File size: 6,888 Bytes
ed93606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
"""
ํ•œ๊ตญ์–ด CLIP ๋ชจ๋ธ ๊ตฌํ˜„
์ด ๋ชจ๋“ˆ์€ HuggingFace์˜ ํ•œ๊ตญ์–ด CLIP ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€์˜ ์ž„๋ฒ ๋”ฉ์„ ์ƒ์„ฑ
"""
import os
import sys
import logging
import torch
from transformers import CLIPProcessor, CLIPModel
from PIL import Image
import requests
from io import BytesIO
import numpy as np

# ์บ์‹œ ๋””๋ ‰ํ† ๋ฆฌ ์„ค์ •
os.environ['TRANSFORMERS_CACHE'] = '/tmp/huggingface_cache'
os.environ['HF_HOME'] = '/tmp/huggingface_cache'

# ๋””๋ ‰ํ† ๋ฆฌ ์ƒ์„ฑ
os.makedirs('/tmp/huggingface_cache', exist_ok=True)
os.makedirs('/tmp/uploads', exist_ok=True)

# ๋กœ๊น… ์„ค์ •
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# ๋ชจ๋ธ ์„ค์ • - ํ™˜๊ฒฝ ๋ณ€์ˆ˜์—์„œ ๊ฐ€์ ธ์˜ค๊ฑฐ๋‚˜ ๊ธฐ๋ณธ๊ฐ’ ์‚ฌ์šฉ
CLIP_MODEL_NAME = os.getenv('CLIP_MODEL_NAME', 'Bingsu/clip-vit-large-patch14-ko')
DEVICE = "cuda" if torch.cuda.is_available() and os.getenv('USE_GPU', 'True').lower() == 'true' else "cpu"

class KoreanCLIPModel:
    """
    ํ•œ๊ตญ์–ด CLIP ๋ชจ๋ธ ํด๋ž˜์Šค
    ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€๋ฅผ ์ž„๋ฒ ๋”ฉํ•˜๊ณ  ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ธฐ๋Šฅ ์ œ๊ณต
    """
    
    def __init__(self, model_name=CLIP_MODEL_NAME, device=DEVICE):
        """
        CLIP ๋ชจ๋ธ ์ดˆ๊ธฐํ™”
        
        Args:
            model_name (str): ์‚ฌ์šฉํ•  CLIP ๋ชจ๋ธ ์ด๋ฆ„ ๋˜๋Š” ๊ฒฝ๋กœ
            device (str): ์‚ฌ์šฉํ•  ์žฅ์น˜ ('cuda' ๋˜๋Š” 'cpu')
        """
        self.device = device
        self.model_name = model_name
        
        logger.info(f"CLIP ๋ชจ๋ธ '{model_name}' ๋กœ๋“œ ์ค‘ (device: {device})...")
        
        try:
            # ์บ์‹œ ๋””๋ ‰ํ† ๋ฆฌ ์„ค์ •
            os.environ["TRANSFORMERS_CACHE"] = "/tmp/transformers_cache"
            os.makedirs("/tmp/transformers_cache", exist_ok=True)
            
            self.model = CLIPModel.from_pretrained(model_name).to(device)
            self.processor = CLIPProcessor.from_pretrained(model_name)
            logger.info("CLIP ๋ชจ๋ธ ๋กœ๋“œ ์™„๋ฃŒ")
        except Exception as e:
            logger.error(f"CLIP ๋ชจ๋ธ ๋กœ๋“œ ์‹คํŒจ: {str(e)}")
            raise
            
    def encode_text(self, text):
        """
        ํ…์ŠคํŠธ๋ฅผ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜
        
        Args:
            text (str or list): ์ธ์ฝ”๋”ฉํ•  ํ…์ŠคํŠธ ๋˜๋Š” ํ…์ŠคํŠธ ๋ฆฌ์ŠคํŠธ
            
        Returns:
            numpy.ndarray: ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ
        """
        if isinstance(text, str):
            text = [text]
            
        try:
            with torch.no_grad():
                # ํ…์ŠคํŠธ ์ธ์ฝ”๋”ฉ
                inputs = self.processor(text=text, return_tensors="pt", padding=True, truncation=True).to(self.device)
                text_features = self.model.get_text_features(**inputs)
                
                # ํ…์ŠคํŠธ ํŠน์„ฑ ์ •๊ทœํ™”
                text_embeddings = text_features / text_features.norm(dim=1, keepdim=True)
                
            return text_embeddings.cpu().numpy()
        except Exception as e:
            logger.error(f"ํ…์ŠคํŠธ ์ธ์ฝ”๋”ฉ ์ค‘ ์˜ค๋ฅ˜ ๋ฐœ์ƒ: {str(e)}")
            return np.zeros((len(text), self.model.text_embed_dim))
            
    def encode_image(self, image_source):
        """
        ์ด๋ฏธ์ง€๋ฅผ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜
        
        Args:
            image_source: ์ธ์ฝ”๋”ฉํ•  ์ด๋ฏธ์ง€ (PIL Image, URL ๋˜๋Š” ์ด๋ฏธ์ง€ ๊ฒฝ๋กœ)
            
        Returns:
            numpy.ndarray: ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ
        """
        try:
            # ์ด๋ฏธ์ง€ ๋กœ๋“œ (URL, ํŒŒ์ผ ๊ฒฝ๋กœ, PIL ์ด๋ฏธ์ง€ ๊ฐ์ฒด ๋˜๋Š” Base64)
            if isinstance(image_source, str):
                if image_source.startswith('http'):
                    # URL์—์„œ ์ด๋ฏธ์ง€ ๋กœ๋“œ
                    response = requests.get(image_source)
                    image = Image.open(BytesIO(response.content)).convert('RGB')
                else:
                    # ๋กœ์ปฌ ํŒŒ์ผ์—์„œ ์ด๋ฏธ์ง€ ๋กœ๋“œ
                    image = Image.open(image_source).convert('RGB')
            else:
                # ์ด๋ฏธ PIL ์ด๋ฏธ์ง€ ๊ฐ์ฒด์ธ ๊ฒฝ์šฐ
                image = image_source.convert('RGB')
                
            with torch.no_grad():
                # ์ด๋ฏธ์ง€ ์ธ์ฝ”๋”ฉ
                inputs = self.processor(images=image, return_tensors="pt").to(self.device)
                image_features = self.model.get_image_features(**inputs)
                
                # ์ด๋ฏธ์ง€ ํŠน์„ฑ ์ •๊ทœํ™”
                image_embeddings = image_features / image_features.norm(dim=1, keepdim=True)
                
            return image_embeddings.cpu().numpy()
        except Exception as e:
            logger.error(f"์ด๋ฏธ์ง€ ์ธ์ฝ”๋”ฉ ์ค‘ ์˜ค๋ฅ˜ ๋ฐœ์ƒ: {str(e)}")
            return np.zeros((1, self.model.vision_embed_dim))
    
    def calculate_similarity(self, text_embedding, image_embedding=None):
        """
        ํ…์ŠคํŠธ์™€ ์ด๋ฏธ์ง€ ์ž„๋ฒ ๋”ฉ ๊ฐ„์˜ ์œ ์‚ฌ๋„ ๊ณ„์‚ฐ
        
        Args:
            text_embedding (numpy.ndarray): ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ
            image_embedding (numpy.ndarray, optional): ์ด๋ฏธ์ง€ ์ž„๋ฒ ๋”ฉ (์—†์œผ๋ฉด ํ…์ŠคํŠธ๋งŒ ๋น„๊ต)
            
        Returns:
            float: ์œ ์‚ฌ๋„ ์ ์ˆ˜ (0~1 ์‚ฌ์ด)
        """
        if image_embedding is None:
            # ํ…์ŠคํŠธ-ํ…์ŠคํŠธ ์œ ์‚ฌ๋„ ๊ณ„์‚ฐ (์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„)
            similarity = np.dot(text_embedding, text_embedding.T)[0, 0]
        else:
            # ํ…์ŠคํŠธ-์ด๋ฏธ์ง€ ์œ ์‚ฌ๋„ ๊ณ„์‚ฐ (์ฝ”์‚ฌ์ธ ์œ ์‚ฌ๋„)
            similarity = np.dot(text_embedding, image_embedding.T)[0, 0]
            
        # ์œ ์‚ฌ๋„๋ฅผ 0~1 ๋ฒ”์œ„๋กœ ์ •๊ทœํ™”
        similarity = (similarity + 1) / 2
        return float(similarity)
        
    def encode_batch_texts(self, texts):
        """
        ์—ฌ๋Ÿฌ ํ…์ŠคํŠธ๋ฅผ ํ•œ ๋ฒˆ์— ์ž„๋ฒ ๋”ฉ
        
        Args:
            texts (list): ํ…์ŠคํŠธ ๋ชฉ๋ก
            
        Returns:
            numpy.ndarray: ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ ๋ฐฐ์—ด
        """
        # ๋ฐฐ์น˜ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ์ฝ”๋“œ
        # ์‹ค์ œ ๊ตฌํ˜„์—์„œ๋Š” ๋ฉ”๋ชจ๋ฆฌ ํฌ๊ธฐ์— ๋”ฐ๋ผ ์ ์ ˆํ•œ ๋ฐฐ์น˜ ํฌ๊ธฐ ์กฐ์ • ํ•„์š”
        return self.encode_text(texts)

# ๋ชจ๋“ˆ ํ…Œ์ŠคํŠธ์šฉ ์ฝ”๋“œ
if __name__ == "__main__":
    # ๋ชจ๋ธ ์ดˆ๊ธฐํ™”
    clip_model = KoreanCLIPModel()
    
    # ์ƒ˜ํ”Œ ํ…์ŠคํŠธ ์ธ์ฝ”๋”ฉ
    sample_text = "๊ฒ€์€์ƒ‰ ์ง€๊ฐ‘์„ ์žƒ์–ด๋ฒ„๋ ธ์Šต๋‹ˆ๋‹ค. ํ˜„๊ธˆ๊ณผ ์นด๋“œ๊ฐ€ ๋“ค์–ด์žˆ์–ด์š”."
    text_embedding = clip_model.encode_text(sample_text)
    
    print(f"ํ…์ŠคํŠธ ์ž„๋ฒ ๋”ฉ shape: {text_embedding.shape}")
    
    # ์œ ์‚ฌ๋„ ๊ณ„์‚ฐ (ํ…์ŠคํŠธ๋งŒ)
    sample_text2 = "๊ฒ€์€์ƒ‰ ์ง€๊ฐ‘์„ ์ฐพ์•˜์Šต๋‹ˆ๋‹ค. ์•ˆ์— ํ˜„๊ธˆ๊ณผ ์นด๋“œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค."
    text_embedding2 = clip_model.encode_text(sample_text2)
    
    similarity = clip_model.calculate_similarity(text_embedding, text_embedding2)
    print(f"ํ…์ŠคํŠธ ๊ฐ„ ์œ ์‚ฌ๋„: {similarity:.4f}")