class

XceptionForImageClassification

extendsPretrainedModelClassificationHeadMixin

XceptionForImageClassification(config: XceptionConfig)

source edit

Xception with dropout + linear classification head.

Combines an Xception backbone with a dropout layer (probability config.dropout, default 0.5 per the paper) and a linear projection to config.num_classes logits. When labels are supplied to forward, a cross-entropy loss is computed and returned alongside the logits inside an XceptionOutput dataclass.

Parameters

configXceptionConfig

Architecture spec. Use the xception_cls factory for the paper-cited configuration.

Attributes

configXceptionConfig

Stored copy of the config that built this model.

conv1, conv2, bn1, bn2, block1, …, block12, conv3, conv4, bn3, bn4, avgpool

Same backbone components as on Xception; see that class for shape semantics.

classifiernn.Module

Built by ClassificationHeadMixin._build_classifier — a lucid.nn.Sequential wrapping lucid.nn.Dropout and lucid.nn.Linear.

Notes

The classification flow is

\text{logits} = W \,\operatorname{Drop}\!\left( \operatorname{GAP}(\,\mathrm{backbone}(x)\,) \right) + b.

Chollet (2017) reports 79.0% top-1 ImageNet validation accuracy with this configuration at the 299×299 input resolution, edging out Inception-v3's 78.2% at the same parameter count (~22.9M).

Examples

Run inference at the native 299×299 input resolution:
>>> import lucid
>>> from lucid.models.vision.xception import xception_cls
>>> model = xception_cls()
>>> x = lucid.randn(4, 3, 299, 299)
>>> out = model(x)
>>> out.logits.shape
(4, 1000)

Used by 2

Constructors

dunder

init

→None

__init__(config: XceptionConfig)

source edit

Instance methods

forward

→XceptionOutput

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

source edit

Run inference at the native 299×299 input resolution: >>> import lucid >>> from lucid.models.vision.xception import xception_cls >>> model = xception_cls() >>> x = lucid.randn(4, 3, 299, 299) >>> out = model(x) >>> out.logits.shape (4, 1000)