feat: End-to-end training pipeline with RuVector signal intelligence #49

2026-03-01T12:59:37+08:00

ruvnet commented

2026-03-01 12:59:37 +08:00

(Migrated from github.com)

Summary

Wire up end-to-end training pipeline (Issue #44): Connect graph transformer, trainer, dataset, and sparse inference into a working --train → --model flow
Weight serialization: flatten_weights()/unflatten_weights() on CsiToPoseTransformer for checkpoint and RVF save/load roundtrip
Trainer↔Transformer bridge: from_dataset_sample() conversion, Trainer::with_transformer() constructor delegates predict_keypoints to CsiToPoseTransformer::forward()
Real INT8 quantization: apply_quantization() now quantizes per-layer weights and uses dequantize-on-the-fly during forward pass
CLI training mode: --train --dataset PATH --epochs N --save-rvf model.rvf runs full training, saves best-epoch weights to RVF container
Best-epoch snapshotting: Trainer saves best validation loss weights and restores them before checkpoint/export
Performance fix: CsiToPoseTransformer::zeros() skips Xavier init for gradient estimation (avoids ~44K wasted constructions per batch)
ADR-021: Vital sign detection + RVF container format (closes #45)
ADR-022: Windows WiFi enhanced fidelity with multi-BSSID pipeline
ADR-023: Full DensePose training pipeline (Phases 1-8)
Docker: Reorganized into docker/ directory with multi-stage Rust and Python images

New Crates

Crate	Lines	Description
`wifi-densepose-vitals`	1,863	Breathing/HR detection, anomaly scoring, RVF storage
`wifi-densepose-wifiscan`	4,829	Multi-BSSID scanning, netsh/WlanAPI adapters, 8-stage pipeline

New Modules (sensing-server)

Module	Lines	Description
`graph_transformer.rs`	855	Cross-attention + GNN transformer with weight serialization
`trainer.rs`	881	6-term loss, SGD+momentum, cosine LR, early stopping, checkpoints
`dataset.rs`	850	MM-Fi + Wi-Pose loaders, windowing, normalization
`sparse_inference.rs`	753	Neuron profiling, sparse forward, INT8/FP16 quantization
`sona.rs`	639	LoRA adapters, EWC++ memory, environment profiles
`rvf_container.rs`	914	Binary container format (manifest, weights, SONA, vitals)
`rvf_pipeline.rs`	1,027	Progressive loading, HNSW index, overlay graph
`vital_signs.rs`	774	Bandpass + FFT breathing/HR detection

Test Results

239 tests passing (157 lib + 48 RVF + 16 container + 18 vitals)
Zero compilation errors
10 new integration tests for weight roundtrip, trainer+transformer, quantization accuracy

Docker

No Dockerfile changes required — the existing docker/Dockerfile.rust builds the same binary with new CLI flags automatically. Docker Hub images can be rebuilt at next release.

Known Limitations (noted by review, not blocking)

Asymmetric quantization zero_point stored as i8 (pre-existing, only affects Int8Asymmetric mode which is not the default)
Body part / UV loss terms use placeholder logic when transformer is active (scheduled for Phase 9)
graph_edge_loss not computed in training loop (loss weight has no effect)

Test plan

cargo check -p wifi-densepose-sensing-server — compiles cleanly
cargo test -p wifi-densepose-sensing-server — 239 tests pass
Weight roundtrip: flatten_weights() → unflatten_weights() produces identical forward output
Quantized forward within 5% MSE of f32 forward
Trainer with transformer produces finite loss across epochs
Best-epoch params restored after run_training()
cargo run -- --train --epochs 5 --save-rvf test.rvf (manual)
cargo run -- --model test.rvf --source simulate (manual)

🤖 Generated with claude-flow

## Summary - **Wire up end-to-end training pipeline** (Issue #44): Connect graph transformer, trainer, dataset, and sparse inference into a working `--train` → `--model` flow - **Weight serialization**: `flatten_weights()`/`unflatten_weights()` on `CsiToPoseTransformer` for checkpoint and RVF save/load roundtrip - **Trainer↔Transformer bridge**: `from_dataset_sample()` conversion, `Trainer::with_transformer()` constructor delegates `predict_keypoints` to `CsiToPoseTransformer::forward()` - **Real INT8 quantization**: `apply_quantization()` now quantizes per-layer weights and uses dequantize-on-the-fly during forward pass - **CLI training mode**: `--train --dataset PATH --epochs N --save-rvf model.rvf` runs full training, saves best-epoch weights to RVF container - **Best-epoch snapshotting**: Trainer saves best validation loss weights and restores them before checkpoint/export - **Performance fix**: `CsiToPoseTransformer::zeros()` skips Xavier init for gradient estimation (avoids ~44K wasted constructions per batch) - **ADR-021**: Vital sign detection + RVF container format (closes #45) - **ADR-022**: Windows WiFi enhanced fidelity with multi-BSSID pipeline - **ADR-023**: Full DensePose training pipeline (Phases 1-8) - **Docker**: Reorganized into `docker/` directory with multi-stage Rust and Python images ## New Crates | Crate | Lines | Description | |-------|-------|-------------| | `wifi-densepose-vitals` | 1,863 | Breathing/HR detection, anomaly scoring, RVF storage | | `wifi-densepose-wifiscan` | 4,829 | Multi-BSSID scanning, netsh/WlanAPI adapters, 8-stage pipeline | ## New Modules (sensing-server) | Module | Lines | Description | |--------|-------|-------------| | `graph_transformer.rs` | 855 | Cross-attention + GNN transformer with weight serialization | | `trainer.rs` | 881 | 6-term loss, SGD+momentum, cosine LR, early stopping, checkpoints | | `dataset.rs` | 850 | MM-Fi + Wi-Pose loaders, windowing, normalization | | `sparse_inference.rs` | 753 | Neuron profiling, sparse forward, INT8/FP16 quantization | | `sona.rs` | 639 | LoRA adapters, EWC++ memory, environment profiles | | `rvf_container.rs` | 914 | Binary container format (manifest, weights, SONA, vitals) | | `rvf_pipeline.rs` | 1,027 | Progressive loading, HNSW index, overlay graph | | `vital_signs.rs` | 774 | Bandpass + FFT breathing/HR detection | ## Test Results - **239 tests passing** (157 lib + 48 RVF + 16 container + 18 vitals) - Zero compilation errors - 10 new integration tests for weight roundtrip, trainer+transformer, quantization accuracy ## Docker No Dockerfile changes required — the existing `docker/Dockerfile.rust` builds the same binary with new CLI flags automatically. Docker Hub images can be rebuilt at next release. ## Known Limitations (noted by review, not blocking) - Asymmetric quantization `zero_point` stored as `i8` (pre-existing, only affects `Int8Asymmetric` mode which is not the default) - Body part / UV loss terms use placeholder logic when transformer is active (scheduled for Phase 9) - `graph_edge_loss` not computed in training loop (loss weight has no effect) ## Test plan - [x] `cargo check -p wifi-densepose-sensing-server` — compiles cleanly - [x] `cargo test -p wifi-densepose-sensing-server` — 239 tests pass - [x] Weight roundtrip: `flatten_weights()` → `unflatten_weights()` produces identical forward output - [x] Quantized forward within 5% MSE of f32 forward - [x] Trainer with transformer produces finite loss across epochs - [x] Best-epoch params restored after `run_training()` - [ ] `cargo run -- --train --epochs 5 --save-rvf test.rvf` (manual) - [ ] `cargo run -- --model test.rvf --source simulate` (manual) 🤖 Generated with [claude-flow](https://github.com/ruvnet/claude-flow)

github-advanced-security[bot] (Migrated from github.com) reviewed 2026-03-01 13:00:54 +08:00

docker/Dockerfile.python

						
				@@ -0,0 +26,4 @@

				ENV PYTHONUNBUFFERED=1

				CMD ["python", "-m", "v1.src.sensing.ws_server"]

github-advanced-security[bot] (Migrated from github.com) commented

@@ -0,0 +43,4 @@
 ENV RUST_LOG=info
 ENTRYPOINT ["/app/sensing-server"]
 CMD ["--source", "simulated", "--tick-ms", "100", "--ui-path", "/app/ui"]

@@ -0,0 +1,26 @@
 version: "3.9"
 services:
   sensing-server:

@@ -0,0 +6,4 @@
       context: ..
       dockerfile: docker/Dockerfile.rust
     image: ruvnet/wifi-densepose:latest
     ports:

@@ -0,0 +14,4 @@
       - RUST_LOG=info
     command: ["--source", "simulated", "--tick-ms", "100", "--ui-path", "/app/ui"]
   python-sensing:

@@ -0,0 +19,4 @@
       context: ..
       dockerfile: docker/Dockerfile.python
     image: ruvnet/wifi-densepose:python
     ports:

feat: End-to-end training pipeline with RuVector signal intelligence #49

Summary

New Crates

New Modules (sensing-server)

Test Results

Docker

Known Limitations (noted by review, not blocking)

Test plan

Ensure that HEALTHCHECK instructions have been added to container images

Ensure that a user for the container has been created

Ensure that HEALTHCHECK instructions have been added to container images

Ensure that a user for the container has been created

Container Capabilities Unrestricted

Healthcheck Not Set

Memory Not Limited

Security Opt Not Set

Cpus Not Limited

Container Traffic Not Bound To Host Interface