CvT-W24 wide backbone (Wu et al., 2021).
Builds the largest reference configuration CvT-W24:
dims=(192, 768, 1024), depths=(2, 2, 20),
num_heads=(3, 12, 16). Roughly an order of magnitude wider
than CvT-13. Approximately 277.2M parameters — typically used
only with large-scale (ImageNet-22k) pretraining.
Model Size
Parameters
pretrainedbool= FalseIf
True, loads ImageNet-22k pretrained weights when
available. Defaults to False.**overridesobject= {}Keyword overrides on top of the canonical CvT-W24 config.
Returns
CvTA CvT backbone returning a flat
mean-pooled feature.
Notes
CvT-W24 reaches 87.7% top-1 on ImageNet-1k at 384x384 fine-tune resolution after ImageNet-22k pretraining (Wu et al., 2021, Table 4) — the headline result of the paper.
Examples
>>> import lucid
>>> from lucid.models.vision.cvt import cvt_w24
>>> model = cvt_w24()
>>> x = lucid.randn(1, 3, 224, 224)
>>> model.forward_features(x).shape
(1, 1024)