File size: 13,189 Bytes
76e08e0 64c57fb f16cb1a f89c606 76e08e0 69a39af 28a7334 7ce0e26 dcb7540 26e52a1 7a2cdb5 f89c606 115fae5 72114b8 63c2769 e53f54d ab48ca2 dcb7540 ab48ca2 26e52a1 ab48ca2 dcb7540 ab48ca2 dcb7540 26e52a1 ae810a8 26e52a1 7a2cdb5 ab48ca2 63c2769 ab48ca2 7fabc42 ab48ca2 e53f54d 459e392 98b216f 459e392 98b216f 459e392 dcb7540 0e6a905 63c2769 459e392 7a2cdb5 63c2769 dcb7540 7a2cdb5 63c2769 28a7334 7a2cdb5 0e6a905 7a2cdb5 dcb7540 ab48ca2 63c2769 d5aff0d 0e6a905 d5aff0d 0e6a905 d5aff0d 0e6a905 d5aff0d 0e6a905 d5aff0d 0e6a905 d5aff0d 0e6a905 d5aff0d 0e6a905 ab48ca2 d5aff0d ab48ca2 d5aff0d 63c2769 d5aff0d ab48ca2 64c57fb 63c2769 0e6a905 64c57fb 0e6a905 64c57fb 92bb45b 64c57fb 92bb45b 64c57fb 92bb45b 64c57fb 98b216f 7ce0e26 98b216f 64c57fb e53f54d ad98685 64c57fb 3cec557 64c57fb 3cec557 64c57fb ad98685 7ce0e26 0e6a905 7ce0e26 0e6a905 64c57fb 0e6a905 ad98685 39446f7 0e6a905 39446f7 0e6a905 39446f7 0e6a905 6d9d526 7ce0e26 6d9d526 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 7ce0e26 98b216f 6d9d526 7ce0e26 0e6a905 6d9d526 98b216f 6d9d526 98b216f 7ce0e26 0e6a905 98b216f 0e6a905 98b216f 7ce0e26 0e6a905 6d9d526 7ce0e26 5f1a404 0e6a905 63c2769 f89c606 7ce0e26 5f1a404 7ce0e26 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 |
import json
import re
from typing import List, Optional, Tuple
import numpy as np
import os
import gradio as gr
import spaces
import torch
from PIL import Image
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login, snapshot_download
from paddleocr import PaddleOCR
# Hugging Face ํ ํฐ์ผ๋ก ๋ก๊ทธ์ธ (Spaces Secret์์ ๊ฐ์ ธ์ด)
HF_TOKEN = os.getenv("HF_TOKEN")
if HF_TOKEN:
login(token=HF_TOKEN.strip())
# ์ฝ ์ ๋ณด ๋ถ์ ๋ชจ๋ธ ID (๋น ๋ฅธ ์ถ๋ก ์ ์ํด ๊ฒฝ๋ ๋ชจ๋ธ ์ฌ์ฉ)
MED_MODEL_ID = "google/gemma-2-2b-it"
# ์ ์ญ ๋ชจ๋ธ ๋ณ์ (ํ ๋ฒ๋ง ๋ก๋)
OCR_READER = None
MED_MODEL = None
MED_TOKENIZER = None
OCR_MODEL_REPO_ID = "PaddlePaddle/korean_PP-OCRv5_mobile_rec"
def load_models():
"""๋ชจ๋ธ๋ค์ ํ ๋ฒ๋ง ๋ก๋"""
global OCR_READER, MED_MODEL, MED_TOKENIZER
if OCR_READER is None:
print("๐ Loading PaddleOCR (Korean PP-OCRv5 mobile recognition)...")
rec_model_dir = snapshot_download(
OCR_MODEL_REPO_ID,
allow_patterns=[
"*.pdmodel",
"*.pdiparams",
"*.pdparams",
"*.json",
"*.yml",
],
)
OCR_READER = PaddleOCR(
lang='korean',
use_textline_orientation=True,
text_recognition_model_dir=rec_model_dir,
)
print("โ
PaddleOCR loaded!")
if MED_MODEL is None:
print("๐ Loading Gemma-2-2B for medical analysis (8bit quantization)...")
MED_MODEL = AutoModelForCausalLM.from_pretrained(
MED_MODEL_ID,
torch_dtype=torch.bfloat16,
device_map="auto",
load_in_8bit=True
)
MED_TOKENIZER = AutoTokenizer.from_pretrained(MED_MODEL_ID)
print("โ
Medical model loaded!")
# ์ฑ ์์ ์ ๋ชจ๋ธ ๋ก๋
load_models()
def _extract_assistant_content(decoded: str) -> str:
"""์ด์์คํดํธ ์๋ต ์ถ์ถ"""
if "<|im_start|>assistant" in decoded:
content = decoded.split("<|im_start|>assistant")[-1]
content = content.replace("<|im_end|>", "").strip()
return content
return decoded.strip()
def _extract_json_block(text: str) -> Optional[str]:
"""JSON ๋ธ๋ก ์ถ์ถ"""
match = re.search(r"\{.*\}", text, re.DOTALL)
if not match:
return None
return match.group(0)
@spaces.GPU(duration=120)
def analyze_medication_image(image: Image.Image) -> Tuple[str, str]:
"""์ด๋ฏธ์ง์์ OCR ์ถ์ถ ํ ์ฝ ์ ๋ณด ๋ถ์"""
import time
try:
# Step 1: OCR - PaddleOCR๋ก ํ๊ธ ํ
์คํธ ์ถ์ถ
start_time = time.time()
img_array = np.array(image)
ocr_results = OCR_READER.ocr(img_array, cls=True)
ocr_time = time.time() - start_time
print(f"โฑ๏ธ OCR took {ocr_time:.2f}s")
if not ocr_results or not ocr_results[0]:
return "ํ
์คํธ๋ฅผ ์ฐพ์ ์ ์์ต๋๋ค.", ""
# ํ
์คํธ ์ถ์ถ
texts = [line[1][0] for line in ocr_results[0]]
ocr_text = "\n".join(texts)
# Step 2: ์ฝ ์ ๋ณด ๋ถ์ - MedGemma๋ก ์๋ฃ ์ ๋ณด ์ ๊ณต
analysis_start = time.time()
analysis_prompt = f"""๋ค์์ ์ฝ ๋ดํฌ๋ ์ฒ๋ฐฉ์ ์์ ์ถ์ถํ ํ
์คํธ์
๋๋ค:
{ocr_text}
์ ํ
์คํธ์์ ์ฝ ์ด๋ฆ์ ์ฐพ์์, ๊ฐ ์ฝ์ ๋ํด **๋
ธ์ธ๊ณผ ์ด๋ฆฐ์ด ๋ชจ๋ ์ฝ๊ฒ ์ดํดํ ์ ์๋๋ก** ์ฌ๋ฏธ์๊ณ ์น๊ทผํ๊ฒ ์ค๋ช
ํด์ฃผ์ธ์:
๐ **๊ฐ ์ฝ๋ง๋ค ๋ค์ ์ ๋ณด๋ฅผ ํฌํจํด์ฃผ์ธ์:**
1. ๐ **์ฝ ์ด๋ฆ**: ์ ํํ ์ฝ ์ด๋ฆ
2. ๐ฏ **ํจ๋ฅ**: ์ด ์ฝ์ด ๋ฌด์์ ์น๋ฃํ๊ณ ์ด๋ป๊ฒ ๋์์ด ๋๋์ง
3. โ ๏ธ **๋ถ์์ฉ**: ์ฃผ์ํด์ผ ํ ๋ถ์์ฉ๋ค
4. ๐ก **๋ณต์ฉ ๋ฐฉ๋ฒ**: ์ธ์ , ์ด๋ป๊ฒ ๋จน์ด์ผ ํ๋์ง (์์ /์ํ, ํ๋ฃจ ๋ช ๋ฒ ๋ฑ)
5. ๐ซ **์ฃผ์์ฌํญ**: ์ด ์ฝ๊ณผ ํจ๊ป ๋จน์ผ๋ฉด ์ ๋๋ ๊ฒ๋ค (์์, ๋ค๋ฅธ ์ฝ ๋ฑ)
**์คํ์ผ ๊ฐ์ด๋:**
- ์ด๋ชจ์ง๋ฅผ ์ ๊ทน ํ์ฉํ์ฌ ์ฌ๋ฏธ์๊ฒ ์์ฑ
- ํ ๋จธ๋ ํ ์๋ฒ์ง๋ ์ด๋ฑํ์๋ ์ดํดํ ์ ์๋ ์ฌ์ด ๋จ์ด ์ฌ์ฉ
- ๊ฐ ์ฝ๋ง๋ค ๊ตฌ๋ถ์ ์ผ๋ก ๊ตฌ๋ถ
- ์น๊ทผํ๊ณ ๋ฐ๋ปํ ๋งํฌ ์ฌ์ฉ
- ๋งํฌ๋ค์ด ํ์์ผ๋ก ์์ฑ
์์ํด์ฃผ์ธ์!"""
messages = [
{"role": "user", "content": analysis_prompt}
]
input_text = MED_TOKENIZER.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = MED_TOKENIZER(input_text, return_tensors="pt").to(MED_MODEL.device)
with torch.no_grad():
outputs = MED_MODEL.generate(
**inputs,
max_new_tokens=768,
temperature=0.7,
top_p=0.9,
do_sample=True
)
analysis_text = MED_TOKENIZER.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
analysis_time = time.time() - analysis_start
total_time = time.time() - start_time
print(f"โฑ๏ธ Medical analysis took {analysis_time:.2f}s")
print(f"โฑ๏ธ Total processing time: {total_time:.2f}s")
return ocr_text.strip(), analysis_text.strip()
except Exception as e:
raise Exception(f"๋ถ์ ์ค๋ฅ: {str(e)}")
def extract_medications_from_text(text: str) -> List[str]:
"""Stage 2: Qwen2.5๋ก ํ
์คํธ์์ ์ฝ ์ด๋ฆ๋ง ์ถ์ถ"""
try:
messages = [
{
"role": "system",
"content": "You are a medical text analyzer. Extract only medication names from the given text and return them as a JSON array. Return ONLY valid JSON format."
},
{
"role": "user",
"content": f"Extract all medication names from this text:\n\n{text}\n\nReturn format: {{\"medications\": [\"name1\", \"name2\"]}}"
}
]
prompt = LLM_TOKENIZER.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
inputs = LLM_TOKENIZER(prompt, return_tensors="pt").to(LLM_MODEL.device)
with torch.no_grad():
outputs = LLM_MODEL.generate(
**inputs,
max_new_tokens=512,
temperature=0.3,
top_p=0.9,
do_sample=True,
pad_token_id=LLM_TOKENIZER.eos_token_id,
)
response = LLM_TOKENIZER.decode(outputs[0], skip_special_tokens=True)
# Extract assistant response (Qwen format)
if "<|im_start|>assistant" in response:
response = response.split("<|im_start|>assistant")[-1]
response = response.replace("<|im_end|>", "").strip()
# Parse JSON
json_match = re.search(r'\{.*?\}', response, re.DOTALL)
if json_match:
data = json.loads(json_match.group(0))
medications = data.get("medications", [])
if isinstance(medications, list) and medications:
return [str(m).strip() for m in medications if str(m).strip()]
return ["์ฝ ์ด๋ฆ์ ์ฐพ์ง ๋ชปํ์ต๋๋ค."]
except Exception as e:
raise Exception(f"LLM ๋ถ์ ์ค๋ฅ: {str(e)}")
@spaces.GPU(duration=120)
def extract_medication_names(image: Image.Image) -> Tuple[str, List[str]]:
"""2๋จ๊ณ ํ์ดํ๋ผ์ธ: OCR โ LLM ๋ถ์"""
try:
# Stage 1: OCR๋ก ํ
์คํธ ์ถ์ถ
extracted_text = extract_text_from_image(image)
if not extracted_text:
return "", ["ํ
์คํธ๋ฅผ ์ถ์ถํ์ง ๋ชปํ์ต๋๋ค."]
# Stage 2: LLM์ผ๋ก ์ฝ ์ด๋ฆ ์ถ์ถ
medications = extract_medications_from_text(extracted_text)
return extracted_text, medications
except Exception as e:
return "", [f"์ค๋ฅ ๋ฐ์: {str(e)}"]
def format_results(extracted_text: str, medications: List[str]) -> Tuple[str, str]:
"""๊ฒฐ๊ณผ๋ฅผ ํฌ๋งทํ
"""
# ์ถ์ถ๋ ์ ์ฒด ํ
์คํธ
text_output = f"### ๐ ์ถ์ถ๋ ํ
์คํธ\n\n```\n{extracted_text}\n```"
# ์ฝ ์ด๋ฆ ๋ฆฌ์คํธ
if not medications or medications[0].startswith("์ค๋ฅ") or medications[0].startswith("์ฝ ์ด๋ฆ์ ์ฐพ์ง") or medications[0].startswith("ํ
์คํธ๋ฅผ"):
med_output = f"### โ ๏ธ {medications[0] if medications else '์ฝ ์ด๋ฆ์ ์ฐพ์ง ๋ชปํ์ต๋๋ค.'}"
else:
med_output = f"### ๐ ๊ฒ์ถ๋ ์ฝ๋ฌผ ({len(medications)}๊ฐ)\n\n"
for idx, med_name in enumerate(medications, 1):
med_output += f"{idx}. **{med_name}**\n"
return text_output, med_output
def run_analysis(image: Optional[Image.Image], progress=gr.Progress()):
"""๋ฉ์ธ ๋ถ์ ํ์ดํ๋ผ์ธ: OCR + ์ฝ ์ ๋ณด ๋ถ์"""
if image is None:
return "๐ท ์ฝ ๋ดํฌ๋ ์ฒ๋ฐฉ์ ์ฌ์ง์ ์
๋ก๋ํด์ฃผ์ธ์.", ""
progress(0.3, desc="๐ธ 1๋จ๊ณ: OCR ํ
์คํธ ์ถ์ถ ์ค...")
progress(0.6, desc="๐ค 2๋จ๊ณ: ์ฝ ์ ๋ณด ๋ถ์ ์ค...")
try:
ocr_text, analysis = analyze_medication_image(image)
progress(1.0, desc="โ
์๋ฃ!")
ocr_output = f"### ๐ ์ถ์ถ๋ ํ
์คํธ\n\n```\n{ocr_text}\n```"
analysis_output = f"### ๐ ์ฝ ์ ๋ณด ์ค๋ช
\n\n{analysis}"
return ocr_output, analysis_output
except Exception as e:
return f"### โ ๏ธ ์ค๋ฅ ๋ฐ์\n\n{str(e)}", ""
# ์ฌํํ CSS
CUSTOM_CSS = """
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap');
:root {
--primary: #6366f1;
--secondary: #8b5cf6;
}
body {
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
font-family: 'Inter', -apple-system, BlinkMacSystemFont, sans-serif;
}
.gradio-container {
max-width: 900px !important;
margin: auto;
background: rgba(255, 255, 255, 0.98);
border-radius: 24px;
box-shadow: 0 25px 50px -12px rgba(0, 0, 0, 0.3);
padding: 40px;
}
.hero {
text-align: center;
padding: 30px 20px;
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
border-radius: 20px;
color: white;
margin-bottom: 30px;
}
.hero h1 {
font-size: 2.5rem;
font-weight: 700;
margin-bottom: 10px;
}
.hero p {
font-size: 1.1rem;
opacity: 0.95;
}
.upload-section {
background: white;
border-radius: 16px;
padding: 30px;
box-shadow: 0 4px 6px rgba(0, 0, 0, 0.07);
margin-bottom: 20px;
}
.result-section {
background: white;
border-radius: 16px;
padding: 30px;
box-shadow: 0 4px 6px rgba(0, 0, 0, 0.07);
min-height: 200px;
}
.analyze-btn button {
background: linear-gradient(135deg, var(--primary), var(--secondary)) !important;
color: white !important;
font-weight: 600 !important;
font-size: 1.1rem !important;
padding: 18px 40px !important;
border-radius: 12px !important;
border: none !important;
box-shadow: 0 10px 20px -5px rgba(99, 102, 241, 0.5) !important;
transition: all 0.3s ease !important;
}
.analyze-btn button:hover {
transform: translateY(-2px) !important;
box-shadow: 0 15px 30px -5px rgba(99, 102, 241, 0.6) !important;
}
.gr-image {
border-radius: 12px !important;
}
"""
HERO_HTML = """
<div class="hero">
<h1>๐ ์ฐ๋ฆฌ ๊ฐ์กฑ ์ฝ ๋์ฐ๋ฏธ</h1>
<p>์ฝ๋ดํฌ/์ฒ๋ฐฉ์ ์ฌ์ง์์ ์ฝ ์ ๋ณด๋ฅผ ์ฝ๊ณ ์ฌ๋ฏธ์๊ฒ ์๋ ค๋๋ ค์!</p>
</div>
"""
# Gradio ์ธํฐํ์ด์ค
with gr.Blocks(theme=gr.themes.Soft(), css=CUSTOM_CSS) as demo:
gr.HTML(HERO_HTML)
with gr.Column(elem_classes=["upload-section"]):
gr.Markdown("### ๐ธ ์ฌ์ง ์
๋ก๋")
image_input = gr.Image(type="pil", label="์ฝ๋ดํฌ ๋๋ ์ฒ๋ฐฉ์ ์ฌ์ง", height=350)
analyze_button = gr.Button("๐ ์ฝ ์ ๋ณด ๋ถ์ํ๊ธฐ", elem_classes=["analyze-btn"], size="lg")
with gr.Row():
with gr.Column(elem_classes=["result-section"]):
gr.Markdown("### ๐ 1๋จ๊ณ: ์ถ์ถ๋ ํ
์คํธ")
ocr_output = gr.Markdown("OCR๋ก ์ถ์ถ๋ ํ
์คํธ๊ฐ ์ฌ๊ธฐ ํ์๋ฉ๋๋ค.")
with gr.Column(elem_classes=["result-section"]):
gr.Markdown("### ๐ 2๋จ๊ณ: ์ฌ์ด ์ฝ ์ค๋ช
")
analysis_output = gr.Markdown("๋
ธ์ธ๊ณผ ์ด๋ฆฐ์ด๋ ์ดํดํ๊ธฐ ์ฌ์ด ์ฝ ์ ๋ณด๊ฐ ์ฌ๊ธฐ ํ์๋ฉ๋๋ค.")
analyze_button.click(
run_analysis,
inputs=image_input,
outputs=[ocr_output, analysis_output],
)
gr.Markdown("""
---
**โน๏ธ ์ฌ์ฉ ๋ฐฉ๋ฒ**
1. ์ฝ ๋ดํฌ๋ ์ฒ๋ฐฉ์ ์ฌ์ง์ ์
๋ก๋ํ์ธ์
2. '์ฝ ์ ๋ณด ๋ถ์ํ๊ธฐ' ๋ฒํผ์ ํด๋ฆญํ์ธ์
3. ์ผ์ชฝ์๋ ์ถ์ถ๋ ํ
์คํธ, ์ค๋ฅธ์ชฝ์๋ ์ฌ์ด ์ค๋ช
์ด ๋ํ๋ฉ๋๋ค!
**โ ๏ธ ์ฃผ์์ฌํญ**
- ์ด ์ฑ์ ์ฐธ๊ณ ์ฉ์ด๋ฉฐ, ์ค์ ๋ณต์ฝ์ ๋ฐ๋์ ์์ฌ๋ ์ฝ์ฌ์ ์ง์๋ฅผ ๋ฐ๋ฅด์ธ์
- AI๊ฐ ์์ฑํ ์ ๋ณด์ด๋ฏ๋ก ์ ํํ์ง ์์ ์ ์์ต๋๋ค
**๐ค ๊ธฐ์ ์คํ**
- EasyOCR (ํ๊ธ+์์ด, ์ด๊ณ ์ OCR)
- Google Gemma-2-2B-IT (8bit ์์ํ, ๋น ๋ฅธ ์๋ฃ ์ ๋ณด ๋ถ์)
**๐ ์ค์ ๋ฐฉ๋ฒ**
- Hugging Face Spaces์ Settings โ Repository secrets์์ `HF_TOKEN` ์ถ๊ฐ ํ์
""")
if __name__ == "__main__":
demo.queue().launch()
|