src/inference.py

import torch
from paddleocr import PaddleOCR

# ── Load model ───────────────────────────────────────────
_model = None

def get_model(checkpoint: str = "best.pt"):
    global _model
    if _model is None:
        print(f"[INFO] Loading model from {checkpoint}...")
        _model = RTDETR(checkpoint)
    return _model

_orig_load = torch.load
def _safe_load(*args, **kwargs):
    kwargs.setdefault("weights_only", False)
    return _orig_load(*args, **kwargs)
torch.load = _safe_load
# ─────────────────────────────────────────────────────────

import cv2, json, os
from pathlib import Path
from ultralytics import RTDETR

# ── Device: M1 dùng MPS ──────────────────────────────────
DEVICE = (
    "mps" if torch.backends.mps.is_available()
    else "cpu"
)
print(f"[INFO] Device: {DEVICE}")

# ── Class config ─────────────────────────────────────────
CLASS_NAMES = ['note', 'part-drawing', 'table']

# Map sang tên chuẩn theo đề bài
CLASS_DISPLAY = {
    'note':         'Note',
    'part-drawing': 'PartDrawing',
    'table':        'Table',
}

COLORS = {
    'note':         (0,  165, 255),   # cam
    'part-drawing': (0,  200,   0),   # xanh lá
    'table':        (220,  0,   0),   # đỏ
}

# ================== OCR MỚI - HOẠT ĐỘNG TRÊN MAC M1 + PP-OCRv5 ==================
from paddleocr import PaddleOCR, PPStructureV3   # ← SỬA Ở ĐÂY: PPStructure → PPStructureV3
import cv2

_ocr_engine = None
_table_engine = None

def get_ocr():
    """OCR thường cho Note"""
    global _ocr_engine
    if _ocr_engine is None:
        _ocr_engine = PaddleOCR(
            use_textline_orientation=True,   # thay cho use_angle_cls cũ
            lang="vi"
        )
    return _ocr_engine

def get_table_engine():
    """Table structure recognition (giữ rows/columns)"""
    global _table_engine
    if _table_engine is None:
        _table_engine = PPStructureV3()      # ← DÙNG PPStructureV3
    return _table_engine

def ocr_note(img_path):
    """OCR cho Note"""
    ocr = get_ocr()
    result = ocr.ocr(img_path)               # KHÔNG dùng cls=True nữa
    if result and result[0]:
        return "\n".join([line[1][0] for line in result[0]])
    return ""

def ocr_table(img_path):
    """OCR cho Table - ưu tiên giữ cấu trúc bảng"""
    try:
        engine = get_table_engine()
        img = cv2.imread(img_path)
        result = engine(img)
        return str(result)                   # Expected output thường chấp nhận dạng này
    except Exception as e:
        print(f"[WARN] Table structure failed: {e}, fallback to plain OCR")
        return ocr_note(img_path)

# ── Main pipeline ─────────────────────────────────────────
def run_pipeline(
    image_path: str,
    output_dir: str = "outputs",
    checkpoint: str = "best.pt",
    conf: float = 0.3,
) -> tuple[dict, str]:
    """
    Chạy full pipeline: detect → crop → OCR → JSON.
    Returns: (result_dict, visualized_image_path)
    """
    image_path = str(image_path)
    img_name   = Path(image_path).name
    stem       = Path(image_path).stem
    crop_dir   = Path(output_dir) / stem / "crops"
    crop_dir.mkdir(parents=True, exist_ok=True)

    # 1. Detect
    model   = get_model(checkpoint)
    results = model(
        image_path,
        imgsz=1024,
        conf=conf,
        iou=0.5,
        device=DEVICE,
        verbose=False,
    )

    img_bgr = cv2.imread(image_path)
    if img_bgr is None:
        raise ValueError(f"Không đọc được ảnh: {image_path}")

    objects = []

    for i, box in enumerate(results[0].boxes):
        x1, y1, x2, y2 = map(int, box.xyxy[0].tolist())
        cls_idx  = int(box.cls[0])
        conf_val = round(float(box.conf[0]), 4)
        cls_raw  = CLASS_NAMES[cls_idx]
        cls_show = CLASS_DISPLAY[cls_raw]

        # 2. Crop
        pad = 4   # padding nhỏ quanh bbox
        cx1 = max(0, x1 - pad)
        cy1 = max(0, y1 - pad)
        cx2 = min(img_bgr.shape[1], x2 + pad)
        cy2 = min(img_bgr.shape[0], y2 + pad)
        crop = img_bgr[cy1:cy2, cx1:cx2]
        crop_path = str(crop_dir / f"{cls_show}_{i+1}.jpg")
        cv2.imwrite(crop_path, crop, [cv2.IMWRITE_JPEG_QUALITY, 95])

        # 3. OCR
        ocr_content = None
        if cls_raw == 'note':
            ocr_content = ocr_note(crop_path)
        elif cls_raw == 'table':
            ocr_content = ocr_table(crop_path)

        objects.append({
            "id":         i + 1,
            "class":      cls_show,
            "confidence": conf_val,
            "bbox":       {"x1": x1, "y1": y1, "x2": x2, "y2": y2},
            "crop_path":  crop_path,
            "ocr_content": ocr_content,
        })

        # 4. Vẽ bbox lên ảnh
        color = COLORS[cls_raw]
        cv2.rectangle(img_bgr, (x1, y1), (x2, y2), color, 2)
        label = f"{cls_show} {conf_val:.2f}"
        (tw, th), _ = cv2.getTextSize(
            label, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
        cv2.rectangle(img_bgr,
                      (x1, y1 - th - 8), (x1 + tw + 4, y1),
                      color, -1)
        cv2.putText(img_bgr, label,
                    (x1 + 2, y1 - 4),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.6,
                    (255, 255, 255), 2)

    # 5. Lưu ảnh visualize
    vis_path = str(Path(output_dir) / stem / "result_vis.jpg")
    cv2.imwrite(vis_path, img_bgr)

    # 6. Lưu JSON
    result = {"image": img_name, "objects": objects}
    json_path = str(Path(output_dir) / stem / "result.json")
    with open(json_path, "w", encoding="utf-8") as f:
        json.dump(result, f, ensure_ascii=False, indent=2)

    print(f"[✓] {img_name}: {len(objects)} objects → {json_path}")
    return result, vis_path



# ── CLI test nhanh ────────────────────────────────────────
if __name__ == "__main__":
    import sys
    img = sys.argv[1] if len(sys.argv) > 1 else "test.jpg"
    result, vis = run_pipeline(img)
    print(json.dumps(result, ensure_ascii=False, indent=2))