MaxViT-XLarge backbone (Tu et al., 2022).
Builds the canonical MaxViT-XLarge configuration:
depths=(2, 6, 14, 2), dims=(192, 384, 768, 1536).
471.1M parameters headless (475.0M with the classifier head,
which is the paper's 475M figure) — the largest variant in the
paper.
Model Size
Parameters
pretrainedbool= FalseNo checkpoint is published for
maxvit_xlarge, so True raises
NotImplementedError rather than return randomly initialised weights.**overridesobject= {}Keyword overrides on top of the canonical MaxViT-XLarge config.
Returns
MaxViTA MaxViT backbone returning a flat
feature.
Notes
MaxViT-XLarge reaches 85.5% top-1 on ImageNet-1k at 224x224 (Tu et al., 2022, Table 6) and 88.5% when fine-tuned at 512x512 after JFT-300M pretraining (Table 7) — the headline result of the paper.
Examples
>>> import lucid
>>> from lucid.models.vision.maxvit import maxvit_xlarge
>>> model = maxvit_xlarge()
>>> x = lucid.randn(1, 3, 224, 224)
>>> model.forward_features(x).shape
(1, 1536)