layers01

class packnet_sfm.networks.layers.packnet.layers01.Conv2D(in_channels, out_channels, kernel_size, stride)[source]

Bases: torch.nn.modules.module.Module

2D convolution with GroupNorm and ELU

Parameters

  • in_channels (int) – Number of input channels

  • out_channels (int) – Number of output channels

  • kernel_size (int) – Kernel size

  • stride (int) – Stride

forward(x)[source]

Runs the Conv2D layer.

class packnet_sfm.networks.layers.packnet.layers01.InvDepth(in_channels, out_channels=1, min_depth=0.5)[source]

Bases: torch.nn.modules.module.Module

Inverse depth layer

forward(x)[source]

Runs the InvDepth layer.

class packnet_sfm.networks.layers.packnet.layers01.PackLayerConv2d(in_channels, kernel_size, r=2)[source]

Bases: torch.nn.modules.module.Module

Packing layer with 2d convolutions. Takes a [B,C,H,W] tensor, packs it into [B,(r^2)C,H/r,W/r] and then convolves it to produce [B,C,H/r,W/r].

forward(x)[source]

Runs the PackLayerConv2d layer.

class packnet_sfm.networks.layers.packnet.layers01.PackLayerConv3d(in_channels, kernel_size, r=2, d=8)[source]

Bases: torch.nn.modules.module.Module

Packing layer with 3d convolutions. Takes a [B,C,H,W] tensor, packs it into [B,(r^2)C,H/r,W/r] and then convolves it to produce [B,C,H/r,W/r].

forward(x)[source]

Runs the PackLayerConv3d layer.

packnet_sfm.networks.layers.packnet.layers01.ResidualBlock(in_channels, out_channels, num_blocks, stride, dropout=None)[source]

Returns a ResidualBlock with various ResidualConv layers.

Parameters

  • in_channels (int) – Number of input channels

  • out_channels (int) – Number of output channels

  • num_blocks (int) – Number of residual blocks

  • stride (int) – Stride

  • dropout (float) – Dropout value

class packnet_sfm.networks.layers.packnet.layers01.ResidualConv(in_channels, out_channels, stride, dropout=None)[source]

Bases: torch.nn.modules.module.Module

2D Convolutional residual block with GroupNorm and ELU

forward(x)[source]

Runs the ResidualConv layer.

class packnet_sfm.networks.layers.packnet.layers01.UnpackLayerConv2d(in_channels, out_channels, kernel_size, r=2)[source]

Bases: torch.nn.modules.module.Module

Unpacking layer with 2d convolutions. Takes a [B,C,H,W] tensor, convolves it to produce [B,(r^2)C,H,W] and then unpacks it to produce [B,C,rH,rW].

forward(x)[source]

Runs the UnpackLayerConv2d layer.

class packnet_sfm.networks.layers.packnet.layers01.UnpackLayerConv3d(in_channels, out_channels, kernel_size, r=2, d=8)[source]

Bases: torch.nn.modules.module.Module

Unpacking layer with 3d convolutions. Takes a [B,C,H,W] tensor, convolves it to produce [B,(r^2)C,H,W] and then unpacks it to produce [B,C,rH,rW].

forward(x)[source]

Runs the UnpackLayerConv3d layer.

packnet_sfm.networks.layers.packnet.layers01.packing(x, r=2)[source]

Takes a [B,C,H,W] tensor and returns a [B,(r^2)C,H/r,W/r] tensor, by concatenating neighbor spatial pixels as extra channels. It is the inverse of nn.PixelShuffle (if you apply both sequentially you should get the same tensor)

Parameters

  • x (torch.Tensor [B,C,H,W]) – Input tensor

  • r (int) – Packing ratio

Returns

out – Packed tensor

Return type

torch.Tensor [B,(r^2)C,H/r,W/r]