maxvit_tiny¶

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

The maxvit_tiny function returns a preconfigured instance of the MaxViT architecture with a lightweight “tiny” setup. This configuration is suitable for tasks requiring a small model footprint with effective performance on visual tasks.

Total Parameters: 25,081,416

Function Signature¶

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

Parameters¶

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

Model Configuration¶

This preset initializes a MaxViT model with the following configuration:

depths: (2, 2, 5, 2) — Number of blocks per stage.
channels: (64, 128, 256, 512) — Channel widths per stage.
embed_dim: 64 — Initial embedding dimension.

Example¶

import lucid
from lucid.models.transformer import maxvit_tiny

model = maxvit_tiny()
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.