CvT-21 backbone (Wu et al., 2021).
Builds the canonical CvT-21 configuration: same widths and head
counts as CvT-13 (dims=(64, 192, 384), num_heads=(1, 3, 6))
but deeper — depths=(1, 4, 16). Approximately 31.6M
parameters.
Model Size
Parameters
pretrainedbool= FalseIf
True, loads ImageNet-1k pretrained weights when
available. Defaults to False.**overridesobject= {}Keyword overrides on top of the canonical CvT-21 config.
Returns
CvTA CvT backbone returning a flat
mean-pooled feature.
Notes
CvT-21 reaches 82.5% top-1 on ImageNet-1k at 224x224 (Wu et al., 2021, Table 1).
Examples
>>> import lucid
>>> from lucid.models.vision.cvt import cvt_21
>>> model = cvt_21()
>>> x = lucid.randn(1, 3, 224, 224)
>>> model.forward_features(x).shape
(1, 384)