InceptionResNetConfig

Configuration for Inception-ResNet v2.

This is the TF-Slim Inception-ResNet-v2 — the network the released checkpoint was trained as — which differs from the paper's figures in three documented ways. All three are dictated by the weights: the tensors in the checkpoint have these shapes and these counts.

Sequential stem, not Figure 3's parallel-branch stem Fig. 3 draws three filter-concat junctions ({MaxPool | Conv} → {1x1→3x3 | 1x1→7x1→1x7→3x3} → {Conv | MaxPool}) reaching 35×35×384. TF-Slim instead runs a plain sequential conv/pool chain into Mixed_5b.

Repeat counts 10 / 20 / 9+1, not Figure 15's 5 / 10 / 5 TF-Slim doubles every stage.

Residual widths 320 / 1088 / 2080, not Figures 16/17/19's 384 / 1154 / 2048 The three Linear 1×1 projections that close each residual block are sized to TF-Slim's widths.

Layout as built:

• Sequential stem (TF-Slim)
  • Mixed_5b (192 → 320)
  • 10× Block35 (scale_a, default 0.17)
  • Mixed_6a / Reduction-A (320 → 1088)
  • 20× Block17 (scale_b, default 0.10)
  • Mixed_7a / Reduction-B (1088 → 2080)
  • 9× Block8 (scale_c, default 0.20) + 1× Block8 (no ReLU)
  • Conv2d_7b 1×1 projection (2080 → 1536)
  • AdaptiveAvgPool → Dropout → FC(1536, num_classes)

scale_a — residual scale for Block35 (paper suggests 0.17). scale_b — residual scale for Block17 (paper suggests 0.10). scale_c — residual scale for Block8 (paper suggests 0.20). dropout — head dropout rate (0.2 in the paper).