class

GoogLeNetForImageClassification

extendsPretrainedModelClassificationHeadMixin

GoogLeNetForImageClassification(config: GoogLeNetConfig)

source edit

GoogLeNet image classifier with optional auxiliary classifiers.

Combines the GoogLeNet backbone with a global-average-pool + lucid.nn.Dropout (config.dropout, default 0.4) + lucid.nn.Linear head producing config.num_classes logits, plus two optional auxiliary classifiers attached at Inception 4a (512-channel feature map) and Inception 4d (528-channel feature map). Auxiliary heads are evaluated only in training mode and only when config.aux_logits=True; their contributions to the loss are weighted 0.3 each, following the original paper.

Parameters

configGoogLeNetConfig

Architecture spec. Use googlenet_cls for the paper-cited configuration (auxiliary heads enabled). Pass aux_logits=False for an inference-only variant.

Attributes

configGoogLeNetConfig

Stored copy of the config that built this model.

conv1, conv2, conv3, maxpool1, maxpool2, inception3a, ...,

inception5b, maxpool3, maxpool4

Same backbone components as on GoogLeNet.

avgpoolnn.AdaptiveAvgPool2d

Final global average pool to

1\times1

dropnn.Dropout

Dropout applied before the main classifier (p=config.dropout, 0.4 in the paper).

classifiernn.Module

Final linear projection 1024 → num_classes, built by ClassificationHeadMixin._build_classifier.

aux1, aux2nn.Module

Auxiliary classifiers (instances of _AuxClassifier) when config.aux_logits=True; otherwise lucid.nn.Identity placeholders.

Notes

From Szegedy et al., "Going Deeper with Convolutions", CVPR 2015, §5. The full training loss with auxiliary heads is

\mathcal{L} = \mathcal{L}_{\text{main}} + 0.3 \cdot \mathcal{L}_{\text{aux1}} + 0.3 \cdot \mathcal{L}_{\text{aux2}},

each term being the standard categorical cross-entropy $-\frac{1}{N} \sum_n \log \operatorname{softmax}(\cdot)_{y_n}$ . The $0.3$ weighting compensates for the auxiliary heads being much shallower than the main classifier; at inference only the main branch is evaluated.

Examples

Inference path (no labels, auxiliary heads inactive):
>>> import lucid
>>> from lucid.models.vision.googlenet import googlenet_cls
>>> model = googlenet_cls().eval()
>>> x = lucid.randn(4, 3, 224, 224)
>>> out = model(x)
>>> out.logits.shape
(4, 1000)
>>> out.aux_logits1 is None
True
Training path (auxiliary heads active, labels provided):
>>> model = model.train()
>>> labels = lucid.tensor([0, 1, 2, 3], dtype=lucid.int64)
>>> out = model(x, labels=labels)
>>> out.aux_logits1.shape   # available during training
(4, 1000)

Used by 2

Constructors

dunder

init

→None

__init__(config: GoogLeNetConfig)

source edit

Instance methods

forward

→GoogLeNetOutput

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

source edit

Inference path (no labels, auxiliary heads inactive): >>> import lucid >>> from lucid.models.vision.googlenet import googlenet_cls >>> model = googlenet_cls().eval() >>> x = lucid.randn(4, 3, 224, 224) >>> out = model(x) >>> out.logits.shape (4, 1000) >>> out.aux_logits1 is None True Training path (auxiliary heads active, labels provided): >>> model = model.train() >>> labels = lucid.tensor([0, 1, 2, 3], dtype=lucid.int64) >>> out = model(x, labels=labels) >>> out.aux_logits1.shape # available during training (4, 1000)