class

InceptionResNetV2

extendsPretrainedModelBackboneMixin

InceptionResNetV2(config: InceptionResNetConfig)

source edit

Inception-ResNet v2 feature-extracting backbone.

Implements the residual-Inception topology from Szegedy et al., "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning", AAAI 2017. The body consists of three residual Inception block families — _Block35 (×10), _Block17 (×20), and _Block8 (×9 + 1 no-ReLU block) — separated by Reduction-A (_Mixed6a) and Reduction-B (_Mixed7a) blocks, with a leading Mixed_5b pre-block (_Mixed5b) and a final $1\times1$ projection (conv2d_7b) to 1536 channels. Designed for $299\times299$ RGB inputs. State-dict naming matches the canonical timm / TensorFlow-Slim implementation so weight transfer is straightforward.

Parameters

configInceptionResNetConfig

Frozen architecture spec. Use inception_resnet_v2 for the paper-cited configuration with the recommended residual scale factors (0.17 / 0.10 / 0.20).

Attributes

configInceptionResNetConfig

Stored copy of the config that built this model.

conv2d_1a, conv2d_2a, conv2d_2b, conv2d_3b, conv2d_4ann.Module

Stem ConvBnReLU layers (TensorFlow-Slim naming preserved).

mixed_5b_Mixed5b

Pre-Block35 Inception block (192 → 320 channels) at

35\times35

repeatnn.Sequential

Stack of 10 _Block35 residual Inception-A blocks.

mixed_6a_Mixed6a

Reduction-A block,

35\times35 \to 17\times17

, widening 320 → 1088 channels.

repeat_1nn.Sequential

Stack of 20 _Block17 residual Inception-B blocks.

mixed_7a_Mixed7a

Reduction-B block,

17\times17 \to 8\times8

, widening 1088 → 2080 channels.

repeat_2nn.Sequential

Stack of 9 _Block8 residual Inception-C blocks.

block8_Block8

Final no-ReLU residual block (scale 1.0).

conv2d_7b_ConvBnReLU

1\times1

projection 2080 → 1536.

avgpoolnn.AdaptiveAvgPool2d

Global average pool to

1\times1

feature_infolist[FeatureInfo]

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

Notes

Each residual Inception block computes

y = \sigma\!\left(x + \alpha \cdot \mathcal{F}_{\text{inception}}(x)\right),

with $\alpha$ a small fixed scale factor (0.10–0.20). The residual-scale trick was introduced specifically for Inception-ResNet because unscaled residuals caused training to diverge on networks with many filters per Inception module. Approximately 55.8 M parameters; achieves a top-5 ImageNet validation error of 3.08% in the original paper.

Examples

>>> import lucid
>>> from lucid.models.vision.inception_resnet import inception_resnet_v2
>>> backbone = inception_resnet_v2()
>>> x = lucid.randn(1, 3, 299, 299)
>>> out = backbone(x)
>>> out.last_hidden_state.shape   # (B, 1536, 1, 1)
(1, 1536, 1, 1)

Used by 2

Constructors

dunder

init

→None

__init__(config: InceptionResNetConfig)

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_resnet import inception_resnet_v2 >>> backbone = inception_resnet_v2() >>> x = lucid.randn(1, 3, 299, 299) >>> out = backbone(x) >>> out.last_hidden_state.shape # (B, 1536, 1, 1) (1, 1536, 1, 1)