YOLOv5 common.py中DetectMultiBackend类中文注释

为多种后端（如 PyTorch、TensorFlow、ONNX 等）提供统一接口的类
原代码：

class DetectMultiBackend(nn.Module):
    # YOLOv5多后端类，用于在不同后端上进行Python推断
    #DetectMultiBackend:       （line279）
  #      w                             判断weights是否为list，若是取出第一个值作为传入路径

   #    model_type             判断模型格式（pt、jit等），执行相应加载模式
    #    fp16                         半精度推算

     #   if data                      加载传入文件，获取names
##
    def __init__(self, weights='yolov5s.pt', device=torch.device('cpu'), dnn=False, data=None, fp16=False, fuse=True):
        # Usage:
        #   PyTorch:              weights = *.pt
        #   TorchScript:                    *.torchscript
        #   ONNX Runtime:                   *.onnx
        #   ONNX OpenCV DNN:                *.onnx --dnn
        #   OpenVINO:                       *_openvino_model
        #   CoreML:                         *.mlmodel
        #   TensorRT:                       *.engine
        #   TensorFlow SavedModel:          *_saved_model
        #   TensorFlow GraphDef:            *.pb
        #   TensorFlow Lite:                *.tflite
        #   TensorFlow Edge TPU:            *_edgetpu.tflite
        #   PaddlePaddle:                   *_paddle_model
        from models.experimental import attempt_download, attempt_load  # 避免循环导入
        super().__init__()
        # 将weights参数处理成字符串，如果是列表则取第一个元素
        w = str(weights[0] if isinstance(weights, list) else weights)
        # 根据权重文件的扩展名，判断模型类型
        pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle, triton = self._model_type(w)
          # 确定是否使用半精度计算，根据模型类型来决定
        fp16 &= pt or jit or onnx or engine or triton  # FP16
         # 确定是否使用NHWC数据格式，这取决于模型类型
        nhwc = coreml or saved_model or pb or tflite or edgetpu  # BHWC formats (vs torch BCWH)
        stride = 32  # default stride# 设置默认的模型stride
        cuda = torch.cuda.is_available() and device.type != 'cpu'  # use CUDA # 判断是否使用CUDA加速
        if not (pt or triton):
            w = attempt_download(w)  # download if not local# 如果不是PyTorch或Triton模型，尝试下载模型文件

        if pt:  # PyTorch
            model = attempt_load(weights if isinstance(weights, list) else w, device=device, inplace=True, fuse=fuse)
            stride = max(int(model.stride.max()), 32)  # model stride
            names = model.module.names if hasattr(model, 'module') else model.names  # get class names
            model.half() if fp16 else model.float()
            self.model = model  # explicitly assign for to(), cpu(), cuda(), half()
        elif jit:  # TorchScript
            LOGGER.info(f'Loading {w} for TorchScript inference...')
            extra_files = {'config.txt': ''}  # model metadata
            model = torch.jit.load(w, _extra_files=extra_files, map_location=device)
            model.half() if fp16 else model.float()
            if extra_files['config.txt']:  # load metadata dict
                d = json.loads(extra_files['config.txt'],
                               object_hook=lambda d: {
                                   int(k) if k.isdigit() else k: v
                                   for k, v in d.items()})
                stride, names = int(d['stride']), d['names']
        elif dnn:  # ONNX OpenCV DNN
            LOGGER.info(f'Loading {w} for ONNX OpenCV DNN inference...')
            check_requirements('opencv-python>=4.5.4')
            net = cv2.dnn.readNetFromONNX(w)
        elif onnx:  # ONNX Runtime
            LOGGER.info(f'Loading {w} for ONNX Runtime inference...')
            check_requirements(('onnx', 'onnxruntime-gpu' if cuda else 'onnxruntime'))
            import onnxruntime
            providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider']
            session = onnxruntime.InferenceSession(w, providers=providers)
            output_names = [x.name for x in session.get_outputs()]
            meta = session.get_modelmeta().custom_metadata_map  # metadata
            if 'stride' in meta:
                stride, names = int(meta['stride']), eval(meta['names'])
        elif xml:  # OpenVINO
            LOGGER.info(f'Loading {w} for OpenVINO inference...')
            check_requirements('openvino>=2023.0')  # requires openvino-dev: https://pypi.org/project/openvino-dev/
            from openvino.runtime import Core, Layout, get_batch
            core = Core()
            if not Path(w).is_file():  # if not *.xml
                w = next(Path(w).glob('*.xml'))  # get *.xml file from *_openvino_model dir
            ov_model = core.read_model(model=w, weights=Path(w).with_suffix('.bin'))
            if ov_model.get_parameters()[0].get_layout().empty:
                ov_model.get_parameters()[0].set_layout(Layout('NCHW'))
            batch_dim = get_batch(ov_model)
            if batch_dim.is_static:
                batch_size = batch_dim.get_length()
            ov_compiled_model = core.compile_model(ov_model, device_name='AUTO')  # AUTO selects best available device
            stride, names = self._load_metadata(Path(w).with_suffix('.yaml'))  # load metadata
        elif engine:  # TensorRT
            LOGGER.info(f'Loading {w} for TensorRT inference...')
            import tensorrt as trt  # https://developer.nvidia.com/nvidia-tensorrt-download
            check_version(trt.__version__, '7.0.0', hard=True)  # require tensorrt>=7.0.0
            if device.type == 'cpu':
                device = torch.device('cuda:0')
            Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
            logger = trt.Logger(trt.Logger.INFO)
            with open(w, 'rb') as f, trt.Runtime(logger) as runtime:
                model = runtime.deserialize_cuda_engine(f.read())
            context = model.create_execution_context()
            bindings = OrderedDict()
            output_names = []
            fp16 = False  # default updated below
            dynamic = False
            for i in range(model.num_bindings):
                name = model.get_binding_name(i)
                dtype = trt.nptype(model.get_binding_dtype(i))
                if model.binding_is_input(i):
                    if -1 in tuple(model.get_binding_shape(i)):  # dynamic
                        dynamic = True
                        context.set_binding_shape(i, tuple(model.get_profile_shape(0, i)[2]))
                    if dtype == np.float16:
                        fp16 = True
                else:  # output
                    output_names.append(name)
                shape = tuple(context.get_binding_shape(i))
                im = torch.from_numpy(np.empty(shape, dtype=dtype)).to(device)
                bindings[name] = Binding(name, dtype, shape, im, int(im.data_ptr()))
            binding_addrs = OrderedDict((n, d.ptr) for n, d in bindings.items())
            batch_size = bindings['images'].shape[0]  # if dynamic, this is instead max batch size
        elif coreml:  # CoreML
            LOGGER.info(f'Loading {w} for CoreML inference...')
            import coremltools as ct
            model = ct.models.MLModel(w)
        elif saved_model:  # TF SavedModel
            LOGGER.info(f'Loading {w} for TensorFlow SavedModel inference...')
            import tensorflow as tf
            keras = False  # assume TF1 saved_model
            model = tf.keras.models.load_model(w) if keras else tf.saved_model.load(w)
        elif pb:  # GraphDef https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt
            LOGGER.info(f'Loading {w} for TensorFlow GraphDef inference...')
            import tensorflow as tf

            def wrap_frozen_graph(gd, inputs, outputs):
                x = tf.compat.v1.wrap_function(lambda: tf.compat.v1.import_graph_def(gd, name=''), [])  # wrapped
                ge = x.graph.as_graph_element
                return x.prune(tf.nest.map_structure(ge, inputs), tf.nest.map_structure(ge, outputs))

            def gd_outputs(gd):
                name_list, input_list = [], []
                for node in gd.node:  # tensorflow.core.framework.node_def_pb2.NodeDef
                    name_list.append(node.name)
                    input_list.extend(node.input)
                return sorted(f'{x}:0' for x in list(set(name_list) - set(input_list)) if not x.startswith('NoOp'))

            gd = tf.Graph().as_graph_def()  # TF GraphDef
            with open(w, 'rb') as f:
                gd.ParseFromString(f.read())
            frozen_func = wrap_frozen_graph(gd, inputs='x:0', outputs=gd_outputs(gd))
        elif tflite or edgetpu:  # https://www.tensorflow.org/lite/guide/python#install_tensorflow_lite_for_python
            try:  # https://coral.ai/docs/edgetpu/tflite-python/#update-existing-tf-lite-code-for-the-edge-tpu
                from tflite_runtime.interpreter import Interpreter, load_delegate
            except ImportError:
                import tensorflow as tf
                Interpreter, load_delegate = tf.lite.Interpreter, tf.lite.experimental.load_delegate,
            if edgetpu:  # TF Edge TPU https://coral.ai/software/#edgetpu-runtime
                LOGGER.info(f'Loading {w} for TensorFlow Lite Edge TPU inference...')
                delegate = {
                    'Linux': 'libedgetpu.so.1',
                    'Darwin': 'libedgetpu.1.dylib',
                    'Windows': 'edgetpu.dll'}[platform.system()]
                interpreter = Interpreter(model_path=w, experimental_delegates=[load_delegate(delegate)])
            else:  # TFLite
                LOGGER.info(f'Loading {w} for TensorFlow Lite inference...')
                interpreter = Interpreter(model_path=w)  # load TFLite model
            interpreter.allocate_tensors()  # allocate
            input_details = interpreter.get_input_details()  # inputs
            output_details = interpreter.get_output_details()  # outputs
            # load metadata
            with contextlib.suppress(zipfile.BadZipFile):
                with zipfile.ZipFile(w, 'r') as model:
                    meta_file = model.namelist()[0]
                    meta = ast.literal_eval(model.read(meta_file).decode('utf-8'))
                    stride, names = int(meta['stride']), meta['names']
        elif tfjs:  # TF.js
            raise NotImplementedError('ERROR: YOLOv5 TF.js inference is not supported')
        elif paddle:  # PaddlePaddle
            LOGGER.info(f'Loading {w} for PaddlePaddle inference...')
            check_requirements('paddlepaddle-gpu' if cuda else 'paddlepaddle')
            import paddle.inference as pdi
            if not Path(w).is_file():  # if not *.pdmodel
                w = next(Path(w).rglob('*.pdmodel'))  # get *.pdmodel file from *_paddle_model dir
            weights = Path(w).with_suffix('.pdiparams')
            config = pdi.Config(str(w), str(weights))
            if cuda:
                config.enable_use_gpu(memory_pool_init_size_mb=2048, device_id=0)
            predictor = pdi.create_predictor(config)
            input_handle = predictor.get_input_handle(predictor.get_input_names()[0])
            output_names = predictor.get_output_names()
        elif triton:  # NVIDIA Triton Inference Server
            LOGGER.info(f'Using {w} as Triton Inference Server...')
            check_requirements('tritonclient[all]')
            from utils.triton import TritonRemoteModel
            model = TritonRemoteModel(url=w)
            nhwc = model.runtime.startswith('tensorflow')
        else:
            raise NotImplementedError(f'ERROR: {w} is not a supported format')

        # class names
        if 'names' not in locals():
            names = yaml_load(data)['names'] if data else {i: f'class{i}' for i in range(999)}
        if names[0] == 'n01440764' and len(names) == 1000:  # ImageNet
            names = yaml_load(ROOT / 'data/ImageNet.yaml')['names']  # human-readable names

        self.__dict__.update(locals())  # assign all variables to self

    def forward(self, im, augment=False, visualize=False):
          # 定义模型的前向传播函数，根据不同后端进行不同的处理
        # 获取输入图像的尺寸信息，并根据需要转换数据类型和格式

        # 根据模型类型执行对应的推断操作，处理不同后端的推断逻辑
        # 返回推断结果，结果可能是单个张量或张量列表
        # YOLOv5 MultiBackend inference
        b, ch, h, w = im.shape  # batch, channel, height, width
        if self.fp16 and im.dtype != torch.float16:
            im = im.half()  # to FP16
        if self.nhwc:
            im = im.permute(0, 2, 3, 1)  # torch BCHW to numpy BHWC shape(1,320,192,3)

        if self.pt:  # PyTorch
            y = self.model(im, augment=augment, visualize=visualize) if augment or visualize else self.model(im)
        elif self.jit:  # TorchScript
            y = self.model(im)
        elif self.dnn:  # ONNX OpenCV DNN
            im = im.cpu().numpy()  # torch to numpy
            self.net.setInput(im)
            y = self.net.forward()
        elif self.onnx:  # ONNX Runtime
            im = im.cpu().numpy()  # torch to numpy
            y = self.session.run(self.output_names, {self.session.get_inputs()[0].name: im})
        elif self.xml:  # OpenVINO
            im = im.cpu().numpy()  # FP32
            y = list(self.ov_compiled_model(im).values())
        elif self.engine:  # TensorRT
            if self.dynamic and im.shape != self.bindings['images'].shape:
                i = self.model.get_binding_index('images')
                self.context.set_binding_shape(i, im.shape)  # reshape if dynamic
                self.bindings['images'] = self.bindings['images']._replace(shape=im.shape)
                for name in self.output_names:
                    i = self.model.get_binding_index(name)
                    self.bindings[name].data.resize_(tuple(self.context.get_binding_shape(i)))
            s = self.bindings['images'].shape
            assert im.shape == s, f"input size {im.shape} {'>' if self.dynamic else 'not equal to'} max model size {s}"
            self.binding_addrs['images'] = int(im.data_ptr())
            self.context.execute_v2(list(self.binding_addrs.values()))
            y = [self.bindings[x].data for x in sorted(self.output_names)]
        elif self.coreml:  # CoreML
            im = im.cpu().numpy()
            im = Image.fromarray((im[0] * 255).astype('uint8'))
            # im = im.resize((192, 320), Image.BILINEAR)
            y = self.model.predict({'image': im})  # coordinates are xywh normalized
            if 'confidence' in y:
                box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]])  # xyxy pixels
                conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float)
                y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1)
            else:
                y = list(reversed(y.values()))  # reversed for segmentation models (pred, proto)
        elif self.paddle:  # PaddlePaddle
            im = im.cpu().numpy().astype(np.float32)
            self.input_handle.copy_from_cpu(im)
            self.predictor.run()
            y = [self.predictor.get_output_handle(x).copy_to_cpu() for x in self.output_names]
        elif self.triton:  # NVIDIA Triton Inference Server
            y = self.model(im)
        else:  # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU)
            im = im.cpu().numpy()
            if self.saved_model:  # SavedModel
                y = self.model(im, training=False) if self.keras else self.model(im)
            elif self.pb:  # GraphDef
                y = self.frozen_func(x=self.tf.constant(im))
            else:  # Lite or Edge TPU
                input = self.input_details[0]
                int8 = input['dtype'] == np.uint8  # is TFLite quantized uint8 model
                if int8:
                    scale, zero_point = input['quantization']
                    im = (im / scale + zero_point).astype(np.uint8)  # de-scale
                self.interpreter.set_tensor(input['index'], im)
                self.interpreter.invoke()
                y = []
                for output in self.output_details:
                    x = self.interpreter.get_tensor(output['index'])
                    if int8:
                        scale, zero_point = output['quantization']
                        x = (x.astype(np.float32) - zero_point) * scale  # re-scale
                    y.append(x)
            y = [x if isinstance(x, np.ndarray) else x.numpy() for x in y]
            y[0][..., :4] *= [w, h, w, h]  # xywh normalized to pixels

        if isinstance(y, (list, tuple)):
            return self.from_numpy(y[0]) if len(y) == 1 else [self.from_numpy(x) for x in y]
        else:
            return self.from_numpy(y)

    def from_numpy(self, x):
        return torch.from_numpy(x).to(self.device) if isinstance(x, np.ndarray) else x
# 将numpy数组转换为torch张量，并移到相应设备上
    def warmup(self, imgsz=(1, 3, 640, 640)):
     # 对模型进行预热，通过一次推断操作来初始化模型状态
        # Warmup model by running inference once
        warmup_types = self.pt, self.jit, self.onnx, self.engine, self.saved_model, self.pb, self.triton
        if any(warmup_types) and (self.device.type != 'cpu' or self.triton):
            im = torch.empty(*imgsz, dtype=torch.half if self.fp16 else torch.float, device=self.device)  # input
            for _ in range(2 if self.jit else 1):  #
                self.forward(im)  # warmup

    @staticmethod
    def _model_type(p='path/to/model.pt'):
     # 根据模型文件路径确定模型类型
        # 这里列出了支持的所有模型格式，并根据文件扩展名进行判断
        # Return model type from model path, i.e. path='path/to/model.onnx' -> type=onnx
        # types = [pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle]
        from export import export_formats
        from utils.downloads import is_url
        sf = list(export_formats().Suffix)  # export suffixes
        if not is_url(p, check=False):
            check_suffix(p, sf)  # checks
        url = urlparse(p)  # if url may be Triton inference server
        types = [s in Path(p).name for s in sf]
        types[8] &= not types[9]  # tflite &= not edgetpu
        triton = not any(types) and all([any(s in url.scheme for s in ['http', 'grpc']), url.netloc])
        return types + [triton]

    @staticmethod
    def _load_metadata(f=Path('path/to/meta.yaml')):
    # 如果存在，从meta.yaml文件加载元数据
        # 返回stride和names信息
        # Load metadata from meta.yaml if it exists
        if f.exists():
            d = yaml_load(f)
            return d['stride'], d['names']  # assign stride, names
        return None, None