The FULL EfficientNet-B0 — all 16 MBConv blocks, batched forward graph + faithfulness #

Scales EfficientNetStagesPC's representative (stem + 3 MBConv + head) to the real B0 [t,c,n,s,k] spec — 16 MBConv layers — at the batched index N·(c·h·w). Pure enumeration + chaining of the generic per-block machinery; the only genuinely-new piece is the 4th block shape (mbExp: expand + stride-1 + no residual, used by stage-5/stage-7 first blocks where ic ≠ oc).

B0 stage spec [t,c,n,s,k]: s1 (1,16,1,1,3) s2 (6,24,2,2,3) s3 (6,40,2,2,5) s4 (6,80,3,2,3) s5 (6,112,3,1,5) s6 (6,192,4,2,5) s7 (6,320,1,1,3); stem 3×3-s2 (3→32), head 1×1 (320→1280)→GAP→dense. Per-block (ic, mid=t·ic, oc, r=⌈ic/4⌉, k, spatial, kind): b1 32→16 mid32 r8 k3 @112 noExp b9 80→112 mid480 r20 k5 @14 exp(no-resid) b2 16→24 mid96 r4 k3 112→56 strided b10 112→112 mid672 r28 k5 @14 resid b3 24→24 mid144 r6 k3 @56 resid b11 112→112 mid672 r28 k5 @14 resid b4 24→40 mid144 r6 k5 56→28 strided b12 112→192 mid672 r28 k5 14→7 strided b5 40→40 mid240 r10 k5 @28 resid b13 192→192 mid1152 r48 k5 @7 resid b6 40→80 mid240 r10 k3 28→14 strided b14 192→192 mid1152 r48 k5 @7 resid b7 80→80 mid480 r20 k3 @14 resid b15 192→192 mid1152 r48 k5 @7 resid b8 80→80 mid480 r20 k3 @14 resid b16 192→320 mid1152 r48 k3 @7 exp(no-resid)

Padding: the 3×3/s2 stem AND the four strided depthwises (b2, b4, b6, b12) pad at the XLA-SAME phase — TF's SAME at every strided conv, so the spec is tf_efficientnet_b0's. (Until D3, 2026-10-12, the four depthwises padded symmetrically, a hybrid neither timm net computes; scripts/parity/enet_timm_parity.py --pad hybrid still measures that distance.)

Precision is a flag on the graph. The block builders (EfficientNetStagesPC, mbExpGraphB here) choose every conv and depthwise constructor through the renderers' switch (StableHLO.PrecisionSwitch), so one graph describes the f32 artifacts (bf16 := false) and the bf16 ones (bf16 := true, the kinds efficientnetin_emarmsdp64dropdowxeps0001bf16 emits); the squeeze-excite denses, BatchNorm, swish, GAP and the classifier carry no flag because no render switches them. efficientnetFwdGraphBFull_faithful holds at either value through Bf16Erasure (the bf16 kind at the identity rounding is its f32 peer); the rounding itself is not modelled, as it is not at f32: den is over ℝ at both values.

References #

• Hu, Shen, Sun 2018, Squeeze-and-Excitation Networks. https://arxiv.org/abs/1709.01507
• Tan & Le 2019, EfficientNet: Rethinking Model Scaling for CNNs. https://arxiv.org/abs/1905.11946