resnext_101_32x4d

→ResNeXt

resnext_101_32x4d(pretrained: bool = False, overrides: object = {})

source edit

ResNeXt-101 (32x4d) feature-extracting backbone.

Builds a ResNeXt with the paper-cited ResNeXt-101 (32x4d) topology: per-stage block counts (3, 4, 23, 3) (same as ResNet-101), cardinality $C = 32$ , width per group $d = 4$ . 42.1 M parameters without a classifier head (resnext_101_32x4d_cls adds 2.05 M more, reaching the 44.2 M the paper quotes); reaches 78.8% ImageNet top-1 in Xie et al., 2017.

Model Size

Parameters

pretrainedbool= False

Reserved for future pretrained-weight loading. Currently ignored.

**overridesobject= {}

Keyword overrides forwarded into ResNeXtConfig.

Returns

ResNeXt

Backbone with the ResNeXt-101 (32x4d) configuration applied (or with overrides merged on top of it).

Notes

See Xie et al., CVPR 2017, Table 1. Same depth as ResNet-101 with the $3\times3$ middle convolution split into 32 groups.

Examples

>>> import lucid
>>> from lucid.models.vision.resnext import resnext_101_32x4d
>>> model = resnext_101_32x4d()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape   # (B, 2048, 7, 7)
(1, 2048, 7, 7)

Used by 1

lucid.models.vision.resnext

>>> import lucid >>> from lucid.models.vision.resnext import resnext_101_32x4d >>> model = resnext_101_32x4d() >>> x = lucid.randn(1, 3, 224, 224) >>> out = model(x) >>> out.last_hidden_state.shape # (B, 2048, 7, 7) (1, 2048, 7, 7)