class

InceptionNeXt

extendsPretrainedModelBackboneMixin

InceptionNeXt(config: InceptionNeXtConfig)

source edit

InceptionNeXt backbone (Yu et al., 2024).

A drop-in replacement for the ConvNeXt backbone that factorizes the single large $7 \times 7$ depthwise conv into four parallel Inception-style branches operating on disjoint channel splits:

\mathrm{IDWConv}(x) = \mathrm{Concat}\bigl( x^{(1)},\; \mathrm{DW}_{3 \times 3}(x^{(2)}),\; \mathrm{DW}_{1 \times K}(x^{(3)}),\; \mathrm{DW}_{K \times 1}(x^{(4)})\bigr),

with $K = \texttt{band\_kernel}$ (default 11). The four branches together preserve a large effective receptive field while cutting depthwise FLOPs and memory traffic relative to a single $7 \times 7$ depthwise conv. Each MetaNeXtBlock further applies BatchNorm, a 1x1 Conv-MLP, and a layer-scale parameter on the residual branch.

forward_features returns the global-average-pooled $(B, \texttt{dims[-1]})$ feature.

Parameters

configInceptionNeXtConfig

Frozen dataclass specifying depths, dims, band_kernel, mlp_ratios, in_channels, and num_classes. See InceptionNeXtConfig.

Attributes

stemnn.Sequential

Patchify stem:

4 \times 4

stride-4 Conv2d + BatchNorm2d.

stagesnn.ModuleList

Four _Stage modules, each containing an optional downsampler and a sequence of MetaNeXt blocks.

feature_infolist[FeatureInfo]

Four-stage feature description with reductions

(4, 8, 16, 32)

Notes

Reference: Weihao Yu et al., "InceptionNeXt: When Inception Meets ConvNeXt", CVPR 2024, arXiv:2303.16900.

Examples

Build an InceptionNeXt-T backbone and run a forward pass:
>>> import lucid
>>> from lucid.models.vision.inception_next import (
...     InceptionNeXt, InceptionNeXtConfig,
... )
>>> model = InceptionNeXt(InceptionNeXtConfig())
>>> x = lucid.randn(1, 3, 224, 224)
>>> feat = model.forward_features(x)
>>> feat.shape                       # (B, dims[-1])
(1, 768)

Used by 2

Constructors

dunder

init

→None

__init__(config: InceptionNeXtConfig)

source edit

Properties

prop

feature_info

→list[FeatureInfo]

feature_info: list[FeatureInfo]

source edit

Instance methods

forward

→BaseModelOutput

forward(x: Tensor)

source edit

forward_features

→Tensor

forward_features(x: Tensor)

source edit

Build an InceptionNeXt-T backbone and run a forward pass: >>> import lucid >>> from lucid.models.vision.inception_next import ( ... InceptionNeXt, InceptionNeXtConfig, ... ) >>> model = InceptionNeXt(InceptionNeXtConfig()) >>> x = lucid.randn(1, 3, 224, 224) >>> feat = model.forward_features(x) >>> feat.shape # (B, dims[-1]) (1, 768)