• 特征金字塔
• 多尺度融合
• 特征金字塔的网络原理
  
• 这里是基于resnet网络与Fpn做的结合，主要把resnet中的特征层利用FPN的思想一起结合，实现resnet_fpn。增强目标检测backone的有效性。
• 代码实现如下：

import torch
from torch import Tensor
from collections import OrderedDict
import torch.nn.functional as F
from torch import nn
from torch.jit.annotations import Tuple, List, Dict


class Bottleneck(nn.Module):
    expansion = 4
    def __init__(self, in_channel, out_channel, stride=1, downsample=None, norm_layer=None):
        super(Bottleneck, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
                               kernel_size=(1,1), stride=(1,1), bias=False)  # squeeze channels
        self.bn1 = norm_layer(out_channel)
        # -----------------------------------------
        self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
                               kernel_size=(3,3), stride=(stride,stride), bias=False, padding=(1,1))
        self.bn2 = norm_layer(out_channel)
        # -----------------------------------------
        self.conv3 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel * self.expansion,
                               kernel_size=(1,1), stride=(1,1), bias=False)  # unsqueeze channels
        self.bn3 = norm_layer(out_channel * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)
        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        out += identity
        out = self.relu(out)

        return out


class ResNet(nn.Module):

    def  __init__(self, block, blocks_num, num_classes=1000, include_top=True, norm_layer=None):
        '''
        :param block:块
        :param blocks_num:块数
        :param num_classes: 分类数
        :param include_top:
        :param norm_layer: BN
        '''
        super(ResNet, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.include_top = include_top
        self.in_channel = 64

        self.conv1 = nn.Conv2d(in_channels=3, out_channels=self.in_channel, kernel_size=(7,7), stride=(2,2),
                               padding=(3,3), bias=False)
        self.bn1 = norm_layer(self.in_channel)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, blocks_num[0])
        self.layer2 = self._make_layer(block, 128, blocks_num[1], stride=2)
        self.layer3 = self._make_layer(block, 256, blocks_num[2], stride=2)
        self.layer4 = self._make_layer(block, 512, blocks_num[3], stride=2)

        if self.include_top:
            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))  # output size = (1, 1)
            self.fc = nn.Linear(512 * block.expansion, num_classes)

        '''
        初始化
        '''
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')

    def _make_layer(self, block, channel, block_num, stride=1):
        norm_layer = self._norm_layer
        downsample = None
        if stride != 1 or self.in_channel != channel * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channel, channel * block.expansion, kernel_size=(1,1), stride=(stride,stride), bias=False),
                norm_layer(channel * block.expansion))

        layers = []
        layers.append(block(self.in_channel, channel, downsample=downsample,
                            stride=stride, norm_layer=norm_layer))
        self.in_channel = channel * block.expansion

        for _ in range(1, block_num):
            layers.append(block(self.in_channel, channel, norm_layer=norm_layer))

        return nn.Sequential(*layers)

    def forward(self, x):

        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        if self.include_top:
            x = self.avgpool(x)
            x = torch.flatten(x, 1)
            x = self.fc(x)

        return x


class IntermediateLayerGetter(nn.ModuleDict):
    """
    Module wrapper that returns intermediate layers from a models
    It has a strong assumption that the modules have been registered
    into the models in the same order as they are used.
    This means that one should **not** reuse the same nn.Module
    twice in the forward if you want this to work.
    Additionally, it is only able to query submodules that are directly
    assigned to the models. So if `models` is passed, `models.feature1` can
    be returned, but not `models.feature1.layer2`.
    Arguments:
        model (nn.Module): models on which we will extract the features
        return_layers (Dict[name, new_name]): a dict containing the names
            of the modules for which the activations will be returned as
            the key of the dict, and the value of the dict is the name
            of the returned activation (which the user can specify).
    """
    __annotations__ = {
        "return_layers": Dict[str, str],
    }

    def __init__(self, model, return_layers):

        if not set(return_layers).issubset([name for name, _ in model.named_children()]):
            raise ValueError("return_layers are not present in models")
        # {'layer1': '0', 'layer2': '1', 'layer3': '2', 'layer4': '3'}
        orig_return_layers = return_layers
        return_layers = {k: v for k, v in return_layers.items()}
        layers = OrderedDict()

        # 遍历模型子模块按顺序存入有序字典
        # 只保存layer4及其之前的结构，舍去之后不用的结构
        for name, module in model.named_children():
            layers[name] = module
            if name in return_layers:
                del return_layers[name]
            if not return_layers:
                break

        super(IntermediateLayerGetter, self).__init__(layers)
        self.return_layers = orig_return_layers

    def forward(self, x):
        out = OrderedDict()
        # 依次遍历模型的所有子模块，并进行正向传播，
        # 收集layer1, layer2, layer3, layer4的输出
        for name, module in self.named_children():
            x = module(x)
            if name in self.return_layers:
                out_name = self.return_layers[name]
                out[out_name] = x
        return out


class FeaturePyramidNetwork(nn.Module):
    """
    Module that adds a FPN from on top of a set of feature maps. This is based on
    `"Feature Pyramid Network for Object Detection" <https://arxiv.org/abs/1612.03144>`_.
    The feature maps are currently supposed to be in increasing depth
    order.
    The input to the models is expected to be an OrderedDict[Tensor], containing
    the feature maps on top of which the FPN will be added.
    Arguments:
        in_channels_list (list[int]): number of channels for each feature map that
            is passed to the module
        out_channels (int): number of channels of the FPN representation
        extra_blocks (ExtraFPNBlock or None): if provided, extra operations will
            be performed. It is expected to take the fpn features, the original
            features and the names of the original features as input, and returns
            a new list of feature maps and their corresponding names
    """

    def __init__(self, in_channels_list, out_channels, extra_blocks=None):
        super(FeaturePyramidNetwork, self).__init__()
        # 用来调整resnet特征矩阵(layer1,2,3,4)的channel（kernel_size=1）
        self.inner_blocks = nn.ModuleList()
        # 对调整后的特征矩阵使用3x3的卷积核来得到对应的预测特征矩阵
        self.layer_blocks = nn.ModuleList()
        for in_channels in in_channels_list:
            if in_channels == 0:
                continue
            inner_block_module = nn.Conv2d(in_channels, out_channels, (1,1))
            layer_block_module = nn.Conv2d(out_channels, out_channels, (3,3), padding=(1,1))
            self.inner_blocks.append(inner_block_module)
            self.layer_blocks.append(layer_block_module)

        # initialize parameters now to avoid modifying the initialization of top_blocks
        for m in self.children():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_uniform_(m.weight, a=1)
                nn.init.constant_(m.bias, 0)

        self.extra_blocks = extra_blocks

    def get_result_from_inner_blocks(self, x, idx):
        # type: (Tensor, int) -> Tensor
        """
        This is equivalent to self.inner_blocks[idx](x),
        but torchscript doesn't support this yet
        """
        num_blocks = len(self.inner_blocks)
        if idx < 0:
            idx += num_blocks
        i = 0
        out = x
        for module in self.inner_blocks:
            if i == idx:
                out = module(x)
            i += 1
        return out

    def get_result_from_layer_blocks(self, x, idx):
        # type: (Tensor, int) -> Tensor
        """
        This is equivalent to self.layer_blocks[idx](x),
        but torchscript doesn't support this yet
        """
        num_blocks = len(self.layer_blocks)
        if idx < 0:
            idx += num_blocks
        i = 0
        out = x
        for module in self.layer_blocks:
            if i == idx:
                out = module(x)
            i += 1
        return out

    def forward(self, x):
        # type: (Dict[str, Tensor]) -> Dict[str, Tensor]
        """
        Computes the FPN for a set of feature maps.
        Arguments:
            x (OrderedDict[Tensor]): feature maps for each feature level.
        Returns:
            results (OrderedDict[Tensor]): feature maps after FPN layers.
                They are ordered from highest resolution first.
        """
        # unpack OrderedDict into two lists for easier handling
        names = list(x.keys())
        x = list(x.values())

        # 将resnet layer4的channel调整到指定的out_channels
        # last_inner = self.inner_blocks[-1](x[-1])
        last_inner = self.get_result_from_inner_blocks(x[-1], -1)
        # result中保存着每个预测特征层
        results = []
        # 将layer4调整channel后的特征矩阵，通过3x3卷积后得到对应的预测特征矩阵
        # results.append(self.layer_blocks[-1](last_inner))
        results.append(self.get_result_from_layer_blocks(last_inner, -1))

        # 倒序遍历resenet输出特征层，以及对应inner_block和layer_block
        # layer3 -> layer2 -> layer1 （layer4已经处理过了）
        # for feature, inner_block, layer_block in zip(
        #         x[:-1][::-1], self.inner_blocks[:-1][::-1], self.layer_blocks[:-1][::-1]
        # ):
        #     if not inner_block:
        #         continue
        #     inner_lateral = inner_block(feature)
        #     feat_shape = inner_lateral.shape[-2:]
        #     inner_top_down = F.interpolate(last_inner, size=feat_shape, mode="nearest")
        #     last_inner = inner_lateral + inner_top_down
        #     results.insert(0, layer_block(last_inner))

        for idx in range(len(x) - 2, -1, -1):
            inner_lateral = self.get_result_from_inner_blocks(x[idx], idx)
            feat_shape = inner_lateral.shape[-2:]
            inner_top_down = F.interpolate(last_inner, size=feat_shape, mode="nearest")
            last_inner = inner_lateral + inner_top_down
            results.insert(0, self.get_result_from_layer_blocks(last_inner, idx))

        # 在layer4对应的预测特征层基础上生成预测特征矩阵5
        if self.extra_blocks is not None:
            results, names = self.extra_blocks(results, names)

        # make it back an OrderedDict
        out = OrderedDict([(k, v) for k, v in zip(names, results)])

        return out


class LastLevelMaxPool(torch.nn.Module):
    """
    Applies a max_pool2d on top of the last feature map
    """

    def forward(self, x, names):
        # type: (List[Tensor], List[str]) -> Tuple[List[Tensor], List[str]]
        names.append("pool")
        x.append(F.max_pool2d(x[-1], 1, 2, 0))
        return x, names


class BackboneWithFPN(nn.Module):
    """
    Adds a FPN on top of a models.
    Internally, it uses torchvision.models._utils.IntermediateLayerGetter to
    extract a submodel that returns the feature maps specified in return_layers.
    The same limitations of IntermediatLayerGetter apply here.
    Arguments:
        backbone (nn.Module)
        return_layers (Dict[name, new_name]): a dict containing the names
            of the modules for which the activations will be returned as
            the key of the dict, and the value of the dict is the name
            of the returned activation (which the user can specify).
        in_channels_list (List[int]): number of channels for each feature map
            that is returned, in the order they are present in the OrderedDict
        out_channels (int): number of channels in the FPN.
    Attributes:
        out_channels (int): the number of channels in the FPN
    """

    def __init__(self, backbone, return_layers, in_channels_list, out_channels):
        '''

        :param backbone: 特征层
        :param return_layers: 返回的层数
        :param in_channels_list: 输入通道数
        :param out_channels: 输出通道数
        '''
        super(BackboneWithFPN, self).__init__()
        '返回有序字典模型'
        self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)

        self.fpn = FeaturePyramidNetwork(
            in_channels_list=in_channels_list,
            out_channels=out_channels,
            extra_blocks=LastLevelMaxPool(),
            )
        # super(BackboneWithFPN, self).__init__(OrderedDict(
        #     [("body", body), ("fpn", fpn)]))
        self.out_channels = out_channels

    def forward(self, x):
        x = self.body(x)
        x = self.fpn(x)
        return x


def resnet50_fpn_backbone():
    # FrozenBatchNorm2d的功能与BatchNorm2d类似，但参数无法更新
    # norm_layer=misc.FrozenBatchNorm2d
    resnet_backbone = ResNet(Bottleneck, [3, 4, 6, 3],
                             include_top=False)

    # freeze layers
    # 冻结layer1及其之前的所有底层权重（基础通用特征）
    for name, parameter in resnet_backbone.named_parameters():
        if 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
            '''
            冻结权重，不参与训练
            '''
            parameter.requires_grad_(False)
    # 字典名字
    return_layers = {'layer1': '0', 'layer2': '1', 'layer3': '2', 'layer4': '3'}

    # in_channel 为layer4的输出特征矩阵channel = 2048
    in_channels_stage2 = resnet_backbone.in_channel // 8

    in_channels_list = [
        in_channels_stage2,  # layer1 out_channel=256
        in_channels_stage2 * 2,  # layer2 out_channel=512
        in_channels_stage2 * 4,  # layer3 out_channel=1024
        in_channels_stage2 * 8,  # layer4 out_channel=2048
    ]
    out_channels = 256
    return BackboneWithFPN(resnet_backbone, return_layers, in_channels_list, out_channels)


if __name__ == '__main__':
    net = resnet50_fpn_backbone()
    x = torch.randn(1,3,224,224)
    for key,value in net(x).items():
        print(key,value.shape)

• 测试结果