def __init__(self,
in_channels,
num_classes,
kernel_size,
n_pool=4,
base_channels=64,
output_activation=None):
"""__init__ Build a vanilla Unet model
https://arxiv.org/abs/1505.04597
The model however uses instance normalization before each activation
and uses Average pooling over max pooling.
Args:
in_channels ([type]): [description]
num_classes ([type]): [description]
kernel_size ([type]): [description]
n_pool (int, optional): [description]. Defaults to 4.
base_channels (int, optional): [description]. Defaults to 64.
output_activation ([type], optional): [description]. Defaults to None.
"""
layers = [1] * n_pool
encoder = Models.VGGNet(in_channels,
base_channels,
kernel_size,
layers=layers)
super(VGGLinkNet, self).__init__(encoder, num_classes, kernel_size)
if output_activation is not None:
self.output_activation = output_activation
def __init__(self,
in_channels,
n_classes,
kernel_size,
layers=[3, 3, 5, 2],
base_channels=64,
output_activation=None,
**kwargs):
super(PSPNetBuilder, self).__init__()
self.encoder = Models.ResNet(in_channels,
base_channels,
kernel_size,
layers=layers,
**kwargs)
encoder_channels = self.encoder.out_channels
self.connector = self.encoder.connector_constructor(self.encoder)
self.bottleneck = nn.PyramidPool(encoder_channels,
encoder_channels * 2)
self.upsample = torch.nn.UpsamplingBilinear2d(scale_factor=2)
self.upsample_layers = []
n_upsamples = len(layers) - 1
in_channels = self.bottleneck.out_channels + self.encoder.out_channels
for i in range(n_upsamples):
self.upsample_layers.append(
nn.UpsampleConv2d(in_channels, in_channels // 2, kernel_size))
setattr(self, 'upsample_%d' % i, self.upsample_layers[-1])
in_channels = in_channels // 2
self.pw_conv = torch.nn.Conv2d(self.upsample_layers[-1].out_channels,
n_classes,
1,
bias=False)
self.output_activation = output_activation
self.out_channels = n_classes
def __init__(self,
in_channels,
num_classes,
kernel_size=3,
layers=[3, 3, 5, 2],
base_channels=64,
output_activation=None,
**kwargs):
""" Build a Resnet Unet model
https://towardsdatascience.com/u-nets-with-resnet-encoders-and-cross-connections-d8ba94125a2c
The model however uses instance normalization before each activation
and uses Average pooling over max pooling.
Args:
in_channels ([int])
num_classes ([int]): the number of output channels
kernel_size ([int])]
Kwargs:
base_channels ([int]): The total number
of channels the networks is initialized with
output_activation([_nn.Layer,_torch.nn.functional]):
activation function of form f(x)
"""
encoder_layers = layers[:-1]
encoder = Models.ResNet(in_channels,
base_channels,
kernel_size,
layers=encoder_layers)
n_bottleneck = layers[-1]
class ResBottle(nn.RepeatedLayers):
layer_constructor = encoder.layer_constructor
_layer_dict = dict(stride=[[1, 1]] * n_bottleneck, )
super(ResnetFPN, self).__init__(encoder,
num_classes,
kernel_size,
bottleneck=ResBottle,
**kwargs)
if output_activation is not None:
self.output_activation = output_activation
https://arxiv.org/abs/1505.04597
The model however uses instance normalization before each activation
and uses Average pooling over max pooling.
Args:
in_channels ([type]): [description]
num_classes ([type]): [description]
kernel_size ([type]): [description]
n_pool (int, optional): [description]. Defaults to 4.
base_channels (int, optional): [description]. Defaults to 64.
output_activation ([type], optional): [description]. Defaults to
None.
"""
layers = [layer_repeats] * n_pool
encoder = Models.CSESEVGGNet(in_channels,
base_channels,
kernel_size,
layers=layers)
super(CSESEVGGFPN, self).__init__(encoder, num_classes, kernel_size,
**kwargs)
if output_activation is not None:
self.output_activation = output_activation
if __name__ == "__main__":
image = torch.zeros(1, 3, 256, 256)
encoder = Models.ResNet(image, 1, 3, layers=[2, 3, 5, 2])
model = ResnetFPN(3, 1, 3)
output = model(image)
print('done')