csp_resnet_50

lucid.models.csp_resnet_50(num_classes: int = 1000, split_ratio: float = 0.5, stem_channels: int = 64, **kwargs) → CSPNet

The csp_resnet_50 function constructs a Cross Stage Partial ResNet-50 model, which integrates the CSPNet topology with the ResNet bottleneck block as the core transformation unit. This design significantly reduces duplicate gradient flow and enhances computational efficiency while preserving representational capacity.

It follows the CSPNet paradigm where the feature maps are split, partially transformed, and then merged. This version uses standard ResNet-50 stage configurations adapted for CSP integration.

Total Parameters: 22,463,016

Function Signature

@register_model
def csp_resnet_50(
    num_classes: int = 1000,
    split_ratio: float = 0.5,
    stem_channels: int = 64,
    **kwargs
) -> CSPNet

Parameters

  • num_classes (int, optional): Number of output classes for classification. Default is 1000.

  • split_ratio (float, optional): The fraction of channels sent through the residual transformation path in each CSP stage. Default is 0.5 (even split).

  • stem_channels (int, optional): Number of channels used in the stem (initial) convolution. Default is 64.

  • kwargs (dict, optional): Additional keyword arguments passed to the base CSPNet class for further customization.

Returns

  • CSPNet: An instance of the generalized CSPNet using ResNet-50 style bottlenecks.

Examples

Basic Usage

from lucid.models import csp_resnet_50

# Instantiate a CSP-ResNet-50 model
model = csp_resnet_50(num_classes=1000)

# Dummy input tensor (1 image, 3 channels, 224x224)
x = lucid.random.randn(1, 3, 224, 224)

# Forward pass
y = model(x)

print(y.shape)  # Output: torch.Size([1, 1000])