def __init__(self, k, anchors=9, depth=4, activation=F.relu):
super(SubNet, self).__init__()
self.anchors = anchors
self.activation = activation
self.base = nn.ModuleList(
[conv3x3x3(256, 256, padding=1) for _ in range(depth)])
self.output = nn.Conv3d(256, k * anchors, kernel_size=3, padding=1)
self.output.weight = nn.init.xavier_normal(self.output.weight)
def __init__(self, resnet):
super(FeaturePyramid_v1, self).__init__()
self.resnet = resnet
# applied in a pyramid
self.pyramid_transformation_1 = conv3x3x3(64, 256, padding=1)
self.pyramid_transformation_2 = conv1x1x1(256, 256)
self.pyramid_transformation_3 = conv1x1x1(512, 256)
self.pyramid_transformation_4 = conv1x1x1(1024, 256)
self.pyramid_transformation_5 = conv1x1x1(2048, 256)
# both based around resnet_feature_5
#self.pyramid_transformation_6 = conv3x3x3(2048, 256, padding=1, stride=2)
#self.pyramid_transformation_7 = conv3x3x3(256, 256, padding=1, stride=2)
# applied after upsampling
self.upsample_transform_1 = conv3x3x3(256, 256, padding=1)
self.upsample_transform_2 = conv3x3x3(256, 256, padding=1)
self.upsample_transform_3 = conv3x3x3(256, 256, padding=1)
self.upsample_transform_4 = conv3x3x3(256, 256, padding=1)