inception_next_tiny

→InceptionNeXt

inception_next_tiny(pretrained: bool = False, overrides: object = {})

source edit

InceptionNeXt-Tiny backbone (Yu et al., 2024).

Builds the canonical InceptionNeXt-T configuration: depths=(3, 3, 9, 3), dims=(96, 192, 384, 768), band_kernel=11. Approximately 28M parameters — ConvNeXt-T compatible, but with each $7 \times 7$ depthwise conv replaced by the four-branch InceptionDWConv mixer.

Model Size

Parameters

pretrainedbool= False

If True, loads ImageNet-1k pretrained weights when available in the model zoo. Defaults to False.

**overridesobject= {}

Keyword overrides applied on top of the canonical InceptionNeXt-T config. Each override must match a field of InceptionNeXtConfig.

Returns

InceptionNeXt

An InceptionNeXt backbone returning a flat $(B, 768)$ feature.

Notes

InceptionNeXt-T matches ConvNeXt-T accuracy (82.3% top-1 on ImageNet-1k, Yu et al., 2024, Table 2) at noticeably lower latency. See arXiv:2303.16900.

Examples

>>> import lucid
>>> from lucid.models.vision.inception_next import inception_next_tiny
>>> model = inception_next_tiny()
>>> x = lucid.randn(1, 3, 224, 224)
>>> feat = model.forward_features(x)
>>> feat.shape
(1, 768)

Used by 1

lucid.models.vision.inception_next

>>> import lucid >>> from lucid.models.vision.inception_next import inception_next_tiny >>> model = inception_next_tiny() >>> x = lucid.randn(1, 3, 224, 224) >>> feat = model.forward_features(x) >>> feat.shape (1, 768)