SKNet

ConvNet

class lucid.models.SKNet(config: SKNetConfig)

The SKNet class extends the ResNet family with Selective Kernel blocks, which adaptively fuse multiple receptive fields inside each residual unit. Model structure is defined through SKNetConfig, which captures the residual stage depths together with the Selective Kernel hyperparameters and shared ResNet stem options.

        %%{init: {"flowchart":{"curve":"monotoneX","nodeSpacing":50,"rankSpacing":50}} }%%
flowchart LR
  linkStyle default stroke-width:2.0px
  subgraph sg_m0["<span style='font-size:20px;font-weight:700'>sk_resnet_50</span>"]
  style sg_m0 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
    subgraph sg_m1["stem"]
      direction TB;
    style sg_m1 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
      m2["Conv2d<br/><span style='font-size:11px;color:#c53030;font-weight:400'>(1,3,224,224) → (1,64,112,112)</span>"];
      m3["BatchNorm2d"];
      m4["ReLU"];
    end
    m5["MaxPool2d<br/><span style='font-size:11px;color:#b7791f;font-weight:400'>(1,64,112,112) → (1,64,56,56)</span>"];
    subgraph sg_m6["layer1 x 4"]
      direction TB;
    style sg_m6 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
      m6_in(["Input"]);
      m6_out(["Output"]);
  style m6_in fill:#e2e8f0,stroke:#64748b,stroke-width:1px;
  style m6_out fill:#e2e8f0,stroke:#64748b,stroke-width:1px;
      subgraph sg_m7["_SKResNetBottleneck"]
        direction TB;
      style sg_m7 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
        m8["ConvBNReLU2d"];
        m9["SelectiveKernel"];
        m10["BatchNorm2d"];
        m11["Sequential<br/><span style='font-size:11px;font-weight:400'>(1,64,56,56) → (1,256,56,56)</span>"];
        m12["ReLU"];
        m13["Sequential<br/><span style='font-size:11px;font-weight:400'>(1,64,56,56) → (1,256,56,56)</span>"];
      end
      subgraph sg_m14["_SKResNetBottleneck x 2"]
        direction TB;
      style sg_m14 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
        m14_in(["Input"]);
        m14_out(["Output"]);
  style m14_in fill:#e2e8f0,stroke:#64748b,stroke-width:1px;
  style m14_out fill:#e2e8f0,stroke:#64748b,stroke-width:1px;
        m15["ConvBNReLU2d<br/><span style='font-size:11px;font-weight:400'>(1,256,56,56) → (1,64,56,56)</span>"];
        m16["SelectiveKernel"];
        m17["BatchNorm2d"];
        m18["Sequential<br/><span style='font-size:11px;font-weight:400'>(1,64,56,56) → (1,256,56,56)</span>"];
        m19["ReLU"];
      end
    end
    m20["AdaptiveAvgPool2d<br/><span style='font-size:11px;color:#b7791f;font-weight:400'>(1,2048,7,7) → (1,2048,1,1)</span>"];
    m21["Linear<br/><span style='font-size:11px;color:#2b6cb0;font-weight:400'>(1,2048) → (1,1000)</span>"];
  end
  input["Input<br/><span style='font-size:11px;color:#a67c00;font-weight:400'>(1,3,224,224)</span>"];
  output["Output<br/><span style='font-size:11px;color:#a67c00;font-weight:400'>(1,1000)</span>"];
  style input fill:#fff3cd,stroke:#a67c00,stroke-width:1px;
  style output fill:#fff3cd,stroke:#a67c00,stroke-width:1px;
  style m2 fill:#ffe8e8,stroke:#c53030,stroke-width:1px;
  style m3 fill:#e6fffa,stroke:#2c7a7b,stroke-width:1px;
  style m4 fill:#faf5ff,stroke:#6b46c1,stroke-width:1px;
  style m5 fill:#fefcbf,stroke:#b7791f,stroke-width:1px;
  style m10 fill:#e6fffa,stroke:#2c7a7b,stroke-width:1px;
  style m12 fill:#faf5ff,stroke:#6b46c1,stroke-width:1px;
  style m17 fill:#e6fffa,stroke:#2c7a7b,stroke-width:1px;
  style m19 fill:#faf5ff,stroke:#6b46c1,stroke-width:1px;
  style m20 fill:#fefcbf,stroke:#b7791f,stroke-width:1px;
  style m21 fill:#ebf8ff,stroke:#2b6cb0,stroke-width:1px;
  input --> m2;
  m10 -.-> m12;
  m11 --> m13;
  m12 --> m11;
  m12 -.-> m15;
  m13 -.-> m12;
  m14_in -.-> m15;
  m14_out -.-> m6_in;
  m15 --> m16;
  m16 --> m17;
  m17 -.-> m19;
  m18 -.-> m19;
  m19 --> m14_in;
  m19 --> m14_out;
  m19 --> m18;
  m19 --> m6_out;
  m2 --> m3;
  m20 --> m21;
  m21 --> output;
  m3 --> m4;
  m4 --> m5;
  m5 -.-> m8;
  m6_in -.-> m8;
  m6_out --> m20;
  m6_out -.-> m6_in;
  m8 --> m9;
  m9 --> m10;

Class Signature

class SKNet(ResNet):
    def __init__(self, config: SKNetConfig)

Parameters

  • config (SKNetConfig): Configuration object describing the residual stage depths, Selective Kernel hyperparameters, classifier size, stem settings, and any additional block keyword arguments.

Attributes

  • config (SKNetConfig): The configuration used to construct the model.

  • kernel_sizes (tuple[int, …]): Kernel sizes used by each Selective Kernel branch.

  • base_width (int): Width scaling factor used by the SK blocks.

  • cardinality (int): Group count used inside the Selective Kernel branches.

  • stem, layer1, layer2, layer3, layer4, avgpool, fc: Inherited ResNet submodules assembled from the configuration.

Forward Calculation

The forward pass of the SKNet model includes:

  1. Stem: Initial convolutional layers for feature extraction.

  2. Selective Kernel Stages: Residual stages built from SK blocks.

  3. Global Pooling: A global average pooling layer reduces spatial dimensions.

  4. Classifier: A fully connected layer maps the features to class scores.

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

Examples

>>> import lucid
>>> import lucid.models as models
>>> config = models.SKNetConfig(
...     block="basic",
...     layers=[2, 2, 2, 2],
...     kernel_sizes=[3, 7],
...     num_classes=10,
...     in_channels=1,
...     stem_type="deep",
...     stem_width=32,
... )
>>> model = models.SKNet(config)
>>> output = model(lucid.zeros(1, 1, 224, 224))
>>> print(output.shape)
(1, 10)

Note

  • SKNet adapts its effective receptive field by weighting multiple kernel branches.

  • Factory helpers such as sk_resnet_50 and sk_resnext_50_32x4d provide standard presets.