Models API Reference¶

Factory Function¶

`create_model`¶

def create_model(model_name: str, ms_bands: int = 4, **kwargs) -> nn.Module

Create a pansharpening model by name.

Parameters:

Parameter	Type	Default	Description
`model_name`	str	required	Model identifier
`ms_bands`	int	4	Number of multispectral bands
`**kwargs`	dict	{}	Additional model-specific arguments

Returns: PyTorch nn.Module

Example:

from models import create_model

model = create_model('panformer_lite', ms_bands=4)

CNN Models¶

PNN¶

class PNN(nn.Module):
    def __init__(self, ms_bands: int = 4)

Basic 3-layer CNN for pansharpening.

PanNet¶

class PanNet(nn.Module):
    def __init__(
        self,
        ms_bands: int = 4,
        n_blocks: int = 4,
        n_features: int = 64
    )

ResNet-style pansharpening with high-pass filtering.

DRPNN¶

class DRPNN(nn.Module):
    def __init__(
        self,
        ms_bands: int = 4,
        n_blocks: int = 8,
        n_features: int = 64
    )

Deep residual pansharpening network.

PanNetCBAM¶

class PanNetCBAM(nn.Module):
    def __init__(
        self,
        ms_bands: int = 4,
        n_blocks: int = 4,
        n_features: int = 64,
        reduction: int = 16
    )

PanNet with CBAM attention modules.

MultiScalePanNet¶

class MultiScalePanNet(nn.Module):
    def __init__(
        self,
        ms_bands: int = 4,
        base_features: int = 64
    )

Multi-scale feature pyramid network.

Transformer Models¶

PanFormer¶

class PanFormer(nn.Module):
    def __init__(
        self,
        ms_bands: int = 4,
        embed_dim: int = 128,
        depth: int = 4,
        num_heads: int = 8,
        patch_size: int = 4,
        mlp_ratio: float = 4.0
    )

Full transformer with cross-attention fusion.

PanFormerLite¶

class PanFormerLite(nn.Module):
    def __init__(
        self,
        ms_bands: int = 4,
        embed_dim: int = 64,
        depth: int = 2,
        num_heads: int = 4,
        window_size: int = 8,
        patch_size: int = 4
    )

Lightweight transformer with window attention.

Forward Pass¶

All models have the same forward signature:

def forward(self, ms: Tensor, pan: Tensor) -> Tensor