|
|
import json |
|
|
import re |
|
|
from typing import Any, Dict, List, Optional, Sequence |
|
|
|
|
|
import easyocr |
|
|
import gradio as gr |
|
|
import numpy as np |
|
|
import torch |
|
|
from PIL import Image, ImageDraw |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline |
|
|
|
|
|
|
|
|
|
|
|
OCR_LANGS = ["ko", "en"] |
|
|
LLM_MODEL_ID = "Qwen/Qwen2.5-0.5B-Instruct" |
|
|
|
|
|
|
|
|
def _load_ocr(): |
|
|
use_gpu = torch.cuda.is_available() |
|
|
return easyocr.Reader(OCR_LANGS, gpu=use_gpu) |
|
|
|
|
|
|
|
|
ocr_reader = _load_ocr() |
|
|
|
|
|
|
|
|
def _load_llm(): |
|
|
device_map = "auto" if torch.cuda.is_available() else None |
|
|
dtype = torch.float16 if torch.cuda.is_available() else torch.float32 |
|
|
model = AutoModelForCausalLM.from_pretrained(LLM_MODEL_ID, device_map=device_map, torch_dtype=dtype) |
|
|
if device_map is None: |
|
|
model = model.to(torch.device("cpu")) |
|
|
tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL_ID) |
|
|
return model, tokenizer |
|
|
|
|
|
|
|
|
LLM_MODEL, LLM_TOKENIZER = _load_llm() |
|
|
|
|
|
|
|
|
TIME_KEYWORDS = [ |
|
|
"์์นจ", |
|
|
"์ ์ฌ", |
|
|
"์ ๋
", |
|
|
"์ทจ์นจ", |
|
|
"์๊ธฐ", |
|
|
"์์ ", |
|
|
"์ํ", |
|
|
"์๊ฐ", |
|
|
"๊ธฐ์", |
|
|
] |
|
|
|
|
|
|
|
|
MED_KNOWLEDGE: Sequence[Dict[str, Any]] = [ |
|
|
{ |
|
|
"keywords": ["ํ์ด๋ ๋", "์์ธํธ์๋ฏธ๋
ธํ", "acetaminophen"], |
|
|
"category": "์งํตยทํด์ด์ ", |
|
|
"what_it_does": "๋ชธ์ด์ด๋ ๊ฐ๊ธฐ๋ก ์ด์ด ๋๊ฑฐ๋ ๋จธ๋ฆฌ๊ฐ ์ํ ๋ ํต์ฆ๊ณผ ์ด์ ๋ฎ์ถฐ ์ค๋๋ค.", |
|
|
"example": "์: ์ํ์ํ ์ค๋น๋ก ๊ธด์ฅํ๋๋ฐ ๋จธ๋ฆฌ๊ฐ ์ง๋๊ฑฐ๋ฆด ๋, ํ ์ ๋ณต์ฉํ๋ฉด ํต์ฆ์ด ์ค์ด๋ญ๋๋ค.", |
|
|
"tip": "์์ ๋ถ๋ด์ ์ค์ด๊ธฐ ์ํด ๊ฐ๋จํ ๊ฐ์๊ณผ ํจ๊ป ๋ฌผ๊ณผ ๋ณต์ฉํ๊ณ , ํ๋ฃจ ์ด ๋ณต์ฉ ํ์(์ผ๋ฐ์ ์ผ๋ก 4ํ ์ดํ)๋ฅผ ๋๊ธฐ์ง ๋ง์ธ์.", |
|
|
}, |
|
|
{ |
|
|
"keywords": ["์ด๋ถํ๋กํ", "๋ถ๋ฃจํ", "ibuprofen"], |
|
|
"category": "์งํตยท์์ผ์ ", |
|
|
"what_it_does": "๋ชธ์ ์ผ์ฆ์ ๊ฐ๋ผ์ํ๊ณ ํต์ฆ์ ์ํํด์ ๊ทผ์กํต์ด๋ ์นํต์ ์์ฃผ ์ฌ์ฉ๋ฉ๋๋ค.", |
|
|
"example": "์: ์ฒด์ก ์๊ฐ์ ๋ฌด๋ฆ์ ์ด์ง ์์์ ๋ ๋ถ๊ธฐ์ ์ํ์ ์ค์ฌ ์ค๋๋ค.", |
|
|
"tip": "์ํ์ ๋ณต์ฉํ๋ฉด ์ ์ฐ๋ฆผ์ ์ค์ผ ์ ์๊ณ , ๋ค๋ฅธ ์์ผ์งํต์ ์๋ ์๊ฐ ๊ฐ๊ฒฉ์ ๋์ธ์.", |
|
|
}, |
|
|
{ |
|
|
"keywords": ["์์", "์ธํฐ๋ฆฌ์ง", "cetirizine", "์ง๋ฅดํ
"], |
|
|
"category": "์๋ ๋ฅด๊ธฐ ์ํ์ ", |
|
|
"what_it_does": "์ฝ๊ฐ ๊ฐ์ง๊ฑฐ๋ฆฌ๊ฑฐ๋ ํผ๋ถ๊ฐ ๊ฐ๋ ค์ธ ๋ ์๋ ๋ฅด๊ธฐ ๋ฐ์์ ๊ฐ๋ผ์ํ ์ค๋๋ค.", |
|
|
"example": "์: ๋ด์ฒ ๊ฝ๊ฐ๋ฃจ ๋๋ฌธ์ ๊ธฐ์นจ๊ณผ ์ฝง๋ฌผ์ด ๋์ฌ ๋ ์ฆ์์ ์ค์ฌ ์ค๋๋ค.", |
|
|
"tip": "์กธ๋ฆด ์ ์์ผ๋ ์ฒซ ๋ณต์ฉ ํ์๋ ์ด์ ์ด๋ ์ง์ค์ด ํ์ํ ํ๋์ ํผํ์ธ์.", |
|
|
}, |
|
|
{ |
|
|
"keywords": ["ํผ์คํ", "pancreatin", "์์ฅ", "์ํ์ "], |
|
|
"category": "์ํ์ ", |
|
|
"what_it_does": "๊ธฐ๋ฆ์ง ์์์ ๋จน๊ณ ๋ฐฐ๊ฐ ๋๋ถ๋ฃฉํ ๋ ์ํ๋ฅผ ๋์ ์์ ํธํ๊ฒ ํด ์ค๋๋ค.", |
|
|
"example": "์: ์นํจ์ ๋ง์ด ๋จน์ด ์์ด ๋๋ถ๋ฃฉํ ๋ ์์ ๊ฐ๋ณ๊ฒ ํด ์ค๋๋ค.", |
|
|
"tip": "์ํ์ ๋ณต์ฉํ๋ฉด ํจ๊ณผ๊ฐ ์ข์ผ๋ฉฐ, ๋ณตํต์ด ๊ณ์๋๋ฉด ๋ณ์์ ๋ฐฉ๋ฌธํ์ธ์.", |
|
|
}, |
|
|
{ |
|
|
"keywords": ["๋นํ๋ฏผ", "multivitamin", "vitamin"], |
|
|
"category": "์์์ ", |
|
|
"what_it_does": "๋ชธ์ ํ์ํ ๋นํ๋ฏผ์ ์ฑ์ ํผ๊ณคํจ์ ์ค์ด๊ณ ๋ฉด์ญ๋ ฅ์ ๋์ต๋๋ค.", |
|
|
"example": "์: ์ํ ์ค๋น๋ก ์ ์ ์ค์์ ๋ ๋ชธ์ด ์ง์น์ง ์๋๋ก ๋์์ค๋๋ค.", |
|
|
"tip": "ํ๋ฃจ ๊ถ์ฅ๋์ ์ง์ผ ๊พธ์คํ ๋ณต์ฉํ๋ฉด ๋ ํจ๊ณผ์ ์ด๋ฉฐ, ๋ฌผ๊ณผ ํจ๊ป ์ผํค์ธ์.", |
|
|
}, |
|
|
] |
|
|
|
|
|
|
|
|
def _extract_time_slots(text: str) -> List[str]: |
|
|
slots = [] |
|
|
for kw in TIME_KEYWORDS: |
|
|
if kw in text: |
|
|
slots.append(kw) |
|
|
|
|
|
for match in re.findall(r"(\d{1,2}[:์]\d{0,2})", text): |
|
|
norm = match.replace("์", ":") |
|
|
if norm.endswith(":"): |
|
|
norm += "00" |
|
|
if norm not in slots: |
|
|
slots.append(norm) |
|
|
return slots |
|
|
|
|
|
|
|
|
STOPWORDS = {"์ฉ๋ฒ", "์ฉ๋", "๋ณต์ฉ", "๋ฐฉ๋ฒ", "์ฝ", "์ "} |
|
|
|
|
|
|
|
|
def _extract_medications(text: str) -> List[Dict[str, Optional[str]]]: |
|
|
meds: List[Dict[str, Optional[str]]] = [] |
|
|
pattern = re.compile( |
|
|
r"([๊ฐ-ํฃA-Za-z]{2,})[\sยท]*(\d+[\./]?\d*\s*(?:mg|mL|ML|ml|์ |์บก์))?" |
|
|
) |
|
|
seen: set[str] = set() |
|
|
for match in pattern.finditer(text): |
|
|
name = match.group(1) |
|
|
if name in STOPWORDS or len(name) <= 1: |
|
|
continue |
|
|
if any(sw in name for sw in STOPWORDS): |
|
|
continue |
|
|
name_norm = name.strip() |
|
|
if name_norm in seen: |
|
|
continue |
|
|
seen.add(name_norm) |
|
|
dose = match.group(2).strip() if match.group(2) else None |
|
|
meds.append({"name": name_norm, "dose": dose}) |
|
|
return meds |
|
|
|
|
|
|
|
|
def parse_fields(raw: str) -> Dict[str, Any]: |
|
|
"""Extract drug name and dosage information from OCR text.""" |
|
|
collapsed = raw.replace("\n", " ") |
|
|
collapsed = re.sub(r"\s+", " ", collapsed) |
|
|
|
|
|
medications = _extract_medications(collapsed) |
|
|
|
|
|
first = medications[0] if medications else {"name": None, "dose": None} |
|
|
drug_name = first.get("name") |
|
|
dose_per_intake = first.get("dose") |
|
|
|
|
|
times_per_day: Optional[int] = None |
|
|
times_match = re.search(r"(?:1์ผ|ํ๋ฃจ)\s*(\d+)\s*ํ", collapsed) |
|
|
if times_match: |
|
|
times_per_day = int(times_match.group(1)) |
|
|
|
|
|
time_slots = _extract_time_slots(collapsed) |
|
|
|
|
|
return { |
|
|
"drug_name": drug_name, |
|
|
"dose_per_intake": dose_per_intake, |
|
|
"times_per_day": times_per_day, |
|
|
"time_slots": time_slots or None, |
|
|
"medications": medications, |
|
|
} |
|
|
|
|
|
|
|
|
def ocr_and_parse(image: Image.Image) -> Dict[str, Any]: |
|
|
np_img = np.array(image.convert("RGB")) |
|
|
results = ocr_reader.readtext(np_img, detail=1, paragraph=False) |
|
|
|
|
|
segments: List[Dict[str, Any]] = [] |
|
|
lines: List[str] = [] |
|
|
for bbox, text, confidence in results: |
|
|
cleaned = text.strip() |
|
|
if not cleaned: |
|
|
continue |
|
|
lines.append(cleaned) |
|
|
segments.append({ |
|
|
"text": cleaned, |
|
|
"confidence": float(confidence), |
|
|
"bbox": bbox, |
|
|
}) |
|
|
|
|
|
raw_text = "\n".join(lines) |
|
|
fields = parse_fields(raw_text) |
|
|
|
|
|
warnings: List[str] = [] |
|
|
if not fields["drug_name"]: |
|
|
warnings.append("์ฝ ์ด๋ฆ ์ธ์์ด ๋ถํ์คํฉ๋๋ค.") |
|
|
if not fields["times_per_day"]: |
|
|
warnings.append("1์ผ ํ์๋ฅผ ์ฐพ์ง ๋ชปํ์ต๋๋ค (์: 1์ผ 3ํ).") |
|
|
|
|
|
return { |
|
|
"raw_text": raw_text, |
|
|
"fields": fields, |
|
|
"warnings": warnings, |
|
|
"segments": segments, |
|
|
} |
|
|
|
|
|
|
|
|
def render_card(fields: Dict[str, Any]) -> Image.Image: |
|
|
width, height = 720, 400 |
|
|
img = Image.new("RGB", (width, height), "white") |
|
|
draw = ImageDraw.Draw(img) |
|
|
|
|
|
header_text = "์ค๋ ๋ณต์ฉ ์ผ์ " |
|
|
draw.rectangle((0, 0, width, 60), fill=(230, 240, 255)) |
|
|
draw.text((24, 18), header_text, fill=(0, 0, 0)) |
|
|
|
|
|
y = 90 |
|
|
|
|
|
def add_line(label: str, value: Optional[str]): |
|
|
nonlocal y |
|
|
draw.text((24, y), label, fill=(60, 60, 60)) |
|
|
display = value if value else "-" |
|
|
draw.text((180, y), f": {display}", fill=(0, 0, 0)) |
|
|
y += 34 |
|
|
|
|
|
add_line("์ฝ ์ด๋ฆ", fields.get("drug_name")) |
|
|
add_line("1ํ ์ฉ๋", fields.get("dose_per_intake")) |
|
|
add_line("1์ผ ํ์", str(fields.get("times_per_day") or "")) |
|
|
|
|
|
slots = fields.get("time_slots") or [] |
|
|
add_line("์๊ฐ๋", ", ".join(slots) if slots else None) |
|
|
|
|
|
footer = "โป ์๋ฃ์ง ์ฒ๋ฐฉ์ด ์ฐ์ ์ด๋ฉฐ, ๋ณธ ์ฑ์ ์ฐธ๊ณ ์ฉ์
๋๋ค." |
|
|
draw.text((24, height - 60), footer, fill=(120, 120, 120)) |
|
|
return img |
|
|
|
|
|
|
|
|
def to_csv_row(output: Dict[str, Any]) -> str: |
|
|
fields = output["fields"] |
|
|
row = [ |
|
|
fields.get("drug_name") or "", |
|
|
fields.get("dose_per_intake") or "", |
|
|
str(fields.get("times_per_day") or ""), |
|
|
";".join(fields.get("time_slots") or []), |
|
|
] |
|
|
return ",".join(row) |
|
|
|
|
|
|
|
|
def _match_knowledge(name: str) -> Optional[Dict[str, Any]]: |
|
|
lowered = name.lower() |
|
|
for info in MED_KNOWLEDGE: |
|
|
for kw in info["keywords"]: |
|
|
if kw.lower() in lowered or lowered in kw.lower(): |
|
|
return info |
|
|
return None |
|
|
|
|
|
|
|
|
def build_kb_explanations(output: Dict[str, Any]) -> str: |
|
|
meds = output["fields"].get("medications") or [] |
|
|
if not meds: |
|
|
return ( |
|
|
"### ์ฝ ์ค๋ช
\n" |
|
|
"- ์ฝ ์ด๋ฆ์ ์ ํํ ์ธ์ํ์ง ๋ชปํ์ด์. ์ฌ์ง์ ๋ค์ ์ฐ๊ฑฐ๋ ์ฝ์ฌ์๊ฒ ์ง์ ํ์ธํด ์ฃผ์ธ์.\n" |
|
|
"\n> โ ๏ธ ์๋ฃ์ง ์ฒ๋ฐฉ๊ณผ ๋ณต์ฝ ์ง์๊ฐ ๊ฐ์ฅ ์ฐ์ ์
๋๋ค." |
|
|
) |
|
|
|
|
|
lines = ["### ์ฝ๊ฒ ์์๋ณด๋ ์ฝ ์ค๋ช
"] |
|
|
for med in meds: |
|
|
name = med.get("name") or "์ด๋ฆ ๋ฏธํ์ธ" |
|
|
info = _match_knowledge(name) if name else None |
|
|
dose = med.get("dose") |
|
|
if info: |
|
|
lines.append( |
|
|
f"- **{name}** ({info['category']})" |
|
|
) |
|
|
if dose: |
|
|
lines.append(f" - ์ฝ ๋ดํฌ์ ์ ํ ์ฉ๋: `{dose}`") |
|
|
lines.append(f" - ํ๋ ์ผ: {info['what_it_does']}") |
|
|
lines.append(f" - ์คํ์ ์์: {info['example']}") |
|
|
lines.append(f" - ๋ณต์ฉ ํ: {info['tip']}") |
|
|
else: |
|
|
lines.append(f"- **{name}**") |
|
|
if dose: |
|
|
lines.append(f" - ์ฝ ๋ดํฌ ์ฉ๋: `{dose}`") |
|
|
lines.append( |
|
|
" - ์์ง ๋ฐ์ดํฐ๊ฐ ์์ด์. ์ฝ ์ด๋ฆ์ ๋ค์ ํ์ธํ๊ฑฐ๋ ์ฝ์ฌ์๊ฒ ๋ฌผ์ด๋ณด์ธ์." |
|
|
) |
|
|
|
|
|
lines.append("\n> โ ๏ธ ์ค์ ๋ณต์ฝ์ ์์ฌยท์ฝ์ฌ์ ์ง์์ ๋ฐ๋์ ๋ฐ๋ฅด์ธ์.") |
|
|
return "\n".join(lines) |
|
|
|
|
|
|
|
|
def generate_llm_explanations(output: Dict[str, Any]) -> str: |
|
|
meds = output["fields"].get("medications") or [] |
|
|
if not meds: |
|
|
return ( |
|
|
"์ฝ ์ด๋ฆ์ ์ ๋๋ก ์ธ์ํ์ง ๋ชปํ์ด์. ์ฌ์ง์ ๋ค์ ์ฐ๊ฑฐ๋ ์ฝ์ฌ์๊ฒ ์ง์ ํ์ธํด ์ฃผ์ธ์." |
|
|
) |
|
|
|
|
|
med_lines = [] |
|
|
for idx, med in enumerate(meds, 1): |
|
|
name = med.get("name") or "์ด๋ฆ ๋ฏธํ์ธ" |
|
|
dose = med.get("dose") or "์ฉ๋ ์ ๋ณด ์์" |
|
|
med_lines.append(f"{idx}. {name} โ {dose}") |
|
|
|
|
|
context = "\n".join(med_lines) |
|
|
raw_text = output.get("raw_text", "") |
|
|
|
|
|
system_prompt = ( |
|
|
"๋น์ ์ ์ฝ์ฌ ์ ์๋์
๋๋ค. ์ด๋ ค์ด ์ํ ์ฉ์ด๋ฅผ ์ฐ์ง ๋ง๊ณ , ์คํ์๋ ์ดํดํ ์ ์๋ ๋งํฌ๋ก ์น์ ํ๊ฒ ์ค๋ช
ํ์ธ์." |
|
|
) |
|
|
user_prompt = ( |
|
|
"๋ค์์ ์ฝ๋ดํฌ์์ OCR๋ก ์ถ์ถํ ์ ์ฒด ํ
์คํธ์
๋๋ค. ์ฝ ์ด๋ฆ๊ณผ ๋ณต์ฉ ์ง์๋ฅผ ๊ธฐ๋ฐ์ผ๋ก ๊ฐ ์ฝ์ ์ ๋ณด๋ฅผ ์์ฃผ ์ฝ๊ฒ ์ ๋ฆฌํด ์ฃผ์ธ์.\n" |
|
|
"์๊ตฌ ์ฌํญ:\n" |
|
|
"1. ๊ฐ ์ฝ๋ง๋ค ์๋ ํญ๋ชฉ์ bullet ํ์์ผ๋ก ์์ฑํฉ๋๋ค.\n" |
|
|
" - ์ฝ ์ด๋ฆ: (๊ฐ๋ฅํ๋ฉด ํ๊ธ/์๋ฌธ ๋ณ๊ธฐ)\n" |
|
|
" - ์ด๋ค ์ฝ์ธ์ง ํ ์ค ์ค๋ช
\n" |
|
|
" - ๋ณต์ฉ ์์: ์ธ์ , ์ด๋ค ์ํฉ์์ ๋ณต์ฉํ๋ฉด ์ข์์ง ์์\n" |
|
|
" - ๋ณต์ฉ ๋ฐฉ๋ฒ ์์: 1ํ ์ฉ๋/ํ๋ฃจ ํ์๊ฐ ์๋ค๋ฉด ์ธ๊ธ\n" |
|
|
" - ๋ถ์์ฉ ๋๋ ์ฃผ์์ฌํญ: ํํ ๋ถ์์ฉ, ํผํด์ผ ํ ํ๋\n" |
|
|
"2. ์ด๋ ค์ด ์ํ ์ฉ์ด๋ ํผํ๊ณ , ์คํ์๋ ์ดํดํ ์ ์๋ ๋งํฌ๋ก ์์ฑํฉ๋๋ค.\n" |
|
|
"3. ์ฝ ์ด๋ฆ์ ํ์คํ ๋ชจ๋ฅด๋ฉด โ์ด๋ฆ ๋ฏธํ์ธโ์ด๋ผ๊ณ ์ฐ๊ณ , ์ฝ์ฌ์๊ฒ ํ์ธํ๋ผ๊ณ ์๋ดํฉ๋๋ค.\n" |
|
|
"4. ๋ง์ง๋ง ๋ฌธ๋จ์ ๋ฐ๋์ โ์ค์ ๋ณต์ฝ์ ์์ฌยท์ฝ์ฌ์ ์ง์๋ฅผ ๋ฐ๋ฅด์ธ์โ ๋ฌธ์ฅ์ ํฌํจํ์ธ์.\n" |
|
|
f"\n์ฝ ๋ชฉ๋ก(์ถ์ถ ์์ฝ):\n{context}\n\nOCR ์๋ฌธ ์ ์ฒด:\n{raw_text}\n" |
|
|
) |
|
|
|
|
|
messages = [ |
|
|
{"role": "system", "content": system_prompt}, |
|
|
{"role": "user", "content": user_prompt}, |
|
|
] |
|
|
|
|
|
input_ids = LLM_TOKENIZER.apply_chat_template( |
|
|
messages, |
|
|
add_generation_prompt=True, |
|
|
return_tensors="pt", |
|
|
) |
|
|
input_ids = input_ids.to(LLM_MODEL.device) |
|
|
|
|
|
with torch.no_grad(): |
|
|
output_ids = LLM_MODEL.generate( |
|
|
input_ids, |
|
|
max_new_tokens=480, |
|
|
temperature=0.7, |
|
|
top_p=0.9, |
|
|
do_sample=True, |
|
|
eos_token_id=LLM_TOKENIZER.eos_token_id, |
|
|
) |
|
|
|
|
|
generated_ids = output_ids[0][input_ids.shape[1]:] |
|
|
text = LLM_TOKENIZER.decode(generated_ids, skip_special_tokens=True).strip() |
|
|
return text |
|
|
|
|
|
|
|
|
def build_explanations(output: Dict[str, Any]) -> str: |
|
|
try: |
|
|
llm_text = generate_llm_explanations(output) |
|
|
if llm_text: |
|
|
return llm_text |
|
|
except Exception as err: |
|
|
print(f"[WARN] LLM generation failed: {err}", flush=True) |
|
|
return build_kb_explanations(output) |
|
|
|
|
|
|
|
|
def format_warnings(warnings: List[str]) -> str: |
|
|
if not warnings: |
|
|
return "โ
์ธ์๋ ์ ๋ณด๊ฐ ์ถฉ๋ถํด์. ๋ณต์ฝ ์๊ฐ๋ง ์ ์ง์ผ ์ฃผ์ธ์." |
|
|
lines = ["### ํ์ธํด ์ฃผ์ธ์"] |
|
|
for warn in warnings: |
|
|
lines.append(f"- {warn}") |
|
|
lines.append("\n> ์๋ฃ์ง์ ์ง์๊ฐ ๊ฐ์ฅ ์ ํํฉ๋๋ค.") |
|
|
return "\n".join(lines) |
|
|
|
|
|
|
|
|
def run_pipeline(image: Optional[Image.Image]): |
|
|
if image is None: |
|
|
return ( |
|
|
"์ด๋ฏธ์ง๋ฅผ ์
๋ก๋ํ์ธ์.", |
|
|
None, |
|
|
None, |
|
|
"์ด๋ฏธ์ง๋ฅผ ๋จผ์ ์
๋ก๋ํด ์ฃผ์ธ์.", |
|
|
"๐ท ์ฝ ๋ดํฌ ์ฌ์ง์ ์ฌ๋ฆฌ๋ฉด ์ธ์์ด ์์๋ผ์.", |
|
|
"", |
|
|
) |
|
|
|
|
|
output = ocr_and_parse(image) |
|
|
card = render_card(output["fields"]) |
|
|
csv_row = to_csv_row(output) |
|
|
json_text = json.dumps(output, ensure_ascii=False, indent=2) |
|
|
explanations = build_explanations(output) |
|
|
warnings_md = format_warnings(output.get("warnings", [])) |
|
|
return json_text, card, csv_row, explanations, warnings_md, output.get("raw_text", "") |
|
|
|
|
|
|
|
|
CUSTOM_CSS = """ |
|
|
body {background: radial-gradient(circle at top left, #f5f0ff 0%, #fff7ec 60%, #ffffff 100%);} |
|
|
.gradio-container {max-width: 1180px !important; margin: auto; font-family: 'Noto Sans KR', sans-serif;} |
|
|
.hero { |
|
|
background: linear-gradient(120deg, rgba(123, 97, 255, 0.12), rgba(255, 207, 117, 0.18)); |
|
|
border-radius: 28px; |
|
|
padding: 36px 44px; |
|
|
box-shadow: 0 20px 40px rgba(66, 46, 138, 0.08); |
|
|
margin-bottom: 32px; |
|
|
} |
|
|
.hero h1 {font-size: 2.4rem; font-weight: 700; color: #1f1c3b; margin-bottom: 12px;} |
|
|
.hero p {color: #514c7b; font-size: 1.05rem; line-height: 1.6; max-width: 640px;} |
|
|
.glass-panel {background: rgba(255, 255, 255, 0.72); backdrop-filter: blur(18px); border-radius: 26px; padding: 28px; box-shadow: 0 12px 32px rgba(80, 60, 160, 0.12);} |
|
|
.panel-title {font-weight: 700; font-size: 1.2rem; margin-bottom: 18px; color: #2f2355;} |
|
|
.primary-btn button {background: linear-gradient(120deg, #7c62ff, #ffa74d); border: none; color: white; font-weight: 600; border-radius: 999px; padding: 12px 22px; box-shadow: 0 12px 24px rgba(124, 98, 255, 0.25);} |
|
|
.primary-btn button:hover {opacity: 0.95; transform: translateY(-1px);} |
|
|
.output-card {background: rgba(255, 255, 255, 0.88); border-radius: 22px; padding: 24px; box-shadow: inset 0 0 0 1px rgba(124, 98, 255, 0.08), 0 14px 30px rgba(49, 32, 114, 0.12);} |
|
|
.notice {background: rgba(255, 247, 226, 0.9); border-radius: 18px; padding: 18px; color: #7a4b00; box-shadow: inset 0 0 0 1px rgba(255, 193, 96, 0.3);} |
|
|
.csv-box textarea {font-family: 'JetBrains Mono', monospace;} |
|
|
.gr-image {border-radius: 20px !important; box-shadow: 0 10px 20px rgba(60, 40, 120, 0.15);} |
|
|
.accordion {border-radius: 20px !important;} |
|
|
""" |
|
|
|
|
|
HERO_HTML = """ |
|
|
<div class="hero"> |
|
|
<h1>MedCard-KR ยท ์ฝ๋ดํฌ ํ ์ปท์ผ๋ก ์ดํดํ๋ ๋ณต์ฉ ์๋ด</h1> |
|
|
<p>์ฌ์ง ์ ์ฝ ์ด๋ฆ์ OCR๋ก ์ฝ์ด ๋ค์ด๊ณ , Qwen LLM์ด ์คํ์๋ ์ดํดํ ์ ์๋ ๋งํฌ๋ก ์ฝ์ ์ค๋ช
ํด ๋๋ฆฝ๋๋ค. |
|
|
๋ณต์ฉ ์ผ์ ์นด๋์ CSV๊น์ง ํ ๋ฒ์ ๋ฐ์ ๋ณด์ธ์.</p> |
|
|
</div> |
|
|
""" |
|
|
|
|
|
|
|
|
with gr.Blocks(theme=gr.themes.Soft(), css=CUSTOM_CSS) as demo: |
|
|
gr.HTML(HERO_HTML) |
|
|
with gr.Row(): |
|
|
with gr.Column(scale=4, elem_classes=["glass-panel"]): |
|
|
gr.Markdown("### 1. ์ฝ ๋ดํฌ ์ฌ์ง์ ์
๋ก๋ํ์ธ์") |
|
|
img_in = gr.Image(type="pil", label="์ฝ ๋ดํฌ/๋ผ๋ฒจ ์ฌ์ง", height=360) |
|
|
warn_md = gr.Markdown("๐ท ์ฝ ๋ดํฌ ์ฌ์ง์ ์ฌ๋ฆฌ๋ฉด ์ธ์์ด ์์๋ผ์.", elem_classes=["notice"]) |
|
|
btn = gr.Button("์ธ์ & ์ค๋ช
์์ฑ", elem_classes=["primary-btn"]) |
|
|
with gr.Column(scale=6, elem_classes=["glass-panel"]): |
|
|
gr.Markdown("### 2. ๊ฒฐ๊ณผ๋ฅผ ํ์ธํ์ธ์") |
|
|
explain_md = gr.Markdown("์ฌ๊ธฐ์ ์ฝ ์ค๋ช
์ด ํ์๋ฉ๋๋ค.", elem_classes=["output-card"]) |
|
|
raw_box = gr.Textbox(label="OCR ์๋ฌธ ํ
์คํธ", lines=5, interactive=False) |
|
|
card_out = gr.Image(type="pil", label="์ผ์ ์นด๋(๋ฏธ๋ฆฌ๋ณด๊ธฐ)") |
|
|
csv_box = gr.Textbox(label="CSV(์ฝ๋ช
,1ํ์ฉ๋,1์ผํ์,์๊ฐ๋)", lines=2, elem_classes=["csv-box"]) |
|
|
with gr.Accordion("์ธ๋ถ JSON ๊ฒฐ๊ณผ", open=False, elem_classes=["accordion"]): |
|
|
json_out = gr.Code(label="์ธ์ ๊ฒฐ๊ณผ(JSON)") |
|
|
|
|
|
btn.click( |
|
|
run_pipeline, |
|
|
inputs=img_in, |
|
|
outputs=[json_out, card_out, csv_box, explain_md, warn_md, raw_box], |
|
|
) |
|
|
|
|
|
gr.Markdown( |
|
|
""" |
|
|
> โน๏ธ **์ฃผ์**: ์ด ์๋น์ค๋ ์ฐธ๊ณ ์ฉ ๋๊ตฌ์ด๋ฉฐ, ์ค์ ๋ณต์ฝ์ ๋ฐ๋์ ์์ฌยท์ฝ์ฌ์ ์ง์์ ๋ฐ๋ผ ์ฃผ์ธ์. |
|
|
""" |
|
|
) |
|
|
|
|
|
|
|
|
if __name__ == "__main__": |
|
|
demo.queue().launch() |
|
|
|