SKNet¶

ConvNet Image Classification

class lucid.models.SKNet(block: Module, layers: list[int], num_classes: int = 1000, kernel_sizes: list[int] = [3, 5], base_width: int = 64, cardinality: int = 1, **resnet_args: Any)¶

The SKNet class extends the ResNet architecture by incorporating Selective Kernel (SK) blocks, which dynamically adjust receptive fields via attention mechanisms. This enables the network to adaptively fuse multi-scale features, improving performance on tasks involving objects of varying scales.

Class Signature¶

class lucid.nn.SKNet(
    block: nn.Module,
    layers: list[int],
    num_classes: int = 1000,
    kernel_sizes: list[int] = [3, 5],
    base_width: int = 64,
    cardinality: int = 1,
)

Parameters¶

block (nn.Module): The building block module used for the SKNet layers. Typically an SKBlock or compatible block type.
layers (list[int]): Specifies the number of blocks in each stage of the network.
num_classes (int, optional): Number of output classes for the final fully connected layer. Default: 1000.
kernel_sizes (list[int], optional): Specifies the sizes of kernels to be used in the SK blocks for multi-scale processing. Default: [3, 5].
base_width (int, optional): Base width of the feature maps in the SK blocks. Default: 64.
cardinality (int, optional): The number of parallel convolutional groups (grouped convolutions) in the SK blocks. Default: 1.

Attributes¶

kernel_sizes (list[int]): Stores the kernel sizes used in the SK blocks.
base_width (int): Stores the base width of feature maps.
cardinality (int): Stores the number of groups for grouped convolutions.
layers (list[nn.Module]): A list of stages, each containing a sequence of SK blocks.

Forward Calculation¶

The forward pass of the SKNet model includes:

Stem: Initial convolutional layers for feature extraction.
Selective Kernel Stages: Each stage applies a series of SK blocks configured via layers.
Global Pooling: A global average pooling layer reduces spatial dimensions.
Classifier: A fully connected layer maps the features to class scores.

\[\text{output} = \text{FC}(\text{GAP}(\text{SKBlocks}(\text{Stem}(\text{input}))))\]

Note

The SKNet is well-suited for tasks requiring multi-scale feature representation.
Increasing the kernel_sizes parameter allows the model to capture features over larger receptive fields.