Documentation

Param shapes owned by a .fpnDetect layer, in the ONE canonical order that emitTrainStepSig, the optimizer, the grad-clip list, the backward's grad-binding, NetSpec.paramShapes and heInitLayer must all agree on:

Wn3 [oc,c3], Wn4 [oc,c4], Wn5 [oc,c5]           -- neck laterals
(W [oc,oc,3,3], b [oc]) × tower, per level P3 → P4 → P5
Wh3, Wh4, Wh5 [A·15, oc]                        -- head convs
bh3, bh4, bh5 [A·15]                            -- head biases, LAST

The biases MUST stay last: NetSpec.applyDetPriorBias installs the RetinaNet prior by splicing the final 3·A·15 floats of the buffer.

Defined here (rather than in the codegen) because SpecHelpers and MlirCodegen both need it and neither should own it — a mismatch between any two of these consumers is a silent parameter-aliasing bug, not a compile error.

Squeeze width of an .mbConv block's SE: a quarter of the block's input channels (timm's se_ratio taken over the block input, not the expanded width). Every parameter-layout writer and emitter reads it from here.

The one writer of a trainable layer's parameter layout — the tensors in signature order, each with its initialization. NetSpec.paramShapes (the packed [θ|m|v] blob), heInitLayer (the initial weights) and Layer.nParams (hence totalParams) all read it, so they cannot disagree. none for layers the codegen does not train (their count is nParamsUntrained). The MLIR signatures (MlirCodegen.fwdSigParts, emitTrainStepSig) render it too.

The displayed parameter count of a layer the codegen does NOT train (Layer.paramSlots is none): the Bestiary's architectures, some counted approximately (.mambaBlock).

Output channels of a layer. Returns 0 for structural layers (pool, flatten, GAP).

Input channels expected by a layer. Returns 0 for layers that accept any input.

Validate that channel dimensions chain correctly through the spec. Returns none if valid, or some errorMessage describing the first mismatch.

Validate at build time. Use as: #eval mySpec.validate!

The shared Bestiary summary block, printed by every Bestiary/*.lean entry. The first line's label, the params abbreviation (unit) and an optional parenthetical on the validate : OK line (okNote) are the parts that vary.

The ── {name} ── and params<pad>: {n} lines are parsed, not decoration: tests/test_bestiary_params.py parses exactly those two shapes out of each binary's stdout to pin every variant against tests/bestiary_params.yml (the CI drift guard in .github/workflows/certs.yml). Keep both shapes if you touch this.