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):
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)
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
@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):
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)
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
self.norm3_name, norm3 = build_norm_layer(
norm_cfg, planes * self.expansion, postfix=3)
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
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(
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(
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
if self.with_gcb:
gcb_inplanes = planes * self.expansion
self.context_block = ContextBlock(inplanes=gcb_inplanes, **gcb)
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,
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)
三、class ResNet(nn.Module)
class ResNet(nn.Module):
"""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)),
34: (BasicBlock, (3, 4, 6, 3)),
50: (Bottleneck, (3, 4, 6, 3)),
101: (Bottleneck, (3, 4, 23, 3)),
152: (Bottleneck, (3, 8, 36, 3))
}
def __init__(self,
depth,
in_channels=3,
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(in_channels)
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)
@property
def norm1(self):
return getattr(self, self.norm1_name)
def _make_stem_layer(self, in_channels):
self.conv1 = build_conv_layer(
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)
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
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):
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():
if isinstance(m, _BatchNorm):
m.eval()
四、ResNet-50详细网络结构图(图源网络)