ResNeSt¶

ConvNet Image Classification

class lucid.models.ResNeSt(block, layers, num_classes=1000, base_width: int = 64, stem_width: int = 32, cardinality: int = 1, radix: int = 2, avd: bool = True)¶

The ResNeSt class extends the ResNet architecture by integrating Split-Attention blocks, providing improved representational power for a variety of vision tasks.

Class Signature¶

class ResNeSt(ResNet):
    def __init__(
        self,
        block,
        layers,
        num_classes=1000,
        base_width: int = 64,
        stem_width: int = 32,
        cardinality: int = 1,
        radix: int = 2,
        avd: bool = True,
    ) -> None

Parameters¶

block (type): The block class to be used in constructing the network (e.g., _ResNeStBottleneck).
layers (list[int]): A list specifying the number of blocks at each stage of the network.
num_classes (int, optional): Number of output classes for the classification task. Defaults to 1000.
base_width (int, optional): Base width of the network for scaling the number of channels. Defaults to 64.
stem_width (int, optional): Width of the initial stem convolution. Defaults to 32.
cardinality (int, optional): Number of groups for grouped convolutions. Defaults to 1.
radix (int, optional): Number of splits for the Split-Attention block. Defaults to 2.
avd (bool, optional): Whether to use Adaptive Average Pooling for downsampling. Defaults to True.

Examples¶

from lucid.models import ResNeSt

# Define a ResNeSt model with 50 layers
model = ResNeSt(
    block=_ResNeStBottleneck,
    layers=[3, 4, 6, 3],  # ResNet-50 configuration
    num_classes=1000,
    base_width=64,
    stem_width=32,
    cardinality=1,
    radix=2,
    avd=True
)

# Forward pass with a sample input
input_tensor = lucid.random.randn(1, 3, 224, 224)
output = model(input_tensor)
print(output.shape)  # Output: (1, 1000)

Note

The ResNeSt class enables flexibility in defining variants of ResNeSt models (e.g., ResNeSt-50, ResNeSt-101, etc.) by customizing parameters such as layers and block.