def __init__(self,
ver_dim,
seg_dim,
im_dim,
fcdim=256,
s8dim=128,
s4dim=64,
s2dim=32,
raw_dim=32):
super(ImageEncoder, self).__init__()
# Load the pretrained weights, remove avg pool
# layer and get the output stride of 8
resnet18_8s = resnet18(fully_conv=True,
pretrained=True,
output_stride=8,
remove_avg_pool_layer=True)
self.ver_dim = ver_dim
self.seg_dim = seg_dim
# Randomly initialize the 1x1 Conv scoring layer
resnet18_8s.fc = nn.Sequential(
nn.Conv2d(resnet18_8s.inplanes, fcdim, 3, 1, 1, bias=False),
nn.BatchNorm2d(fcdim), nn.ReLU(True))
resnet18_8s.conv1 = nn.Conv2d(im_dim,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.resnet18_8s = resnet18_8s
def __init__(self):
super(Resnet18_8s_detector, self).__init__()
# Load the pretrained weights, remove avg pool
# layer and get the output stride of 8
self.resnet18_8s = resnet18(fully_conv=True,
pretrained=True,
output_stride=8,
remove_avg_pool_layer=True)
self.resnet18_8s.fc = nn.Conv2d(self.resnet18_8s.inplanes, 1, 3, 1, 1)
def __init__(self,
ver_dim,
seg_dim,
fcdim=256,
s8dim=128,
s4dim=64,
s2dim=32,
raw_dim=32):
super(Resnet18_8s, self).__init__()
# Load the pretrained weights, remove avg pool
# layer and get the output stride of 8
resnet18_8s = resnet18(fully_conv=True,
pretrained=True,
output_stride=8,
remove_avg_pool_layer=True)
self.ver_dim = ver_dim
self.seg_dim = seg_dim
# Randomly initialize the 1x1 Conv scoring layer
resnet18_8s.fc = nn.Sequential(
nn.Conv2d(resnet18_8s.inplanes, fcdim, 3, 1, 1, bias=False),
nn.BatchNorm2d(fcdim), nn.ReLU(True))
self.resnet18_8s = resnet18_8s
# x8s->128
self.conv8s = nn.Sequential(
nn.Conv2d(128 + fcdim, s8dim, 3, 1, 1, bias=False),
nn.BatchNorm2d(s8dim), nn.LeakyReLU(0.1, True))
self.up8sto4s = nn.UpsamplingBilinear2d(scale_factor=2)
# x4s->64
self.conv4s = nn.Sequential(
nn.Conv2d(64 + s8dim, s4dim, 3, 1, 1, bias=False),
nn.BatchNorm2d(s4dim), nn.LeakyReLU(0.1, True))
self.up4sto2s = nn.UpsamplingBilinear2d(scale_factor=2)
# x2s->64
self.conv2s = nn.Sequential(
nn.Conv2d(64 + s4dim, s2dim, 3, 1, 1, bias=False),
nn.BatchNorm2d(s2dim), nn.LeakyReLU(0.1, True))
self.up2storaw = nn.UpsamplingBilinear2d(scale_factor=2)
self.convraw = nn.Sequential(
nn.Conv2d(3 + s2dim, raw_dim, 3, 1, 1, bias=False),
nn.BatchNorm2d(raw_dim), nn.LeakyReLU(0.1, True),
nn.Conv2d(raw_dim, seg_dim + ver_dim, 1, 1))
def __init__(self,
ver_dim,
seg_dim,
im_dim,
fcdim=256,
s8dim=128,
s4dim=64,
s2dim=32,
raw_dim=32,
pre_trained=True):
super(EstimateEncoder, self).__init__()
# Load the pretrained weights, remove avg pool
# layer and get the output stride of 8
resnet18_8s = resnet18(fully_conv=True,
pretrained=True,
output_stride=8,
remove_avg_pool_layer=True)
self.conv1 = nn.Conv2d(18,
64,
kernel_size=7,
stride=2,
padding=3,
bias=False)
self.bn1 = nn.BatchNorm2d(64)
self.relu = nn.ReLU(inplace=True)
self.mpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.l1 = nn.Sequential(*list(resnet18_8s.children())[4:5])
self.l2 = nn.Sequential(*list(resnet18_8s.children())[5:6])
self.l3 = nn.Sequential(*list(resnet18_8s.children())[6:7])
self.l4 = nn.Sequential(*list(resnet18_8s.children())[7:8])
self.apool = nn.AvgPool2d(7)
self.fc = nn.Sequential(
nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=1,
bias=False), nn.BatchNorm2d(256), nn.ReLU(inplace=True))
self.ver_dim = ver_dim
self.seg_dim = seg_dim
