def __init__(self,
feature_levels=(3, 4),
pretrained=True,
l2_norm=True,
**kwargs):
super().__init__()
assert feature_levels == (3, 4)
assert pretrained, "Only pretrained VGG16 is provided."
_check_levels(feature_levels)
self.forward_levels = tuple(range(1, feature_levels[-1] + 1))
self.feature_levels = feature_levels
from torchvision.models import vgg16
backbone = vgg16(pretrained=True)
f = backbone.features
f[4].ceil_mode = True
f[9].ceil_mode = True
f[16].ceil_mode = True
f[23].ceil_mode = True
f[30].kernel_size = (3, 3)
f[30].stride = (1, 1)
f[30].padding = (1, 1)
conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
fc6 = backbone.classifier[0]
fc6_weight = fc6.weight.data.view(4096, 512, 7, 7)
fc6_bias = fc6.bias.data
conv6.weight.data = decimate(fc6_weight, m=[4, None, 3, 3])
conv6.bias.data = decimate(fc6_bias, m=[4])
conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
fc7 = backbone.classifier[3]
fc7_weight = fc7.weight.data.view(4096, 4096, 1, 1)
fc7_bias = fc7.bias.data
conv7.weight.data = decimate(fc7_weight, m=[4, 4, None, None])
conv7.bias.data = decimate(fc7_bias, m=[4])
self.layer1 = f[:9]
self.layer2 = f[9:16]
self.layer3 = f[16:23]
self.l2_norm = L2Norm(512, 20) if l2_norm else nn.Identity()
self.layer4 = nn.Sequential(
*f[23:],
conv6,
nn.ReLU(inplace=True),
conv7,
nn.ReLU(inplace=True),
)
self.out_channels = [512, 1024]
def __init__(self, feature_levels=(3, 4), pretrained=True):
super().__init__()
assert pretrained, "Only pretrained VGG16 is provided."
self.feature_levels = feature_levels
from torchvision.models import vgg16_bn
backbone = vgg16_bn(pretrained=True)
f = backbone.features
f[6].ceil_mode = True
f[13].ceil_mode = True
f[23].ceil_mode = True
f[33].ceil_mode = True
f[43].kernel_size = (3, 3)
f[43].stride = (1, 1)
f[43].padding = (1, 1)
conv6 = nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6)
fc6 = backbone.classifier[0]
fc6_weight = fc6.weight.data.view(4096, 512, 7, 7)
fc6_bias = fc6.bias.data
conv6.weight.data = decimate(fc6_weight, m=[4, None, 3, 3])
conv6.bias.data = decimate(fc6_bias, m=[4])
conv7 = nn.Conv2d(1024, 1024, kernel_size=1)
fc7 = backbone.classifier[3]
fc7_weight = fc7.weight.data.view(4096, 4096, 1, 1)
fc7_bias = fc7.bias.data
conv7.weight.data = decimate(fc7_weight, m=[4, 4, None, None])
conv7.bias.data = decimate(fc7_bias, m=[4])
self.stage0 = f[:6]
self.stage1 = f[6:13]
self.stage2 = f[13:23]
self.stage3 = f[23:33]
self.stage4 = nn.Sequential(
*f[33:],
conv6,
nn.BatchNorm2d(1024),
nn.ReLU(inplace=True),
conv7,
nn.BatchNorm2d(1024),
nn.ReLU(inplace=True),
)
self.out_channels = [
get_out_channels(getattr(self, ("stage%d" % i)))
for i in feature_levels
]