zfnet

→ZFNet

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

source edit

ZFNet feature-extracting backbone (no fully-connected head).

Builds a ZFNet with the paper-cited Zeiler & Fergus 2014 topology: five convolutional blocks with a tightened $7\times7$ stride-2 first convolution and a $5\times5$ stride-2 second convolution (versus AlexNet's $11\times11$ stride-4 and $5\times5$ stride-1). Approximately 62 M parameters in the full classifier variant — the topology difference versus AlexNet does not change the headline parameter count.

Model Size

Parameters

pretrainedbool= False

Reserved for future pretrained-weight loading. Currently ignored — the returned model is randomly initialised.

**overridesobject= {}

Keyword overrides forwarded into ZFNetConfig (e.g. in_channels=1 for grayscale inputs).

Returns

ZFNet

Backbone with the ZFNet configuration applied (or with overrides merged on top of it).

Notes

See Zeiler & Fergus, "Visualizing and Understanding Convolutional Networks", ECCV 2014. The paper reports a top-5 ImageNet validation error of 11.7% — an improvement over AlexNet's 15.3% achieved with essentially the same parameter count by retuning the first two conv layers based on deconvolutional-network visualisations. ZFNet is a single architecture; there are no paper-cited size variants (H11).

Examples

>>> import lucid
>>> from lucid.models.vision.zfnet import zfnet
>>> model = zfnet()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape   # (B, 256, 6, 6)
(1, 256, 6, 6)

Used by 1

lucid.models.vision.zfnet

>>> import lucid >>> from lucid.models.vision.zfnet import zfnet >>> model = zfnet() >>> x = lucid.randn(1, 3, 224, 224) >>> out = model(x) >>> out.last_hidden_state.shape # (B, 256, 6, 6) (1, 256, 6, 6)