Xception feature-extracting backbone (no classification head).
Builds an Xception with the paper-cited "Extreme
Inception" topology from Chollet, 2017: a 2-conv stem, three
entry-flow blocks (1×1 strided residuals with channel
progression 64 → 128 → 256 → 728), eight middle-flow blocks at
728 channels, one exit-flow transition block (728 → 1024),
and two final separable convolutions expanding to 1536 → 2048
channels. Approximately 22.9M parameters — the same budget as
Inception-v3 but with depthwise separable convolutions
throughout.
Model Size
Parameters
pretrainedbool= False**overridesobject= {}XceptionConfig
(e.g. in_channels=1 for grayscale input).Returns
XceptionBackbone with the canonical Xception configuration applied
(or with overrides merged on top of it).
Notes
See Chollet, "Xception: Deep Learning with Depthwise Separable Convolutions", CVPR 2017 (arXiv:1610.02357). Designed for 299×299 input — the same crop as Inception-v3 — so the spatial-reduction schedule isolates the depthwise-separable factorisation as the sole source of accuracy improvement.
Examples
>>> import lucid
>>> from lucid.models.vision.xception import xception
>>> model = xception()
>>> x = lucid.randn(1, 3, 299, 299)
>>> out = model(x)
>>> out.last_hidden_state.shape
(1, 2048, 1, 1)