SE-ResNet-101 feature-extracting backbone (no classification head).
Builds an SENet with ResNet-101 topology
(_SEBottleneck blocks stacked [3, 4, 23, 3]) plus
squeeze-and-excitation modules. Approximately 49.3M
parameters and 77.62% ImageNet-1k top-1 accuracy in Hu et al.,
2018 (Table 2: 22.38% top-1 error) versus 76.83% for their
re-implemented ResNet-101 baseline. (78.32% is this family's
shipped-checkpoint metric, recorded in _weights.py — not a
paper figure.)
Model Size
Parameters
pretrainedbool= FalseNo checkpoint is published for
se_resnet_101, so True raises
NotImplementedError rather than return randomly initialised weights.
se_resnet_101_cls is the checkpointed variant.**overridesobject= {}Keyword overrides forwarded into
SENetConfig.Returns
SENetBackbone with the SE-ResNet-101 configuration applied (or
with overrides merged on top of it).
Notes
See Hu et al., "Squeeze-and-Excitation Networks", CVPR 2018 (arXiv:1709.01507), Table 2.
Examples
>>> import lucid
>>> from lucid.models.vision.senet import se_resnet_101
>>> model = se_resnet_101()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape
(1, 2048, 7, 7)