SR-Forge

Graph Interpolation

ConvolutionalUpscaling

Bases: Module

Graph-based convolutional upscaling using spline convolutions.

Constructs a high-resolution point cloud from low-resolution graph data by placing target positions on a regular grid and performing message passing from neighbouring source nodes via radius-based connectivity.

Parameters:

Name	Type	Description	Default
scale	int	Spatial upsampling factor.	None
in_channels	int	Number of input feature channels.	None
out_channels	int	Number of output feature channels.	None
nn	Module	Custom message-passing network.	None
transform	optional	Edge-attribute transform (defaults to Cartesian).	None
legacy_mode	bool	If True, uses legacy position handling.	False
kernel_size	int	Spline kernel size (defaults to scale ** 2).	None

forward_hr_pos(in_data: Batch, x: torch.Tensor, r=3)

Upscale using pre-defined high-resolution positions from the input data.

forward_org(data: Batch, x: torch.Tensor, scale=None, pos_mean=True)

Upscale by constructing a regular HR grid from original node shapes.