class

InceptionV3

extendsPretrainedModelBackboneMixin

InceptionV3(config: InceptionConfig)

source edit

Inception v3 feature-extracting backbone.

Implements the topology from Szegedy et al., "Rethinking the Inception Architecture for Computer Vision", CVPR 2016: a deep pre-Inception stem (3×Conv → MaxPool → 2×Conv → MaxPool), three _InceptionA blocks at $35\times35$ , a Reduction-A block (_InceptionB) down to $17\times17$ , four _InceptionC blocks with factorised $1\times7$ and $7\times1$ convolutions, a Reduction-B block (_InceptionD) down to $8\times8$ , two _InceptionE blocks with the expanded $1\times3$ / $3\times1$ parallel branches, and a final lucid.nn.AdaptiveAvgPool2d to $1\times1$ . Designed for $299\times299$ RGB inputs.

Parameters

configInceptionConfig

Frozen architecture spec. Use inception_v3 for the paper-cited configuration. aux_logits only affects the classifier variant.

Attributes

configInceptionConfig

Stored copy of the config that built this model.

stemnn.Sequential

The pre-Inception conv stack.

inception_a0, inception_a1, inception_a2_InceptionA

Three Inception-A blocks at

35\times35

(factorised

5\times5

into two

3\times3

reduction_a_InceptionB

Reduction-A block reducing spatial size

35\times35 \to 17\times17

inception_c0, inception_c1, inception_c2, inception_c3_InceptionC

Four Inception-C blocks at

17\times17

with

1\times7

7\times1

factorisation.

reduction_b_InceptionD

Reduction-B block reducing spatial size

17\times17 \to 8\times8

inception_e0, inception_e1_InceptionE

Two Inception-E blocks at

8\times8

with expanded

1\times3

3\times1

parallel branches.

avgpoolnn.AdaptiveAvgPool2d

Global average pool to

1\times1

feature_infolist[FeatureInfo]

Per-stage descriptor (channels + reduction factor) exposed via BackboneMixin.

Notes

The headline factorisation idea is to replace an $n\times n$ convolution with an $n\times1$ followed by a $1\times n$ convolution: for $n = 7$ this trades $49 C^2$ parameters for $14 C^2$ while adding an extra ReLU and matching the original receptive field along a separable manifold. Inception v3 uses three block topologies (A, B, C) at three spatial resolutions (35×35, 17×17, 8×8), each tailored to the receptive-field budget at that stage. Approximately 23.8 M parameters (without auxiliary head); achieves a top-5 ImageNet error of 3.5% on the validation set.

Examples

>>> import lucid
>>> from lucid.models.vision.inception import inception_v3
>>> backbone = inception_v3()
>>> x = lucid.randn(1, 3, 299, 299)
>>> out = backbone(x)
>>> out.last_hidden_state.shape   # (B, 2048, 1, 1)
(1, 2048, 1, 1)

Used by 2

Constructors

dunder

init

→None

__init__(config: InceptionConfig)

source edit

Properties

prop

feature_info

→list[FeatureInfo]

feature_info: list[FeatureInfo]

source edit

Instance methods

forward

→BaseModelOutput

forward(x: Tensor)

source edit

forward_features

→Tensor

forward_features(x: Tensor)

source edit

>>> import lucid >>> from lucid.models.vision.inception import inception_v3 >>> backbone = inception_v3() >>> x = lucid.randn(1, 3, 299, 299) >>> out = backbone(x) >>> out.last_hidden_state.shape # (B, 2048, 1, 1) (1, 2048, 1, 1)