MobileNetV4-Conv-M — every parameter gradient node IS the loss's derivative in that parameter #
mnv4_net_tiedB says each of the 233 parameter gradient nodes denotes its layer's parameter
Jacobian contracted with the cotangent the emitted backward chain threads to it, from a loss
cotangent g; the *CotIn_eq_vjp lemmas say each block's input cotangent is its CertLayer's
certified backward. mnv4_net_lossGrad composes them: for any loss L of the logits whose gradient
at the net's output is g, every node is ∂L/∂θ of the WHOLE net with that one parameter varied.
mnv4_net_lossGrad_smoothedCE discharges hL for the label-smoothed loss the artifacts ship.
How. ResNet50ParamGrad's shape, with two things MNv4 adds:
- The depthwise slots. A UIB body's pre- and post-depthwise are
if k = 0 then id' else …layers read off the row, and the depthwise weights live in aDWSlotthat is aDWBnParamsonly atk > 0. So each family bundle (mnv4ExtraDWLossTiedB,mnv4ConvNeXtLossTiedB,mnv4FfnLossTiedB,mnv4StridedLossTiedB) takes its row'sk ≠ 0/k = 0facts, the slots' forwards reduce bymnv4PreDWSlot_fwd_of_ne_zeroand friends, and a depthwise parameter is varied throughUibParams.withPre/withPost(the slot record with one field replaced). - The layer-group trunk. The forward nests the five resolution groups, not the 21 blocks, so
each block's factor lemma peels its group with
CertLayer.comp_fwd_apply, andmnv4Pre{k}meetsmnv4Blk{k}at the group boundaries (mnv4Pre2_eq_blk…). At the net's literal widths these checks stay cheap because every peel is a named rewrite, never an unfolding.
Hypotheses. Mnv4SmoothAt (every relu off its kink at the real activations — the stem's
clause and each group's .ok); the BN ε > 0 facts live in the weights. For the smoothed loss,
every example's target summing to one and 0 < nCls.
The record with its pre-depthwise slot's parameters replaced.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The record with its post-depthwise slot's parameters replaced.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Stem, every parameter node a loss derivative — the three nodes mnv4StemTiedB ties, Φ
the loss as a function of the stem's (W, γ, β).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Fused stage, every parameter node a loss derivative — the six nodes mnv4FusedTiedB
ties, Φ the loss as a function of (Wc, γc, βc, Wp, γp, βp).
Equations
- One or more equations did not get rendered due to their size.
Instances For
The head as one plain function of its eight trained parameters (the ε's and conv biases held
at the given values): mnv4Head's forward.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Head, every parameter node a loss derivative — the eight nodes mnv4HeadTiedB ties, Φ
the loss as a function of (W1, γ1, β1, W2, γ2, β2, Wd, bd).
Equations
- One or more equations did not get rendered due to their size.
Instances For
ExtraDW body, every parameter node a loss derivative — the twelve nodes
mnv4ExtraDWTiedB ties, Φ the loss at the body output as a function of the row's weight
record. A depthwise slot's parameter is varied through withPre / withPost.
Equations
- One or more equations did not get rendered due to their size.
Instances For
ConvNeXt-like body, every parameter node a loss derivative — the nine nodes
mnv4ConvNeXtTiedB ties.
Equations
- One or more equations did not get rendered due to their size.
Instances For
FFN body, every parameter node a loss derivative — the six nodes mnv4FfnTiedB ties.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Strided body, every parameter node a loss derivative — the twelve nodes
mnv4StridedTiedB ties; the post-DW is the symmetric strided depthwise.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Pull a gradient back through any certified layer, at a point it certifies.
A skip block's step: the emitted fan-in body dx + dyOut is the residual layer's backward.
At a skip block the loss read at the BODY output is u ↦ G (u + v): the skip is a constant
once a body parameter varies, so its gradient there is still the block-output cotangent.
The group prefixes meet the block prefixes at every group boundary.
The fused stage's output, as the plain stage functions.
The net after block 21 — the head.
Equations
Instances For
The net after block 20: block 21, then the rest.
Equations
Instances For
The net after block 19: block 20, then the rest.
Equations
Instances For
The net after block 18: block 19, then the rest.
Equations
Instances For
The net after block 17: block 18, then the rest.
Equations
Instances For
The net after block 16: block 17, then the rest.
Equations
Instances For
The net after block 15: block 16, then the rest.
Equations
Instances For
The net after block 14: block 15, then the rest.
Equations
Instances For
The net after block 13: block 14, then the rest.
Equations
Instances For
The net after block 12: block 13, then the rest.
Equations
Instances For
The net after block 11: block 12, then the rest.
Equations
Instances For
The net after block 10: block 11, then the rest.
Equations
Instances For
The net after block 9: block 10, then the rest.
Equations
Instances For
The net after block 8: block 9, then the rest.
Equations
Instances For
The net after block 7: block 8, then the rest.
Equations
Instances For
The net after block 6: block 7, then the rest.
Equations
Instances For
The net after block 5: block 6, then the rest.
Equations
Instances For
The net after block 4: block 5, then the rest.
Equations
Instances For
The net after block 3: block 4, then the rest.
Equations
Instances For
The net after block 2: block 3, then the rest.
Equations
Instances For
The net after block 1: block 2, then the rest.
Equations
Instances For
The net after the fused stage: block 1, then the rest.
Equations
Instances For
The net after the stem: the fused stage, then the rest.
Equations
- Proofs.Mnv4TieB.mnv4SufStem N w y = Proofs.Mnv4TieB.mnv4Suf0 N w ((Proofs.StableHLO.mnv4FusedStack N w).fwd y)
Instances For
The net with the stem's trained parameters varied is the suffix after the stem at the varied stem.
The net with the fused stage's trained parameters varied.
The net with block 1's weights varied is the suffix after block 1 at the varied block.
The net with block 2's weights varied is the suffix after block 2 at the varied block.
The net with block 3's weights varied is the suffix after block 3 at the varied block.
The net with block 4's weights varied is the suffix after block 4 at the varied block.
The net with block 5's weights varied is the suffix after block 5 at the varied block.
The net with block 6's weights varied is the suffix after block 6 at the varied block.
The net with block 7's weights varied is the suffix after block 7 at the varied block.
The net with block 8's weights varied is the suffix after block 8 at the varied block.
The net with block 9's weights varied is the suffix after block 9 at the varied block.
The net with block 10's weights varied is the suffix after block 10 at the varied block.
The net with block 11's weights varied is the suffix after block 11 at the varied block.
The net with block 12's weights varied is the suffix after block 12 at the varied block.
The net with block 13's weights varied is the suffix after block 13 at the varied block.
The net with block 14's weights varied is the suffix after block 14 at the varied block.
The net with block 15's weights varied is the suffix after block 15 at the varied block.
The net with block 16's weights varied is the suffix after block 16 at the varied block.
The net with block 17's weights varied is the suffix after block 17 at the varied block.
The net with block 18's weights varied is the suffix after block 18 at the varied block.
The net with block 19's weights varied is the suffix after block 19 at the varied block.
The net with block 20's weights varied is the suffix after block 20 at the varied block.
The net with block 21's weights varied is the suffix after block 21 at the varied block.
The net with the head's trained parameters varied is the head at them.
Every MobileNetV4-Conv-M parameter gradient node is the derivative of L in that
parameter, for a loss L of the logits and g the cotangent the chain starts from: the 233
nodes mnv4_net_tiedB ties, each at the cotangent the emitted chain threads to it, stated
against L of mobilenetv4ForwardBFull with that one parameter varied.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Every MobileNetV4-Conv-M parameter gradient node is the derivative of the loss in that
parameter. For any loss L of the logits with gradient g at the net's output, each of the
233 nodes mnv4_net_tiedB ties — at the same cotangent — is ∂L/∂θ of the WHOLE net,
mobilenetv4ForwardBFull with that one parameter varied (a stem, fused-stage or head field,
or a block's weight record w.bk := p with one slot changed — a depthwise slot through
withPre / withPost).
The only hypothesis is Mnv4SmoothAt (the stem's relu clause and each group's .ok); the BN
ε > 0 facts are fields of the weights. The loss enters only through hL;
mnv4_net_lossGrad_smoothedCE discharges it for the loss the artifacts ship.
The loss the artifacts ship: every node is the derivative of the batched label-smoothed
cross-entropy smoothedBatchLoss, g the six-op cotangent the render emits.