data

MaxViTConfig

extendsModelConfig

MaxViTConfig(num_classes: int = 1000, in_channels: int = 3, depths: tuple[int, ...] = (2, 2, 5, 2), dims: tuple[int, ...] = (64, 128, 256, 512), window_size: int = 7, num_heads: int = 32, mlp_ratio: float = 4.0, stem_width: int | None = None)

source edit

Configuration dataclass for every MaxViT variant.

MaxViTConfig is an immutable container that fully specifies the architecture of a MaxViT (Tu et al., 2022). MaxViT combines an MBConv convolutional stage with two complementary sparse self-attention mechanisms — block attention and grid attention — in every MaxViT block, realising dense local + dense global self-attention at linear complexity in the number of tokens.

Parameters

num_classesint= 1000

Number of output classes for the classification head. Defaults to 1000 (ImageNet-1k).

in_channelsint= 3

Number of input image channels. Defaults to 3 (RGB).

depthstuple of int= (2, 2, 5, 2)

Number of MaxViT blocks in each of the four stages. Defaults to (2, 2, 5, 2) (MaxViT-Tiny).

dimstuple of int= (64, 128, 256, 512)

Output channel width of each of the four stages. Defaults to (64, 128, 256, 512) (MaxViT-Tiny).

window_sizeint= 7

Side length

P

of the partition window for block and grid attention. Defaults to 7.

num_headsint= 32

Number of attention heads used by both block and grid attention. In practice the reference recipe fixes head_dim to 32 and derives num_heads = dim / 32 per stage; this field is informational. Defaults to 32.

mlp_ratiofloat= 4.0

MLP expansion ratio inside the partition-attention blocks. Defaults to 4.0.

stem_widthint or None= None

Output width of the convolutional stem. The reference MaxViT decouples this from dims[0]: Tiny/Large keep stem_width == dims[0] while Small/Base use a narrower stem_width = 64 and let the first block's shortcut expand to dims[0]. None (the default) falls back to dims[0].

Attributes

model_typeClassVar[str]

Constant string "maxvit" used by the model registry.

Notes

The canonical variants registered as factory functions in lucid.models.vision.maxvit are:

================ ================== ==================== Variant depths dims ================ ================== ==================== MaxViT-Tiny (2, 2, 5, 2) (64, 128, 256, 512) MaxViT-Small (2, 2, 5, 2) (96, 192, 384, 768) MaxViT-Base (2, 6, 14, 2) (96, 192, 384, 768) MaxViT-Large (2, 6, 14, 2) (128, 256, 512, 1024) MaxViT-XLarge (2, 6, 14, 2) (192, 384, 768, 1536) ================ ================== ====================

Reference: Zhengzhong Tu et al., "MaxViT: Multi-Axis Vision Transformer", ECCV 2022, arXiv:2204.01697.

Examples

Build a MaxViT-Tiny configuration with a 100-class head:
>>> from lucid.models.vision.maxvit import MaxViTConfig
>>> cfg = MaxViTConfig(num_classes=100)
>>> cfg.depths, cfg.dims, cfg.window_size, cfg.num_classes
((2, 2, 5, 2), (64, 128, 256, 512), 7, 100)

Used by 3

Constructors

dunder

init

→None

__init__(num_classes: int = 1000, in_channels: int = 3, depths: tuple[int, ...] = (2, 2, 5, 2), dims: tuple[int, ...] = (64, 128, 256, 512), window_size: int = 7, num_heads: int = 32, mlp_ratio: float = 4.0, stem_width: int | None = None)

source edit

MaxViTConfig(num_classes: int = 1000, in_channels: int = 3, depths: tuple[int, ...] = (2, 2, 5, 2), dims: tuple[int, ...] = (64, 128, 256, 512), window_size: int = 7, num_heads: int = 32, mlp_ratio: float = 4.0, stem_width: int | None = None)

Build a MaxViT-Tiny configuration with a 100-class head: >>> from lucid.models.vision.maxvit import MaxViTConfig >>> cfg = MaxViTConfig(num_classes=100) >>> cfg.depths, cfg.dims, cfg.window_size, cfg.num_classes ((2, 2, 5, 2), (64, 128, 256, 512), 7, 100)

__init__(num_classes: int = 1000, in_channels: int = 3, depths: tuple[int, ...] = (2, 2, 5, 2), dims: tuple[int, ...] = (64, 128, 256, 512), window_size: int = 7, num_heads: int = 32, mlp_ratio: float = 4.0, stem_width: int | None = None)