vathos.model.resunet

Residual - Unet - Enhanced

class ResUNet

A ResNet - Unet inspired custom model for monocular depth estimation

class ResDoubleConv(in_channels, out_channels)

Basic DoubleConv of a ResNetV2

Performs basic Pre Activated ResNet Double Convolution

Parameters:

• in_channels – input channels
• out_channels – output channels

class ResDownBlock(in_channels, out_channels)

Basic DownBlock of a ResNetV2

Performs a Residual Down operation

Parameters:

• in_channels – input channels
• out_channels – output channels

output: \((N, C, H/2, W/2)\)

class ResUpBlock(in_channels, out_channels, skip_channels, dense_channels=None)

Basic UpBlock of a ResNetV2

Performs Residual Up Convolution on the input, uses PixelShuffle to produce checkerboard-free outputs

Parameters:

• in_channels – input channels
• out_channels – output channels
• skip_channels – skip input channels
• dense_channels – dense input channels (from another decoder)

Note

The input is applied with a 1x1 convolution and then pixel shuffle to keep the channels constant and also produce checkerboard-free outputs, the rest is then followed by double convolution