Xception

Xception feature-extracting backbone (no classification head).

Implements the "Extreme Inception" topology from Chollet, "Xception: Deep Learning with Depthwise Separable Convolutions", CVPR 2017 (arXiv:1610.02357). The architecture replaces every Inception module of Inception-v3 with a depthwise separable convolution — the limiting case of an Inception block in which every output channel of the pointwise convolution receives its own independent spatial filter. This makes the cross-channel and spatial correlations fully decoupled, dramatically reducing parameters and FLOPs:

The body is organised into three phases — entry flow (3 blocks, with 1×1 strided residuals and channel doubling from 64 → 728), middle flow (8 identical 728-ch blocks with additive identity shortcuts), and exit flow (1 strided transition block plus two final separable convolutions expanding to 1536 → 2048 channels). Designed for a 299×299 input — the same crop size as Inception-v3 — so the spatial-reduction schedule isolates the depthwise-separable factorisation as the sole source of accuracy improvement.