VGG-16 feature-extracting backbone (config D, no BatchNorm).
Builds a VGG with the paper-cited VGG-16 topology:
per-block convolution counts (2, 2, 3, 3, 3) — adding a third
convolution to each of the deeper blocks 3/4/5.
Approximately 138.4 M parameters in the full classifier variant
(≈14.7 M in the convolutional trunk alone). Reaches a top-5
ImageNet validation error of 8.8% (Simonyan & Zisserman, ICLR 2015,
Table 3, configuration D, single test scale S = Q = 256) — VGG-16
was the most widely deployed variant of the family. The often-quoted
7.3% is the ILSVRC-2014 submission test error of an ensemble of
seven nets (Table 6), not a single-model number.
Model Size
Parameters
pretrainedbool= Falsevgg_16, so True raises
NotImplementedError rather than return randomly initialised weights.
vgg_16_cls is the checkpointed variant.**overridesobject= {}VGGConfig.Returns
VGGBackbone with the VGG-16 configuration applied (or with
overrides merged on top of it).
Notes
See Simonyan & Zisserman, ICLR 2015, Table 1 (configuration D).
Examples
>>> import lucid
>>> from lucid.models.vision.vgg import vgg_16
>>> model = vgg_16()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape # (B, 512, 7, 7)
(1, 512, 7, 7)