def __init__(self, output_size, filter_size=5):
super(network, self).__init__()
size = filter_size
bias = True
self.conv1 = nn.Sequential(
nn.Conv3d(1, 64, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.Conv3d(64, 64, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.BatchNorm3d(64))
self.conv2 = nn.Sequential(
nn.Conv3d(64, 128, kernel_size=size, padding=size//2, stride=2, bias=bias),
nn.ReLU(inplace=True),
nn.Conv3d(128, 128, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.BatchNorm3d(128))
self.conv3 = nn.Sequential(
nn.Conv3d(128, 256, kernel_size=size, padding=size//2, stride=2, bias=bias),
nn.ReLU(inplace=True),
nn.Conv3d(256, 256, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.BatchNorm3d(256))
self.up1 = nn.Sequential(
nn.Upsample(scale_factor=2, mode="nearest"),
nn.Conv3d(256, 128, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True))
self.merge1 = Merge()
self.conv4 = nn.Sequential(
nn.Conv3d(256, 128, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.Conv3d(128, 128, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.BatchNorm3d(128))
self.up2 = nn.Sequential(
nn.Upsample(scale_factor=2, mode="nearest"),
nn.Conv3d(128, 64, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True))
self.merge2 = Merge()
self.conv5 = nn.Sequential(
nn.Conv3d(128, 64, kernel_size=size, padding=size // 2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.Conv3d(64, 64, kernel_size=size, padding=size//2, stride=1, bias=bias),
nn.ReLU(inplace=True),
nn.BatchNorm3d(64))
self.conv_final = nn.Conv3d(64, output_size, kernel_size=1, padding=0, stride=1, bias=bias)
def __init__(self, output_size, filter_size=5):
super(network, self).__init__()
size = filter_size
common_params = {
'radial_window':
partial(basis_kernels.gaussian_window_fct_convenience_wrapper,
mode='compromise',
border_dist=0,
sigma=0.6),
'batch_norm_momentum':
0.01,
}
features = [
(1, ),
# (12, 12, 12),
# (24, 24, 24),
# (48, 48, 48),
# (24, 24, 24),
# (12, 12, 12),
(8, 8, 8, 4),
(16, 16, 16, 8),
(32, 32, 32, 16),
(16, 16, 16, 8),
(8, 8, 8, 4),
(output_size, )
]
# TODO: do padding using ReplicationPad3d?
# TODO: on validation - use overlapping patches and only use center of patch
self.conv1 = nn.Sequential(
GatedBlock(features[0],
features[1],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params),
GatedBlock(features[1],
features[1],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.conv2 = nn.Sequential(
GatedBlock(features[1],
features[2],
size=size,
padding=size // 2,
stride=2,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params),
GatedBlock(features[2],
features[2],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.conv3 = nn.Sequential(
GatedBlock(features[2],
features[3],
size=size,
padding=size // 2,
stride=2,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params),
GatedBlock(features[3],
features[3],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.up1 = nn.Sequential(
nn.Upsample(scale_factor=2, mode="nearest"),
GatedBlock(features[3],
features[4],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.merge1 = Merge()
self.conv4 = nn.Sequential(
GatedBlock(features[3],
features[4],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params),
GatedBlock(features[4],
features[4],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.up2 = nn.Sequential(
nn.Upsample(scale_factor=2, mode="nearest"),
GatedBlock(features[4],
features[5],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.merge2 = Merge()
self.conv5 = nn.Sequential(
GatedBlock(features[4],
features[5],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params),
GatedBlock(features[5],
features[5],
size=size,
padding=size // 2,
stride=1,
activation=(F.relu, F.sigmoid),
normalization="instance",
**common_params))
self.conv_final = GatedBlock(features[5],
features[6],
size=1,
padding=0,
stride=1,
activation=None,
normalization=None,
**common_params)