data

SKNetConfig

extendsModelConfig

SKNetConfig(num_classes: int = 1000, in_channels: int = 3, layers: tuple[int, ...] = (3, 4, 6, 3), block_type: str = 'bottleneck', cardinality: int = 1, base_width: int = 64, split_input: bool = True, rd_ratio: float = 1.0 / 16, rd_divisor: int = 8)

source edit

Unified config for all SK-ResNet variants (Li et al., 2019).

Paper: "Selective Kernel Networks"

Architecture is identical to ResNet-50 (expansion=4, stages output 256/512/1024/2048 channels) except each 3×3 conv in the bottleneck is replaced by a SelectiveKernel unit with two parallel branches (3×3 + 3×3 dilated-2, mimicking 5×5 receptive field).

Key hyper-parameters:

layers Number of bottleneck blocks per stage (default ResNet-50 = 3/4/6/3).

cardinality Number of groups for the SK branch convolutions (G in the paper). Also used in the ResNeXt-style width formula: width = int(planes * (base_width / 64)) * cardinality Set to 1 for plain SK-ResNet (default).

base_width Base channel multiplier for the ResNeXt width formula. 64 → plain ResNet widths (64/128/256/512 at each stage). 4 with cardinality=32 → SK-ResNeXt-50 32×4d (SKNet-50 from paper).

split_input If True (timm default), each SK branch receives half the input channels, keeping the param count similar to a single grouped conv.

rd_ratio Reduction ratio for the SelectiveKernelAttn bottleneck.

rd_divisor Divisor for rounding the attention channel count.

block_type "bottleneck" (default, expansion=4) for SK-ResNet-50/101 or "basic" (expansion=1) for SK-ResNet-18/34.

Used by 3

Constructors

dunder

init

→None

__init__(num_classes: int = 1000, in_channels: int = 3, layers: tuple[int, ...] = (3, 4, 6, 3), block_type: str = 'bottleneck', cardinality: int = 1, base_width: int = 64, split_input: bool = True, rd_ratio: float = 1.0 / 16, rd_divisor: int = 8)

source edit

data

SKNetConfig

extendsModelConfig

SKNetConfig(num_classes: int = 1000, in_channels: int = 3, layers: tuple[int, ...] = (3, 4, 6, 3), block_type: str = 'bottleneck', cardinality: int = 1, base_width: int = 64, split_input: bool = True, rd_ratio: float = 1.0 / 16, rd_divisor: int = 8)

source edit

Unified config for all SK-ResNet variants (Li et al., 2019).

Paper: "Selective Kernel Networks"

Key hyper-parameters:

layers Number of bottleneck blocks per stage (default ResNet-50 = 3/4/6/3).

base_width Base channel multiplier for the ResNeXt width formula. 64 → plain ResNet widths (64/128/256/512 at each stage). 4 with cardinality=32 → SK-ResNeXt-50 32×4d (SKNet-50 from paper).

split_input If True (timm default), each SK branch receives half the input channels, keeping the param count similar to a single grouped conv.

rd_ratio Reduction ratio for the SelectiveKernelAttn bottleneck.

rd_divisor Divisor for rounding the attention channel count.

block_type "bottleneck" (default, expansion=4) for SK-ResNet-50/101 or "basic" (expansion=1) for SK-ResNet-18/34.

Used by 3

Constructors

dunder

init

→None

__init__(num_classes: int = 1000, in_channels: int = 3, layers: tuple[int, ...] = (3, 4, 6, 3), block_type: str = 'bottleneck', cardinality: int = 1, base_width: int = 64, split_input: bool = True, rd_ratio: float = 1.0 / 16, rd_divisor: int = 8)

source edit