Wide ResNet-50-2 feature-extracting backbone (no classification head).
Builds a ResNet with ResNet-50's [3, 4, 6, 3]
_Bottleneck layout but with a bottleneck_width_mult=2
multiplier on the inner 3×3 convolutions. The per-stage output
channel count is unchanged from ResNet-50 (256 / 512 / 1024 /
2048), but the inner bottleneck width doubles — yielding roughly
68.9M parameters. Defined in Zagoruyko & Komodakis, "Wide
Residual Networks", BMVC 2016 (arXiv:1605.07146).
Model Size
Parameters
pretrainedbool= False**overridesobject= {}ResNetConfig.Returns
ResNetBackbone with the Wide ResNet-50-2 configuration applied (or
with overrides merged on top of it).
Notes
Widening the inner channels lets shallower networks recover (and surpass) the accuracy of much deeper canonical ResNets while being easier to parallelise on modern accelerators because the FLOPs are concentrated in a smaller number of wider layers.
Examples
>>> import lucid
>>> from lucid.models.vision.resnet import wide_resnet_50
>>> model = wide_resnet_50()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape # (B, 2048, H/32, W/32)
(1, 2048, 7, 7)