 생성
|
51 |
embedded_dataset = []
|
52 |
-
for entry in dataset:
|
53 |
try:
|
|
|
54 |
anchor_embedding = embed_product_data(entry["anchor"]["product"])
|
55 |
positive_embedding = embed_product_data(entry["positive"]["product"])
|
56 |
negative_embedding = embed_product_data(entry["negative"]["product"])
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
"
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
except Exception as e:
|
63 |
-
print(f"Error embedding data: {e}")
|
64 |
-
|
65 |
return TripletDataset(embedded_dataset)
|
66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
67 |
# Triplet Loss를 학습시키는 함수
|
68 |
-
def train_triplet_model(
|
|
|
|
|
69 |
optimizer = Adam(product_model.parameters(), lr=learning_rate)
|
70 |
|
71 |
for epoch in range(num_epochs):
|
@@ -83,7 +126,9 @@ def train_triplet_model(product_model, train_loader, num_epochs=10, learning_rat
|
|
83 |
# Triplet loss 계산
|
84 |
positive_distance = F.pairwise_distance(anchor_vec, positive_vec)
|
85 |
negative_distance = F.pairwise_distance(anchor_vec, negative_vec)
|
86 |
-
triplet_loss = torch.clamp(
|
|
|
|
|
87 |
|
88 |
# 역전파와 최적화
|
89 |
triplet_loss.backward()
|
@@ -91,17 +136,20 @@ def train_triplet_model(product_model, train_loader, num_epochs=10, learning_rat
|
|
91 |
|
92 |
total_loss += triplet_loss.item()
|
93 |
|
94 |
-
print(
|
|
|
|
|
95 |
|
96 |
return product_model
|
97 |
|
|
|
98 |
# 모델 학습 파이프라인
|
99 |
def main():
|
100 |
# 모델 초기화 (예시 모델)
|
101 |
product_model = torch.nn.Sequential(
|
102 |
-
torch.nn.Linear(768, 256), # 768:
|
103 |
torch.nn.ReLU(),
|
104 |
-
torch.nn.Linear(256, 128)
|
105 |
)
|
106 |
|
107 |
# 데이터 준비
|
@@ -114,6 +162,8 @@ def main():
|
|
114 |
# 학습된 모델 저장
|
115 |
torch.save(trained_model.state_dict(), "product_model.pth")
|
116 |
print("Model training completed and saved.")
|
|
|
|
|
117 |
|
118 |
if __name__ == "__main__":
|
119 |
main()
|
|
|
7 |
import numpy as np
|
8 |
|
9 |
# MongoDB Atlas 연결 설정
|
10 |
+
client = MongoClient(
|
11 |
+
"mongodb+srv://waseoke:[email protected]/test?retryWrites=true&w=majority"
|
12 |
+
)
|
13 |
db = client["two_tower_model"]
|
14 |
train_dataset = db["train_dataset"]
|
15 |
|
16 |
+
# KoBERT 모델 및 토크나이저 로드
|
17 |
+
tokenizer = BertTokenizer.from_pretrained('monologg/kobert')
|
18 |
+
model = BertModel.from_pretrained('monologg/kobert')
|
19 |
+
|
20 |
|
21 |
# 상품 임베딩 함수
|
22 |
def embed_product_data(product):
|
23 |
"""
|
24 |
+
상품 데이터를 KoBERT로 임베딩하는 함수.
|
25 |
"""
|
26 |
+
text = (
|
27 |
+
product.get("product_name", "") + " " + product.get("product_description", "")
|
28 |
+
)
|
29 |
+
inputs = tokenizer(
|
30 |
+
text, return_tensors="pt", truncation=True, padding=True, max_length=128
|
31 |
+
)
|
32 |
+
outputs = model(**inputs)
|
33 |
+
embedding = (
|
34 |
+
outputs.last_hidden_state.mean(dim=1).detach().numpy().flatten()
|
35 |
+
) # 평균 풀링
|
36 |
return embedding
|
37 |
|
38 |
+
|
39 |
# PyTorch Dataset 정의
|
40 |
class TripletDataset(Dataset):
|
41 |
def __init__(self, dataset):
|
|
|
51 |
negative = torch.tensor(data["negative_embedding"], dtype=torch.float32)
|
52 |
return anchor, positive, negative
|
53 |
|
54 |
+
|
55 |
# MongoDB에서 데이터셋 로드 및 임베딩 변환
|
56 |
+
def prepare_training_data(verbose=False):
|
57 |
+
dataset = list(train_dataset.find())
|
58 |
if not dataset:
|
59 |
raise ValueError("No training data found in MongoDB.")
|
60 |
|
61 |
# Anchor, Positive, Negative 임베딩 생성
|
62 |
embedded_dataset = []
|
63 |
+
for idx, entry in enumerate(dataset):
|
64 |
try:
|
65 |
+
# Anchor, Positive, Negative 데이터 임베딩
|
66 |
anchor_embedding = embed_product_data(entry["anchor"]["product"])
|
67 |
positive_embedding = embed_product_data(entry["positive"]["product"])
|
68 |
negative_embedding = embed_product_data(entry["negative"]["product"])
|
69 |
+
|
70 |
+
# 임베딩 확인 (옵션으로 출력)
|
71 |
+
if verbose:
|
72 |
+
print(f"Sample {idx + 1}:")
|
73 |
+
print(
|
74 |
+
f"Anchor Embedding: {anchor_embedding[:5]}... (shape: {anchor_embedding.shape})"
|
75 |
+
)
|
76 |
+
print(
|
77 |
+
f"Positive Embedding: {positive_embedding[:5]}... (shape: {positive_embedding.shape})"
|
78 |
+
)
|
79 |
+
print(
|
80 |
+
f"Negative Embedding: {negative_embedding[:5]}... (shape: {negative_embedding.shape})"
|
81 |
+
)
|
82 |
+
|
83 |
+
# 임베딩 결과 저장
|
84 |
+
embedded_dataset.append(
|
85 |
+
{
|
86 |
+
"anchor_embedding": anchor_embedding,
|
87 |
+
"positive_embedding": positive_embedding,
|
88 |
+
"negative_embedding": negative_embedding,
|
89 |
+
}
|
90 |
+
)
|
91 |
except Exception as e:
|
92 |
+
print(f"Error embedding data at sample {idx + 1}: {e}")
|
93 |
+
|
94 |
return TripletDataset(embedded_dataset)
|
95 |
|
96 |
+
|
97 |
+
# 데이터셋 검증용 함수
|
98 |
+
def validate_embeddings():
|
99 |
+
"""
|
100 |
+
데이터셋 임베딩을 생성하고 각 임베딩의 일부를 출력하여 확인.
|
101 |
+
"""
|
102 |
+
print("Validating embeddings...")
|
103 |
+
triplet_dataset = prepare_training_data(verbose=True)
|
104 |
+
print(f"Total samples: {len(triplet_dataset)}")
|
105 |
+
return triplet_dataset
|
106 |
+
|
107 |
+
|
108 |
# Triplet Loss를 학습시키는 함수
|
109 |
+
def train_triplet_model(
|
110 |
+
product_model, train_loader, num_epochs=10, learning_rate=0.001, margin=0.05
|
111 |
+
):
|
112 |
optimizer = Adam(product_model.parameters(), lr=learning_rate)
|
113 |
|
114 |
for epoch in range(num_epochs):
|
|
|
126 |
# Triplet loss 계산
|
127 |
positive_distance = F.pairwise_distance(anchor_vec, positive_vec)
|
128 |
negative_distance = F.pairwise_distance(anchor_vec, negative_vec)
|
129 |
+
triplet_loss = torch.clamp(
|
130 |
+
positive_distance - negative_distance + margin, min=0
|
131 |
+
).mean()
|
132 |
|
133 |
# 역전파와 최적화
|
134 |
triplet_loss.backward()
|
|
|
136 |
|
137 |
total_loss += triplet_loss.item()
|
138 |
|
139 |
+
print(
|
140 |
+
f"Epoch {epoch + 1}/{num_epochs}, Loss: {total_loss / len(train_loader):.4f}"
|
141 |
+
)
|
142 |
|
143 |
return product_model
|
144 |
|
145 |
+
|
146 |
# 모델 학습 파이프라인
|
147 |
def main():
|
148 |
# 모델 초기화 (예시 모델)
|
149 |
product_model = torch.nn.Sequential(
|
150 |
+
torch.nn.Linear(768, 256), # 768: KoBERT 임베딩 차원
|
151 |
torch.nn.ReLU(),
|
152 |
+
torch.nn.Linear(256, 128),
|
153 |
)
|
154 |
|
155 |
# 데이터 준비
|
|
|
162 |
# 학습된 모델 저장
|
163 |
torch.save(trained_model.state_dict(), "product_model.pth")
|
164 |
print("Model training completed and saved.")
|
165 |
+
print(validate_embeddings())
|
166 |
+
|
167 |
|
168 |
if __name__ == "__main__":
|
169 |
main()
|