resnext_101_64x4d(pretrained: bool = False, overrides: object = {})ResNeXt-101 (64x4d) feature-extracting backbone — the paper's headline model.
Builds a ResNeXt with per-stage block counts
(3, 4, 23, 3), cardinality and width per group
. This is Table 4's 2x-complexity configuration, the
one §5.1's "reduces the top-1 error to 20.4%" refers to, and the
basis of the 2nd-place ILSVRC-2016 submission.
Model Size
Parameters
pretrainedbool= FalseReserved for future pretrained-weight loading. Currently
ignored.
**overridesobject= {}Keyword overrides forwarded into
ResNeXtConfig.Returns
ResNeXtBackbone with the ResNeXt-101 (64x4d) configuration applied
(or with overrides merged on top of it).
Notes
Xie et al., CVPR 2017, Table 4: 20.4% top-1 / 5.3% top-5 error at 224x224. The paper also reports 19.1 / 4.4 for this configuration at 320x320, in its comparison against state-of-the-art models.
Examples
>>> import lucid
>>> from lucid.models.vision.resnext import resnext_101_64x4d
>>> model = resnext_101_64x4d()
>>> model(lucid.randn(1, 3, 224, 224)).last_hidden_state.shape
(1, 2048, 7, 7)