vgg_19

→VGG20M

vgg_19(pretrained: bool = False, overrides: object = {})

source edit

VGG-19 feature-extracting backbone (config E, no BatchNorm).

Builds a VGG with the paper-cited VGG-19 topology: per-block convolution counts (2, 2, 4, 4, 4) — four $3\times3$ convolutions in each of the deeper blocks. Approximately 143.7 M parameters in the full classifier variant. Reaches a top-5 ImageNet validation error of 9.0% (Simonyan & Zisserman, ICLR 2015, Table 3, configuration E, single test scale S = Q = 256). 7.5% is Table 4's multi-scale figure, which D and E share. E is not worse than D: at multiple scales they tie, and the paper's best single model is E at 7.1% (Table 5, multi-crop and dense evaluation).

Model Size

Parameters

pretrainedbool= False

No checkpoint is published for vgg_19, so True raises NotImplementedError rather than return randomly initialised weights. vgg_19_cls is the checkpointed variant.

**overridesobject= {}

Keyword overrides forwarded into VGGConfig.

Returns

VGG

Backbone with the VGG-19 configuration applied (or with overrides merged on top of it).

Notes

See Simonyan & Zisserman, ICLR 2015, Table 1 (configuration E).

Examples

>>> import lucid
>>> from lucid.models.vision.vgg import vgg_19
>>> model = vgg_19()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape   # (B, 512, 7, 7)
(1, 512, 7, 7)

Used by 1

lucid.models.vision.vgg

>>> import lucid >>> from lucid.models.vision.vgg import vgg_19 >>> model = vgg_19() >>> x = lucid.randn(1, 3, 224, 224) >>> out = model(x) >>> out.last_hidden_state.shape # (B, 512, 7, 7) (1, 512, 7, 7)