models.py

"""
UNet
The main UNet model implementation
"""

import torch
import torch.nn as nn

# Utility Functions
''' when filter kernel= 3x3, padding=1 makes in&out matrix same size'''


def conv_bn_leru(in_channels, out_channels, kernel_size=3, stride=1, padding=1):
    return nn.Sequential(
        nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True),
        nn.Conv2d(out_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True),
    )


def down_pooling():
    return nn.MaxPool2d(2)


def up_pooling(in_channels, out_channels, kernel_size=2, stride=2):
    return nn.Sequential(
        nn.ConvTranspose2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True)
    )


# UNet class
class UNet(nn.Module):
    def __init__(self, n_channels, n_classes):
        super().__init__()
        # go down
        self.conv1 = conv_bn_leru(n_channels, 64)
        self.conv2 = conv_bn_leru(64, 128)
        self.conv3 = conv_bn_leru(128, 256)
        self.conv4 = conv_bn_leru(256, 512)
        self.conv5 = conv_bn_leru(512, 1024)
        self.down_pooling = nn.MaxPool2d(2)

        # dropout
        self.dropout = nn.Dropout(0.5)

        # go up
        self.up_pool6 = up_pooling(1024, 512)
        self.conv6 = conv_bn_leru(1024, 512)
        self.up_pool7 = up_pooling(512, 256)
        self.conv7 = conv_bn_leru(512, 256)
        self.up_pool8 = up_pooling(256, 128)
        self.conv8 = conv_bn_leru(256, 128)
        self.up_pool9 = up_pooling(128, 64)
        self.conv9 = conv_bn_leru(128, 64)

        # output
        self.conv10 = nn.Conv2d(64, n_classes, 1)

        # test weight init
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight.data, a=0, mode='fan_out')
                if m.bias is not None:
                    m.bias.data.zero_()

    def forward(self, x):
        # go down
        x1 = self.conv1(x)
        p1 = self.down_pooling(x1)
        x2 = self.conv2(p1)
        p2 = self.down_pooling(x2)
        x3 = self.conv3(p2)
        p3 = self.down_pooling(x3)
        x4 = self.conv4(p3)
        p4 = self.down_pooling(x4)
        x5 = self.conv5(p4)

        x5 = self.dropout(x5)

        # go up
        p6 = self.up_pool6(x5)
        x6 = torch.cat([p6, x4], dim=1)
        x6 = self.conv6(x6)

        x6 = self.dropout(x6)

        p7 = self.up_pool7(x6)
        x7 = torch.cat([p7, x3], dim=1)
        x7 = self.conv7(x7)

        p8 = self.up_pool8(x7)
        x8 = torch.cat([p8, x2], dim=1)
        x8 = self.conv8(x8)

        p9 = self.up_pool9(x8)
        x9 = torch.cat([p9, x1], dim=1)
        x9 = self.conv9(x9)

        output = self.conv10(x9)

        return output