Closes the prototype's "iter 3 = plan + wiring documented" item (ADR-027 §2.0):
- scripts/pretrain-mae-gcloud.sh — GCloud GPU driver for the MAE pre-train: a
thin, reviewable mirror of scripts/gcloud-train.sh that provisions a VM in
cognitum-20260110, builds wifi-densepose-train --features tch-backend,cuda,
runs the `pretrain-mae` binary, downloads the .ot variable store, tears the
VM down. Currently drives SyntheticCsiDataset (the smoke path); the one TODO
is the --data-dir/--datasets plumbing for the real heterogeneous corpus.
NOT run as part of this prototype. Also supports --dry-run (local synthetic
pre-train, needs LibTorch).
- ADR-027 §2.0 — added the "Iteration 3 plan" subsection: heterogeneous-CSI
ingest (own recordings + MM-Fi + Wi-Pose + multi-band virtual sub-carriers,
normalised to 56 sub-carriers), the GPU run, lifting the v0 limits
(per-sample masking, transformer blocks, circular phase loss), the fine-tune
handoff (load the CsiMae encoder into WiFiDensePoseModel via a
`--init-encoder <mae.ot>` flag, then train the §2.x heads as regularisers),
cross-domain eval (§4.6 protocol), and shipping the encoder as an RVF segment.
- wifi-densepose-train/README.md — new "MERIDIAN-MAE" section pointing at the
csi_mae module, the pretrain-mae binary, the gcloud script, and ADR-027 §2.0.
- csi_mae.rs module doc — updated the iteration-status block.
cargo test -p wifi-densepose-train --no-default-features → 121 lib tests pass.
This completes the MERIDIAN CSI-MAE *prototype* (iter 1 masking pipeline +
iter 2 tch model/pretrain loop/bin + iter 3 plan/wiring). Real cross-domain
results need the heterogeneous ingest + a GPU pre-train run (iter 3 execution),
out of scope for the prototype.
Co-Authored-By: claude-flow <ruv@ruv.net>
Adds §2.0 — the primary MERIDIAN path is now a three-stage pipeline:
1. pre-train a CIG-MAE-style dual-stream (amplitude+phase) masked autoencoder
on heterogeneous CSI (data breadth > pose-net capacity — arXiv:2511.18792);
2. fine-tune the existing §2.1–§2.6 heads (17-kpt/DensePose, AETHER, domain-
adversarial, geometry-conditioned) on top of the pre-trained encoder;
3. adapt per-room with source-free unsupervised domain adaptation behind
coherence_gate.rs::Recalibrate (separate ADR).
§2.1+ is retained but re-framed as the fine-tune-stage head, not a from-scratch
design. Adds the supporting references (2511.18792, 2512.04723, 2605.01369,
2506.12052, ACM TOSN 10.1145/3715130) and points at the 2026-Q2 SOTA survey.
Co-Authored-By: claude-flow <ruv@ruv.net>
Deep SOTA research into WiFi sensing domain gap problem (2024-2026).
Proposes 7-phase implementation: hardware normalization, domain-adversarial
training with gradient reversal, geometry-conditioned FiLM inference,
virtual environment augmentation, few-shot rapid adaptation, and
cross-domain evaluation protocol.
Cites 10 papers: PerceptAlign, AdaPose, Person-in-WiFi 3D (CVPR 2024),
DGSense, CAPC, X-Fi (ICLR 2025), AM-FM, LatentCSI, Ganin GRL, FiLM.
Addresses the single biggest deployment blocker: models trained in one
room lose 40-70% accuracy in another room. MERIDIAN adds ~12K params
(67K total, still fits ESP32) for cross-layout + cross-hardware
generalization with zero-shot and few-shot adaptation paths.
Co-Authored-By: claude-flow <ruv@ruv.net>
ruv