c9fde3cba5
ADR-096 train integration. Additive — does NOT modify model.rs. The
existing WiFiDensePoseModel forward stays bit-equivalent for back-compat.
New code lives in temporal_aether.rs behind the `aether-sparse-temporal`
feature flag (which itself requires `tch-backend`).
Architecture:
tch::Tensor [T, in_dim] ──── tch nn::Linear (q/k/v projections)
↓
[T, q_heads*head_dim] etc
↓
tch_to_tensor3 (CPU, f32, 1× copy)
↓
ruvllm_sparse_attention::Tensor3
↓
AetherTemporalHead::forward()
↓
Tensor3 [T, q_heads, head_dim]
↓
tensor3_to_tch (1× copy)
↓
tch::Tensor [T, q_heads*head_dim]
↓
tch nn::Linear (output projection)
↓
tch::Tensor [T, in_dim]
Why additive rather than swapping `apply_antenna_attention` /
`apply_spatial_attention` in model.rs: those are over antenna and
spatial axes, not temporal — ADR-096 §8.1 was right that AETHER
doesn't currently HAVE a temporal-axis attention. This commit adds
that path without disturbing the others, so the §5 validation gate
can A/B the two options before flipping the production default.
Scope notes:
- B=1 prefill only this version. Multi-batch lands when §5 turns
green and we need to take perf seriously. The forward expects
`[T, in_dim]` not `[B, T, in_dim]`; documented in the file.
- Streaming step() bridge deferred — KvCache lifecycle ties to
PoseTrack per ADR-096 §8.5, which is signal-side not train-side.
- Two CPU memory copies per call (in + out). For training-rate
forwards (~100/sec at batch 16) this is negligible vs the actual
attention work; for inference-rate streaming it'd be the
bottleneck and a zero-copy path is the natural follow-up.
Build verification:
- Source compiles cleanly with cargo check on the host crate
(`-p wifi-densepose-temporal`, 21/21 tests still passing).
- The train crate's tch-backend build is environmentally blocked
on this Windows machine — torch-sys fails to link against the
system PyTorch 2.11 + MSVC 14.50 toolchain. This predates this
commit and affects all tch-bound code paths in the workspace.
CI runners with working libtorch will verify the new module
builds; the source follows the same nn::Linear / Module patterns
the existing model.rs uses.
Feature gating ensures default builds are byte-equivalent. Off by
default; enable with `--features aether-sparse-temporal`.
Co-Authored-By: claude-flow <ruv@ruv.net>
|
||
|---|---|---|
| .. | ||
| benches | ||
| src | ||
| tests | ||
| Cargo.toml | ||
| README.md | ||
README.md
wifi-densepose-train
Complete training pipeline for WiFi-DensePose, integrated with all five ruvector crates.
Overview
wifi-densepose-train provides everything needed to train the WiFi-to-DensePose model: dataset
loading, subcarrier interpolation, loss functions, evaluation metrics, and the training loop
orchestrator. It supports both the MM-Fi dataset (NeurIPS 2023) and deterministic synthetic data
for reproducible experiments.
Without the tch-backend feature the crate still provides the dataset, configuration, and
subcarrier interpolation APIs needed for data preprocessing and proof verification.
Features
- MM-Fi dataset loader -- Reads the MM-Fi multimodal dataset (NeurIPS 2023) from disk with
memory-mapped
.npyfiles. - Synthetic dataset -- Deterministic, fixed-seed CSI generation for unit tests and proofs.
- Subcarrier interpolation -- 114 -> 56 subcarrier compression via
ruvector-solversparse interpolation with variance-based selection. - Loss functions (
tch-backend) -- Pose estimation losses including MSE, OKS, and combined multi-task loss. - Metrics (
tch-backend) -- PCKh, OKS-AP, and per-keypoint evaluation withruvector-mincut-based person matching. - Training orchestrator (
tch-backend) -- Full training loop with learning rate scheduling, gradient clipping, checkpointing, and reproducible proofs. - All 5 ruvector crates --
ruvector-mincut,ruvector-attn-mincut,ruvector-temporal-tensor,ruvector-solver, andruvector-attentionintegrated across dataset loading, metrics, and model attention.
Feature flags
| Flag | Default | Description |
|---|---|---|
tch-backend |
no | Enable PyTorch training via tch-rs |
cuda |
no | CUDA GPU acceleration (implies tch) |
Binaries
| Binary | Description |
|---|---|
train |
Main training entry point |
verify-training |
Proof verification (requires tch-backend) |
Quick Start
use wifi_densepose_train::config::TrainingConfig;
use wifi_densepose_train::dataset::{SyntheticCsiDataset, SyntheticConfig, CsiDataset};
// Build and validate config
let config = TrainingConfig::default();
config.validate().expect("config is valid");
// Create a synthetic dataset (deterministic, fixed-seed)
let syn_cfg = SyntheticConfig::default();
let dataset = SyntheticCsiDataset::new(200, syn_cfg);
// Load one sample
let sample = dataset.get(0).unwrap();
println!("amplitude shape: {:?}", sample.amplitude.shape());
Architecture
wifi-densepose-train/src/
lib.rs -- Re-exports, VERSION
config.rs -- TrainingConfig, hyperparameters, validation
dataset.rs -- CsiDataset trait, MmFiDataset, SyntheticCsiDataset, DataLoader
error.rs -- TrainError, ConfigError, DatasetError, SubcarrierError
subcarrier.rs -- interpolate_subcarriers (114->56), variance-based selection
losses.rs -- (tch) MSE, OKS, multi-task loss [feature-gated]
metrics.rs -- (tch) PCKh, OKS-AP, person matching [feature-gated]
model.rs -- (tch) Model definition with attention [feature-gated]
proof.rs -- (tch) Deterministic training proofs [feature-gated]
trainer.rs -- (tch) Training loop orchestrator [feature-gated]
Related Crates
| Crate | Role |
|---|---|
wifi-densepose-signal |
Signal preprocessing consumed by dataset loaders |
wifi-densepose-nn |
Inference engine that loads trained models |
ruvector-mincut |
Person matching in metrics |
ruvector-attn-mincut |
Attention-weighted graph cuts |
ruvector-temporal-tensor |
Compressed CSI buffering in datasets |
ruvector-solver |
Sparse subcarrier interpolation |
ruvector-attention |
Spatial attention in model |
License
MIT OR Apache-2.0