SR-Forge

MagNAt Model

MagNAt_vanilla

Bases: Model

A vanilla implementation of the MagNAt model.

This version appears to be a baseline or earlier version of the MagNAt model for multi-image super-resolution, built using PyTorch Geometric. It includes a registration network (RegNet) to estimate shifts between input images.

__init__(in_channels: int = 1, d=56, s=16, scale: int = 3, processing_layers: int = 4, attention_heads: int = 1, radius: float = 1.0, filter_size: int = 9, regnet_bias: bool = True, remove_masked=False, magnet_legacy: bool = False, **kwargs)

Initializes the MagNAt_vanilla model.

Parameters:

Name	Type	Description	Default
in_channels	int	Number of input channels for each image.	1
d	int	Base feature dimension size.	56
s	int	Shrinking layer feature dimension size.	16
scale	int	The target upscaling factor.	3
processing_layers	int	Number of non-linear mapping layers.	4
attention_heads	int	Number of heads in EdgeAttentionEmbedding.	1
radius	float	Radius for graph construction.	1.0
filter_size	int	Kernel size for the dynamic filter generation in RegNet.	9
regnet_bias	bool	Whether to use a bias in the final RegNet layer.	True
remove_masked	bool	If True, removes masked nodes before processing.	False
**kwargs	Any	Additional keyword arguments.	{}

MagNAt

Bases: Model

The MagNAt model for Multi-Image Super-Resolution using Graph Attention.

This model registers a set of low-resolution images by treating them as nodes in a graph, estimates their alignment, and fuses their features using a graph neural network to produce a super-resolved image. This version includes edge dropout for regularization.

__init__(in_channels: int = 1, d=56, s=16, scale: int = 3, processing_layers: int = 4, attention_heads: int = 1, radius: float = 1.0, filter_size: int = 9, regnet_bias: bool = True, remove_masked=False, edge_dropout: float = 0.3, **kwargs)

Initializes the MagNAt model.

Parameters:

Name	Type	Description	Default
in_channels	int	Number of input channels for each image.	1
d	int	Base feature dimension size.	56
s	int	Shrinking layer feature dimension size.	16
scale	int	The target upscaling factor.	3
processing_layers	int	Number of non-linear mapping layers.	4
attention_heads	int	Number of heads in EdgeAttentionEmbedding.	1
radius	float	Radius for graph construction.	1.0
filter_size	int	Kernel size for the dynamic filter generation in RegNet.	9
regnet_bias	bool	Whether to use a bias in the final RegNet layer.	True
remove_masked	bool	If True, removes masked nodes before processing.	False
edge_dropout	float	Dropout probability for graph edges.	0.3
**kwargs	Any	Additional keyword arguments.	{}