CSPNetConfig

class lucid.models.CSPNetConfig(stage_specs: tuple[tuple[object, ...], ...] | list[tuple[object, ...]] | list[list[object]], stack_type: Literal['resnet', 'resnext', 'darknet'], in_channels: int = 3, stem_channels: int = 64, num_classes: int = 1000, norm: Callable[..., lucid.nn.module.Module] = <class 'lucid.nn.modules.norm.BatchNorm2d'>, act: Callable[..., lucid.nn.module.Module] = <class 'lucid.nn.modules.activation.ReLU'>, split_ratio: float = 0.5, global_pool: Literal['avg', 'max'] = 'avg', dropout: float = 0.0, feature_channels: int | None = None, pre_kernel_size: int = 1, groups: int = 1, base_width: int = 64)

CSPNetConfig stores the stage layout and classifier settings used by lucid.models.CSPNet. It defines the CSP stage specs, backbone stack style, normalization and activation choices, pooling mode, and head widths.

Class Signature

@dataclass
class CSPNetConfig:
    stage_specs: tuple[tuple[object, ...], ...] | list[tuple[object, ...]] | list[list[object]]
    stack_type: Literal["resnet", "resnext", "darknet"]
    in_channels: int = 3
    stem_channels: int = 64
    num_classes: int = 1000
    norm: Callable[..., nn.Module] = nn.BatchNorm2d
    act: Callable[..., nn.Module] = nn.ReLU
    split_ratio: float = 0.5
    global_pool: Literal["avg", "max"] = "avg"
    dropout: float = 0.0
    feature_channels: int | None = None
    pre_kernel_size: int = 1
    groups: int = 1
    base_width: int = 64

Parameters

• stage_specs: Sequence of stage specs in the form (stage_width, num_layers, downsample).

• stack_type: Backbone stack style: “resnet”, “resnext”, or “darknet”.

• in_channels (int): Number of input channels.

• stem_channels (int): Output width of the stem.

• num_classes (int): Number of output classes.

• norm: Callable normalization layer factory.

• act: Callable activation layer factory.

• split_ratio (float): Fraction of stage channels routed through one side of the CSP split.

• global_pool: Final pooling mode, “avg” or “max”.

• dropout (float): Dropout probability before the classifier.

• feature_channels (int or None): Optional pre-head projection width.

• pre_kernel_size (int): Kernel size of the pre-stage convolution.

• groups (int): Group count used by ResNeXt-style stacks.

• base_width (int): Base width used by ResNet/ResNeXt bottlenecks.

Validation

• stage_specs must contain at least one three-value stage spec.

• Stage widths and layer counts must be positive integers, and downsample must be boolean.

• stack_type must be one of “resnet”, “resnext”, or “darknet”.

• in_channels, stem_channels, num_classes, pre_kernel_size, groups, and base_width must be greater than 0.

• norm and act must be callable.

• split_ratio must be in the range (0, 1).

• global_pool must be either “avg” or “max”.

• dropout must be in the range [0, 1).

• feature_channels must be greater than 0 when provided.

Usage

from lucid.models import CSPNet, CSPNetConfig

config = CSPNetConfig(
    stage_specs=((32, 1, False), (64, 1, True)),
    stack_type="resnet",
    in_channels=3,
    stem_channels=16,
    num_classes=10,
    feature_channels=32,
)
model = CSPNet(config)