inception_v3_cls

→InceptionV3ForImageClassification

inception_v3_cls(pretrained: bool | str = False, weights: InceptionV3Weights | None = None, overrides: object = {})

source edit

Inception v3 image classifier (no auxiliary head by default).

Builds an InceptionV3ForImageClassification with the paper-cited Szegedy 2016 topology. Default configuration matches the standard reference-framework setting of aux_logits=False — approximately 23.8 M parameters; enable the auxiliary classifier by passing aux_logits=True (adds ≈4 M parameters used only during training). Designed for $299\times299$ RGB inputs; achieves a top-5 ImageNet validation error of 3.5%.

Model Size

Parameters

pretrainedbool or str= False

Pretrained-weight selector. False → random init; True → the DEFAULT tag (InceptionV3Weights.IMAGENET1K_V1); a tag string (e.g. "IMAGENET1K_V1") → that specific checkpoint. Mutually exclusive with weights (which wins if both are given).

weights(InceptionV3Weights, optional, keyword - only)= None

Explicit weights enum member, e.g. InceptionV3Weights.IMAGENET1K_V1. Takes precedence over pretrained.

**overridesobject= {}

Keyword overrides forwarded into InceptionConfig. Common picks: num_classes=N to retarget the head, aux_logits=True to enable the auxiliary classifier, dropout=p to adjust regularisation strength. Note: overriding num_classes (or aux_logits) away from the checkpoint topology makes pretrained loading fail the strict key/shape check.

Returns

InceptionV3ForImageClassification

Classifier with the Inception v3 configuration applied (or with overrides merged on top of it), optionally initialised from pretrained weights.

Notes

See Szegedy et al., "Rethinking the Inception Architecture for Computer Vision", CVPR 2016, §6. Training with the auxiliary head adds a 0.4-weighted cross-entropy term as

\mathcal{L} = \mathcal{L}_{\text{main}} + 0.4 \cdot \mathcal{L}_{\text{aux}}.

Pretrained weights are converted from torchvision's Inception_V3_Weights (auxiliary head dropped) and hosted on the Hugging Face Hub under lucid-dl/inception-v3. They evaluate at a 299-pixel centre crop resized from 342.

Examples

>>> import lucid
>>> from lucid.models.vision.inception import inception_v3_cls
>>> model = inception_v3_cls().eval()
>>> x = lucid.randn(2, 3, 299, 299)
>>> out = model(x)
>>> out.logits.shape
(2, 1000)
Load ImageNet-pretrained weights:
>>> model = inception_v3_cls(pretrained=True)            # DEFAULT tag
>>> from lucid.models.vision.inception import InceptionV3Weights
>>> model = inception_v3_cls(weights=InceptionV3Weights.IMAGENET1K_V1)

Used by 1

lucid.models.vision.inception

>>> import lucid >>> from lucid.models.vision.inception import inception_v3_cls >>> model = inception_v3_cls().eval() >>> x = lucid.randn(2, 3, 299, 299) >>> out = model(x) >>> out.logits.shape (2, 1000) Load ImageNet-pretrained weights: >>> model = inception_v3_cls(pretrained=True) # DEFAULT tag >>> from lucid.models.vision.inception import InceptionV3Weights >>> model = inception_v3_cls(weights=InceptionV3Weights.IMAGENET1K_V1)