def __init__(self, n_comps, mem_shape):
"""
Args:
mem_shape: (C, H, W) tuple (3-elem)
"""
super().__init__()
self.shape = mem_shape
self.bias = nn.Parameter(torch.randn(n_comps, *mem_shape))
C = mem_shape[0]
self.hypernet = nn.Sequential(ConvBlock(C, C), ConvBlock(C, C),
ConvBlock(C, C))
def __init__(self, n_channel_in=1, n_channel_out=1):
super(BabyUnet, self).__init__()
self.pool1 = nn.MaxPool2d(kernel_size=2)
self.pool2 = nn.MaxPool2d(kernel_size=2)
self.up1 = lambda x: F.interpolate(x, mode='bilinear', scale_factor=2)
self.up2 = lambda x: F.interpolate(x, mode='bilinear', scale_factor=2)
self.conv1 = ConvBlock(n_channel_in, 16)
self.conv2 = ConvBlock(16, 32)
self.conv3 = ConvBlock(32, 32)
self.conv4 = ConvBlock(64, 32)
self.conv5 = ConvBlock(48, 16)
self.conv6 = nn.Conv2d(16, n_channel_out, 1)
def __init__(self, n_channel_in=1, n_channel_out=1, width=16):
super(BabyUnet, self).__init__()
self.pool1 = nn.MaxPool2d(kernel_size=2)
self.pool2 = nn.MaxPool2d(kernel_size=2)
self.up1 = lambda x: F.interpolate(
x, mode='bilinear', scale_factor=2, align_corners=False)
self.up2 = lambda x: F.interpolate(
x, mode='bilinear', scale_factor=2, align_corners=False)
self.conv1 = ConvBlock(n_channel_in, width)
self.conv2 = ConvBlock(width, 2 * width)
self.conv3 = ConvBlock(2 * width, 2 * width)
self.conv4 = ConvBlock(4 * width, 2 * width)
self.conv5 = ConvBlock(3 * width, width)
self.conv6 = nn.Conv2d(width, n_channel_out, 1)
def __init__(
self,
n_channel_in=1,
n_channel_out=1,
residual=False,
down="conv",
up="tconv",
activation="selu",
):
super(Unet, self).__init__()
self.residual = residual
if down == "maxpool":
self.down1 = nn.MaxPool2d(kernel_size=2)
self.down2 = nn.MaxPool2d(kernel_size=2)
self.down3 = nn.MaxPool2d(kernel_size=2)
self.down4 = nn.MaxPool2d(kernel_size=2)
elif down == "avgpool":
self.down1 = nn.AvgPool2d(kernel_size=2)
self.down2 = nn.AvgPool2d(kernel_size=2)
self.down3 = nn.AvgPool2d(kernel_size=2)
self.down4 = nn.AvgPool2d(kernel_size=2)
elif down == "conv":
self.down1 = nn.Conv2d(32, 32, kernel_size=2, stride=2, groups=32)
self.down2 = nn.Conv2d(64, 64, kernel_size=2, stride=2, groups=64)
self.down3 = nn.Conv2d(128, 128, kernel_size=2, stride=2, groups=128)
self.down4 = nn.Conv2d(256, 256, kernel_size=2, stride=2, groups=256)
self.down1.weight.data = 0.01 * self.down1.weight.data + 0.25
self.down2.weight.data = 0.01 * self.down2.weight.data + 0.25
self.down3.weight.data = 0.01 * self.down3.weight.data + 0.25
self.down4.weight.data = 0.01 * self.down4.weight.data + 0.25
self.down1.bias.data = 0.01 * self.down1.bias.data + 0
self.down2.bias.data = 0.01 * self.down2.bias.data + 0
self.down3.bias.data = 0.01 * self.down3.bias.data + 0
self.down4.bias.data = 0.01 * self.down4.bias.data + 0
if up == "bilinear" or up == "nearest":
self.up1 = lambda x: nn.functional.interpolate(x, mode=up, scale_factor=2)
self.up2 = lambda x: nn.functional.interpolate(x, mode=up, scale_factor=2)
self.up3 = lambda x: nn.functional.interpolate(x, mode=up, scale_factor=2)
self.up4 = lambda x: nn.functional.interpolate(x, mode=up, scale_factor=2)
elif up == "tconv":
self.up1 = nn.ConvTranspose2d(256, 256, kernel_size=2, stride=2, groups=256)
self.up2 = nn.ConvTranspose2d(128, 128, kernel_size=2, stride=2, groups=128)
self.up3 = nn.ConvTranspose2d(64, 64, kernel_size=2, stride=2, groups=64)
self.up4 = nn.ConvTranspose2d(32, 32, kernel_size=2, stride=2, groups=32)
self.up1.weight.data = 0.01 * self.up1.weight.data + 0.25
self.up2.weight.data = 0.01 * self.up2.weight.data + 0.25
self.up3.weight.data = 0.01 * self.up3.weight.data + 0.25
self.up4.weight.data = 0.01 * self.up4.weight.data + 0.25
self.up1.bias.data = 0.01 * self.up1.bias.data + 0
self.up2.bias.data = 0.01 * self.up2.bias.data + 0
self.up3.bias.data = 0.01 * self.up3.bias.data + 0
self.up4.bias.data = 0.01 * self.up4.bias.data + 0
self.conv1 = ConvBlock(n_channel_in, 32, residual, activation)
self.conv2 = ConvBlock(32, 64, residual, activation)
self.conv3 = ConvBlock(64, 128, residual, activation)
self.conv4 = ConvBlock(128, 256, residual, activation)
self.conv5 = ConvBlock(256, 256, residual, activation)
self.conv6 = ConvBlock(2 * 256, 128, residual, activation)
self.conv7 = ConvBlock(2 * 128, 64, residual, activation)
self.conv8 = ConvBlock(2 * 64, 32, residual, activation)
self.conv9 = ConvBlock(2 * 32, n_channel_out, residual, activation)
if self.residual:
self.convres = ConvBlock(n_channel_in, n_channel_out, residual, activation)