def __init__(self, in_channels_dict, out_channels=256, mode="concat"):
super().__init__()
assert mode in ["concat", "add"]
self.mode = mode
self.net = nn.ModuleDict()
self.reduce = nn.ModuleDict()
keys = sorted(list(in_channels_dict.keys()))[::-1] # [P7,P6,P5,...P2]
for i, key in enumerate(keys):
in_channels = in_channels_dict[key]
if i == 0:
self.reduce[key] = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 3, 1, 1),
nn.BatchNorm2d(out_channels), nn.ReLU(inplace=True))
else:
self.net[key] = nn.Sequential(
nn.Conv2d(out_channels, out_channels, 3, 1, 1),
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True),
nn.ConvTranspose2d(out_channels, out_channels, 3, 2, 1,
1), # 2,2,0,0
nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True))
if self.mode == "concat":
self.reduce[key] = nn.Sequential(
nn.Conv2d(in_channels + out_channels, out_channels, 3,
1, 1), nn.BatchNorm2d(out_channels),
nn.ReLU(inplace=True))
else:
self.reduce[key] = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 3, 1, 1),
nn.BatchNorm2d(out_channels), nn.ReLU(inplace=True))
_initParmasV2(self, self.modules())
def __init__(self, in_channels_dict, out_channels=256):
super().__init__()
self.inner_blocks = nn.ModuleDict()
self.layer_blocks = nn.ModuleDict()
for name, in_channels in in_channels_dict.items():
if in_channels == 0:
continue
inner_block_module = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 1),
nn.BatchNorm2d(out_channels), nn.ReLU(inplace=True))
layer_block_module = nn.Sequential(
nn.Conv2d(out_channels, out_channels, 3, padding=1),
nn.BatchNorm2d(out_channels), nn.ReLU(inplace=True))
self.inner_blocks[name] = inner_block_module
self.layer_blocks[name] = layer_block_module
# initialize parameters now to avoid modifying the initialization of top_blocks
# _initParmas(self, self.modules())
_initParmasV2(self, self.modules())
def __init__(self,
model_name="vgg16_bn",
num_classes=21,
freeze_at=["res2", "res3", "res4", "res5"],
use_shortcut=False,
pretrained=False,
return_indices=False):
super().__init__()
self.return_indices = return_indices
self.num_classes = num_classes
self.use_shortcut = use_shortcut
self.inplanes = 64
features = nn.Sequential(
nn.Conv2d(3, self.inplanes, 3, padding=1),
nn.BatchNorm2d(self.inplanes), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes, self.inplanes, 3, padding=1),
nn.BatchNorm2d(self.inplanes), nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2, return_indices=return_indices),
nn.Conv2d(self.inplanes, self.inplanes * 2, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 2), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 2, self.inplanes * 2, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 2), nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2, return_indices=return_indices),
nn.Conv2d(self.inplanes * 2, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 4, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 4, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2, return_indices=return_indices),
nn.Conv2d(self.inplanes * 4, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2, return_indices=return_indices),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=True),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=True),
nn.MaxPool2d(2, 2, return_indices=return_indices))
# self._initialize_weights()
if pretrained:
# 加载预训练的模型
state_dict = load_state_dict_from_url(model_urls["vgg16_bn"],
progress=True)
features.load_state_dict({
k[9:]: v
for k, v in state_dict.items()
if k[9:] in features.state_dict()
})
else:
_initialize_weights(self, features.modules())
self.backbone = nn.ModuleDict(
OrderedDict([ # nn.Sequential
("res1", features[0:7]),
("res2", features[7:14]),
("res3", features[14:24]),
("res4", features[24:34]),
("res5", features[34:]),
]))
# 参数冻结
for name in freeze_at:
for parameter in self.backbone[name].parameters():
parameter.requires_grad_(False)
# 统计所有可更新梯度的变量
print("只有以下变量做梯度更新:")
for name, parameter in self.backbone.named_parameters():
if parameter.requires_grad:
print("name:", name)
self.decode = nn.ModuleDict()
self.decode["res5"] = nn.Sequential(
# nn.MaxUnpool2d(2,2),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8),
nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8),
nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8),
nn.ReLU(inplace=False))
self.decode["res4"] = nn.Sequential(
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=False))
self.decode["res3"] = nn.Sequential(
nn.Conv2d(self.inplanes * 4, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 4, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 4, self.inplanes * 2, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 2), nn.ReLU(inplace=False))
self.decode["res2"] = nn.Sequential(
nn.Conv2d(self.inplanes * 2, self.inplanes * 2, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 2), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 2, self.inplanes, 3, padding=1),
nn.BatchNorm2d(self.inplanes), nn.ReLU(inplace=False))
self.decode["res1"] = nn.Sequential(
nn.Conv2d(self.inplanes, self.inplanes, 3, padding=1),
nn.BatchNorm2d(self.inplanes),
nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes, num_classes, 3, padding=1),
# nn.BatchNorm2d(num_classes),
# nn.ReLU(inplace=False)
# nn.Softmax(1)
)
_initParmasV2(self, self.decode.modules())
def __init__(self,
model_name="vgg16_bn",
num_classes=21,
freeze_at=["res2", "res3", "res4", "res5"],
use_shortcut=False,
pretrained=False,
return_indices=False):
super().__init__()
self.num_classes = num_classes
self.use_shortcut = use_shortcut
# self.backbone = Backbone_vgg16bn(pretrained=pretrained,freeze_at=freeze_at,return_indices=return_indices)
self.backbone = Backbone_vgg(model_name,
pretrained=pretrained,
freeze_at=freeze_at)
self.inplanes = self.backbone.inplanes
self.decode = nn.ModuleDict()
self.decode["res5"] = nn.Sequential(
# nn.MaxUnpool2d(2,2),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8),
nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8),
nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8),
nn.ReLU(inplace=False))
self.decode["res4"] = nn.Sequential(
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 8, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 8), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 8, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=False))
self.decode["res3"] = nn.Sequential(
nn.Conv2d(self.inplanes * 4, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 4, self.inplanes * 4, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 4), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 4, self.inplanes * 2, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 2), nn.ReLU(inplace=False))
self.decode["res2"] = nn.Sequential(
nn.Conv2d(self.inplanes * 2, self.inplanes * 2, 3, padding=1),
nn.BatchNorm2d(self.inplanes * 2), nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes * 2, self.inplanes, 3, padding=1),
nn.BatchNorm2d(self.inplanes), nn.ReLU(inplace=False))
self.decode["res1"] = nn.Sequential(
nn.Conv2d(self.inplanes, self.inplanes, 3, padding=1),
nn.BatchNorm2d(self.inplanes),
nn.ReLU(inplace=False),
nn.Conv2d(self.inplanes, num_classes, 3, padding=1),
# nn.BatchNorm2d(num_classes),
# nn.ReLU(inplace=False)
# nn.Softmax(1)
)
_initParmasV2(self, self.decode.modules())