class

ConvNeXt

extendsPretrainedModelBackboneMixin

ConvNeXt(config: ConvNeXtConfig)

source edit

ConvNeXt backbone (Liu et al., 2022).

ConvNeXt is a pure-convolutional backbone built by systematically modernising a ResNet using design choices borrowed from the Swin Transformer. The trunk is a four-stage pyramid that starts with a non-overlapping $4 \times 4$ patchify stem and processes the feature map through stacks of identical ConvNeXt blocks:

x \leftarrow x + \gamma \odot \mathrm{Linear}_{4d \to d}\!\bigl( \mathrm{GELU}\!\bigl(\mathrm{Linear}_{d \to 4d}\!\bigl( \mathrm{LN}(\mathrm{DWConv}_{7 \times 7}(x)) \bigr)\bigr)\bigr),

where $\gamma \in \mathbb{R}^d$ is a per-channel layer-scale parameter initialised to $10^{-6}$ (Touvron et al., 2021). Between stages a $2 \times 2$ stride-2 LN + Conv2d downsampler halves the resolution and doubles the channel width.

forward_features returns a global-average-pooled $(B, C_{\text{out}})$ feature where $C_{\text{out}} = \texttt{dims[-1]}$ . Use this backbone when you want features for transfer learning or dense prediction; for end-to-end classification use ConvNeXtForImageClassification.

Parameters

configConvNeXtConfig

Frozen dataclass specifying depths, dims, layer_scale_init, dropout, and in_channels. See ConvNeXtConfig.

Attributes

stem_StemWithNorm

Patchify stem:

4 \times 4

stride-4 Conv2d followed by a channel-last LayerNorm.

stagesnn.ModuleList

Four stages of stacked _ConvNeXtBlock modules.

downsamplersnn.ModuleList

Three between-stage LN + stride-2 Conv2d downsamplers.

head_normnn.LayerNorm

Final LayerNorm applied (channel-last) to the pooled feature.

avgpoolnn.AdaptiveAvgPool2d

1 \times 1

adaptive average pool over spatial dims.

feature_infolist[FeatureInfo]

Four-stage feature description with reductions

(4, 8, 16, 32)

and channel counts from config.dims.

Notes

Reference: Zhuang Liu et al., "A ConvNet for the 2020s", CVPR 2022, arXiv:2201.03545.

Examples

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

Used by 2

Constructors

dunder

init

→None

__init__(config: ConvNeXtConfig)

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 a ConvNeXt-T backbone and run a forward pass: >>> import lucid >>> from lucid.models.vision.convnext import ConvNeXt, ConvNeXtConfig >>> model = ConvNeXt(ConvNeXtConfig()) >>> x = lucid.randn(1, 3, 224, 224) >>> feat = model.forward_features(x) >>> feat.shape # (B, dims[-1]) (1, 768) >>> out = model(x) >>> out.last_hidden_state.shape # (B, 1, dims[-1]) (1, 1, 768)