def __init__(self,
feature_scale=4,
n_classes=21,
is_deconv=True,
in_channels=3,
is_batchnorm=True):
super(unetoriginal, self).__init__()
self.is_deconv = is_deconv
self.in_channels = in_channels
self.is_batchnorm = is_batchnorm
self.feature_scale = feature_scale
filters = [64, 128, 256, 512, 1024]
# filters = [128, 256, 512, 1024, 2048]
# filters = [256, 512, 1024, 2048,4096]
filters = [int(x / self.feature_scale) for x in filters]
# downsampling
self.conv1 = unetConv2(self.in_channels, filters[0], self.is_batchnorm)
self.maxpool1 = nn.MaxPool2d(kernel_size=2)
self.conv2 = unetConv2(filters[0], filters[1], self.is_batchnorm)
self.maxpool2 = nn.MaxPool2d(kernel_size=2)
self.conv3 = unetConv2(filters[1], filters[2], self.is_batchnorm)
self.maxpool3 = nn.MaxPool2d(kernel_size=2)
self.conv4 = unetConv2(filters[2], filters[3], self.is_batchnorm)
self.maxpool4 = nn.MaxPool2d(kernel_size=2)
self.center = unetConv2(filters[3], filters[4], self.is_batchnorm)
# upsampling
self.up_concat4 = unetUp(filters[4], filters[3], self.is_deconv)
self.up_concat3 = unetUp(filters[3], filters[2], self.is_deconv)
self.up_concat2 = unetUp(filters[2], filters[1], self.is_deconv)
self.up_concat1 = unetUp(filters[1], filters[0], self.is_deconv)
# final conv (without any concat)
self.final = nn.Conv2d(filters[0], n_classes, 1)
self.drop = nn.Dropout(p=0.25)
def __init__(self,
feature_scale=4,
n_classes=21,
is_deconv=True,
in_channels=3,
is_batchnorm=True):
super(unet, self).__init__()
self.is_deconv = is_deconv
self.in_channels = in_channels
self.is_batchnorm = is_batchnorm
self.feature_scale = feature_scale
filters = [64, 128, 256, 512, 1024]
# filters = [128,256,512,1024,2048]
filters = [int(x / self.feature_scale) for x in filters]
print(filters)
#print("unet_initial_mark1")
# downsampling
#TODO: DOUBLE THE BASE FILTER NUMBER
#TODO: RESIDUAL CONNECTION
#TODO: MOVE DROPOUT LAYERS
#TODO: TUNE DOWN THE LEARNING RATE WHEN TRAINING LOSS FLUCTUATES
# print("unet_initial_mark1.2")
self.conv1 = unetfirstConv2(self.in_channels, filters[0],
self.is_batchnorm) # channel 3-16
self.maxpool1 = nn.MaxPool2d(kernel_size=2)
self.half1 = nn.Conv2d(filters[0], filters[0], 2, 2, 0)
# print("unet_initial_mark1.3")
self.conv2 = unetConv2(filters[0], filters[1],
self.is_batchnorm) #channel 16-32
# print("unet_initial_mark1.4")
self.maxpool2 = nn.MaxPool2d(kernel_size=2)
self.half2 = nn.Conv2d(filters[1], filters[1], 2, 2, 0)
self.conv3 = unetConv2(filters[1], filters[2],
self.is_batchnorm) #channel 32-64
self.maxpool3 = nn.MaxPool2d(kernel_size=2)
self.half3 = nn.Conv2d(filters[2], filters[2], 2, 2, 0)
self.conv4 = unetConv2(filters[2], filters[3],
self.is_batchnorm) #channel 64-128
self.maxpool4 = nn.MaxPool2d(kernel_size=2)
self.half4 = nn.Conv2d(filters[3], filters[3], 2, 2, 0)
self.center = unetConv2(filters[3], filters[4],
self.is_batchnorm) #channel 128-256
# upsampling
self.up_concat4 = unetUp(filters[4], filters[3],
self.is_deconv) #channel 256-128,problem
self.up_concat3 = unetUp(filters[3], filters[2],
self.is_deconv) #channel 128-64
self.up_concat2 = unetUp(filters[2], filters[1],
self.is_deconv) #channel 64-32
self.up_concat1 = unetUp(filters[1], filters[0],
self.is_deconv) #channel 32-16
# final conv (without any concat)
self.final = nn.Conv2d(filters[0], n_classes,
1) #channel 16-21,kernel 1x1
#print("unet_initial_mark1.5")
self.drop = nn.Dropout(p=0.3)
self.shortcut1 = nn.Sequential(
nn.Conv2d(self.in_channels, filters[0], 1, 1, 0),
nn.BatchNorm2d(filters[0]))
self.shortcut2 = nn.Sequential(
nn.Conv2d(filters[0], filters[1], 1, 2, 0),
nn.BatchNorm2d(filters[1]))
self.shortcut3 = nn.Sequential(
nn.Conv2d(filters[1], filters[2], 1, 2, 0),
nn.BatchNorm2d(filters[2]))
self.shortcut4 = nn.Sequential(
nn.Conv2d(filters[2], filters[3], 1, 2, 0),
nn.BatchNorm2d(filters[3]))
self.shortcut_center = nn.Sequential(
nn.Conv2d(filters[3], filters[4], 1, 2, 0),
nn.BatchNorm2d(filters[4]))