unet.py

# Code taken (and modified a bit) from
# https://github.com/PyTorchLightning/pytorch-lightning/blob/master/pl_examples/models/unet.py

import torch
import torch.nn as nn
import torch.nn.functional as F


class UNet(nn.Module):
    """
    Architecture based on U-Net: Convolutional Networks for Biomedical Image Segmentation
    Link - https://arxiv.org/abs/1505.04597
    Parameters:
        num_classes: Number of output classes required (default 19 for KITTI dataset)
        num_layers: Number of layers in each side of U-net
        features_start: Number of features in first layer
    """

    def __init__(
        self,
        num_classes: int = 1,
        num_layers: int = 4,
        features_start: int = 64,
        dropout=False,
    ):
        super().__init__()
        self.num_layers = num_layers
        self.dropout = dropout

        layers = [DoubleConv(192, features_start)]

        feats = features_start
        for _ in range(num_layers - 1):
            layers.append(Down(feats, feats * 2, self.dropout))
            feats *= 2

        for _ in range(num_layers - 1):
            layers.append(Up(feats, feats // 2))
            feats //= 2

        layers.append(nn.Conv2d(feats, num_classes, kernel_size=1))

        self.layers = nn.ModuleList(layers)

    def forward(self, x):
        xi = [self.layers[0](x)]
        # Down path
        for layer in self.layers[1 : self.num_layers]:
            xi.append(layer(xi[-1]))
        # Up path
        for i, layer in enumerate(self.layers[self.num_layers : -1]):
            xi[-1] = layer(xi[-1], xi[-2 - i])
        return self.layers[-1](xi[-1])


class DoubleConv(nn.Module):
    """
    Double Convolution and BN and ReLU
    (3x3 conv -> BN -> ReLU) ** 2
    """

    def __init__(self, in_ch: int, out_ch: int):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, kernel_size=3, padding=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True),
            nn.Conv2d(out_ch, out_ch, kernel_size=3, padding=1),
            nn.BatchNorm2d(out_ch),
            nn.ReLU(inplace=True),
        )

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


class Down(nn.Module):
    """
    Combination of MaxPool2d and DoubleConv in series
    """

    def __init__(self, in_ch: int, out_ch: int, dropout=False):
        super().__init__()
        self.net = nn.Sequential(
            nn.MaxPool2d(kernel_size=2, stride=2), DoubleConv(in_ch, out_ch),
        )
        if dropout:
            self.net.add_module("dropout", nn.Dropout2d(dropout))

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


class Up(nn.Module):
    """
    Upsampling (by transpose convolutions)
    followed by concatenation of feature map from contracting path,
    followed by double 3x3 convolution.
    """

    def __init__(self, in_ch: int, out_ch: int):
        super().__init__()
        self.upsample = nn.ConvTranspose2d(in_ch, in_ch // 2, kernel_size=2, stride=2)
        self.conv = DoubleConv(in_ch, out_ch)

    def forward(self, x1, x2):
        x1 = self.upsample(x1)

        # Pad x1 to the size of x2
        diff_h = x2.shape[2] - x1.shape[2]
        diff_w = x2.shape[3] - x1.shape[3]

        x1 = F.pad(
            x1, [diff_w // 2, diff_w - diff_w // 2, diff_h // 2, diff_h - diff_h // 2]
        )

        # Concatenate along the channels axis
        x = torch.cat([x2, x1], dim=1)
        return self.conv(x)