ResNeSt-269 feature-extracting backbone (no classification head).
Builds a ResNeSt with the very deep topology
[3, 30, 48, 8], radix = 2, and stem_width = 64 —
the deepest variant from Zhang et al., 2022. Approximately
110.9M parameters and 84.5% ImageNet-1k top-1 accuracy
(Table 4) at 416×416 input resolution.
Model Size
Parameters
pretrainedbool= FalseReserved for future pretrained-weight loading. Currently
ignored.
**overridesobject= {}Keyword overrides forwarded into
ResNeStConfig.Returns
ResNeStBackbone with the ResNeSt-269 configuration applied (or
with overrides merged on top of it).
Notes
See Zhang et al., "ResNeSt: Split-Attention Networks", CVPR
Workshops 2022 (arXiv:2004.08955), Table 4. Memory footprint
is substantial; consider zero_init_residual=True for
stable large-batch training.
Examples
>>> import lucid
>>> from lucid.models.vision.resnest import resnest_269
>>> model = resnest_269()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape
(1, 2048, 7, 7)