class

InceptionResNetV2ForImageClassification

extendsImageClassificationModel

InceptionResNetV2ForImageClassification(config: InceptionResNetConfig)

source edit

Inception-ResNet v2 image classifier (backbone + GAP + dropout + linear).

Combines an InceptionResNetV2 backbone with a global-average-pool, optional lucid.nn.Dropout (p=config.dropout, default 0.2), and a single lucid.nn.Linear projection producing config.num_classes logits. No auxiliary classifier — the residual shortcuts already provide effective gradient flow throughout the network. The classifier attribute is named classif (not classifier) to match the canonical timm / TensorFlow-Slim state-dict key.

Parameters

configInceptionResNetConfig

Architecture spec. Use inception_resnet_v2_cls for the paper-cited configuration.

Attributes

configInceptionResNetConfig

Stored copy of the config that built this model.

conv2d_1a … conv2d_7b, mixed_5b, repeat, mixed_6a, repeat_1,

mixed_7a, repeat_2, block8 Same backbone components as on InceptionResNetV2.

avgpoolnn.AdaptiveAvgPool2d

Final global average pool to

1\times1

dropoutnn.Module

lucid.nn.Dropout when config.dropout > 0 (default 0.2), otherwise lucid.nn.Identity.

classifnn.Linear

Final classifier projecting 1536 → num_classes — named classif for state-dict compatibility.

Notes

From Szegedy et al., "Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning", AAAI 2017. Loss is the standard categorical cross-entropy

\mathcal{L} = -\frac{1}{N} \sum_{n=1}^{N} \log \operatorname{softmax}(\text{logits}_n)_{\,y_n},

computed only when labels is not None. Approximately 55.8 M parameters; top-5 ImageNet validation error 3.08%. The key empirical message of the paper is that residual connections do not raise the final accuracy ceiling beyond a comparably-sized non-residual Inception v4, but they dramatically accelerate training convergence.

Examples

>>> import lucid
>>> from lucid.models.vision.inception_resnet import inception_resnet_v2_cls
>>> model = inception_resnet_v2_cls()
>>> x = lucid.randn(2, 3, 299, 299)
>>> out = model(x)
>>> out.logits.shape
(2, 1000)

Used by 2

Constructors

dunder

init

→None

__init__(config: InceptionResNetConfig)

source edit

Instance methods

forward

→InceptionResNetOutput

forward(x: Tensor, labels: Tensor | None = None)

source edit

reset_classifier

→None

reset_classifier(num_classes: int)

source edit

Replace the classification head with a freshly initialised one.

The head keeps the reference checkpoint's classif attribute name, so this cannot come from ClassificationHeadMixin (which looks for self.classifier). The class docstring documented the method without anything implementing it.

>>> import lucid >>> from lucid.models.vision.inception_resnet import inception_resnet_v2_cls >>> model = inception_resnet_v2_cls() >>> x = lucid.randn(2, 3, 299, 299) >>> out = model(x) >>> out.logits.shape (2, 1000)