mmDetection源码分析（四）：Faster R-CNN模块解读（二）

发表于   |   分类于   |  

Faster R-CNN模块解读（二） Backbone的构建

上一篇博客介绍了检测器的构建，是通过build_from_cfg(cfg.model, DETECTORS, dict(train_cfg=train_cfg, test_cfg=test_cfg))调用得到，而构建检测器时，必须要构建Backbone，在TwoStageDetector类初始化时，就需要构建self.backbone = builder.build_backbone(backbone)。而在

mmdetection/mmdet/models/builder.py

1
2
3
4
5
6
7
8
9
10
11
12
13

def build(cfg, registry, default_args=None):
    if isinstance(cfg, list):
        modules = [
            build_from_cfg(cfg_, registry, default_args) for cfg_ in cfg
        ]
        return nn.Sequential(*modules)
    else:
        return build_from_cfg(cfg, registry, default_args)


def build_backbone(cfg):
    return build(cfg, BACKBONES)

可以看到同样是调用build_from_cfg(cfg.model.backbone, BACKBONES, dict(train_cfg=train_cfg, test_cfg=test_cfg))来构建Backbone的。

在mmdetection/configs/faster_rcnn_r50_fpn_1x.py中，backbone的字典参数如下：

1
2
3
4
5
6
7
backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        style='pytorch'),

此处type为ResNet，则将会去构造mmdetection/mmdet/models/backbones/resnet.py这个类。

该类则是大名鼎鼎的ResNet，下面将会与ResNet论文 中的网络结构一起来学习ResNet的实现。主要介绍ResNet-18与ResNet-50。

对网络结构不是很熟悉的同学，可能还要借助caffe的prototxt可视化一起，会有更深层的理解，也会对编写该代码的作者更加崇拜。

ResNet-18 prototxt：https://raw.githubusercontent.com/marvis/pytorch-caffe-darknet-convert/master/cfg/resnet-18.prototxt

可视化网页：
https://ethereon.github.io/netscope/#/editor

将prototxt复制进左边的输入框，shift+enter即可以看到可视化的网络结构。

ResNet-50，何恺明大大已经可视化了：http://ethereon.github.io/netscope/#/gist/db945b393d40bfa26006

准备工作做好后，我们便正式开始学习Backbone（ResNet）的构建了。

先跳过BasicBlock，Bottleneck这两个类，看ResNet这个类。

在__init__函数之前，

1
2
3
4
5
6
7
arch_settings = {
    18: (BasicBlock, (2, 2, 2, 2)),
    34: (BasicBlock, (3, 4, 6, 3)),
    50: (Bottleneck, (3, 4, 6, 3)),
    101: (Bottleneck, (3, 4, 23, 3)),
    152: (Bottleneck, (3, 8, 36, 3))
}

代码中很清晰的将不同层数的ResNet的区别，以及不同层数的ResNet的配置都一一告知。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84

def __init__(self,
             depth,
             num_stages=4,
             strides=(1, 2, 2, 2),
             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]
    self.inplanes = 64

    self._make_stem_layer()

    self.res_layers = []
    for i, num_blocks in enumerate(self.stage_blocks):
        stride = strides[i]
        dilation = dilations[i]
        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
        res_layer = make_res_layer(
            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
        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)

从__init__函数中，我们可以看到，首先是将一系列的卷积层的变量传递到该类当中，比如：stride（步长），dilation（扩张率）等，之后调用_make_stem_layer()。

1
2
3
4
5
6
7
8
9
10
11
12
13
def _make_stem_layer(self):
    self.conv1 = build_conv_layer(
        self.conv_cfg,
        3,
        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)

从这个类，我们可以看到定义了conv1是一个7x7、步长为2，padding为3，输入channel为3，输出的channel为64，之后再接BatchNorm层与ReLU层，最后接一个kernel_size为3，步长为2的maxPooling 层

可以看到与可视化的第一部分一致，经过conv1与maxPooling之后，Feature map的尺寸变成原来的1/4。

之后便是一个循环，构造每个stage的卷积层。通过调用make_res_layer，构造每个stage的卷积层。make_res_layer有个参数block，用于指定使用的是BasicBlock还是Bottleneck。通过__init__中的
self.block, stage_blocks = self.arch_settings[depth]设置，可以看到ResNet-18使用的是BasicBlock，而ResNet-50使用的是Bottleneck。两者的区别可以从图1不同层数的ResNet结构图看出，BasicBlock是堆叠了两个3x3的卷积核，Bottleneck则是1x1->3x3->1x1。具体结构如下图：

通过代码，我们同样可以看出区别。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
class BasicBlock(nn.Module):
    expansion = 1

        self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
        self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)

        self.conv1 = build_conv_layer(
            conv_cfg,
            inplanes,
            planes,
            3,
            stride=stride,
            padding=dilation,
            dilation=dilation,
            bias=False)
        self.add_module(self.norm1_name, norm1)
        self.conv2 = build_conv_layer(
            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

而Bottleneck，此处的代码忽略了一些判断，dcn，gcb等。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
class Bottleneck(nn.Module):
    expansion = 4
     
     self.conv1 = build_conv_layer(
            conv_cfg,
            inplanes,
            planes,
            kernel_size=1,
            stride=self.conv1_stride,
            bias=False)
        self.add_module(self.norm1_name, norm1)
        fallback_on_stride = False
        self.with_modulated_dcn = False
        if self.with_dcn:
            fallback_on_stride = dcn.get('fallback_on_stride', False)
            self.with_modulated_dcn = dcn.get('modulated', False)
        if not self.with_dcn or fallback_on_stride:
            self.conv2 = build_conv_layer(
                conv_cfg,
                planes,
                planes,
                kernel_size=3,
                stride=self.conv2_stride,
                padding=dilation,
                dilation=dilation,
                bias=False)
        self.conv3 = build_conv_layer(
            conv_cfg,
            planes,
            planes * self.expansion,
            kernel_size=1,
            bias=False)
        self.add_module(self.norm3_name, norm3)

        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

简单了解了下BasicBlock和Bottleneck之后，我们来看下make_res_layer

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def make_res_layer(block,
                   inplanes,
                   planes,
                   blocks,
                   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(
            build_conv_layer(
                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(
            inplanes=inplanes,
            planes=planes,
            stride=stride,
            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
    for i in range(1, blocks):
        layers.append(
            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)

可以看到若stride不等于1或者输入的Feature map的channel不等于64 × expansion（BasicBlock.expansion = 1，Bottleneck.expansion = 4），则downsample设置成1x1x（64 × expansion），在可视化部分如图所示，downsample便指的是这一块的卷积层。

shortcut的操作（恒等映射）在BasicBlock和Bottleneck的forward函数中。

class BasicBlock(nn.Module):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

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

class Bottleneck(nn.Module):

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
def forward(self, x):

    def _inner_forward(x):
        identity = x

        out = self.conv1(x)
        out = self.norm1(out)
        out = self.relu(out)

        if not self.with_dcn:
            out = self.conv2(out)
        elif self.with_modulated_dcn:
            offset_mask = self.conv2_offset(out)
            offset = offset_mask[:, :18, :, :]
            mask = offset_mask[:, -9:, :, :].sigmoid()
            out = self.conv2(out, offset, mask)
        else:
            offset = self.conv2_offset(out)
            out = self.conv2(out, offset)
        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

需要注意的是在每个stage，只有第一次才会传入downsample，之后是不传入的。这和ResNet的可视化结果一致。ResNet-50第一个stage也确实是堆叠了3个Bottleneck，与代码中也一致。

此外还需要说明的是在每个stage之后channel数会发生变化。

另外，在resnet.py中还有init_weights，forward，train函数。

  • init_weights在two_stage.py被调用，将预训练权重加载到backbone中。

  • forward函数则是在two_stage.py中的extract_feat(img)中通过self.backbone(img)调用，这是因为ResNet类继承至nn.Module，而在nn.Module中实现了__call__，在nn.Module的__call__中result = self.forward(*input, **kwargs)，因此可以通过self.backbone(img)直接调用self.backbone.forward(img)。

在本篇博客中，我们详细介绍了mmDetection框架中backbone的构建，此处以ResNet为例，结合论文、caffe的prototxt的可视化和代码，一起讲述了ResNet网络的构建过程。