SKNetConfig¶

class lucid.models.SKNetConfig(block: Literal['basic', 'bottleneck'], layers: tuple[int, int, int, int] | list[int], kernel_sizes: tuple[int, ...] | list[int] = (3, 5), base_width: int = 64, cardinality: int = 1, num_classes: int = 1000, in_channels: int = 3, stem_width: int = 64, stem_type: Literal['deep'] | None = None, avg_down: bool = False, channels: tuple[int, int, int, int] | list[int] = (64, 128, 256, 512), block_args: dict[str, typing.Any] = <factory>)¶

SKNetConfig stores the Selective Kernel hyperparameters used by lucid.models.SKNet. It defines the residual stage depths, Selective Kernel branch settings, classifier size, and the shared ResNet stem options.

Class Signature¶

@dataclass
class SKNetConfig:
    block: Literal["basic", "bottleneck"]
    layers: tuple[int, int, int, int] | list[int]
    kernel_sizes: tuple[int, ...] | list[int] = (3, 5)
    base_width: int = 64
    cardinality: int = 1
    num_classes: int = 1000
    in_channels: int = 3
    stem_width: int = 64
    stem_type: Literal["deep"] | None = None
    avg_down: bool = False
    channels: tuple[int, int, int, int] | list[int] = (64, 128, 256, 512)
    block_args: dict[str, Any] = field(default_factory=dict)

Parameters¶

block (Literal[“basic”, “bottleneck”]): Selects the SK residual block variant.
layers (tuple[int, int, int, int] | list[int]): Number of SK blocks in each of the four residual stages.
kernel_sizes (tuple[int, …] | list[int]): Kernel sizes used by the Selective Kernel branches inside each block.
base_width (int): Width scaling factor used to derive intermediate block widths.
cardinality (int): Group count used by the grouped Selective Kernel branches.
num_classes (int): Number of output classes.
in_channels (int): Number of channels in the input image tensor.
stem_width (int): Width parameter used by the deep stem variant.
stem_type (Literal[“deep”] | None): Stem style. None uses the classic single 7x7 stem, while “deep” uses a three-layer stem.
avg_down (bool): Whether projection shortcuts should downsample using average pooling before the 1x1 projection.
channels (tuple[int, int, int, int] | list[int]): Output width of each residual stage.
block_args (dict[str, Any]): Extra keyword arguments forwarded to each SK block, in addition to the preset kernel_sizes, base_width, and cardinality.

Validation¶

block must be “basic” or “bottleneck”.
layers and channels must each contain exactly four positive integers.
kernel_sizes must contain at least one positive integer.
base_width, cardinality, num_classes, in_channels, and stem_width must be greater than 0.
stem_type must be None or “deep”.
block_args must be a dictionary.

Usage¶

import lucid.models as models

config = models.SKNetConfig(
    block="bottleneck",
    layers=[3, 4, 6, 3],
    kernel_sizes=[3, 5, 7],
    cardinality=32,
    base_width=4,
    num_classes=10,
    avg_down=True,
)

model = models.SKNet(config)