1.configs

我们看下yolox的backbone吧。https://github.com/open-mmlab/mmdetection/blob/master/configs/yolox/yolox_s_8x8_300e_coco.py

    backbone=dict(type='CSPDarknet', deepen_factor=0.33, widen_factor=0.5),

2.具体实现

具体实现，去找CSPDarknet这个类：

https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/backbones/csp_darknet.py#L124

@BACKBONES.register_module()
class CSPDarknet(BaseModule):
    """CSP-Darknet backbone used in YOLOv5 and YOLOX.
    Args:
        arch (str): Architecture of CSP-Darknet, from {P5, P6}.
            Default: P5.
        deepen_factor (float): Depth multiplier, multiply number of
            blocks in CSP layer by this amount. Default: 1.0.
        widen_factor (float): Width multiplier, multiply number of
            channels in each layer by this amount. Default: 1.0.
        out_indices (Sequence[int]): Output from which stages.
            Default: (2, 3, 4).
        frozen_stages (int): Stages to be frozen (stop grad and set eval
            mode). -1 means not freezing any parameters. Default: -1.
        use_depthwise (bool): Whether to use depthwise separable convolution.
            Default: False.
        arch_ovewrite(list): Overwrite default arch settings. Default: None.
        spp_kernal_sizes: (tuple[int]): Sequential of kernel sizes of SPP
            layers. Default: (5, 9, 13).
        conv_cfg (dict): Config dict for convolution layer. Default: None.
        norm_cfg (dict): Dictionary to construct and config norm layer.
            Default: dict(type='BN', requires_grad=True).
        act_cfg (dict): Config dict for activation layer.
            Default: dict(type='LeakyReLU', negative_slope=0.1).
        norm_eval (bool): Whether to set norm layers to eval mode, namely,
            freeze running stats (mean and var). Note: Effect on Batch Norm
            and its variants only.
        init_cfg (dict or list[dict], optional): Initialization config dict.
            Default: None.
    Example:
        >>> from mmdet.models import CSPDarknet
        >>> import torch
        >>> self = CSPDarknet(depth=53)
        >>> self.eval()
        >>> inputs = torch.rand(1, 3, 416, 416)
        >>> level_outputs = self.forward(inputs)
        >>> for level_out in level_outputs:
        ...     print(tuple(level_out.shape))
        ...
        (1, 256, 52, 52)
        (1, 512, 26, 26)
        (1, 1024, 13, 13)
    """
    # From left to right:
    # in_channels, out_channels, num_blocks, add_identity, use_spp
    arch_settings = {
        'P5': [[64, 128, 3, True, False], [128, 256, 9, True, False],
               [256, 512, 9, True, False], [512, 1024, 3, False, True]],
        'P6': [[64, 128, 3, True, False], [128, 256, 9, True, False],
               [256, 512, 9, True, False], [512, 768, 3, True, False],
               [768, 1024, 3, False, True]]
    }

    def __init__(self,
                 arch='P5',
                 deepen_factor=1.0,
                 widen_factor=1.0,
                 out_indices=(2, 3, 4),
                 frozen_stages=-1,
                 use_depthwise=False,
                 arch_ovewrite=None,
                 spp_kernal_sizes=(5, 9, 13),
                 conv_cfg=None,
                 norm_cfg=dict(type='BN', momentum=0.03, eps=0.001),
                 act_cfg=dict(type='Swish'),
                 norm_eval=False,
                 init_cfg=dict(
                     type='Kaiming',
                     layer='Conv2d',
                     a=math.sqrt(5),
                     distribution='uniform',
                     mode='fan_in',
                     nonlinearity='leaky_relu')):
        super().__init__(init_cfg)
        arch_setting = self.arch_settings[arch]
        if arch_ovewrite:
            arch_setting = arch_ovewrite
        assert set(out_indices).issubset(
            i for i in range(len(arch_setting) + 1))
        if frozen_stages not in range(-1, len(arch_setting) + 1):
            raise ValueError('frozen_stages must be in range(-1, '
                             'len(arch_setting) + 1). But received '
                             f'{frozen_stages}')

        self.out_indices = out_indices
        self.frozen_stages = frozen_stages
        self.use_depthwise = use_depthwise
        self.norm_eval = norm_eval
        conv = DepthwiseSeparableConvModule if use_depthwise else ConvModule

        self.stem = Focus(
            3,
            int(arch_setting[0][0] * widen_factor),
            kernel_size=3,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=act_cfg)
        self.layers = ['stem']

        for i, (in_channels, out_channels, num_blocks, add_identity,
                use_spp) in enumerate(arch_setting):
            in_channels = int(in_channels * widen_factor)
            out_channels = int(out_channels * widen_factor)
            num_blocks = max(round(num_blocks * deepen_factor), 1)
            stage = []
            conv_layer = conv(
                in_channels,
                out_channels,
                3,
                stride=2,
                padding=1,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                act_cfg=act_cfg)
            stage.append(conv_layer)
            if use_spp:
                spp = SPPBottleneck(
                    out_channels,
                    out_channels,
                    kernel_sizes=spp_kernal_sizes,
                    conv_cfg=conv_cfg,
                    norm_cfg=norm_cfg,
                    act_cfg=act_cfg)
                stage.append(spp)
            csp_layer = CSPLayer(
                out_channels,
                out_channels,
                num_blocks=num_blocks,
                add_identity=add_identity,
                use_depthwise=use_depthwise,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                act_cfg=act_cfg)
            stage.append(csp_layer)
            self.add_module(f'stage{i + 1}', nn.Sequential(*stage))
            self.layers.append(f'stage{i + 1}')

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            for i in range(self.frozen_stages + 1):
                m = getattr(self, self.layers[i])
                m.eval()
                for param in m.parameters():
                    param.requires_grad = False

    def train(self, mode=True):
        super(CSPDarknet, self).train(mode)
        self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                if isinstance(m, _BatchNorm):
                    m.eval()

    def forward(self, x):
        outs = []
        for i, layer_name in enumerate(self.layers):
            layer = getattr(self, layer_name)
            x = layer(x)
            if i in self.out_indices:
                outs.append(x)
        return tuple(outs)

3.如何调用的？

3.1 注册

注册创建类名字典。

@BACKBONES.register_module()

3.2 调用

即以字典中的类名，进行类的实例化。

由于backbone是包在models里面的，所以去看YOLOX这个类。

https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/detectors/yolox.py#L15

@DETECTORS.register_module()
class YOLOX(SingleStageDetector):
    r"""Implementation of `YOLOX: Exceeding YOLO Series in 2021
    <https://arxiv.org/abs/2107.08430>`_
    Note: Considering the trade-off between training speed and accuracy,
    multi-scale training is temporarily kept. More elegant implementation
    will be adopted in the future.
    Args:
        backbone (nn.Module): The backbone module.
        neck (nn.Module): The neck module.
        bbox_head (nn.Module): The bbox head module.
        train_cfg (obj:`ConfigDict`, optional): The training config
            of YOLOX. Default: None.
        test_cfg (obj:`ConfigDict`, optional): The testing config
            of YOLOX. Default: None.
        pretrained (str, optional): model pretrained path.
            Default: None.
        input_size (tuple): The model default input image size. The shape
            order should be (height, width). Default: (640, 640).
        size_multiplier (int): Image size multiplication factor.
            Default: 32.
        random_size_range (tuple): The multi-scale random range during
            multi-scale training. The real training image size will
            be multiplied by size_multiplier. Default: (15, 25).
        random_size_interval (int): The iter interval of change
            image size. Default: 10.
        init_cfg (dict, optional): Initialization config dict.
            Default: None.
    """

    def __init__(self,
                 backbone,
                 neck,
                 bbox_head,
                 train_cfg=None,
                 test_cfg=None,
                 pretrained=None,
                 input_size=(640, 640),
                 size_multiplier=32,
                 random_size_range=(15, 25),
                 random_size_interval=10,
                 init_cfg=None):
        super(YOLOX, self).__init__(backbone, neck, bbox_head, train_cfg,
                                    test_cfg, pretrained, init_cfg)
        log_img_scale(input_size, skip_square=True)
        self.rank, self.world_size = get_dist_info()
        self._default_input_size = input_size
        self._input_size = input_size
        self._random_size_range = random_size_range
        self._random_size_interval = random_size_interval
        self._size_multiplier = size_multiplier
        self._progress_in_iter = 0

    def forward_train(self,
                      img,
                      img_metas,
                      gt_bboxes,
                      gt_labels,
                      gt_bboxes_ignore=None):
        """
        Args:
            img (Tensor): Input images of shape (N, C, H, W).
                Typically these should be mean centered and std scaled.
            img_metas (list[dict]): A List of image info dict where each dict
                has: 'img_shape', 'scale_factor', 'flip', and may also contain
                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
                For details on the values of these keys see
                :class:`mmdet.datasets.pipelines.Collect`.
            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
                image in [tl_x, tl_y, br_x, br_y] format.
            gt_labels (list[Tensor]): Class indices corresponding to each box
            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
                boxes can be ignored when computing the loss.
        Returns:
            dict[str, Tensor]: A dictionary of loss components.
        """
        # Multi-scale training
        img, gt_bboxes = self._preprocess(img, gt_bboxes)

        losses = super(YOLOX, self).forward_train(img, img_metas, gt_bboxes,
                                                  gt_labels, gt_bboxes_ignore)

        # random resizing
        if (self._progress_in_iter + 1) % self._random_size_interval == 0:
            self._input_size = self._random_resize(device=img.device)
        self._progress_in_iter += 1

        return losses

    def _preprocess(self, img, gt_bboxes):
        scale_y = self._input_size[0] / self._default_input_size[0]
        scale_x = self._input_size[1] / self._default_input_size[1]
        if scale_x != 1 or scale_y != 1:
            img = F.interpolate(
                img,
                size=self._input_size,
                mode='bilinear',
                align_corners=False)
            for gt_bbox in gt_bboxes:
                gt_bbox[..., 0::2] = gt_bbox[..., 0::2] * scale_x
                gt_bbox[..., 1::2] = gt_bbox[..., 1::2] * scale_y
        return img, gt_bboxes

    def _random_resize(self, device):
        tensor = torch.LongTensor(2).to(device)

        if self.rank == 0:
            size = random.randint(*self._random_size_range)
            aspect_ratio = float(
                self._default_input_size[1]) / self._default_input_size[0]
            size = (self._size_multiplier * size,
                    self._size_multiplier * int(aspect_ratio * size))
            tensor[0] = size[0]
            tensor[1] = size[1]

        if self.world_size > 1:
            dist.barrier()
            dist.broadcast(tensor, 0)

        input_size = (tensor[0].item(), tensor[1].item())
        return input_size

发现并不容易看出backbone的部分，然后去找其基类SingleStageDetector中的代码段，通过获取注册字典中的CSPDarknet：

self.backbone = build_backbone(backbone)

https://github.com/open-mmlab/mmdetection/blob/31c84958f54287a8be2b99cbf87a6dcf12e57753/mmdet/models/detectors/single_stage.py#L12

class SingleStageDetector(BaseDetector):
    """Base class for single-stage detectors.
    Single-stage detectors directly and densely predict bounding boxes on the
    output features of the backbone+neck.
    """

    def __init__(self,
                 backbone,
                 neck=None,
                 bbox_head=None,
                 train_cfg=None,
                 test_cfg=None,
                 pretrained=None,
                 init_cfg=None):
        super(SingleStageDetector, self).__init__(init_cfg)
        if pretrained:
            warnings.warn('DeprecationWarning: pretrained is deprecated, '
                          'please use "init_cfg" instead')
            backbone.pretrained = pretrained
        self.backbone = build_backbone(backbone)
        if neck is not None:
            self.neck = build_neck(neck)
        bbox_head.update(train_cfg=train_cfg)
        bbox_head.update(test_cfg=test_cfg)
        self.bbox_head = build_head(bbox_head)
        self.train_cfg = train_cfg
        self.test_cfg = test_cfg

    def extract_feat(self, img):
        """Directly extract features from the backbone+neck."""
        x = self.backbone(img)
        if self.with_neck:
            x = self.neck(x)
        return x

    def forward_dummy(self, img):
        """Used for computing network flops.
        See `mmdetection/tools/analysis_tools/get_flops.py`
        """
        x = self.extract_feat(img)
        outs = self.bbox_head(x)
        return outs

    def forward_train(self,
                      img,
                      img_metas,
                      gt_bboxes,
                      gt_labels,
                      gt_bboxes_ignore=None):
        """
        Args:
            img (Tensor): Input images of shape (N, C, H, W).
                Typically these should be mean centered and std scaled.
            img_metas (list[dict]): A List of image info dict where each dict
                has: 'img_shape', 'scale_factor', 'flip', and may also contain
                'filename', 'ori_shape', 'pad_shape', and 'img_norm_cfg'.
                For details on the values of these keys see
                :class:`mmdet.datasets.pipelines.Collect`.
            gt_bboxes (list[Tensor]): Each item are the truth boxes for each
                image in [tl_x, tl_y, br_x, br_y] format.
            gt_labels (list[Tensor]): Class indices corresponding to each box
            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
                boxes can be ignored when computing the loss.
        Returns:
            dict[str, Tensor]: A dictionary of loss components.
        """
        super(SingleStageDetector, self).forward_train(img, img_metas)
        x = self.extract_feat(img)
        losses = self.bbox_head.forward_train(x, img_metas, gt_bboxes,
                                              gt_labels, gt_bboxes_ignore)
        return losses

    def simple_test(self, img, img_metas, rescale=False):
        """Test function without test-time augmentation.
        Args:
            img (torch.Tensor): Images with shape (N, C, H, W).
            img_metas (list[dict]): List of image information.
            rescale (bool, optional): Whether to rescale the results.
                Defaults to False.
        Returns:
            list[list[np.ndarray]]: BBox results of each image and classes.
                The outer list corresponds to each image. The inner list
                corresponds to each class.
        """
        feat = self.extract_feat(img)
        results_list = self.bbox_head.simple_test(
            feat, img_metas, rescale=rescale)
        bbox_results = [
            bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
            for det_bboxes, det_labels in results_list
        ]
        return bbox_results

    def aug_test(self, imgs, img_metas, rescale=False):
        """Test function with test time augmentation.
        Args:
            imgs (list[Tensor]): the outer list indicates test-time
                augmentations and inner Tensor should have a shape NxCxHxW,
                which contains all images in the batch.
            img_metas (list[list[dict]]): the outer list indicates test-time
                augs (multiscale, flip, etc.) and the inner list indicates
                images in a batch. each dict has image information.
            rescale (bool, optional): Whether to rescale the results.
                Defaults to False.
        Returns:
            list[list[np.ndarray]]: BBox results of each image and classes.
                The outer list corresponds to each image. The inner list
                corresponds to each class.
        """
        assert hasattr(self.bbox_head, 'aug_test'), \
            f'{self.bbox_head.__class__.__name__}' \
            ' does not support test-time augmentation'

        feats = self.extract_feats(imgs)
        results_list = self.bbox_head.aug_test(
            feats, img_metas, rescale=rescale)
        bbox_results = [
            bbox2result(det_bboxes, det_labels, self.bbox_head.num_classes)
            for det_bboxes, det_labels in results_list
        ]
        return bbox_results

    def onnx_export(self, img, img_metas, with_nms=True):
        """Test function without test time augmentation.
        Args:
            img (torch.Tensor): input images.
            img_metas (list[dict]): List of image information.
        Returns:
            tuple[Tensor, Tensor]: dets of shape [N, num_det, 5]
                and class labels of shape [N, num_det].
        """
        x = self.extract_feat(img)
        outs = self.bbox_head(x)
        # get origin input shape to support onnx dynamic shape

        # get shape as tensor
        img_shape = torch._shape_as_tensor(img)[2:]
        img_metas[0]['img_shape_for_onnx'] = img_shape
        # get pad input shape to support onnx dynamic shape for exporting
        # `CornerNet` and `CentripetalNet`, which 'pad_shape' is used
        # for inference
        img_metas[0]['pad_shape_for_onnx'] = img_shape

        if len(outs) == 2:
            # add dummy score_factor
            outs = (*outs, None)
        # TODO Can we change to `get_bboxes` when `onnx_export` fail
        det_bboxes, det_labels = self.bbox_head.onnx_export(
            *outs, img_metas, with_nms=with_nms)

        return det_bboxes, det_labels