maxvit_small¶

lucid.models.maxvit_small(in_channels: int = 3, num_classes: int = 1000, **kwargs) → MaxViT¶

The maxvit_small function constructs a compact yet more expressive variant of the MaxViT model. Compared to maxvit_tiny, it increases channel width, enabling stronger representational capacity while maintaining efficiency.

Total Parameters: 55,757,304

Function Signature¶

def maxvit_small(
    in_channels: int = 3,
    num_classes: int = 1000,
    **kwargs
) -> MaxViT

Parameters¶

in_channels (int, optional): Number of input image channels. Default is 3.
num_classes (int, optional): Number of output classes for classification. Default is 1000.
kwargs (any): Additional keyword arguments passed to the MaxViT constructor.

Model Configuration¶

This preset uses the following setup:

depths: (2, 2, 5, 2)
channels: (96, 192, 384, 768)
embed_dim: 64

Example¶

import lucid
from lucid.models.transformer import maxvit_small

model = maxvit_small()
input_tensor = lucid.randn(1, 3, 224, 224)
output = model(input_tensor)
print(output.shape)  # (1, 1000)

Tip

Use **kwargs to override components such as activation, norm layers, or dropout.