Files
ruvnet--RuView/v2/crates/wifi-densepose-train/src/mae.rs
T
rUv 17471e93ff ADR-152: WiFi-Pose SOTA 2026 intake — WiFlow-STD benchmark, Rust integrations, ADR-153 802.11bf layer, efficiency frontier (#1008)
* feat(calibration): NodeGeometry transceiver-geometry recording (ADR-152 §2.1.1)

PerceptAlign-motivated geometry capture at enrollment: per-node optional
records (position, antenna orientation, inter-node distances, acquisition
method) — recorded when known, never required. Event-sourced via
EnrollmentEvent::GeometryRecorded (latest recording wins); persisted on
SpecialistBank with serde defaults so pre-ADR-152 bank JSON loads cleanly
(fixture-proven, and geometry-free banks serialize byte-shape-identical
to the old schema); threaded through MultiNodeMixture as data only — the
learned geometry embeddings and algorithmic fusion use are §2.1.2,
deliberately deferred until the ADR-151 P6 LoRA heads exist.

Geometry recorded from now on means banks captured today remain usable
for layout-conditioned training later — you can't retroactively add
geometry to data you didn't record.

8 new tests (3 geometry, 2 anchor, 2 bank, 1 multistatic) + full-loop
extension (2-node geometry, one tape-measured + one unknown, surviving
the bank JSON round-trip the runtime loads from). 50/50 calibration
(both feature configs) + 23 CLI tests green.

Co-Authored-By: RuFlo <ruv@ruv.net>

* feat(training): two-checkerboard camera↔room calibration for ADR-079 labels (ADR-152 §2.1.3)

Defends the camera-supervised pipeline against PerceptAlign's
"coordinate overfitting": MediaPipe keypoints were emitted in raw camera
coordinates with no shared frame and no transceiver-geometry metadata —
the exact label shape that memorizes deployment layout and collapses
cross-layout.

- scripts/calibrate-camera-room.py + calibration_lib.py: OpenCV
  two-checkerboard calibration → versioned bundle JSON (intrinsics,
  camera→room extrinsics, checkerboard spec, transceiver geometry,
  sha256 calibration_id). Intrinsics resolve from file > cache >
  multi-view computation > loud-warning 2-view fallback.
- collect-ground-truth.py --calibration <bundle>: every sample gains
  keypoints_room (unit bearing rays from the camera center in the room
  frame — documented projective alignment; raw image coords preserved
  so training chooses), camera_origin_room, calibration_id, and the
  transceiver geometry stamp. Without the flag, output is byte-identical
  to before (tested) + a one-line ADR-152 warning.

Design finding (recorded for ADR-152): a single planar checkerboard's
corner grid is centrosymmetric — the reversed corner ordering fits a
ghost camera pose with IDENTICAL reprojection error, so per-board flip
disambiguation is mathematically ill-posed. solve_two_board_extrinsics
solves the joint wall+floor set over all 4 flip combinations, where the
minimum is unique — an independent reason the TWO-checkerboard method is
required, beyond what PerceptAlign states.

15 headless pytest tests green (synthetic corners: extrinsics recovery
incl. ghost resolution, bundle round-trip + hash stability, ray
transforms w/ distortion + cross-resolution, no-calibration byte
identity).

Co-Authored-By: RuFlo <ruv@ruv.net>

* feat(benchmarks): WiFlow-STD reproduction harness + measurement (a) results (ADR-152 §2.2)

Shipped checkpoint REFUTED (0.08% PCK@20, wrong keypoint normalization);
6 reproducibility defects documented (broken imports, corrupted dataset
tail with float32-max garbage that NaN-poisons fp16 BatchNorm, unreachable
test phase). After repairs, retraining with upstream defaults reproduces
96.09% PCK@20 full-test / 96.61% corruption-free (published 97.25%) on
RTX 5080. Claims graded MEASURED-EQUIVALENT; 2.23M params + ~0.055 GFLOPs
verified. Third-party code/weights/data stay out of tree (gitignored).

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat: ADR-152 Rust integrations + ADR-153 802.11bf protocol model

- calibration: GeometryEmbedding — 32-slot permutation-invariant NodeGeometry
  featurization for future LoRA-head conditioning (ADR-152 §2.1.2); derived
  SpecialistBank::geometry_embedding() accessor; 59 tests
- train: MaePretrainConfig + patchify/random-mask with UNSW measured recipe
  (80% masking, (30,3) patches; ADR-152 §2.3, arXiv 2511.18792); strict
  no-truncate/no-NaN policy; proptest properties
- train: WiFlowStdModel — tch-gated port of the verified ~96%-PCK@20
  WiFlow-STD architecture (ADR-152 §2.2 beyond-SOTA); ungated param formula
  pinned to 2,225,042; 15/17-keypoint support; 239 crate tests
- hardware: ieee80211bf forward-compatibility protocol model (ADR-153):
  SpecProfile gates, SensingCapabilities negotiation, required ConsentMode,
  session FSM, SensingTransport + SimTransport + OpportunisticCsiBridge;
  full acceptance checklist covered; 156+4 tests
- deps: ruvector bumps per ADR-152 §2.6 survey (mincut/solver 2.0.6,
  attention 2.1.0, gnn 2.2.0); vendor/ruvector synced to a083bd77f
- docs: ADR-153 accepted; ADR-152 §2.2 status, §2.4 amendment, §2.6 added

Workspace: 162 test suites green (--no-default-features); Python proof PASS.
Known pre-existing flake: homecore-api env_empty_falls_back_to_defaults
(unserialized env-var mutation) — untouched, follow-up.

Co-Authored-By: claude-flow <ruv@ruv.net>

* docs: CHANGELOG + CLAUDE.md entries for ADR-152 integrations and ADR-153

Co-Authored-By: claude-flow <ruv@ruv.net>

* fix(train): repair tch-backend bit-rot — gated path compiles and tests run again

Mechanical API refresh against current tch: Vec::from(Tensor) -> try_from
(+ explicit flatten), numel() usize cast, Rem/div ops -> remainder() /
divide_scalar_mode(floor) — the latter fixed a silent true-division bug in
heatmap argmax decoding; clamp(1.0, f64::MAX) -> clamp_min (torch 2.x scalar
overflow panic); petgraph EdgeRef import; missing EvalMetrics and
verify_checkpoint_dir APIs that tests documented. wiflow_std roundtrip test
uses safetensors (.pt _save_parameters roundtrip broken in torch 2.11
Windows). Gated: 349 passed (incl. all 20 wiflow_std); ungated: unchanged.
Known pre-existing: gaussian-heatmap convention mismatch (2 tests), proof
seed race under parallel threads — documented, deliberate follow-ups.

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat(train): WiFlow-STD PyTorch->tch weight import + numerical parity proof

export_to_safetensors.py maps the retrained checkpoint (295 tensors -> 248
mapped, param sum exactly 2,225,042; num_batches_tracked dropped) into a
tch-loadable safetensors plus a deterministic parity fixture. Gated #[ignore]
integration test loads it strictly and asserts forward-pass agreement:
max abs diff 1.192e-7 on the seed-42 fixture. dump_variable_names test makes
the tch name layout authoritative. Zero architecture discrepancies found.

Co-Authored-By: claude-flow <ruv@ruv.net>

* fix: workflow-review findings — BN gamma init, ThresholdParams serde, init docs

Concurrent validation workflow (2 review lanes + adversarial verification,
13 agents): 5 confirmed findings, 3 refuted. Fixes:
- wiflow_std: pin BatchNorm gamma to 1.0 (tch default draws Uniform(0,1) —
  silently halves activations in from-scratch training; loaded checkpoints
  unaffected, parity re-verified after the change)
- wiflow_std: document the conv-init divergences vs the reference's
  effective kaiming_normal(fan_out) re-init (from-scratch dynamics only)
- ieee80211bf: ThresholdParams deserialization validates via try_from so
  the <=100 invariant holds for untrusted payloads (+ rejection test)

Benchmarks (release, ruvzen): GeometryEmbedding 1.84us/call (542k/s),
MAE tokenization 7.38us/window (135k/s), 802.11bf FSM 8.9M events/s —
nothing suspicious.

Co-Authored-By: claude-flow <ruv@ruv.net>

* docs(adr): ADR-152 §2.1.4 gate resolved — PerceptAlign repo MIT, dataset on HF

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat(benchmarks): edge optimization measured + measurement (b) blocked + 92.9% retraction

Edge optimization (ADR-152 optimize track): ONNX Runtime fp32 is the CPU
latency win (3.2 ms/window, ~3.4x faster than torch, parity 2.4e-7); ORT
dynamic int8 reaches 2.44 MB (paper's ~2.2 MB claim plausible only via
conv-capable toolchains; -0.16pt PCK@20, +18% MPJPE, 2x slower); torch
dynamic quant converts 0% of this conv-only model; fp16 halves storage free
but is slower on CPU.

Measurement (b) BLOCKED-ON-DATA: only 1,077 paired ESP32 windows exist
(stop rule <2k). Forensic recheck of the surviving April holdout RETRACTS
the ADR-079 '92.9% PCK@20' figure: constant-output model, absolute (not
torso) threshold, 69 near-static frames — mean predictor scores 100% under
that protocol; torso-PCK@20 is 19.1%. Corroborates PR #535. Stale citations
removed from user-guide, readme-details, ADR-152 §2.1.3; no-citation rule
extended to ADR-079 accuracy claims. Unblock: >=2k-window multi-pose paired
session + torso-PCK re-baseline.

Co-Authored-By: claude-flow <ruv@ruv.net>

* docs(user-guide): corrected camera-supervised collection tutorial

Step 0 CSI-rate check + session-length math (window yield = frames/20 —
the May session's 8x under-delivery was a ~12 Hz CSI rate, not an aligner
bug); two-checkerboard calibration step (ADR-152 §2.1.3); pose-variety and
confidence guidance; torso-normalized PCK + temporal-split + pred-variance
eval protocol (lessons from the 92.9% retraction); scale presets re-keyed
to realistic window counts.

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat(benchmarks): static PTQ int8 (calibrated) results + overnight capture script

Conv-only static QDQ beats dynamic int8 on accuracy (PCK@20 96.61-96.63%
vs 96.52%, MPJPE +10% vs +18% over fp32) at ~equal size/latency; all-ops
QDQ strictly worse (int8 activations through attention glue). Entropy
calibration verified bit-identical to MinMax on this data. Deployment:
ONNX fp32 for speed (3.2ms), static conv-only QDQ for smallest (2.53MB).

Also: scripts/overnight-empty-capture.py — segmented UDP CSI recorder for
empty-room baselines (no glob collisions, detach-safe).

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat(benchmarks): measurement (b) MEASURED — optimization transfer only, mean-pose baseline wins

WiFlow-STD fine-tuned on 2,046 fresh single-room ESP32 paired windows
(temporal 70/15/15, 70->540 adapter, K=17): pretrained-init 65% PCK@20 vs
scratch 0% (optimization transfer) but frozen-trunk ~0% (no feature
transfer), and NOTHING beats the mean-pose baseline (95.9% PCK@20 —
single subject, near-static normalized coords). Honesty gates held: pred
std 0.0113 (non-constant model) but mean-baseline dominance means no
citable CSI->pose capability from this data. ADR-152 open question 1
answered partially; definitive answer needs multi-subject/position data.

Two new aligner findings: heterogeneous csi_shape with silent zero-padding
(~20%), and extractCsiMatrix's transposed shape label (frame-major data,
[nSc, nFrames] label) — fixes pending.

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat(benchmarks): efficiency sweep MEASURED — half model dominates full reference

Compact WiFlow-STD variants on the same data/split/protocol: half (843,834
params, 0.38x) strictly dominates the 2.23M reference (PCK@20 96.62 vs
96.61, PCK@50 99.47 vs 99.11, MPJPE 0.00898 vs 0.0094) — the published
architecture is over-parameterized for its own benchmark. quarter (338k)
96.05%; tiny (56,290 params, 1/39.5) holds 94.11% — a ~220KB fp32 edge
candidate. In-domain caveats recorded; cross-domain untested.

Co-Authored-By: claude-flow <ruv@ruv.net>

* feat(train): compact WiFlow-STD presets in Rust + tiny edge artifact (ADR-152)

WiFlowStdConfig gains half()/quarter()/tiny() mirroring the overnight sweep
exactly: TcnGroupsMode (Fixed/Gcd/Depthwise), input_pw_groups, derived
stride schedule and decoder-mid (all default to upstream behavior; legacy
serde JSON unaffected). Param formulas pin to trained ground truth first
try: 843,834 / 338,600 / 56,290; default 2,225,042 pin and 1.192e-7 parity
unchanged. 248 tests green.

Tiny edge artifact (tiny_edge_bench.py): ONNX fp32 = 295 KB, 0.66 ms/win
(~1,500/s CPU), 94.11% PCK@20 (matches sweep clean-test exactly; parity
1.49e-7). Static int8 is a bad trade at this scale (-1.43pt, +19% MPJPE,
-16% size, slower) — recorded as negative result. Export note: width-16
breaks AdaptiveAvgPool((15,1)) TorchScript export; replaced by exact
mean+matmul equivalent, proven by parity.

Co-Authored-By: claude-flow <ruv@ruv.net>

* fix: resolve all 10 confirmed code-review findings (7-angle review, 20/20 verified)

wiflow_std: min_feature_width (default 15) replaces the keypoints->stride
coupling — for_keypoints(17) now provably builds the trained [2,2,2,2]
graph and pools 15->17, matching the validated Python protocol (pinned by
tests); param_count() total on invalid configs; random_mask returns Result
and rejects non-finite/out-of-range ratios; trainer checkpoints switched
to safetensors (.pt VarStore roundtrip broken on Windows torch 2.11).

ieee80211bf: SBP proxy now re-triggers instances and relays reports via
Action::RelaySbpReport -> SensingFrame::SbpReport (clients consume via
their existing path); missed_instances reset on success = consecutive
semantics; SessionTable gains a guarded SBP entry point + unknown-id drop
counter; initiator-role sessions reject inbound setup/SBP requests
(RejectedNotSupported) closing the idle hijack; StartSetup/StartSbp
outside Idle return InvalidStateForCommand; SBP validation unified
through evaluate_setup with a 1:1 SetupStatus->SbpStatus mapping.
events.rs split out to honor the 500-line cap.

calibration/cli: enrollment geometry now actually reaches trained banks —
both production call sites attach .with_geometry; --geometry flag on
train-room and POST /enroll/geometry + train-body geometry on
calibrate-serve give production a recording surface; geometry-free banks
log the ADR-152 §2.1.2 note.

benchmarks: corruption masks committed as ground truth (unregenerable
after in-place cleaning; verified bit-identical regeneration from the
pristine copy) + generate_corruption_masks.py producer; _bench_common.py
dedups the 5x-copied shim/evaluate/seed/remap (post-refactor PCK@20
re-verified equal to the last digit); remote scripts get the mmap patch;
tiny_edge --calib validated multiple-of-64; onnx_bench --help no longer
executes (and overwrote) the export — artifact restored byte-exact.

Workspace: 2,963 tests passed, 0 failed; Python proof PASS.

Co-Authored-By: claude-flow <ruv@ruv.net>

* ci: build workspace tests without debuginfo — runner disk exhaustion

The combined 38-crate debug target exceeds the GitHub runner's disk
('final link failed: No space left on device'); the same tree measured
151GB locally with full debuginfo. CARGO_PROFILE_{DEV,TEST}_DEBUG=0
shrinks the target ~5-10x; debuginfo serves no purpose in CI test runs.

Co-Authored-By: claude-flow <ruv@ruv.net>
2026-06-11 17:02:23 -04:00

397 lines
15 KiB
Rust
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
//! Masked-autoencoder (MAE) pretraining recipe for the ADR-150 RF foundation
//! encoder — ADR-152 §2.3 (amends ADR-150 §2.3).
//!
//! Implements the *measured* tokenization recipe from the UNSW MAE pretraining
//! study (arXiv [2511.18792](https://arxiv.org/abs/2511.18792), Nov 2025), the
//! largest heterogeneous CSI pretraining run to date (1,320,892 samples, 14
//! public datasets, 4 devices, 2.4/5/6 GHz, 20160 MHz):
//!
//! - **80% masking ratio** over the patch grid.
//! - **Small (30, 3) patches** — 30 time steps × 3 subcarriers — measured
//! **+4.7%** over (40, 5) patches by preserving fine temporal dynamics.
//! - Encoder capacity stays **ViT-Small-class (~15M params)**: ViT-Base adds
//! only +0.40.9% over ViT-Small in-study, corroborating ADR-150's own
//! finding that capacity hurts cross-subject transfer.
//! - Unseen-domain performance scales **log-linearly with pretraining data,
//! unsaturated at 1.3M samples** — data aggregation outranks architecture
//! work (ADR-152 §2.3).
//!
//! This module provides the GPU-free half of the recipe: configuration,
//! patchification, and deterministic random masking. The (future, ADR-150)
//! encoder consumes [`PatchGrid`] + [`MaskIndices`] to compute the masked
//! reconstruction loss (`L_masked_csi` in ADR-150 §2.3's loss stack).
//!
//! ## Axis convention
//!
//! A CSI window is `time × subcarriers`, row-major (`index = t * subc + sc`),
//! matching the crate's `[T, …, n_sc]` dataset layout (time first, subcarriers
//! last) and the UNSW "(30 time steps, 3 subcarriers)" patch framing. Patches
//! are indexed row-major over the patch grid (`p = pt * n_patches_subc + ps`),
//! and values within a patch are row-major time-major
//! (`local = lt * patch_subc + lsc`).
//!
//! ## Divisibility policy: error, never truncate
//!
//! Window dimensions **must** be exact multiples of the patch dimensions.
//! Non-divisible shapes return [`MaeError::NotDivisible`] instead of silently
//! truncating trailing samples (this crate never silently drops data). The
//! error names the largest divisible crop; use
//! [`MaePretrainConfig::cropped_window_shape`] to compute it and crop
//! explicitly before calling [`patchify`].
//!
//! ## Example
//!
//! ```rust
//! use wifi_densepose_train::mae::MaePretrainConfig;
//!
//! let cfg = MaePretrainConfig::default(); // 0.80 masking, (30, 3) patches
//! cfg.validate().expect("default recipe is valid");
//!
//! // 90 frames × 54 subcarriers → a 3 × 18 grid of (30, 3) patches.
//! let window = vec![0.25_f32; 90 * 54];
//! let (grid, mask) = cfg.mask_window(&window, 90, 54).unwrap();
//! assert_eq!(grid.n_patches(), 54);
//! assert_eq!(mask.masked.len(), 43); // round(0.80 * 54)
//! assert_eq!(mask.visible.len(), 11);
//! ```
use serde::{Deserialize, Serialize};
use crate::error::{ConfigError, MaeError};
use crate::virtual_aug::Xorshift64;
// ---------------------------------------------------------------------------
// MaePretrainConfig
// ---------------------------------------------------------------------------
/// Hyper-parameters for masked-CSI pretraining (ADR-152 §2.3).
///
/// Defaults are the measured-optimal UNSW recipe (arXiv 2511.18792); change
/// them only with benchmark evidence. Serializable so the recipe is recorded
/// in checkpoint metadata alongside [`crate::config::TrainingConfig`].
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct MaePretrainConfig {
/// Fraction of patches hidden from the encoder, in `(0, 1)`.
///
/// Default: **0.80** (UNSW measured optimum).
pub mask_ratio: f64,
/// Patch extent along the time axis, in frames. Default: **30**.
pub patch_time: usize,
/// Patch extent along the subcarrier axis. Default: **3**.
pub patch_subc: usize,
/// Base seed for the deterministic mask sampler. Default: **42**.
///
/// For per-sample masks derive a child seed (e.g.
/// `seed ^ sample_idx as u64`) and pass it to [`random_mask`]; reusing one
/// seed yields the identical mask for every sample.
pub seed: u64,
}
impl Default for MaePretrainConfig {
fn default() -> Self {
MaePretrainConfig {
mask_ratio: 0.80,
patch_time: 30,
patch_subc: 3,
seed: 42,
}
}
}
impl MaePretrainConfig {
/// Validate the shape-independent fields.
///
/// # Validated invariants
///
/// - `mask_ratio` must be strictly inside `(0, 1)` and finite.
/// - `patch_time` and `patch_subc` must be at least 1.
pub fn validate(&self) -> Result<(), ConfigError> {
if !self.mask_ratio.is_finite() || self.mask_ratio <= 0.0 || self.mask_ratio >= 1.0 {
return Err(ConfigError::invalid_value(
"mask_ratio",
format!("must be in (0.0, 1.0), got {}", self.mask_ratio),
));
}
if self.patch_time == 0 {
return Err(ConfigError::invalid_value("patch_time", "must be >= 1"));
}
if self.patch_subc == 0 {
return Err(ConfigError::invalid_value("patch_subc", "must be >= 1"));
}
Ok(())
}
/// Check this recipe against a concrete `time × subc` window shape.
///
/// Errors if a patch dimension exceeds the window or if either axis is
/// not an exact multiple of the patch extent (divisibility policy above).
pub fn validate_for_window(&self, time: usize, subc: usize) -> Result<(), MaeError> {
check_axis("time", time, self.patch_time)?;
check_axis("subcarrier", subc, self.patch_subc)?;
Ok(())
}
/// Largest `(time, subc)` crop of the given window that is exactly
/// divisible by the patch dimensions. Either component may be 0 when the
/// window is smaller than one patch.
#[must_use]
pub fn cropped_window_shape(&self, time: usize, subc: usize) -> (usize, usize) {
(
(time / self.patch_time) * self.patch_time,
(subc / self.patch_subc) * self.patch_subc,
)
}
/// Number of patches a `time × subc` window yields under this recipe.
pub fn num_patches(&self, time: usize, subc: usize) -> Result<usize, MaeError> {
self.validate_for_window(time, subc)?;
Ok((time / self.patch_time) * (subc / self.patch_subc))
}
/// Exact number of masked patches for a grid of `n_patches`:
/// `round(mask_ratio * n_patches)`, clamped to `[0, n_patches]`.
#[must_use]
pub fn num_masked(&self, n_patches: usize) -> usize {
((self.mask_ratio * n_patches as f64).round() as usize).min(n_patches)
}
/// Patchify `window` and draw the deterministic random mask in one step,
/// using `self.seed`. See [`patchify`] and [`random_mask`].
///
/// # Errors
///
/// Everything [`patchify`] rejects, plus [`MaeError::InvalidMaskRatio`]
/// if `self.mask_ratio` is not finite or outside `(0, 1)` (the
/// [`Self::validate`] rule) — a NaN ratio must never silently mask zero
/// patches.
pub fn mask_window(
&self,
window: &[f32],
time: usize,
subc: usize,
) -> Result<(PatchGrid, MaskIndices), MaeError> {
let grid = patchify(window, time, subc, self)?;
let mask = random_mask(grid.n_patches(), self.mask_ratio, self.seed)?;
Ok((grid, mask))
}
}
// ---------------------------------------------------------------------------
// PatchGrid / MaskIndices
// ---------------------------------------------------------------------------
/// A CSI window decomposed into non-overlapping `patch_time × patch_subc`
/// patches (see the module-level axis convention).
#[derive(Debug, Clone, PartialEq)]
pub struct PatchGrid {
/// Patch extent along the time axis.
pub patch_time: usize,
/// Patch extent along the subcarrier axis.
pub patch_subc: usize,
/// Number of patch rows (`time / patch_time`).
pub n_patches_time: usize,
/// Number of patch columns (`subc / patch_subc`).
pub n_patches_subc: usize,
/// Flattened patches, row-major over the grid; each inner `Vec` is one
/// patch of length `patch_time * patch_subc`, row-major time-major.
pub patches: Vec<Vec<f32>>,
}
impl PatchGrid {
/// Total number of patches in the grid.
#[must_use]
pub fn n_patches(&self) -> usize {
self.n_patches_time * self.n_patches_subc
}
/// Number of scalar values per patch.
#[must_use]
pub fn patch_len(&self) -> usize {
self.patch_time * self.patch_subc
}
/// Window shape `(time, subc)` this grid reconstructs to.
#[must_use]
pub fn window_shape(&self) -> (usize, usize) {
(
self.n_patches_time * self.patch_time,
self.n_patches_subc * self.patch_subc,
)
}
}
/// Sorted, disjoint patch-index sets produced by [`random_mask`]. Together
/// they cover `0..n_patches` exactly.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct MaskIndices {
/// Indices of patches hidden from the encoder (`round(ratio * n)` of them).
pub masked: Vec<usize>,
/// Indices of patches the encoder sees.
pub visible: Vec<usize>,
}
// ---------------------------------------------------------------------------
// patchify / unpatchify
// ---------------------------------------------------------------------------
/// Decompose a row-major `time × subc` CSI window into the patch grid defined
/// by `cfg`.
///
/// # Errors
///
/// - [`MaeError::WindowShapeMismatch`] if `window.len() != time * subc`.
/// - [`MaeError::PatchExceedsWindow`] / [`MaeError::NotDivisible`] per the
/// module-level divisibility policy.
/// - [`MaeError::NonFiniteValue`] on the first NaN/±inf encountered —
/// corrupted CSI must be cleaned upstream, never masked over (cf. the
/// WiFlow-STD NaN-poisoning incident, ADR-152 §2.2).
pub fn patchify(
window: &[f32],
time: usize,
subc: usize,
cfg: &MaePretrainConfig,
) -> Result<PatchGrid, MaeError> {
let expected = time * subc;
if window.len() != expected {
return Err(MaeError::WindowShapeMismatch {
time,
subc,
expected,
actual: window.len(),
});
}
cfg.validate_for_window(time, subc)?;
if let Some(idx) = window.iter().position(|v| !v.is_finite()) {
return Err(MaeError::NonFiniteValue {
row: idx / subc,
col: idx % subc,
value: window[idx],
});
}
let n_patches_time = time / cfg.patch_time;
let n_patches_subc = subc / cfg.patch_subc;
let mut patches = Vec::with_capacity(n_patches_time * n_patches_subc);
for pt in 0..n_patches_time {
for ps in 0..n_patches_subc {
let mut patch = Vec::with_capacity(cfg.patch_time * cfg.patch_subc);
for lt in 0..cfg.patch_time {
let t = pt * cfg.patch_time + lt;
let row_start = t * subc + ps * cfg.patch_subc;
patch.extend_from_slice(&window[row_start..row_start + cfg.patch_subc]);
}
patches.push(patch);
}
}
Ok(PatchGrid {
patch_time: cfg.patch_time,
patch_subc: cfg.patch_subc,
n_patches_time,
n_patches_subc,
patches,
})
}
/// Reassemble the full row-major `time × subc` window from a [`PatchGrid`].
/// Exact inverse of [`patchify`].
#[must_use]
pub fn unpatchify(grid: &PatchGrid) -> Vec<f32> {
unpatchify_select(grid, None, 0.0)
}
/// Reassemble the window keeping only the patches listed in `visible`;
/// every other patch's region is filled with `fill` (the standard MAE
/// "visible tokens + mask token" view of the input).
#[must_use]
pub fn unpatchify_visible(grid: &PatchGrid, visible: &[usize], fill: f32) -> Vec<f32> {
unpatchify_select(grid, Some(visible), fill)
}
fn unpatchify_select(grid: &PatchGrid, keep: Option<&[usize]>, fill: f32) -> Vec<f32> {
let (time, subc) = grid.window_shape();
let mut window = vec![fill; time * subc];
for (p, patch) in grid.patches.iter().enumerate() {
if let Some(keep) = keep {
if !keep.contains(&p) {
continue;
}
}
let pt = p / grid.n_patches_subc;
let ps = p % grid.n_patches_subc;
for lt in 0..grid.patch_time {
let t = pt * grid.patch_time + lt;
let row_start = t * subc + ps * grid.patch_subc;
let local_start = lt * grid.patch_subc;
window[row_start..row_start + grid.patch_subc]
.copy_from_slice(&patch[local_start..local_start + grid.patch_subc]);
}
}
window
}
// ---------------------------------------------------------------------------
// random_mask
// ---------------------------------------------------------------------------
/// Draw a deterministic random mask over `n_patches` patches.
///
/// Exactly `round(mask_ratio * n_patches)` patches (clamped to
/// `[0, n_patches]`) are masked, chosen by a seeded FisherYates shuffle
/// ([`Xorshift64`]), so the same `(n_patches, mask_ratio, seed)` triple always
/// yields the same mask. Both index lists are sorted ascending, disjoint, and
/// together cover `0..n_patches`.
///
/// # Errors
///
/// [`MaeError::InvalidMaskRatio`] if `mask_ratio` is not finite or outside
/// the open interval `(0, 1)` — the same rule as
/// [`MaePretrainConfig::validate`]. Erroring (never clamping) keeps the
/// module's error-not-silent policy: a NaN ratio would otherwise silently
/// mask zero patches and a ratio ≥ 1 would mask everything.
pub fn random_mask(n_patches: usize, mask_ratio: f64, seed: u64) -> Result<MaskIndices, MaeError> {
if !mask_ratio.is_finite() || mask_ratio <= 0.0 || mask_ratio >= 1.0 {
return Err(MaeError::InvalidMaskRatio { ratio: mask_ratio });
}
let n_masked = ((mask_ratio * n_patches as f64).round() as usize).min(n_patches);
let mut order: Vec<usize> = (0..n_patches).collect();
let mut rng = Xorshift64::new(seed);
for i in (1..n_patches).rev() {
let j = (rng.next_u64() % (i as u64 + 1)) as usize;
order.swap(i, j);
}
let mut masked: Vec<usize> = order[..n_masked].to_vec();
let mut visible: Vec<usize> = order[n_masked..].to_vec();
masked.sort_unstable();
visible.sort_unstable();
Ok(MaskIndices { masked, visible })
}
// ---------------------------------------------------------------------------
// helpers
// ---------------------------------------------------------------------------
fn check_axis(axis: &'static str, window: usize, patch: usize) -> Result<(), MaeError> {
if patch > window {
return Err(MaeError::PatchExceedsWindow {
axis,
patch,
window,
});
}
let remainder = window % patch;
if remainder != 0 {
return Err(MaeError::NotDivisible {
axis,
window,
patch,
remainder,
crop: window - remainder,
});
}
Ok(())
}