【mmdetection源码解读(一)】骨干网络--ResNet

以下仅为个人理解，若有不正之处还请指出，欢迎交流！

• 本文解读的源码为mmdet/models/backbones中的resnet.py
• 首先附上ResNet原文地址Deep Residual Learning for Image Recognition
• 其中，ResNet整体网络结构图如下：
  
• 文末附有ResNet-50完整的网络结构图

一、ResNet网络中的两种基本残差块

• 由网络结构图可以看出，ResNet-18和ResNet-34使用的为包含2个3×3卷积层的残差块，在源码中对应BasicBlock
• 而更深层的ResNet-50，ResNet-101，ResNet-152使用的残差块则略微复杂一些，在源码中对应Bottleneck

1.class BasicBlock(nn.Module)

class BasicBlock(nn.Module):   # 对应ResNet-18和ResNet-34中使用的残差块
    expansion = 1
    def __init__(self,
                 inplanes,
                 planes,
                 stride=1,
                 dilation=1,
                 downsample=None,
                 style='pytorch',
                 with_cp=False,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 dcn=None,
                 gcb=None,
                 gen_attention=None):
        super(BasicBlock, self).__init__()
        assert dcn is None, "Not implemented yet."
        assert gen_attention is None, "Not implemented yet."
        assert gcb is None, "Not implemented yet."

        self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)  # 'BN': ('bn', nn.BatchNorm2d)
        self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)  # postfix->后缀
        self.conv1 = build_conv_layer(    # 3×3卷积层
            conv_cfg,                     # 'Conv': nn.Conv2d 默认
            inplanes,
            planes,
            3,
            stride=stride,
            padding=dilation,
            dilation=dilation,
            bias=False)
        self.add_module(self.norm1_name, norm1)
        self.conv2 = build_conv_layer(    # 3×3卷积层
            conv_cfg, planes, planes, 3, padding=1, bias=False)
        self.add_module(self.norm2_name, norm2)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride
        self.dilation = dilation
        assert not with_cp

    @property
    def norm1(self):
        return getattr(self, self.norm1_name)
    @property
    def norm2(self):
        return getattr(self, self.norm2_name)

    def forward(self, x):
        identity = x
        out = self.conv1(x)
        out = self.norm1(out)
        out = self.relu(out)
        out = self.conv2(out)
        out = self.norm2(out)
        if self.downsample is not None:    
            identity = self.downsample(x)
        out += identity
        out = self.relu(out)
        return out

2.class Bottleneck(nn.Module)

class Bottleneck(nn.Module):     # 对应ResNet网络层数大于50的情况
    expansion = 4

    def __init__(self,
                 inplanes,
                 planes,
                 stride=1,
                 dilation=1,
                 downsample=None,
                 style='pytorch',
                 with_cp=False,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN'),
                 dcn=None,
                 gcb=None,
                 gen_attention=None):
        """Bottleneck block for ResNet.
        If style is "pytorch", the stride-two layer is the 3x3 conv layer,
        if it is "caffe", the stride-two layer is the first 1x1 conv layer.
        """
        super(Bottleneck, self).__init__()
        assert style in ['pytorch', 'caffe']
        assert dcn is None or isinstance(dcn, dict)
        assert gcb is None or isinstance(gcb, dict)
        assert gen_attention is None or isinstance(gen_attention, dict)

        self.inplanes = inplanes
        self.planes = planes
        self.stride = stride
        self.dilation = dilation
        self.style = style
        self.with_cp = with_cp
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.dcn = dcn
        self.with_dcn = dcn is not None
        self.gcb = gcb
        self.with_gcb = gcb is not None
        self.gen_attention = gen_attention
        self.with_gen_attention = gen_attention is not None

        if self.style == 'pytorch': 
            self.conv1_stride = 1
            self.conv2_stride = stride
        else:
            self.conv1_stride = stride
            self.conv2_stride = 1

        self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)  # 'BN': ('bn', nn.BatchNorm2d)
        self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)  # postfix->后缀
        self.norm3_name, norm3 = build_norm_layer(
            norm_cfg, planes * self.expansion, postfix=3)
        self.conv1 = build_conv_layer(  #  1×1卷积层
            conv_cfg,                   # 'Conv': nn.Conv2d 默认
            inplanes,
            planes,
            kernel_size=1,
            stride=self.conv1_stride, 
            bias=False)
        self.add_module(self.norm1_name, norm1)
        fallback_on_stride = False
        if self.with_dcn:
            fallback_on_stride = dcn.pop('fallback_on_stride', False)
        if not self.with_dcn or fallback_on_stride:
            self.conv2 = build_conv_layer(   # 3×3卷积层
                conv_cfg,
                planes,
                planes,
                kernel_size=3,
                stride=self.conv2_stride,
                padding=dilation,
                dilation=dilation,
                bias=False)
        else:
            assert self.conv_cfg is None, 'conv_cfg cannot be None for DCN'
            self.conv2 = build_conv_layer(
                dcn,
                planes,
                planes,
                kernel_size=3,
                stride=self.conv2_stride,
                padding=dilation,
                dilation=dilation,
                bias=False)
        self.add_module(self.norm2_name, norm2)
        self.conv3 = build_conv_layer(  # 1×1卷积层
            conv_cfg,
            planes,
            planes * self.expansion,    # 输出通道数扩大四倍
            kernel_size=1,
            bias=False)
        self.add_module(self.norm3_name, norm3)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        if self.with_gcb:
            gcb_inplanes = planes * self.expansion
            self.context_block = ContextBlock(inplanes=gcb_inplanes, **gcb)
        # gen_attention
        if self.with_gen_attention:
            self.gen_attention_block = GeneralizedAttention(
                planes, **gen_attention)
                
    @property
    def norm1(self):
        return getattr(self, self.norm1_name)
    @property
    def norm2(self):
        return getattr(self, self.norm2_name)
    @property
    def norm3(self):
        return getattr(self, self.norm3_name)

    def forward(self, x):
        def _inner_forward(x):
            identity = x
            out = self.conv1(x)
            out = self.norm1(out)
            out = self.relu(out)
            out = self.conv2(out)
            out = self.norm2(out)
            out = self.relu(out)
            if self.with_gen_attention:
                out = self.gen_attention_block(out)
            out = self.conv3(out)
            out = self.norm3(out)
            if self.with_gcb:
                out = self.context_block(out)
            if self.downsample is not None:   
                identity = self.downsample(x)
            out += identity    
            return out
        if self.with_cp and x.requires_grad:
            out = cp.checkpoint(_inner_forward, x)
        else:
            out = _inner_forward(x)
        out = self.relu(out)
        return out

二、构造ResNet网络中的conv2_x、conv3_x、conv4_x、conv5_x---->4个stage

def make_res_layer(block,     # 对应上方两种残差块的类型，下文均以ResNet-50为例进行解释
                   inplanes,
                   planes,
                   blocks,    # 当前stage有几个残差块  50: (Bottleneck, (3, 4, 6, 3))
                   stride=1,
                   dilation=1,
                   style='pytorch',
                   with_cp=False,
                   conv_cfg=None,
                   norm_cfg=dict(type='BN'),
                   dcn=None,
                   gcb=None,
                   gen_attention=None,
                   gen_attention_blocks=[]):
    downsample = None
    if stride != 1 or inplanes != planes * block.expansion:
        downsample = nn.Sequential(     # 通过1×1卷积增加通道数为原来的4倍，为了残差块中相加的项有着相同的维度
            build_conv_layer(    # 实际4个stage中各自第一个残差块中有该部分，其余残差块没有
                conv_cfg,
                inplanes,
                planes * block.expansion,
                kernel_size=1,
                stride=stride,
                bias=False),
            build_norm_layer(norm_cfg, planes * block.expansion)[1],
        )
    layers = []
    layers.append(
        block(                # 当前stage的第一个残差块
            inplanes=inplanes,
            planes=planes,
            stride=stride,    # conv3_1,conv4_1,conv5_1,需要通过s=2的卷积层下采样
            dilation=dilation,
            downsample=downsample,
            style=style,
            with_cp=with_cp,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            dcn=dcn,
            gcb=gcb,
            gen_attention=gen_attention if
            (0 in gen_attention_blocks) else None))
    inplanes = planes * block.expansion   #  ×4
    for i in range(1, blocks):
        layers.append(                   # 当前stage的剩余残差块堆叠起来
            block(                       
                inplanes=inplanes,
                planes=planes,
                stride=1,
                dilation=dilation,
                style=style,
                with_cp=with_cp,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                dcn=dcn,
                gcb=gcb,
                gen_attention=gen_attention if
                (i in gen_attention_blocks) else None))

    return nn.Sequential(*layers)

三、class ResNet(nn.Module)

class ResNet(nn.Module):
    # 所有的ResNet网络都是通过这个类产生，只需要传入不同的参数
    """ResNet backbone.
    Args:
        depth (int): Depth of resnet, from {18, 34, 50, 101, 152}.
        in_channels (int): Number of input image channels. Normally 3.
        num_stages (int): Resnet stages, normally 4.
        strides (Sequence[int]): Strides of the first block of each stage.
        dilations (Sequence[int]): Dilation of each stage.
        out_indices (Sequence[int]): Output from which stages.  # 有几个输出
        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
            layer is the 3x3 conv layer, otherwise the stride-two layer is
            the first 1x1 conv layer.
        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
            -1 means not freezing any parameters.
        norm_cfg (dict): dictionary to construct and config norm layer.
        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.
        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
            memory while slowing down the training speed.
        zero_init_residual (bool): whether to use zero init for last norm layer
            in resblocks to let them behave as identity.
    Example:
        >>> from mmdet.models import ResNet
        >>> import torch
        >>> self = ResNet(depth=18)
        >>> self.eval()
        >>> inputs = torch.rand(1, 3, 32, 32)
        >>> level_outputs = self.forward(inputs)
        >>> for level_out in level_outputs:
        ...     print(tuple(level_out.shape))
        (1, 64, 8, 8)
        (1, 128, 4, 4)
        (1, 256, 2, 2)
        (1, 512, 1, 1)
    """
    arch_settings = {
        18: (BasicBlock, (2, 2, 2, 2)),   # 对应原文中的网络结构
        34: (BasicBlock, (3, 4, 6, 3)),
        50: (Bottleneck, (3, 4, 6, 3)),   # 下文以ResNet-50为例
        101: (Bottleneck, (3, 4, 23, 3)),
        152: (Bottleneck, (3, 8, 36, 3))
    }
    def __init__(self,
                 depth,
                 in_channels=3,
                 num_stages=4,    # conv2_x、conv3_x、conv4_x、conv5_x ---->4个stage
                 strides=(1, 2, 2, 2), # conv2_x中不通过卷积层进行下采样，conv3_1,conv4_1,conv5_1,需要通过s=2的卷积层下采样
                 dilations=(1, 1, 1, 1),
                 out_indices=(0, 1, 2, 3),
                 style='pytorch',
                 frozen_stages=-1,
                 conv_cfg=None,
                 norm_cfg=dict(type='BN', requires_grad=True),
                 norm_eval=True,
                 dcn=None,
                 stage_with_dcn=(False, False, False, False),
                 gcb=None,
                 stage_with_gcb=(False, False, False, False),
                 gen_attention=None,
                 stage_with_gen_attention=((), (), (), ()),
                 with_cp=False,
                 zero_init_residual=True):
        super(ResNet, self).__init__()
        if depth not in self.arch_settings:
            raise KeyError('invalid depth {} for resnet'.format(depth))
        self.depth = depth
        self.num_stages = num_stages
        assert num_stages >= 1 and num_stages <= 4
        self.strides = strides
        self.dilations = dilations
        assert len(strides) == len(dilations) == num_stages
        self.out_indices = out_indices
        assert max(out_indices) < num_stages
        self.style = style
        self.frozen_stages = frozen_stages
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        self.with_cp = with_cp
        self.norm_eval = norm_eval
        self.dcn = dcn
        self.stage_with_dcn = stage_with_dcn
        if dcn is not None:
            assert len(stage_with_dcn) == num_stages  
        self.gen_attention = gen_attention
        self.gcb = gcb
        self.stage_with_gcb = stage_with_gcb
        if gcb is not None:
            assert len(stage_with_gcb) == num_stages
        self.zero_init_residual = zero_init_residual
        self.block, stage_blocks = self.arch_settings[depth]
        self.stage_blocks = stage_blocks[:num_stages]   # 50: (Bottleneck, (3, 4, 6, 3))
        self.inplanes = 64
        self._make_stem_layer(in_channels)
        self.res_layers = []
        for i, num_blocks in enumerate(self.stage_blocks):
            stride = strides[i]      # (1, 2, 2, 2)
            dilation = dilations[i]  # (1, 1, 1, 1)
            dcn = self.dcn if self.stage_with_dcn[i] else None
            gcb = self.gcb if self.stage_with_gcb[i] else None
            planes = 64 * 2**i    # 64 128 256 512
            res_layer = make_res_layer(   # 详见上一节，构建4个stage
                self.block,  
                self.inplanes,
                planes,
                num_blocks,
                stride=stride,
                dilation=dilation,
                style=self.style,
                with_cp=with_cp,
                conv_cfg=conv_cfg,
                norm_cfg=norm_cfg,
                dcn=dcn,
                gcb=gcb,
                gen_attention=gen_attention,
                gen_attention_blocks=stage_with_gen_attention[i])
            self.inplanes = planes * self.block.expansion   # × 4
            layer_name = 'layer{}'.format(i + 1)
            self.add_module(layer_name, res_layer)
            self.res_layers.append(layer_name)
        self._freeze_stages()
        self.feat_dim = self.block.expansion * 64 * 2**(
            len(self.stage_blocks) - 1)
            
    @property
    def norm1(self):
        return getattr(self, self.norm1_name)
    def _make_stem_layer(self, in_channels):
        self.conv1 = build_conv_layer(   # 构建网络结构中的conv1层
            self.conv_cfg,
            in_channels,
            64,
            kernel_size=7,
            stride=2,
            padding=3,
            bias=False)
        self.norm1_name, norm1 = build_norm_layer(self.norm_cfg, 64, postfix=1)
        self.add_module(self.norm1_name, norm1)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) # 网络结构中conv1之后的maxpooling层
    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            self.norm1.eval()
            for m in [self.conv1, self.norm1]:
                for param in m.parameters():
                    param.requires_grad = False    # 不进行梯度下降
        for i in range(1, self.frozen_stages + 1):
            m = getattr(self, 'layer{}'.format(i))
            m.eval()
            for param in m.parameters():
                param.requires_grad = False
    def init_weights(self, pretrained=None):
        if isinstance(pretrained, str):
            from mmdet.apis import get_root_logger
            logger = get_root_logger()
            load_checkpoint(self, pretrained, strict=False, logger=logger)
        elif pretrained is None:
            for m in self.modules():
                if isinstance(m, nn.Conv2d):
                    kaiming_init(m)
                elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
                    constant_init(m, 1)
            if self.dcn is not None:
                for m in self.modules():
                    if isinstance(m, Bottleneck) and hasattr(
                            m, 'conv2_offset'):
                        constant_init(m.conv2_offset, 0)
            if self.zero_init_residual:
                for m in self.modules():
                    if isinstance(m, Bottleneck):
                        constant_init(m.norm3, 0)
                    elif isinstance(m, BasicBlock):
                        constant_init(m.norm2, 0)
        else:
            raise TypeError('pretrained must be a str or None')

    def forward(self, x):    # 完整ResNet网络结构
        x = self.conv1(x)
        x = self.norm1(x)
        x = self.relu(x)
        x = self.maxpool(x)
        outs = []
        for i, layer_name in enumerate(self.res_layers):
            res_layer = getattr(self, layer_name)
            x = res_layer(x)
            if i in self.out_indices:  
                outs.append(x)
        return tuple(outs)

    def train(self, mode=True):
        super(ResNet, self).train(mode)
        self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                # trick: eval have effect on BatchNorm only
                if isinstance(m, _BatchNorm):
                    m.eval()

四、ResNet-50详细网络结构图（图源网络）