The EfficientNet-B0 backward chain — the ℝ map the B0 whole-net tie is about #
The hand-composed reverse of the committed EfficientNet-B0 forward, as a plain def on the
cotangent: the batched sixteen-block paper net efficientnetInputGradB_full (the reverse of
efficientnetForwardB_full, at a variable batch N and class count). The chain keeps its block
backwards, its BatchNorm backwards and its swish backwards as supplied maps and spells only the
endpoints, so that the certified tie (EfficientNetFullWholeBackCertifiedTie) is a statement
about a NAMED chain of the forward's shape.
⚠ The stem is XLA-SAME (flatConvStride2XlaBack, the odd-phase scatter), the TF-origin
convention; the strided depthwises inside the blocks are symmetric and sit in the supplied block
backwards. B0 has no stem pool, so every batched endpoint is StableHLO.batchMap N of a
per-example leaf.
No number is stated about this chain.
The batched whole-net input-gradient backward of the sixteen-block EfficientNet-B0 —
the reverse of efficientnetForwardB_full = head ∘ b16 ∘ … ∘ b1 ∘ stem: classifier-back →
GAP-back → head-conv-bn-swish-back → the sixteen MBConv block backs → stem-conv-bn-swish-back.
The block backs and the stem/head BN+swish backs are supplied; the conv/GAP/dense leaves are
concrete, batchMap-lifted over the N examples, the stem at the XLA-SAME phase.
Equations
- One or more equations did not get rendered due to their size.