torchunet-neural-network

Input Size for UNet3D


First and foremost, I want to precise that I am a begginer in torch. I am struggling with the following code (found on github), I want to know what should be the input size of a such UNet.

"""Adapted from https://github.com/milesial/Pytorch-UNet/tree/master/unet"""
import torch
import torch.nn as nn
import torch.nn.functional as F


class UNet(nn.Module):
    def __init__(self, n_channels, n_classes, width_multiplier=1, trilinear=True, use_ds_conv=False):
        """A simple 3D Unet, adapted from a 2D Unet from https://github.com/milesial/Pytorch-UNet/tree/master/unet
        Arguments:
          n_channels = number of input channels; 3 for RGB, 1 for grayscale input
          n_classes = number of output channels/classes
          width_multiplier = how much 'wider' your UNet should be compared with a standard UNet
                  default is 1;, meaning 32 -> 64 -> 128 -> 256 -> 512 -> 256 -> 128 -> 64 -> 32
                  higher values increase the number of kernels pay layer, by that factor
          trilinear = use trilinear interpolation to upsample; if false, 3D convtranspose layers will be used instead
          use_ds_conv = if True, we use depthwise-separable convolutional layers. in my experience, this is of little help. This
                  appears to be because with 3D data, the vast vast majority of GPU RAM is the input data/labels, not the params, so little
                  VRAM is saved by using ds_conv, and yet performance suffers."""
        super(UNet, self).__init__()
        _channels = (32, 64, 128, 256, 512)
        self.n_channels = n_channels
        self.n_classes = n_classes
        self.channels = [int(c*width_multiplier) for c in _channels]
        self.trilinear = trilinear
        self.convtype = DepthwiseSeparableConv3d if use_ds_conv else nn.Conv3d

        self.inc = DoubleConv(n_channels, self.channels[0], conv_type=self.convtype)
        self.down1 = Down(self.channels[0], self.channels[1], conv_type=self.convtype)
        self.down2 = Down(self.channels[1], self.channels[2], conv_type=self.convtype)
        self.down3 = Down(self.channels[2], self.channels[3], conv_type=self.convtype)
        factor = 2 if trilinear else 1
        self.down4 = Down(self.channels[3], self.channels[4] // factor, conv_type=self.convtype)
        self.up1 = Up(self.channels[4], self.channels[3] // factor, trilinear)
        self.up2 = Up(self.channels[3], self.channels[2] // factor, trilinear)
        self.up3 = Up(self.channels[2], self.channels[1] // factor, trilinear)
        self.up4 = Up(self.channels[1], self.channels[0], trilinear)
        self.outc = OutConv(self.channels[0], n_classes)

    def forward(self, x):
        x1 = self.inc(x)
        x2 = self.down1(x1)
        x3 = self.down2(x2)
        x4 = self.down3(x3)
        x5 = self.down4(x4)
        x = self.up1(x5, x4)
        x = self.up2(x, x3)
        x = self.up3(x, x2)
        x = self.up4(x, x1)
        logits = self.outc(x)
        return logits


class DoubleConv(nn.Module):
    """(convolution => [BN] => ReLU) * 2"""

    def __init__(self, in_channels, out_channels, conv_type=nn.Conv3d, mid_channels=None):
        super().__init__()
        if not mid_channels:
            mid_channels = out_channels
        self.double_conv = nn.Sequential(
            conv_type(in_channels, mid_channels, kernel_size=3, padding=1),
            nn.BatchNorm3d(mid_channels),
            nn.ReLU(inplace=True),
            conv_type(mid_channels, out_channels, kernel_size=3, padding=1),
            nn.BatchNorm3d(out_channels),
            nn.ReLU(inplace=True)
        )

    def forward(self, x):
        return self.double_conv(x)


class Down(nn.Module):
    """Downscaling with maxpool then double conv"""

    def __init__(self, in_channels, out_channels, conv_type=nn.Conv3d):
        super().__init__()
        self.maxpool_conv = nn.Sequential(
            nn.MaxPool3d(2),
            DoubleConv(in_channels, out_channels, conv_type=conv_type)
        )

    def forward(self, x):
        return self.maxpool_conv(x)


class Up(nn.Module):
    """Upscaling then double conv"""

    def __init__(self, in_channels, out_channels, trilinear=True):
        super().__init__()

        # if trilinear, use the normal convolutions to reduce the number of channels
        if trilinear:
            self.up = nn.Upsample(scale_factor=2, mode='trilinear', align_corners=True)
            self.conv = DoubleConv(in_channels, out_channels, mid_channels=in_channels // 2)
        else:
            self.up = nn.ConvTranspose3d(in_channels, in_channels // 2, kernel_size=2, stride=2)
            self.conv = DoubleConv(in_channels, out_channels)


    def forward(self, x1, x2):
        x1 = self.up(x1)
        # input is CHW
        diffY = x2.size()[2] - x1.size()[2]
        diffX = x2.size()[3] - x1.size()[3]

        x1 = F.pad(x1, [diffX // 2, diffX - diffX // 2,
                        diffY // 2, diffY - diffY // 2])
        # if you have padding issues, see
        # https://github.com/HaiyongJiang/U-Net-Pytorch-Unstructured-Buggy/commit/0e854509c2cea854e247a9c615f175f76fbb2e3a
        # https://github.com/xiaopeng-liao/Pytorch-UNet/commit/8ebac70e633bac59fc22bb5195e513d5832fb3bd
        x = torch.cat([x2, x1], dim=1)
        return self.conv(x)


class OutConv(nn.Module):
    def __init__(self, in_channels, out_channels):
        super(OutConv, self).__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size=1)

    def forward(self, x):
        return self.conv(x)


class DepthwiseSeparableConv3d(nn.Module):
    def __init__(self, nin, nout, kernel_size, padding, kernels_per_layer=1):
        super(DepthwiseSeparableConv3d, self).__init__()
        self.depthwise = nn.Conv3d(nin, nin * kernels_per_layer, kernel_size=kernel_size, padding=padding, groups=nin)
        self.pointwise = nn.Conv3d(nin * kernels_per_layer, nout, kernel_size=1)

    def forward(self, x):
        out = self.depthwise(x)
        out = self.pointwise(out)
        return out

The Unet can simply be built with the following command:

model = UNet(n_channels, n_classes, width_multiplier=1, trilinear=True, use_ds_conv=False)

Thank very much beforehand!

PS: The code can be found here

I have tested a lot of input with different size with no result until now.


Solution

  • Let's see the forward function under the UNet class.

    def forward(self, x):
        x1 = self.inc(x)
        x2 = self.down1(x1)
        x3 = self.down2(x2)
        x4 = self.down3(x3)
        x5 = self.down4(x4)
        x = self.up1(x5, x4)
        x = self.up2(x, x3)
        x = self.up3(x, x2)
        x = self.up4(x, x1)
        logits = self.outc(x)
        return logits
    

    We observe that this 3D U-Net employs the self.down module four times. To gain insight, we navigate to the self.down module declared within the __init__ function.

    self.down1 = Down(self.channels[0], self.channels[1], conv_type=self.convtype)
    self.down2 = Down(self.channels[1], self.channels[2], conv_type=self.convtype)
    self.down3 = Down(self.channels[2], self.channels[3], conv_type=self.convtype)
    factor = 2 if trilinear else 1
    self.down4 = Down(self.channels[3], self.channels[4] // factor, conv_type=self.convtype)
    

    We can find that all of the self.down module use Down class.

    class Down(nn.Module):
    """Downscaling with maxpool then double conv"""
    
    def __init__(self, in_channels, out_channels, conv_type=nn.Conv3d):
        super().__init__()
        self.maxpool_conv = nn.Sequential(
            nn.MaxPool3d(2),
            DoubleConv(in_channels, out_channels, conv_type=conv_type)
        )
    
    def forward(self, x):
        return self.maxpool_conv(x)
    

    Here, nn.MaxPool3d(2) represents the 3D max-pooling layer, which reduce the input image size by half. For instance, if your input image size is 256 × 256 × 256, it becomes 128 × 128 × 128 after passing through the 3D max-pooling layers.

    Given the 3D U-Net architecture, which employs 4 sets of 3D max-pooling layers, the input size should be multiples of 2^4. For example, an input size of 256 × 128 × 64 is acceptable since all dimensions can be divided by 2^4. However, an input size of 72 × 128 × 36 would result in an error because 72 and 36 are not divisible by 2^4.