# models/loader.py """ Loads and warms up both ML models at startup. Imported once by app.py — models are module-level singletons. """ import time import numpy as np import torch from torchvision import transforms, models as tv_models from ultralytics import YOLO from config.settings import MODEL_PATH, SKIN_TYPE_MODEL_PATH # ── Device ───────────────────────────────────────────────────────────────────── device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') print(f"[loader] Using device: {device}") # ── YOLO ─────────────────────────────────────────────────────────────────────── print("[loader] Loading YOLO model…") yolo_model = YOLO(MODEL_PATH) # ── Skin-type ResNet-50 ──────────────────────────────────────────────────────── print("[loader] Loading skin-type model…") skin_type_model = tv_models.resnet50(weights=None) num_ftrs = skin_type_model.fc.in_features skin_type_model.fc = torch.nn.Linear(num_ftrs, 3) try: checkpoint = torch.load(SKIN_TYPE_MODEL_PATH, map_location=device) model_dict = skin_type_model.state_dict() pretrained = {} for k, v in checkpoint.items(): if k in model_dict and v.shape == model_dict[k].shape: pretrained[k] = v elif k == 'fc.weight' and model_dict['fc.weight'].shape[1] == v.shape[1]: pretrained['fc.weight'] = v[:3, :] elif k == 'fc.bias' and len(v) >= 3: pretrained['fc.bias'] = v[:3] skin_type_model.load_state_dict(pretrained, strict=False) print("[loader] Skin-type model loaded ✓") except Exception as exc: print(f"[loader] WARNING — skin-type model load error: {exc}") skin_type_model = skin_type_model.to(device) skin_type_model.eval() # ── Image transform ──────────────────────────────────────────────────────────── skin_transform = transforms.Compose([ transforms.ToPILImage(), transforms.Resize((224, 224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), ]) def warm_up(): """Run a dummy forward pass on both models to eliminate first-request latency.""" print("[loader] Warming up models…") dummy = np.zeros((224, 224, 3), dtype=np.uint8) t = time.time() yolo_model(dummy, conf=0.15, iou=0.2) print(f"[loader] YOLO warm-up: {time.time()-t:.2f}s") t = time.time() tensor = skin_transform(dummy).unsqueeze(0).to(device) with torch.no_grad(): skin_type_model(tensor) print(f"[loader] Skin-type warm-up: {time.time()-t:.2f}s") print("[loader] Models ready ✓")