InceptionResNet¶

ConvNet

class lucid.models.InceptionResNet(config: InceptionResNetConfig)¶

Overview¶

The InceptionResNet base class defines a flexible architecture that combines Inception-style modules with residual connections. This approach improves optimization and gradient flow in deep neural networks, making it suitable for a variety of image classification tasks.

This class serves as a foundation for specific versions like Inception-ResNet v1 and v2 by providing essential components such as a stem network, convolutional layers, and fully connected layers. It is configured through InceptionResNetConfig.

        %%{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'>inception_resnet_v2</span>"]
  style sg_m0 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
    subgraph sg_m1["_InceptionStem_V4"]
    style sg_m1 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
      m2["Sequential<br/><span style='font-size:11px;font-weight:400'>(1,3,224,224) → (1,64,109,109)</span>"];
      m3["MaxPool2d<br/><span style='font-size:11px;color:#b7791f;font-weight:400'>(1,64,109,109) → (1,64,54,54)</span>"];
      m4["ConvBNReLU2d<br/><span style='font-size:11px;font-weight:400'>(1,64,109,109) → (1,96,54,54)</span>"];
      m5(["Sequential x 2<br/><span style='font-size:11px;font-weight:400'>(1,160,54,54) → (1,96,52,52)</span>"]);
      m6["ConvBNReLU2d<br/><span style='font-size:11px;font-weight:400'>(1,192,52,52) → (1,192,25,25)</span>"];
      m7["MaxPool2d<br/><span style='font-size:11px;color:#b7791f;font-weight:400'>(1,192,52,52) → (1,192,25,25)</span>"];
    end
    subgraph sg_m8["conv"]
    style sg_m8 fill:#000000,fill-opacity:0.05,stroke:#000000,stroke-opacity:0.75,stroke-width:1px
      m9(["_InceptionResModule_A x 5"]);
      m10["_InceptionReduce_V4A<br/><span style='font-size:11px;font-weight:400'>(1,384,25,25) → (1,1152,12,12)</span>"];
      m11(["_InceptionResModule_B x 10"]);
      m12["_InceptionResReduce<br/><span style='font-size:11px;font-weight:400'>(1,1152,12,12) → (1,2144,5,5)</span>"];
      m13(["_InceptionResModule_C x 5"]);
    end
    m14["AdaptiveAvgPool2d<br/><span style='font-size:11px;color:#b7791f;font-weight:400'>(1,2144,5,5) → (1,2144,1,1)</span>"];
    m15["Dropout"];
    m16["Linear<br/><span style='font-size:11px;color:#2b6cb0;font-weight:400'>(1,2144) → (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 m3 fill:#fefcbf,stroke:#b7791f,stroke-width:1px;
  style m7 fill:#fefcbf,stroke:#b7791f,stroke-width:1px;
  style m14 fill:#fefcbf,stroke:#b7791f,stroke-width:1px;
  style m15 fill:#edf2f7,stroke:#4a5568,stroke-width:1px;
  style m16 fill:#ebf8ff,stroke:#2b6cb0,stroke-width:1px;
  input --> m2;
  m10 --> m11;
  m11 --> m12;
  m12 --> m13;
  m13 --> m14;
  m14 --> m15;
  m15 --> m16;
  m16 --> output;
  m2 --> m3;
  m3 --> m4;
  m4 --> m5;
  m5 --> m7;
  m6 --> m9;
  m7 --> m6;
  m9 --> m10;

Class Signature¶

class InceptionResNet(nn.Module):
    def __init__(self, config: InceptionResNetConfig) -> None

Parameters¶

config (InceptionResNetConfig): A configuration object describing the selected variant, output class count, input channels, and classifier dropout probability.

Attributes¶

stem (nn.Module): The initial stem module that extracts low-level features from the input.
conv (nn.Sequential): A sequential container for the main convolutional and residual blocks.
fc (nn.Sequential): A sequential container for the fully connected layers that perform classification.

Methods¶

forward(x: Tensor) -> Tensor Performs the forward pass through the stem, convolutional blocks, and fully connected layers.

def forward(self, x):
    x = self.stem(x)
    x = self.conv(x)
    x = x.view(x.shape[0], -1)  # Flatten
    x = self.fc(x)
    return x