scoliosis-inference/model_loader.py

import os
import threading
from typing import Optional

import torch
import torch.nn as nn
from torchvision import models
from ultralytics import YOLO

from config import load_config

_LOCK = threading.Lock()
_CACHE = {}
_CFG = load_config()

def _cache_key(kind: str, path: str, device: Optional[str]) -> str:
    return f"{kind}|{os.path.abspath(path)}|{device or 'auto'}"

def _load_shufflenet(path: str, device: Optional[str]):
    """Загружает shufflenet‑классификатор из state_dict"""
    sd = torch.load(path, map_location=device or "cpu")
    num_classes = sd["fc.weight"].shape[0]
    candidates = [
        models.shufflenet_v2_x0_5(weights=None),
        models.shufflenet_v2_x1_0(weights=None),
    ]
    model = None
    for m in candidates:
        m.fc = nn.Linear(m.fc.in_features, num_classes)
        try:
            m.load_state_dict(sd, strict=True)
            model = m
            break
        except Exception:
            continue
    if model is None:
        raise RuntimeError(f"Cannot load state_dict from {path} into supported shufflenet variants")
    model.eval()
    if device:
        model.to(device)
    return model

def _load_mobilenet_v3_large(path: str, device: Optional[str]):
    """Загружает mobilenet_v3_large из чекпойнта, настраивает число классов, переносит на device, ставит eval()"""
    ckpt = torch.load(path, map_location=device or "cpu")
    sd = ckpt.get("model_state", ckpt)
    num_classes = sd["classifier.3.weight"].shape[0]
    model = models.mobilenet_v3_large(weights=None)
    model.classifier[3] = nn.Linear(model.classifier[3].in_features, num_classes)
    model.load_state_dict(sd, strict=True)
    model.eval()
    if device:
        model.to(device)
    return model

def get_detection_model(device: Optional[str] = None, model_path: Optional[str] = None):
    """Получает YOLO‑детектор: берёт путь (из аргумента или конфига), достаёт из кеша или загружает и кэширует"""
    path = model_path or _CFG.models.detection
    key = _cache_key("detection", path, device)
    with _LOCK:
        if key not in _CACHE:
            model = YOLO(path)
            if device:
                try:
                    model.to(device)
                except Exception:
                    if hasattr(model, "model"):
                        model.model.to(device)
            _CACHE[key] = model
        return _CACHE[key]

def get_classification_model(device: Optional[str] = None, model_path: Optional[str] = None):
    """Получает классификатор stitched/single (mobilenet): кеш + загрузка при первом вызове"""
    path = model_path or _CFG.models.classification
    key = _cache_key("classification", path, device)
    with _LOCK:
        if key not in _CACHE:
            _CACHE[key] = _load_mobilenet_v3_large(path, device)
        return _CACHE[key]

def get_projection_model(device: Optional[str] = None, model_path: Optional[str] = None):
    """Получает классификатор проекции (shufflenet): кеш + загрузка при первом вызове"""
    path = model_path or _CFG.models.projection
    key = _cache_key("projection", path, device)
    with _LOCK:
        if key not in _CACHE:
            _CACHE[key] = _load_shufflenet(path, device)
        return _CACHE[key]