""" Inference helpers for the ArcDetectorCNN door orientation model. Supports single-image and batch inference on already-cropped images. Classes (alphabetical ImageFolder order): 0 = double 1 = hinge_left 2 = hinge_right Usage — single image: from inference import load_model, predict model = load_model("runs/blocks3_drop0.2/best_model.pt") label, confidence = predict(img_bgr, model) print(label, confidence) # e.g. "hinge_left", 0.94 Usage — batch: labels, confidences = predict_batch([img1, img2, img3], model) Usage — CLI: python inference.py --model runs/blocks3_drop0.2/best_model.pt image1.png image2.png """ from __future__ import annotations import argparse import sys from pathlib import Path import cv2 from modal import Image import numpy as np import torch _HERE = Path(__file__).parent for _p in (_HERE, _HERE.parent.parent): if str(_p) not in sys.path: sys.path.insert(0, str(_p)) from cnn_dataset_builder import TRANSFORM, preprocess_real_crop # noqa: E402 from cnn_door_orientation_detection import ArcDetectorCNN, CLASS_NAMES # noqa: E402 if torch.cuda.is_available(): _DEVICE = torch.device("cuda") elif torch.backends.mps.is_available(): _DEVICE = torch.device("mps") else: _DEVICE = torch.device("cpu") # ── model loading ────────────────────────────────────────────────────────────── def load_model(model_path: str | Path) -> ArcDetectorCNN: """ Load a trained ArcDetectorCNN from a .pt file. Reads config.json from the same directory to restore n_blocks / dropout. """ import json model_path = Path(model_path) config_path = model_path.parent / "config.json" config = json.loads(config_path.read_text()) if config_path.exists() else {} model = ArcDetectorCNN( n_blocks=config.get("n_blocks", 3), dropout=config.get("dropout", 0.2), ) model.load_state_dict(torch.load(model_path, map_location="cpu", weights_only=True)) model.eval() return model.to(_DEVICE) # ── preprocessing ────────────────────────────────────────────────────────────── def _to_tensor(img: np.ndarray) -> torch.Tensor: """ Convert a single crop (BGR or grayscale uint8) to a normalised (1, 1, H, W) tensor. Applies the same CLAHE preprocessing used during training. """ gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) if img.ndim == 3 else img from PIL import Image processed = preprocess_real_crop(gray) return TRANSFORM(Image.fromarray(processed)).unsqueeze(0) # (1, 1, 128, 128) def _to_batch_tensor(imgs: list[np.ndarray]) -> torch.Tensor: """Stack a list of crops into a (N, 1, H, W) tensor.""" return torch.cat([_to_tensor(img) for img in imgs], dim=0) # ── inference ────────────────────────────────────────────────────────────────── def predict( img: np.ndarray, model: ArcDetectorCNN, confidence_threshold: float = 0.0, ) -> tuple[str, float]: """ Predict the door orientation class for a single crop. Args: img: H×W×3 BGR or H×W grayscale uint8 crop. model: Loaded ArcDetectorCNN (use load_model()). confidence_threshold: If the softmax confidence is below this, return "unknown". Returns: (class_name, confidence) where class_name is one of "double" / "hinge_left" / "hinge_right" / "unknown". """ labels, confidences = predict_batch([img], model, confidence_threshold) return labels[0], confidences[0] def predict_batch( imgs: list[np.ndarray], model: ArcDetectorCNN, confidence_threshold: float = 0.5, ) -> tuple[list[str], list[float]]: """ Predict door orientation classes for a batch of crops. Args: imgs: List of H×W×3 BGR or H×W grayscale uint8 crops. model: Loaded ArcDetectorCNN (use load_model()). confidence_threshold: Images whose top softmax score is below this get "unknown". Returns: (labels, confidences) — parallel lists, one entry per input image. """ model.eval() batch = _to_batch_tensor(imgs).to(_DEVICE) # (N, 1, 128, 128) with torch.no_grad(): probs = batch.float() probs = model(probs).softmax(dim=1) # (N, 3) confidences_t, indices_t = probs.max(dim=1) # (N,), (N,) labels: list[str] = [] confidences: list[float] = [] for conf, idx in zip(confidences_t.cpu().tolist(), indices_t.cpu().tolist()): if conf < confidence_threshold: labels.append("unknown") else: labels.append(CLASS_NAMES[idx]) confidences.append(round(conf, 4)) return labels, confidences # ── CLI ──────────────────────────────────────────────────────────────────────── _DEFAULT_RUNS = _HERE / "runs" def _auto_detect_model(runs_dir: Path) -> Path: import json configs = list(runs_dir.glob("*/config.json")) if not configs: raise FileNotFoundError(f"No trained runs found in {runs_dir}. Run --train first.") best = max(configs, key=lambda p: json.loads(p.read_text()).get("best_val_acc", 0)) return best.parent / "best_model.pt" if __name__ == "__main__": parser = argparse.ArgumentParser(description="ArcDetectorCNN inference") parser.add_argument("images", nargs="+", type=Path, help="Image file(s) to classify") parser.add_argument( "--model", type=Path, default=None, help="Path to best_model.pt (auto-detected from --runs-dir if omitted)", ) parser.add_argument( "--runs-dir", type=Path, default=_DEFAULT_RUNS, help=f"Runs directory for auto-detection (default: {_DEFAULT_RUNS})", ) parser.add_argument( "--threshold", type=float, default=0.0, help="Confidence threshold below which output is 'unknown' (default: 0.0)", ) args = parser.parse_args() model_path = args.model or _auto_detect_model(args.runs_dir) print(f"Model: {model_path}") model = load_model(model_path) imgs = [] paths = [] for p in args.images: img = cv2.imread(str(p)) if img is None: print(f" Warning: could not read {p} — skipping") continue imgs.append(img) paths.append(p) if not imgs: print("No valid images to process.") sys.exit(1) labels, confidences = predict_batch(imgs, model, confidence_threshold=args.threshold) print(f"\n{'Image':<40} {'Label':<12} Confidence") print("-" * 62) for p, label, conf in zip(paths, labels, confidences): print(f"{str(p):<40} {label:<12} {conf:.4f}")