CvT-13 backbone (Wu et al., 2021).
Builds the canonical CvT-13 configuration: dims=(64, 192, 384),
depths=(1, 2, 10), num_heads=(1, 3, 6),
embed_strides=(4, 2, 2). Total spatial downsampling is
, so a 224x224 input yields
a 14x14 final token grid. Approximately 20M parameters.
Model Size
Parameters
pretrainedbool= FalseIf
True, loads ImageNet-1k pretrained weights when
available in the model zoo. Defaults to False.**overridesobject= {}Keyword overrides applied on top of the canonical CvT-13
config. Each override must match a field of
CvTConfig.Returns
CvTA CvT backbone returning a flat
mean-pooled feature.
Notes
CvT-13 reaches 81.6% top-1 on ImageNet-1k at 224x224 (Wu et al., 2021, Table 1). See arXiv:2103.15808.
Examples
>>> import lucid
>>> from lucid.models.vision.cvt import cvt_13
>>> model = cvt_13()
>>> x = lucid.randn(1, 3, 224, 224)
>>> feat = model.forward_features(x)
>>> feat.shape
(1, 384)