def __init__(self, inp, oup, stride, expand_ratio):
super(InvertedResidual, self).__init__()
self.stride = stride
assert stride in [1, 2]
hidden_dim = int(round(inp * expand_ratio))
self.use_res_connect = self.stride == 1 and inp == oup
if expand_ratio == 1:
self.conv = nn.Sequential(
# dw
Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
# pw-linear
Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
BatchNorm2d(oup),
)
else:
self.conv = nn.Sequential(
# pw
Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
# dw
Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
BatchNorm2d(hidden_dim),
nn.ReLU6(inplace=True),
# pw-linear
Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
BatchNorm2d(oup),
)
def __init__(self, C):
super(SEModule, self).__init__()
mid = max(C // self.reduction, 8)
conv1 = Conv2d(C, mid, 1, 1, 0)
conv2 = Conv2d(mid, C, 1, 1, 0)
self.op = nn.Sequential(nn.AdaptiveAvgPool2d(1), conv1,
nn.ReLU(inplace=True), conv2, nn.Sigmoid())
def __init__(self, C_in, C_out, stride):
assert stride in [1, 2]
ops = [
Conv2d(C_in, C_in, 3, stride, 1, bias=False),
BatchNorm2d(C_in),
nn.ReLU(inplace=True),
Conv2d(C_in, C_out, 3, 1, 1, bias=False),
BatchNorm2d(C_out),
]
super(CascadeConv3x3, self).__init__(*ops)
self.res_connect = (stride == 1) and (C_in == C_out)
def make_conv(
in_channels, out_channels, kernel_size, stride=1, dilation=1
):
conv = Conv2d(
in_channels,
out_channels,
kernel_size=kernel_size,
stride=stride,
padding=dilation * (kernel_size - 1) // 2,
dilation=dilation,
bias=False if use_gn else True
)
# Caffe2 implementation uses XavierFill, which in fact
# corresponds to kaiming_uniform_ in PyTorch
nn.init.kaiming_uniform_(conv.weight, a=1)
if not use_gn:
nn.init.constant_(conv.bias, 0)
module = [conv,]
if use_gn:
module.append(group_norm(out_channels))
if use_relu:
module.append(nn.ReLU(inplace=True))
if len(module) > 1:
return nn.Sequential(*module)
return conv
def make_conv3x3(
in_channels,
out_channels,
dilation=1,
stride=1,
use_gn=False,
use_relu=False,
kaiming_init=True
):
conv = Conv2d(
in_channels,
out_channels,
kernel_size=3,
stride=stride,
padding=dilation,
dilation=dilation,
bias=False if use_gn else True
)
if kaiming_init:
nn.init.kaiming_normal_(
conv.weight, mode="fan_out", nonlinearity="relu"
)
else:
torch.nn.init.normal_(conv.weight, std=0.01)
if not use_gn:
nn.init.constant_(conv.bias, 0)
module = [conv,]
if use_gn:
module.append(group_norm(out_channels))
if use_relu:
module.append(nn.ReLU(inplace=True))
if len(module) > 1:
return nn.Sequential(*module)
return conv
def __init__(self, cfg, in_channels):
super(KeypointRCNNFeatureExtractor, self).__init__()
resolution = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_RESOLUTION
scales = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_SCALES
sampling_ratio = cfg.MODEL.ROI_KEYPOINT_HEAD.POOLER_SAMPLING_RATIO
pooler = Pooler(
output_size=(resolution, resolution),
scales=scales,
sampling_ratio=sampling_ratio,
)
self.pooler = pooler
input_features = in_channels
layers = cfg.MODEL.ROI_KEYPOINT_HEAD.CONV_LAYERS
next_feature = input_features
self.blocks = []
for layer_idx, layer_features in enumerate(layers, 1):
layer_name = "conv_fcn{}".format(layer_idx)
module = Conv2d(next_feature, layer_features, 3, stride=1, padding=1)
nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
nn.init.constant_(module.bias, 0)
self.add_module(layer_name, module)
next_feature = layer_features
self.blocks.append(layer_name)
self.out_channels = layer_features
def __init__(self, C_in, C_out, expansion, stride):
assert stride in [1, 2]
self.res_connect = (stride == 1) and (C_in == C_out)
C_mid = _get_divisible_by(C_in * expansion, 8, 8)
ops = [
# pw
Conv2d(C_in, C_mid, 1, 1, 0, bias=False),
BatchNorm2d(C_mid),
nn.ReLU(inplace=True),
# shift
Shift(C_mid, 5, stride, 2),
# pw-linear
Conv2d(C_mid, C_out, 1, 1, 0, bias=False),
BatchNorm2d(C_out),
]
super(ShiftBlock5x5, self).__init__(*ops)
def __init__(self, cfg, norm_func):
super(BaseStem, self).__init__()
out_channels = cfg.MODEL.RESNETS.STEM_OUT_CHANNELS
self.conv1 = Conv2d(
3, out_channels, kernel_size=7, stride=2, padding=3, bias=False
)
self.bn1 = norm_func(out_channels)
for l in [self.conv1,]:
nn.init.kaiming_uniform_(l.weight, a=1)
def __init__(self, in_channels, bottleneck_channels, out_channels,
num_groups, stride_in_1x1, stride, dilation, norm_func,
dcn_config):
"""
Args:
in_channels (int): Number of input channels.
out_channels (int): Number of output channels.
stride (int): Stride for the first conv.
norm (str or callable): normalization for all conv layers.
See :func:`layers.get_norm` for supported format.
"""
super().__init__()
if in_channels != out_channels:
self.shortcut = Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=stride,
bias=False)
self.bn_shortcut = norm_func(out_channels)
else:
self.shortcut = None
self.conv1 = Conv2d(in_channels,
out_channels,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
self.bn1 = norm_func(out_channels)
self.conv2 = Conv2d(out_channels,
out_channels,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn2 = norm_func(out_channels)
for layer in [self.conv1, self.conv2, self.shortcut]:
if layer is not None: # shortcut can be None
nn.init.kaiming_uniform_(layer.weight, a=1)
def __init__(
self,
input_depth,
output_depth,
kernel,
stride,
pad,
no_bias,
use_relu,
bn_type,
group=1,
*args,
**kwargs
):
super(ConvBNRelu, self).__init__()
assert use_relu in ["relu", None]
if isinstance(bn_type, (list, tuple)):
assert len(bn_type) == 2
assert bn_type[0] == "gn"
gn_group = bn_type[1]
bn_type = bn_type[0]
assert bn_type in ["bn", "af", "gn", None]
assert stride in [1, 2, 4]
op = Conv2d(
input_depth,
output_depth,
kernel_size=kernel,
stride=stride,
padding=pad,
bias=not no_bias,
groups=group,
*args,
**kwargs
)
nn.init.kaiming_normal_(op.weight, mode="fan_out", nonlinearity="relu")
if op.bias is not None:
nn.init.constant_(op.bias, 0.0)
self.add_module("conv", op)
if bn_type == "bn":
bn_op = BatchNorm2d(output_depth)
elif bn_type == "gn":
bn_op = nn.GroupNorm(num_groups=gn_group, num_channels=output_depth)
elif bn_type == "af":
bn_op = FrozenBatchNorm2d(output_depth)
if bn_type is not None:
self.add_module("bn", bn_op)
if use_relu == "relu":
self.add_module("relu", nn.ReLU(inplace=True))
def __init__(self, cfg, in_channels):
super(MaskRCNNConv1x1Predictor, self).__init__()
num_classes = cfg.MODEL.ROI_BOX_HEAD.NUM_CLASSES
num_inputs = in_channels
self.mask_fcn_logits = Conv2d(num_inputs, num_classes, 1, 1, 0)
for name, param in self.named_parameters():
if "bias" in name:
nn.init.constant_(param, 0)
elif "weight" in name:
# Caffe2 implementation uses MSRAFill, which in fact
# corresponds to kaiming_normal_ in PyTorch
nn.init.kaiming_normal_(param, mode="fan_out", nonlinearity="relu")
def __init__(self, in_channels, bottleneck_channels, out_channels,
num_groups, stride_in_1x1, stride, dilation, norm_func,
dcn_config):
super(Bottleneck, self).__init__()
self.downsample = None
if in_channels != out_channels:
down_stride = stride if dilation == 1 else 1
self.downsample = nn.Sequential(
Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=down_stride,
bias=False),
norm_func(out_channels),
)
for modules in [
self.downsample,
]:
for l in modules.modules():
if isinstance(l, Conv2d):
nn.init.kaiming_uniform_(l.weight, a=1)
if dilation > 1:
stride = 1 # reset to be 1
# The original MSRA ResNet models have stride in the first 1x1 conv
# The subsequent fb.torch.resnet and Caffe2 ResNe[X]t implementations have
# stride in the 3x3 conv
stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride)
self.conv1 = Conv2d(
in_channels,
bottleneck_channels,
kernel_size=1,
stride=stride_1x1,
bias=False,
)
self.bn1 = norm_func(bottleneck_channels)
# TODO: specify init for the above
# dcn could be none here
with_dcn = dcn_config and dcn_config.get("stage_with_dcn", False)
if with_dcn:
deformable_groups = dcn_config.get("deformable_groups", 1)
with_modulated_dcn = dcn_config.get("with_modulated_dcn", False)
self.conv2 = DFConv2d(bottleneck_channels,
bottleneck_channels,
with_modulated_dcn=with_modulated_dcn,
kernel_size=3,
stride=stride_3x3,
groups=num_groups,
dilation=dilation,
deformable_groups=deformable_groups,
bias=False)
else:
self.conv2 = Conv2d(bottleneck_channels,
bottleneck_channels,
kernel_size=3,
stride=stride_3x3,
padding=dilation,
bias=False,
groups=num_groups,
dilation=dilation)
nn.init.kaiming_uniform_(self.conv2.weight, a=1)
self.bn2 = norm_func(bottleneck_channels)
self.conv3 = Conv2d(bottleneck_channels,
out_channels,
kernel_size=1,
bias=False)
self.bn3 = norm_func(out_channels)
for l in [
self.conv1,
self.conv3,
]:
nn.init.kaiming_uniform_(l.weight, a=1)
def __init__(
self,
in_channels,
bottleneck_channels,
out_channels,
num_groups,
stride_in_1x1,
stride,
dilation,
):
super(Bottleneck, self).__init__()
self.downsample = None
if in_channels != out_channels:
down_stride = stride if dilation == 1 else 1
self.downsample = nn.Sequential(
Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=down_stride,
bias=False), )
for modules in [
self.downsample,
]:
for l in modules.modules():
if isinstance(l, Conv2d):
nn.init.kaiming_uniform_(l.weight, a=1)
if dilation > 1:
stride = 1 # reset to be 1
# The original MSRA ResNet models have stride in the first 1x1 conv
# The subsequent fb.torch.resnet and Caffe2 ResNe[X]t implementations have
# stride in the 3x3 conv
stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride)
# avoid to use name "conv" to avoid error in loading pretrained weights
self.con1 = Conv2d(
in_channels,
bottleneck_channels,
kernel_size=1,
stride=stride_1x1,
)
# TODO: specify init for the above
self.con2 = Conv2d(bottleneck_channels,
bottleneck_channels,
kernel_size=3,
stride=stride_3x3,
padding=dilation,
groups=num_groups,
dilation=dilation)
nn.init.kaiming_uniform_(self.con2.weight, a=1)
self.con3 = Conv2d(
bottleneck_channels,
out_channels,
kernel_size=1,
)
for l in [
self.con1,
self.con3,
]:
nn.init.kaiming_uniform_(l.weight, a=1)
def conv_1x1_bn(inp, oup):
return nn.Sequential(Conv2d(inp, oup, 1, 1, 0, bias=False),
BatchNorm2d(oup), nn.ReLU6(inplace=True))
def conv_bn(inp, oup, stride):
return nn.Sequential(Conv2d(inp, oup, 3, stride, 1, bias=False),
BatchNorm2d(oup), nn.ReLU6(inplace=True))
def __init__(self, cfg, norm_func):
super(BaseStem, self).__init__()
out_channels = cfg.MODEL.RESNETS.STEM_OUT_CHANNELS
self.aa = cfg.MODEL.RESNETS.STEM_AA
self.no_maxpool = cfg.MODEL.RESNETS.STEM_NO_MAXPOOL
if cfg.MODEL.RESNETS.STEM_3X3X3:
self.conv1 = nn.Sequential(
Conv2d(3,
out_channels // 2,
kernel_size=3,
stride=2,
padding=1,
bias=False),
FrozenBatchNorm2d(out_channels // 2, True),
Conv2d(out_channels // 2,
out_channels // 4,
kernel_size=3,
stride=1,
padding=1,
bias=False),
FrozenBatchNorm2d(out_channels // 4, True),
Conv2d(out_channels // 4,
out_channels,
kernel_size=3,
stride=1,
padding=1,
bias=False),
)
for modules in [
self.conv1,
]:
for l in modules.modules():
if isinstance(l, Conv2d):
nn.init.kaiming_uniform_(l.weight, a=1)
else:
self.conv1 = Conv2d(3,
out_channels,
kernel_size=7,
stride=2,
padding=3,
bias=False)
nn.init.kaiming_uniform_(self.conv1.weight, a=1)
self.bn1 = norm_func(out_channels)
if self.aa:
self.maxpool = nn.Sequential(*[
nn.MaxPool2d(kernel_size=2, stride=1),
Downsample(filt_size=3, stride=2, channels=out_channels)
])
if self.no_maxpool:
self.maxpool = nn.Sequential(
Conv2d(out_channels,
out_channels // 2,
kernel_size=3,
stride=1,
padding=1,
bias=False),
FrozenBatchNorm2d(out_channels // 2, True),
Conv2d(out_channels // 2,
out_channels,
kernel_size=3,
stride=2,
padding=1,
bias=False),
)
for i in self.maxpool.modules():
nn.init.kaiming_uniform_(i.weight, a=1)
for l in [
self.conv1,
]:
nn.init.kaiming_uniform_(l.weight, a=1)
def __init__(
self,
in_channels,
bottleneck_channels,
out_channels,
num_groups,
stride_in_1x1,
stride,
dilation,
norm_func,
dcn_config,
cfg=None,
):
super(Bottleneck, self).__init__()
self.downsample = None
self.aa = cfg.MODEL.RESNETS.BOTTLENECK_AA
if in_channels != out_channels:
down_stride = stride if dilation == 1 else 1
if down_stride > 1 and self.aa:
self.downsample = nn.Sequential(
OrderedDict([('aa',
Downsample(filt_size=3,
stride=stride,
channels=in_channels)),
('conv',
Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=1,
bias=False)),
('norm', norm_func(out_channels))]))
else:
self.downsample = nn.Sequential(
Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=down_stride,
bias=False),
norm_func(out_channels),
)
for modules in [
self.downsample,
]:
for l in modules.modules():
if isinstance(l, Conv2d):
nn.init.kaiming_uniform_(l.weight, a=1)
if dilation > 1:
stride = 1 # reset to be 1
# The original MSRA ResNet models have stride in the first 1x1 conv
# The subsequent fb.torch.resnet and Caffe2 ResNe[X]t implementations have
# stride in the 3x3 conv
stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride)
if stride_1x1 > 1 and self.aa:
self.conv1 = nn.Sequential(
OrderedDict([('aa',
Downsample(filt_size=2,
stride=stride_1x1,
channels=in_channels)),
('conv',
Conv2d(
in_channels,
bottleneck_channels,
kernel_size=1,
stride=1,
bias=False,
))]))
for i in self.conv1.modules():
if isinstance(i, Conv2d):
nn.init.kaiming_uniform_(i.weight, a=1)
else:
self.conv1 = Conv2d(
in_channels,
bottleneck_channels,
kernel_size=1,
stride=stride_1x1,
bias=False,
)
nn.init.kaiming_uniform_(self.conv1.weight, a=1)
self.bn1 = norm_func(bottleneck_channels)
# TODO: specify init for the above
with_dcn = dcn_config.get("stage_with_dcn", False)
if with_dcn:
deformable_groups = dcn_config.get("deformable_groups", 1)
with_modulated_dcn = dcn_config.get("with_modulated_dcn", False)
self.conv2 = DFConv2d(bottleneck_channels,
bottleneck_channels,
with_modulated_dcn=with_modulated_dcn,
kernel_size=3,
stride=stride_3x3,
groups=num_groups,
dilation=dilation,
deformable_groups=deformable_groups,
bias=False)
elif stride_3x3 > 1 and self.aa:
self.conv2 = nn.Sequential(
OrderedDict([('conv',
Conv2d(bottleneck_channels,
bottleneck_channels,
kernel_size=3,
stride=1,
padding=dilation,
bias=False,
groups=num_groups,
dilation=dilation)),
('aa',
Downsample(filt_size=3,
stride=stride_3x3,
channels=bottleneck_channels))]))
for i in self.conv2.modules():
if isinstance(i, Conv2d):
nn.init.kaiming_uniform_(i.weight, a=1)
else:
self.conv2 = Conv2d(bottleneck_channels,
bottleneck_channels,
kernel_size=3,
stride=stride_3x3,
padding=dilation,
bias=False,
groups=num_groups,
dilation=dilation)
nn.init.kaiming_uniform_(self.conv2.weight, a=1)
self.bn2 = norm_func(bottleneck_channels)
self.conv3 = Conv2d(bottleneck_channels,
out_channels,
kernel_size=1,
bias=False)
self.bn3 = norm_func(out_channels)
for l in [
self.conv1,
self.conv3,
]:
nn.init.kaiming_uniform_(l.weight, a=1)
def __init__(self, in_channels, bottleneck_channels, out_channels,
num_groups, stride_in_1x1, stride, dilation, norm_func,
dcn_config):
super(GrowResNeXtBottleneck, self).__init__()
self.downsample = None
self.embryo_channels = 8
self.embryo_grow_tic = 0
self.embryo_grow_tic_threshold = 2
self.adults_volume = 32
self.adults_channels = self.embryo_channels * self.adults_volume
self.adults_occupied = torch.zeros((self.adults_volume))
self.adults_bns = [None] * self.adults_volume
if in_channels != out_channels:
down_stride = stride if dilation == 1 else 1
self.downsample = nn.Sequential(
Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=down_stride,
bias=False),
norm_func(out_channels),
)
for modules in [
self.downsample,
]:
for l in modules.modules():
if isinstance(l, Conv2d):
nn.init.kaiming_uniform_(l.weight, a=1)
if dilation > 1:
stride = 1 # reset to be 1
# 策略：Xembryo和adults采用不同的optimizer，embryo长得快，成熟期或掐表或判定。
# embryo成熟后，导入adults空白部分，相应的配套conv1 conv3 也要妥善转入。
# adults的导入要管理好。要测试bn导入前后的影响，不行就强行多adults_bn。
# adults的各个体要用进废退，市占率低的要有序退出，允许embryo再生。
# adults的首尾1*1conv要细心设计其初始化。
# 对不同的类成立其weight库，因为有些weight的被使用与这类物体的出现频率相关。
# stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride)
self.embryo_conv1 = Conv2d(in_channels,
self.embryo_channels,
1,
stride,
bias=False)
self.embryo_bn1 = norm_func(self.embryo_channels)
self.embryo_conv2 = Conv2d(self.embryo_channels,
self.embryo_channels,
3,
1,
padding=dilation,
bias=False,
dilation=dilation)
self.embryo_bn2 = norm_func(self.embryo_channels)
self.embryo_conv3 = Conv2d(self.embryo_channels,
out_channels,
1,
bias=False)
self.embryo_bn3 = norm_func(out_channels)
# TODO: specify init for self.conv1 1*1conv
# TODO: try to init with spercific rules, or just init once then never change.
nn.init.kaiming_uniform_(self.embryo_conv1.weight, a=1)
nn.init.kaiming_uniform_(self.embryo_conv2.weight, a=1)
nn.init.kaiming_uniform_(self.embryo_conv3.weight, a=1)
self.embryo_init_reserve = {
"embryo_conv1": self.embryo_conv1.state_dict(),
"embryo_conv2": self.embryo_conv2.state_dict(),
"embryo_conv3": self.embryo_conv3.state_dict(),
"embryo_bn1": self.embryo_bn1.state_dict(),
"embryo_bn2": self.embryo_bn2.state_dict(),
"embryo_bn3": self.embryo_bn3.state_dict(),
}
self.adults_conv1 = Conv2d(in_channels,
self.adults_channels,
1,
stride,
bias=False)
self.adults_conv2 = Conv2d(self.adults_channels,
self.adults_channels,
3,
1,
padding=dilation,
bias=False,
dilation=dilation)
# padding=dilation, bias=False, dilation=dilation, groups=self.adults_volume)
self.adults_conv3 = Conv2d(self.adults_channels,
out_channels,
1,
bias=False)
# TODO: adults 三层卷积的group和初始化要细心设计
self.adults_conv1.weight.data.zero_()
self.adults_conv2.weight.data.zero_()
self.adults_conv3.weight.data.zero_()
def __init__(
self,
in_channels, # bottleneck的输入channels
bottleneck_channels, # bottleneck压缩后的channels
out_channels, # bottleneck的输出channels
num_groups, # bottleneck分组的num
stride_in_1x1, # 在每个stage的开始的1x1conv中的stride
stride, # 卷积步长
dilation, # 空洞卷积的间隔
norm_func # 用哪一个归一化函数
):
super(Bottleneck, self).__init__()
# downsample: 当 bottleneck 的输入和输出的 channels 不相等时, 则需要采用一定的策略
# 在原文中, 有 A, B, C三种策略, 本文采用的是 B 策略(也是原文推荐的)
# 即只有在输入输出通道数不相等时才使用 projection shortcuts,
# 也就是利用参数矩阵映射使得输入输出的 channels 相等
self.downsample = None
# 当输入输出通道数不同时, 额外添加一个1×1的卷积层使得输入通道数映射成输出通道数
if in_channels != out_channels:
down_stride = stride if dilation == 1 else 1
self.downsample = nn.Sequential(
Conv2d(in_channels,
out_channels,
kernel_size=1,
stride=down_stride,
bias=False),
norm_func(out_channels),
)
for modules in [
self.downsample,
]:
for l in modules.modules():
if isinstance(l, Conv2d):
nn.init.kaiming_uniform_(l.weight, a=1)
if dilation > 1:
stride = 1 # reset to be 1
# The original MSRA ResNet models have stride in the first 1x1 conv
# The subsequent fb.torch.resnet and Caffe2 ResNe[X]t implementations have
# stride in the 3x3 conv
# 这里的意思就是本来论文里的stride=2的卷积用在stage3-5的第一个1x1conv上，现在用在
# 3x3conv里，但是这里因为是原来框架的，我打印出来还是在1x1conv上，系没有删除注释
# 因为下面调用的时候都是stride_in_1x1=True
stride_1x1, stride_3x3 = (stride, 1) if stride_in_1x1 else (1, stride)
self.conv1 = Conv2d(
in_channels,
bottleneck_channels,
kernel_size=1,
stride=stride_1x1,
bias=False,
)
self.bn1 = norm_func(bottleneck_channels)
# TODO: specify init for the above
#dcn_config字典中有键"stage_with_dcn"，则返回对应的值，否则为False
self.conv2 = Conv2d(bottleneck_channels,
bottleneck_channels,
kernel_size=3,
stride=stride_3x3,
padding=dilation,
bias=False,
groups=num_groups,
dilation=dilation)
self.bn2 = norm_func(bottleneck_channels)
# 创建bottleneck的第3层卷积层
self.conv3 = Conv2d(bottleneck_channels,
out_channels,
kernel_size=1,
bias=False)
self.bn3 = norm_func(out_channels)
for l in [
self.conv1,
self.conv2,
self.conv3,
]:
nn.init.kaiming_uniform_(l.weight, a=1)
