这张图其实有问题，一个特征图的来源最多有四个，图中画了5个，绿色和紫色的来源不能同时存在，如果配置文件的paths=['bu'] 则是绿色，否则是紫色。
这是配置文件
in_channels=[256, 512, 1024, 2048],
out_channels=256,
inter_channels=256,
num_outs=5,
stack_times=4,
paths=['bu'] * 4,
same_down_trans=None,
same_up_trans=dict(
type='conv',
kernel_size=3,
stride=2,
padding=1,
norm_cfg=norm_cfg,
inplace=False,
order=('act', 'conv', 'norm')),
across_lateral_trans=dict(
type='conv',
kernel_size=1,
norm_cfg=norm_cfg,
inplace=False,
order=('act', 'conv', 'norm')),
across_down_trans=dict(
type='interpolation_conv',
mode='nearest',
kernel_size=3,
norm_cfg=norm_cfg,
order=('act', 'conv', 'norm'),
inplace=False),
across_up_trans=None,
across_skip_trans=dict(
type='conv',
kernel_size=1,
norm_cfg=norm_cfg,
inplace=False,
order=('act', 'conv', 'norm')),
output_trans=dict(
type='last_conv',
kernel_size=3,
order=('act', 'conv', 'norm'),
inplace=False),
norm_cfg=norm_cfg,
skip_inds=[(0, 1, 2, 3), (0, 1, 2), (0, 1), (0, ), ()]))stack_times=9，这个时一共多少列
paths=['bu'] * 9, 这个时路径，bu是从下而上，大特征图用stride卷积逐步变小，特征图从下往上计算，最后计算最小的特征图
skip_inds=[(0, 1, 2, 3), (0, 1, 2), (0, 1), (0, ), ()]))
这个很难理解，五个元组代表五层输出，一个元组代表一列即一个stack
第0层的0，1，2，3列，跳过，也就是从上一列对应层数直接拿过来，不处理
第1层的第0，1，2 列直接从上一列拿
第2层的第0，1列直接从上一列拿
第3层的第0，0列直接从上一列拿
第4层不从前面拿
一个特征图如果没被跳过，则有4个来源，除了顶层和底层只有三个来源之外。这里只说bu方向，4个来源分别：正下方的特征图下采样，左上的特征图上采样，正左方向的特征图1*1卷积，第0列的对应层的特征图1*1卷积，然后把这几个加起来。如果是顶层和底层就没有左上方上采样和下方下采样。
注意上采样时，先进行一个3*3卷积再进行一个上采样。
最开始的时候
# build all levels from original feature maps
feats = [
lateral_conv(inputs[i + self.start_level])
for i, lateral_conv in enumerate(self.lateral_convs)
]
for downsample in self.extra_downsamples:
feats.append(downsample(feats[-1]))
先变成256通道的，如果层数不够，则上采样，凑出额外层，也就时FPN的extra_layer= 'input'
最后的时候：把最后一列做一个3*3的卷积
我把self.fpn_transitions 网络显示出来，篇幅限制我把bn 和激活去了。
(0): ModuleList(
(0): ModuleDict()
(1): ModuleDict()
(2): ModuleDict()
(3): ModuleDict()
(4): ModuleDict((same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(across_down): None
(across_up): None
(across_skip): ConvModule(
)
)
)
最外层的(0): ModuleList代表一列，里面的(0): ModuleDict() (1): ModuleDict() (2):ModuleDict()
(3): ModuleDict() (4): ModuleDict 代表每一层。
如果为空则他被跳过了，
如果不为空，下面举例，他是顶层所以只有三个部分。
(4): ModuleDict(
正下方下采样，如果来自同一列则用same,不同列用across
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(same_down): None
前一列对应层卷积
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
bu的时候左上角上采样
(across_down): None
tb的时候右下角上采样
(across_up): None
对第0列卷积
(across_skip): ConvModule(
fpn_transitions：
ModuleList(
(0): ModuleList(
(0): ModuleDict()
(1): ModuleDict()
(2): ModuleDict()
(3): ModuleDict()
(4): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): None
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
)
(1): ModuleList(
(0): ModuleDict()
(1): ModuleDict()
(2): ModuleDict()
(3): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): UpInterpolationConv(
(conv): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(4): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): None
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
)
(2): ModuleList(
(0): ModuleDict()
(1): ModuleDict()
(2): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): UpInterpolationConv(
(conv): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(3): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): UpInterpolationConv(
(conv): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(4): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): None
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
)
(3): ModuleList(
(0): ModuleDict()
(1): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): UpInterpolationConv(
(conv): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(2): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): UpInterpolationConv(
(conv): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(3): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): UpInterpolationConv(
(conv): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
(4): ModuleDict(
(same_up): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(same_down): None
(across_lateral): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
(across_down): None
(across_up): None
(across_skip): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activate): ReLU()
)
)
)
)
