UNet with Bottom Layer Map scale parameter (only x1, x1/2, x1/2^2, x1/2^3, x1/2^4 support)

UNet.py

# Based on https://github.com/mateuszbuda/brain-segmentation-pytorch/blob/master/unet.py
# Original version is licensed under MIT License by mateuszbuda.

from collections import OrderedDict

import torch
import torch.nn as nn


class UNet(nn.Module):

    def __init__(self, in_channels=3, out_channels=1, init_features=32, bottom_inv_scale=5):
        super(UNet, self).__init__()
        assert bottom_inv_scale in [ 1, 2, 3, 4, 5 ]

        K1 = 2 if bottom_inv_scale > 4 else 1
        K2 = 2 if bottom_inv_scale > 3 else 1
        K3 = 2 if bottom_inv_scale > 2 else 1
        K4 = 2 if bottom_inv_scale > 1 else 1


        features = init_features
        self.encoder1 = UNet._block(in_channels, features, name="enc1")
        self.pool1 = nn.MaxPool2d(kernel_size=K1, stride=K1)

        self.encoder2 = UNet._block(features, features * 2, name="enc2")
        self.pool2 = nn.MaxPool2d(kernel_size=K2, stride=K2)

        self.encoder3 = UNet._block(features * 2, features * 4, name="enc3")
        self.pool3 = nn.MaxPool2d(kernel_size=K3, stride=K3)

        self.encoder4 = UNet._block(features * 4, features * 8, name="enc4")
        self.pool4 = nn.MaxPool2d(kernel_size=K4, stride=K4)

        self.bottleneck = UNet._block(features * 8, features * 16, name="bottleneck")

        self.upconv4 = nn.ConvTranspose2d(
            features * 16, features * 8, kernel_size=K4, stride=K4
        )
        self.decoder4 = UNet._block((features * 8) * 2, features * 8, name="dec4")
        self.upconv3 = nn.ConvTranspose2d(
            features * 8, features * 4, kernel_size=K3, stride=K3
        )
        self.decoder3 = UNet._block((features * 4) * 2, features * 4, name="dec3")
        self.upconv2 = nn.ConvTranspose2d(
            features * 4, features * 2, kernel_size=K2, stride=K2
        )
        self.decoder2 = UNet._block((features * 2) * 2, features * 2, name="dec2")
        self.upconv1 = nn.ConvTranspose2d(
            features * 2, features, kernel_size=K1, stride=K1
        )
        self.decoder1 = UNet._block(features * 2, features, name="dec1")

        self.conv = nn.Conv2d(
            in_channels=features, out_channels=out_channels, kernel_size=1
        )

    def forward(self, x):
        enc1 = self.encoder1(x)
        enc2 = self.encoder2(self.pool1(enc1))
        enc3 = self.encoder3(self.pool2(enc2))
        enc4 = self.encoder4(self.pool3(enc3))

        print(enc1.shape)
        print(enc2.shape)
        print(enc3.shape)
        print(enc4.shape)

        bottleneck = self.bottleneck(self.pool4(enc4))
        print('Bottom:', bottleneck.shape)

        dec4 = self.upconv4(bottleneck)
        print(dec4.shape)
        dec4 = torch.cat((dec4, enc4), dim=1)
        dec4 = self.decoder4(dec4)
        dec3 = self.upconv3(dec4)
        print(dec3.shape)
        dec3 = torch.cat((dec3, enc3), dim=1)
        dec3 = self.decoder3(dec3)
        dec2 = self.upconv2(dec3)
        print(dec2.shape)
        dec2 = torch.cat((dec2, enc2), dim=1)
        dec2 = self.decoder2(dec2)
        dec1 = self.upconv1(dec2)
        print(dec1.shape)
        dec1 = torch.cat((dec1, enc1), dim=1)
        dec1 = self.decoder1(dec1)
        return torch.sigmoid(self.conv(dec1))

    @staticmethod
    def _block(in_channels, features, name):
        return nn.Sequential(
            OrderedDict(
                [
                    (
                        name + "conv1",
                        nn.Conv2d(
                            in_channels=in_channels,
                            out_channels=features,
                            kernel_size=3,
                            padding=1,
                            bias=False,
                        ),
                    ),
                    (name + "norm1", nn.BatchNorm2d(num_features=features)),
                    (name + "relu1", nn.ReLU(inplace=True)),
                    (
                        name + "conv2",
                        nn.Conv2d(
                            in_channels=features,
                            out_channels=features,
                            kernel_size=3,
                            padding=1,
                            bias=False,
                        ),
                    ),
                    (name + "norm2", nn.BatchNorm2d(num_features=features)),
                    (name + "relu2", nn.ReLU(inplace=True)),
                ]
            )
        )

if __name__ == '__main__':
    model = UNet(
        in_channels=3,
        out_channels=1,
        bottom_inv_scale=2,
    )
    model.eval()

    image = torch.randn((1, 3, 256, 256))
    print('In:', image.shape)

    with torch.no_grad():
        y = model(image)

    print('Out:', y.shape)