def make_layer(layer_cfg):
nonlocal in_channels
# Possible patterns:
# ( 256, 3, {}) -> conv
# ( 256,-2, {}) -> deconv
# (None,-2, {}) -> bilinear interpolate
# ('cat',[],{}) -> concat the subnetworks in the list
#
# You know it would have probably been simpler just to adopt a 'c' 'd' 'u' naming scheme.
# Whatever, it's too late now.
# layer 가능한 패턴이 4가지 있는데 코드가 너무 방대해서 각각 이름 붙이기에는 늦었다..
# 따라서 cfg 앞의 값들을 보고 어떤 종류의 layer인지 판단
if isinstance(layer_cfg[0], str): # 맨 앞 값이 string인지 확인
layer_name = layer_cfg[0]
if layer_name == 'cat': # 'cat'인 경우만 존재하므로 이 경우에는 예외처리가 따로 없음, subnets 합쳐서 처리
nets = [make_net(in_channels, x) for x in layer_cfg[1]
] # subnetwork들의 list에서 하나씩 뽑아서 network로 만들어준다
layer = Concat(
[net[0] for net in nets], layer_cfg[2]
) # make_net의 반환값은 Sequential로 묶인 layers와 출력channel이므로
num_channels = sum([net[1] for net in nets
]) # net[0]들은 같은 layer로 묶어주고,
else:
num_channels = layer_cfg[0]
kernel_size = layer_cfg[1]
if kernel_size > 0: # conv
layer = nn.Conv2d(in_channels, num_channels, kernel_size,
**layer_cfg[2])
else:
if num_channels is None: # bilinear interpolate
layer = InterpolateModule(scale_factor=-kernel_size,
mode='bilinear',
align_corners=False,
**layer_cfg[2])
else: # deconv
layer = nn.ConvTranspose2d(in_channels, num_channels,
-kernel_size, **layer_cfg[2])
in_channels = num_channels if num_channels is not None else in_channels
# Don't return a ReLU layer if we're doing an upsample. This probably doesn't affect anything
# output-wise, but there's no need to go through a ReLU here.
# Commented out for backwards compatibility with previous models
# if num_channels is None:
# return [layer]
# else:
return [layer, nn.ReLU(inplace=True)
] # 일단 layer에는 ReLU 넣어놓고 아래에서 net 리턴할 때 제외
def make_layer(layer_cfg):
nonlocal in_channels
# Possible patterns:
# ( 256, 3, {}) -> conv
# ( 256,-2, {}) -> deconv
# (None,-2, {}) -> bilinear interpolate
# ('cat',[],{}) -> concat the subnetworks in the list
#
# You know it would have probably been simpler just to adopt a 'c' 'd' 'u' naming scheme.
# Whatever, it's too late now.
if isinstance(layer_cfg[0], str):
layer_name = layer_cfg[0]
if layer_name == 'cat':
nets = [make_net(in_channels, x) for x in layer_cfg[1]]
layer = Concat([net[0] for net in nets], layer_cfg[2])
num_channels = sum([net[1] for net in nets])
else:
num_channels = layer_cfg[0]
kernel_size = layer_cfg[1]
if kernel_size > 0:
layer = nn.Conv2d(in_channels, num_channels, kernel_size,
**layer_cfg[2])
else:
if num_channels is None:
layer = InterpolateModule(scale_factor=-kernel_size,
mode='bilinear',
align_corners=False,
**layer_cfg[2])
else:
layer = nn.ConvTranspose2d(in_channels, num_channels,
-kernel_size, **layer_cfg[2])
in_channels = num_channels if num_channels is not None else in_channels
# Don't return a ReLU layer if we're doing an upsample. This probably doesn't affect anything
# output-wise, but there's no need to go through a ReLU here.
# Commented out for backwards compatibility with previous models
if num_channels is None:
return [layer]
else:
# return [layer, nn.BatchNorm2d(in_channels) ,nn.ReLU(inplace=True)]
return [
layer,
nn.GroupNorm(32, in_channels),
nn.ReLU(inplace=True)
]
def forward(self, fpn_outs):
if self.upsample_layer is None:
final_size = (fpn_outs[0].size(2), fpn_outs[0].size(3))
self.upsample_layer = InterpolateModule(size=final_size,
mode='bilinear',
align_corners=False)
fusion_maps = []
for i in range(len(fpn_outs)):
out = self.layer_modules[i](fpn_outs[i])
out = self.upsample_layer(out)
fusion_maps.append(out)
fusion_out = sum(fusion_maps)
return fusion_out
def make_layer(layer_cfg):
#nonlocal to clarify a local_area parameter that share with up-layer
nonlocal in_channels
# Possible patterns:
# ( 256, 3, {}) -> conv
# ( 256,-2, {}) -> deconv
# (None,-2, {}) -> bilinear interpolate
# ('cat',[],{}) -> concat the subnetworks in the list
#
# You know it would have probably been simpler just to adopt a 'c' 'd' 'u' naming scheme.
# Whatever, it's too late now.
if isinstance(layer_cfg[0], str):
layer_name = layer_cfg[0]
if layer_name == 'cat':
#seems no called
print("make_net =",layer_cfg[1],"\n")
nets = [make_net(in_channels, x) for x in layer_cfg[1]]
layer = Concat([net[0] for net in nets], layer_cfg[2])
num_channels = sum([net[1] for net in nets])
else:
num_channels = layer_cfg[0]
kernel_size = layer_cfg[1]
if kernel_size > 0:
# *args 与 **kwargs 的区别，两者都是 python 中的可变参数：
#*args 表示任何多个无名参数，它本质是一个 tuple
#**kwargs 表示关键字参数，它本质上是一个 dict
layer = nn.Conv2d(in_channels, num_channels, kernel_size, **layer_cfg[2])
else:
if num_channels is None:
layer = InterpolateModule(scale_factor=-kernel_size, mode='bilinear', align_corners=False, **layer_cfg[2])
else:
layer = nn.ConvTranspose2d(in_channels, num_channels, -kernel_size, **layer_cfg[2])
in_channels = num_channels if num_channels is not None else in_channels
# Don't return a ReLU layer if we're doing an upsample. This probably doesn't affect anything
# output-wise, but there's no need to go through a ReLU here.
# Commented out for backwards compatibility with previous models
# if num_channels is None:
# return [layer]
# else:
return [layer, nn.ReLU(inplace=True)]
