mirror of
https://github.com/ruvnet/RuView
synced 2026-07-22 17:23:19 +00:00
82d5c73396f10f057a1559bb8609111f04f84f66
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 ruv
|
65da488add |
feat(adr-185): add weight-loading capability to AETHER EmbeddingExtractor (§13.a)
Closes the ADR-185 §13.a follow-up (the tractable, data-independent one of
the three §6.7 gaps): the bound EmbeddingExtractor was random-Xavier-init
only, with NO way to load real weights — so it was structurally untrained.
This adds the *capability* to load weights whenever a trained checkpoint
exists. It does NOT itself produce trained/SOTA embeddings and does NOT
close §6.7 (still no trained checkpoint, no labeled data, no eval harness).
Serialization: not greenfield — EmbeddingExtractor already had
flatten_weights()/unflatten_weights() (flat Vec<f32>). wifi-densepose-aether
is a deliberately dependency-free std-only leaf crate (ADR-185 §13), so
rather than add safetensors/serde/bincode (which would undo the zero-dep
property), the on-disk format is raw little-endian f32 with a 12-byte header
(magic "AETHERW1" + u32 param count) — zero new deps.
Native (v2/crates/wifi-densepose-aether/src/embedding.rs):
- EmbeddingExtractor::save_weights(path) / load_weights(path). load_weights
never panics: errors on unreadable file, short/oversized payload, bad
magic, or a param-count mismatch (delegated to unflatten_weights).
- Default (no weights) construction is UNCHANGED — still random init,
clearly labeled untrained. Purely additive.
Python binding (python/src/bindings/aether.rs, .pyi):
- EmbeddingExtractor.load_weights(path) / save_weights(path) / param_count,
GIL-released, ValueError on bad input.
Tests (both Rust and Python, per the task):
- Rust unit (embedding.rs): load_weights_actually_replaces_weights_and_
round_trips proves the loaded weights MOVE the embedding away from the
random-init baseline (not a silent no-op), match the source extractor
(round-trip), and are bit-identical after the file round-trip; plus a
bad-magic/wrong-count rejection test.
- Cross-language golden (aether_weights_parity.rs + test_aether.py): a
shared deterministic weight formula (w[i]=k/65536-0.5, exact in f32+f64)
written to the AETHER format; both native Rust and the Python binding load
it and must produce the byte-identical embedding SHA-256
(tests/golden/aether_loaded_embedding.sha256) — proving the binding's
load path is bit-identical to native, and (vs the random baseline) that
the loaded weights are actually used.
Verified:
cargo test -p wifi-densepose-aether 98 passed, 0 failed
cargo test --features aether --test aether_parity
--test aether_weights_parity 3 passed, 0 failed
maturin develop --features aether + pytest test_aether.py 13/13 pass
default cargo build (no aether feature) clean
|
||
|
ruv
|
1c9727f9cf |
feat(adr-185): P3 MAT bindings (wifi_densepose.mat) + parity harness
Bind the ADR-024 MAT (Mass Casualty Assessment Tool) disaster-survivor detection + START triage surface into the wheel behind a gated [mat] extra / Cargo `mat` feature, mirroring the upstream disaster/ML gating. Also adds the [sota] superset extra (aether+meridian+mat). Surface (bound against the REAL code at HEAD, not the ADR wishlist): - DisasterType (9 variants) / TriageStatus (5, START) enums - DisasterConfig (builder-backed, continuous_monitoring forced off) - DisasterResponse: initialize_event / add_zone / push_csi_data / scan_once / survivors / survivors_by_triage - Survivor (id, triage_status, confidence, location, latest_vitals) - VitalSignsReading (breathing/heartbeat rate, movement, confidence) - ScanZone.rectangle / ScanZone.circle push_csi_data + scan_once are GIL-released. Honest deviations from ADR section 3.4 (documented in module header): - ADR proposed adding a Rust-side sync scan_once() (section 11.3). That was UNNECESSARY: the public async start_scanning() runs exactly one scan_cycle and returns when continuous_monitoring == false. The binding forces that flag off and drives one cycle on a private current-thread tokio runtime -- NO change to wifi-densepose-mat. - scan_cycle requires an active event + Active zone, which the ADR surface omitted; initialize_event + add_zone are bound as required additions. - Survivor.vital_signs is a *history* in the real code; bound as Survivor.latest_vitals -> Optional[VitalSignsReading]. - DisasterType has 9 variants at HEAD (adds Landslide/MineCollapse/ Industrial/TunnelCollapse); all bound. Parity (section 4.1, release-blocking): committed fixture mat_input.json (synthetic breathing-modulated CSI stream) -> native Rust reference (tests/mat_parity.rs, drives DisasterResponse directly) locks tests/golden/mat_result.sha256 over a canonical `count=<K>;triage_priorities=<sorted>` string (survivor UUIDs/timestamps excluded as non-deterministic); pytest (tests/test_mat.py) runs the same stream through the binding and asserts the identical hash. Both detect exactly 1 survivor, triage Delayed. Honest: synthetic fixture proves binding==native path equality, NOT live detection accuracy. Verified: cargo test --features mat --test mat_parity -> 2/2 pass maturin develop --features mat + pytest tests/test_mat.py -> 7/7 pass default cargo build clean, 0 mat/tokio refs in the default dep graph. WHEEL-SIZE FINDING (ADR-185 section 9): default-features=false drops MAT's `api` (axum) and `ruvector` features, but MAT still carries NON-optional tokio (rt/sync/time), wifi-densepose-nn (ort/ONNX + reqwest/hyper), rustfft, geo, ndarray. So a [mat] wheel exceeds the ADR-117 section 5.4 <=5 MB budget -- same leaf-crate-hoist follow-up as AETHER/MERIDIAN. The default wheel is untouched (feature-gated). |
||
|
ruv
|
189ac9dfb0 |
feat(adr-185): P2 MERIDIAN bindings (wifi_densepose.meridian) + parity harness
Bind the ADR-027 MERIDIAN cross-environment domain-generalization surface
into the wheel behind a gated [meridian] extra / Cargo `meridian` feature.
Inference/adaptation path only (tch-free), per ADR-185 section 3.3.
Surface (bound against the REAL code at HEAD, not the ADR wishlist):
- HardwareType / HardwareNormalizer / CanonicalCsiFrame (from
wifi-densepose-signal::hardware_norm)
- MeridianGeometryConfig / GeometryEncoder (64-dim, permutation-invariant)
- RapidAdaptation / AdaptationResult (push_frame + adapt, LoRA deltas)
- CrossDomainEvaluator + mpjpe (from wifi-densepose-train, NO tch-backend)
All compute paths GIL-released (py.allow_threads).
Honest deviations from ADR section 3.3 (documented in the module header):
- ADR's RapidAdaptation.calibrate(windows) and AdaptationResult.converged
DO NOT EXIST. Real API is push_frame + adapt(); result carries
{lora_weights, final_loss, frames_used, adaptation_epochs}. Bound as-is.
- ADR's HardwareType.detect exposed as a staticmethod delegating to the
real HardwareNormalizer::detect_hardware.
- ADR's normalize(frame: CsiFrame, hw) is really normalize(amplitude,
phase, hw) over f64 vectors returning Result; bound faithfully.
- CanonicalCsiFrame fields are singular amplitude/phase (ADR said plural).
Training-time types (DomainFactorizer, GradientReversalLayer,
VirtualDomainAugmentor) are out of P6 scope (need the libtorch tier).
Parity (section 4.1, release-blocking): committed fixture
meridian_input.json -> native Rust reference (tests/meridian_parity.rs,
calls hardware_norm + geometry + rapid_adapt directly) locks
tests/golden/meridian_output.sha256 over the concatenated f32 outputs
(esp32+intel canonical frames, 64-dim geometry vector, rapid-adapt LoRA
weights); pytest (tests/test_meridian.py) runs the same fixture through
the binding and asserts the identical SHA-256. Both pass.
Verified:
cargo test --features meridian --test meridian_parity -> 2/2 pass
maturin develop --features meridian + pytest tests/test_meridian.py
-> 13/13 pass
default cargo build clean, 0 train/signal/sensing-server refs in the
default dep graph (gate keeps the base wheel lean).
WHEEL-SIZE FINDING (ADR-185 section 9 / section 1.2): the libtorch risk
the ADR feared is AVOIDED -- wifi-densepose-train's `tch` dep is properly
optional (feature tch-backend, OFF), so no libtorch links. BUT train
still carries NON-optional deps: tokio (rt subset), the five ruvector-*
crates, and wifi-densepose-nn (which itself pulls `ort` / ONNX Runtime +
reqwest/hyper). So a [meridian] wheel exceeds the ADR-117 section 5.4
<=5 MB budget (though lighter than AETHER's axum/tokio server tree). The
clean fix is the same leaf-crate hoist: move the pure inference modules
(geometry, rapid_adapt, eval, hardware_norm) into a tch/tokio/ort-free
leaf crate. A required pre-release follow-up, not a functional blocker;
P2 binds real code and proves parity today.
|
||
|
ruv
|
d060998e3b |
feat(adr-185): P1 AETHER bindings (wifi_densepose.aether) + parity harness
Bind the ADR-024 contrastive CSI embedding surface into the wheel behind
a gated [aether] extra / Cargo `aether` feature, per ADR-185 section 3.2.
Surface (bound against the REAL code at HEAD, not the ADR wishlist):
- AetherConfig -> EmbeddingConfig {d_model,d_proj,temperature,normalize}
- CsiAugmenter -> augment_pair(window, seed)
- EmbeddingExtractor -> embed(csi): 128-dim L2-normed, GIL-released
- info_nce_loss, cosine_similarity (module fns)
ADR-185 section 3.2 also names aether_loss/VICReg components,
alignment_metric, uniformity_metric, forward_dual, and vicreg_* config
fields. None exist in embedding.rs at HEAD, so they are intentionally NOT
bound (a Rust-side gap, not a binding gap) rather than fabricated.
Parity (section 4.1, release-blocking): committed fixture
aether_input.json -> native Rust reference (tests/aether_parity.rs, calls
sensing-server EmbeddingExtractor directly) locks
tests/golden/aether_embedding.sha256; pytest (tests/test_aether.py) runs
the same fixture through the binding and asserts the identical SHA-256 of
the LE f32 bytes. Both pass.
Verified: cargo test --features aether --test aether_parity -> 2/2 pass;
maturin develop --features aether + pytest tests/test_aether.py -> 9/9
pass; default cargo build clean with 0 sensing-server refs in the dep
graph (gate keeps the base wheel lean).
HONEST WHEEL-SIZE FINDING (ADR-185 section 9 / Open Q 11.1, unresolved):
wifi-densepose-sensing-server is an Axum/tokio crate whose tokio/axum/
worldgraph/ruvector deps are NON-optional (only mqtt/matter are
features), so default-features=false does NOT drop them. An [aether]
wheel therefore links the full server tree and BREAKS the ADR-117 section
5.4 <=5 MB budget. The fix is the ADR-sanctioned hoist of embedding.rs
(+ its graph_transformer/sona siblings) into a leaf crate so the wheel
links pure compute only -- a change INSIDE wifi-densepose-sensing-server,
owned by another agent this session, so deferred as a required
pre-release follow-up. P1 binds real code and proves parity today; the
hoist is a wheel-size optimization, not a functional blocker.
Note: building required initializing the worldgraph/rufield/ruvector
submodules (absent in the fresh worktree).
|