res_unet_3d(pretrained: bool = False, overrides: object = {})3-D ResUNet — volumetric residual U-Net.
H11 marker: not a paper variant. No single paper defines a 3-D
residual U-Net with this depth and channel schedule; it is the
volumetric counterpart of res_unet_2d, assembled from the
same building blocks rather than reproduced from a publication. It
ships no weights and no accuracy claim. Treat it as a starting
point to configure, not as a reproduction of any published result.
Builds a 3-D U-Net with residual DoubleConv blocks: the volumetric
counterpart to res_unet_2d. Useful for deeper variants on
biomedical volumes where the residual shortcut materially helps
optimisation stability.
Model Size
Parameters
pretrainedbool= Falseres_unet_3d, so True raises
NotImplementedError rather than return randomly initialised weights.**overridesobject= {}UNetConfig.Returns
UNetForSemanticSegmentationSegmentation model with the 3-D ResUNet configuration applied
(or with overrides merged on top of it).
Notes
Combines the 3-D U-Net topology of Cicek et al. (MICCAI 2016) with the residual DoubleConv blocks of Zhang et al. (2018).
Examples
>>> import lucid
>>> from lucid.models.vision.unet import res_unet_3d
>>> model = res_unet_3d()
>>> x = lucid.randn(1, 1, 64, 64, 64)
>>> out = model(x)
>>> out.logits.shape
(1, 2, 64, 64, 64)