def __init__(self, in_channels=3, withSkipConnections=True):
"""
:param in_channels:
:param pretrained:
:param withSkipConnections:
"""
super().__init__()
self.in_channels = in_channels
self.withSkipConnections = withSkipConnections
self.down1 = segnetDown2(self.in_channels,
64,
withFeatureMap=self.withSkipConnections)
self.down2 = segnetDown2(64,
128,
withFeatureMap=self.withSkipConnections)
self.down3 = segnetDown3(128,
256,
withFeatureMap=self.withSkipConnections)
self.down4 = segnetDown3(256,
512,
withFeatureMap=self.withSkipConnections)
self.down5 = segnetDown3(512,
512,
withFeatureMap=self.withSkipConnections)
def __init__(self,
n_classes=21,
in_channels=3,
is_unpooling=True,
pretrained=True,
withSkipConnections=False,
enablePermEq=True):
super().__init__()
self.in_channels = in_channels
self.is_unpooling = is_unpooling
self.withSkipConnections = withSkipConnections
self.enablePermEq = enablePermEq
self.down1 = segnetDown2(self.in_channels,
64,
withFeatureMap=self.withSkipConnections)
self.down2 = segnetDown2(64,
128,
withFeatureMap=self.withSkipConnections)
self.down3 = segnetDown3(128,
256,
withFeatureMap=self.withSkipConnections)
self.down4 = segnetDown3(256,
512,
withFeatureMap=self.withSkipConnections)
self.down5 = segnetDown3(512,
512,
withFeatureMap=self.withSkipConnections)
self.up5 = segnetUp3(512,
512,
withSkipConnections=self.withSkipConnections)
self.up4 = segnetUp3(512,
256,
withSkipConnections=self.withSkipConnections)
self.up3 = segnetUp3(256,
128,
withSkipConnections=self.withSkipConnections)
self.up2 = segnetUp2(128,
64,
withSkipConnections=self.withSkipConnections)
self.up1 = segnetUp2(64,
n_classes,
withSkipConnections=self.withSkipConnections)
if pretrained:
vgg16 = models.vgg16(pretrained=True)
Arch = 'SetSegNet'
if self.withSkipConnections:
Arch = 'SetSegNetSkip'
print(
'[ INFO ]: Using pre-trained weights from VGG16 with {}. Permutation equivariant layers are {}.'
.format(Arch, 'ENABLED' if self.enablePermEq else 'DISABLED'))
self.init_vgg16_params(vgg16)
def __init__(self,
out_channels=8,
in_channels=3,
pretrained=True,
withSkipConnections=True):
super().__init__()
self.in_channels = in_channels
self.withSkipConnections = withSkipConnections
self.down1 = segnetDown2(self.in_channels,
64,
withFeatureMap=self.withSkipConnections)
self.down2 = segnetDown2(64,
128,
withFeatureMap=self.withSkipConnections)
self.down3 = segnetDown3(128,
256,
withFeatureMap=self.withSkipConnections)
self.down4 = segnetDown3(256,
512,
withFeatureMap=self.withSkipConnections)
self.down5 = segnetDown3(512,
512,
withFeatureMap=self.withSkipConnections)
self.up5 = segnetUp3(512,
512,
withSkipConnections=self.withSkipConnections)
self.up4 = segnetUp3(512,
256,
withSkipConnections=self.withSkipConnections)
self.up3 = segnetUp3(256,
128,
withSkipConnections=self.withSkipConnections)
self.up2 = segnetUp2(128,
64,
withSkipConnections=self.withSkipConnections)
self.up1 = segnetUp2(64,
out_channels,
withSkipConnections=self.withSkipConnections)
if pretrained:
vgg16 = models.vgg16(pretrained=True)
Arch = 'SegNet'
if self.withSkipConnections:
Arch = 'SegNetSkip'
print('[ INFO ]: Using pre-trained weights from VGG16 with {}.'.
format(Arch))
self.init_vgg16_params(vgg16)
def __init__(self,
out_channels=8,
in_channels=3,
pretrained=True,
withSkipConnections=True,
new_version=False,
additional=None):
"""
:param out_channels:
:param in_channels:
:param pretrained:
:param withSkipConnections:
:param new_version:
:param additional: all additional output layer are new version
"""
super().__init__()
self.in_channels = in_channels
self.withSkipConnections = withSkipConnections
self.down1 = segnetDown2(self.in_channels,
64,
withFeatureMap=self.withSkipConnections)
self.down2 = segnetDown2(64,
128,
withFeatureMap=self.withSkipConnections)
self.down3 = segnetDown3(128,
256,
withFeatureMap=self.withSkipConnections)
self.down4 = segnetDown3(256,
512,
withFeatureMap=self.withSkipConnections)
self.down5 = segnetDown3(512,
512,
withFeatureMap=self.withSkipConnections)
self.up5 = segnetUp3(512,
512,
withSkipConnections=self.withSkipConnections)
self.up4 = segnetUp3(512,
256,
withSkipConnections=self.withSkipConnections)
self.up3 = segnetUp3(256,
128,
withSkipConnections=self.withSkipConnections)
self.up2 = segnetUp2(128,
64,
withSkipConnections=self.withSkipConnections)
self.up1 = segnetUp2(64,
out_channels,
last_layer=True if new_version else False,
withSkipConnections=self.withSkipConnections)
if additional is not None:
self.additional_last_layer = segnetUp2(
64,
additional,
last_layer=True,
withSkipConnections=self.withSkipConnections)
self.additional = True
else:
self.additional = False
if pretrained:
vgg16 = models.vgg16(pretrained=True)
Arch = 'SegNet'
if self.withSkipConnections:
Arch = 'SegNetSkip'
print('[ INFO ]: Using pre-trained weights from VGG16 with {}.'.
format(Arch))
self.init_vgg16_params(vgg16)