The full EfficientNet-B0 at inference — all 16 MBConv blocks, eval forward + graph + #

faithfulness

The eval twin of EfficientNetFullB0.lean. That file states the sixteen-block [t,c,n,s,k] net at TRAINING BatchNorm (bnBatchLA), the world its VJP and its typed graph live in; this file states the same ladder at INFERENCE BatchNorm — frozen running statistics at all 49 sites, one shared ε, as the shipped efficientnet_fwd_eval does — and proves its typed SHlo graph denotes it: the graph of efficientnet_fwd_eval.mlir and its 1000-class twin.

Pure enumeration and chaining of EfficientNetStagesPCEval's per-block machinery, at the batched index N·(c·h·w) and generic in the class count. The one block shape added here is the fourth at inference — mbExpFwdBEval / mbExpGraphBEval: expand, stride 1, no residual (ic ≠ oc; the stage-5 and stage-7 first blocks b9/b16) — the eval twin of mbExpFwdB (EfficientNetChainClose.lean).

What it is tied to. efficientnet_fwd_eval.mlir is THIS net: %x plus 213 parameters (the render folds each conv bias into the BatchNorm that follows it, so %sb/%b1db/… have no slot) plus 98 statistic slots — 312 inputs at ten classes, and efficientnetin_fwd_eval.mlir its 1000-class twin. The typed graph uses the artifact's names (folded biases as the shared zero-bias constant, %stnmu/%b{k}{e,d,p}nmu/%hnmu statistics, zW1…zb2 SE denses, %Wd/%bd), and FwdGraphTextTies checks the render's stem, block and head text against it per kind.

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) at the XLA-SAME phase, head 1×1 (320→1280) → GAP → dense.