mobilenet_v2_075

→MobileNetV21.4M

mobilenet_v2_075(pretrained: bool = False, overrides: object = {})

source edit

MobileNet-v2 backbone at width multiplier $\alpha = 0.75$ .

Builds a MobileNetV2 with every channel count multiplied by 0.75 — 1.36M parameters headless (2.64M with the classifier head).

Sandler et al., 2018 has no $\alpha = 0.75$ row: Table 4 tabulates only 1.0 and 1.4, and 0.75 appears solely in Figure 5's caption as a point on a curve. The often-quoted 69.8% top-1 is the TF model-zoo figure for mobilenet_v2_0.75_224, not a paper result, and the paper's 300M MAdds is for $\alpha = 1.0$ only — it supports no FLOP ratio for this width.

Model Size

Parameters

pretrainedbool= False

Reserved for future pretrained-weight loading. Currently ignored.

**overridesobject= {}

Keyword overrides forwarded into MobileNetV2Config.

Returns

MobileNetV2

Backbone with the MobileNet-v2 ( $\alpha = 0.75$ ) configuration applied (or with overrides merged on top of it).

Notes

See Sandler et al., "MobileNetV2: Inverted Residuals and Linear Bottlenecks", CVPR 2018 (arXiv:1801.04381), Table 4. The head expansion remains at 1280 channels — narrow variants do not shrink the head below 1280.

Examples

>>> import lucid
>>> from lucid.models.vision.mobilenet_v2 import mobilenet_v2_075
>>> model = mobilenet_v2_075()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape
(1, 1280, 1, 1)

Used by 1

lucid.models.vision.mobilenet_v2

>>> import lucid >>> from lucid.models.vision.mobilenet_v2 import mobilenet_v2_075 >>> model = mobilenet_v2_075() >>> x = lucid.randn(1, 3, 224, 224) >>> out = model(x) >>> out.last_hidden_state.shape (1, 1280, 1, 1)