sk_resnext_50_32x4d

→SKNet

sk_resnext_50_32x4d(pretrained: bool = False, overrides: object = {})

source edit

SK-ResNeXt-50 32×4d feature-extracting backbone (the paper's SKNet-50).

Builds an SKNet with ResNet-50 bottleneck topology and ResNeXt-style grouped widening: cardinality = 32, base_width = 4, split_input = False. The bottleneck width per stage follows the ResNeXt formula

\text{width} = \lfloor \mathrm{planes} \cdot \tfrac{\text{base\_width}}{64} \rfloor \cdot \text{cardinality},

matching the SKNet-50 entry in Li et al., 2019. Approximately 27.5M parameters and 77.5% ImageNet-1k top-1 accuracy in the paper.

Model Size

Parameters

pretrainedbool= False

Reserved for future pretrained-weight loading. Currently ignored.

**overridesobject= {}

Keyword overrides forwarded into SKNetConfig.

Returns

SKNet

Backbone with the SK-ResNeXt-50-32×4d configuration applied (or with overrides merged on top of it).

Notes

See Li et al., "Selective Kernel Networks", CVPR 2019 (arXiv:1903.06586), Table 1 (SKNet-50 row). Combines the cardinality of ResNeXt with the data-dependent receptive-field selection of SK.

Examples

>>> import lucid
>>> from lucid.models.vision.sknet import sk_resnext_50_32x4d
>>> model = sk_resnext_50_32x4d()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape
(1, 2048, 7, 7)

Used by 1

lucid.models.vision.sknet

>>> import lucid >>> from lucid.models.vision.sknet import sk_resnext_50_32x4d >>> model = sk_resnext_50_32x4d() >>> x = lucid.randn(1, 3, 224, 224) >>> out = model(x) >>> out.last_hidden_state.shape (1, 2048, 7, 7)