wifi-densepose/benchmarks/wiflow-std/remote/measb/train_measb.py

375 lines
17 KiB
Python

"""ADR-152 SS2.2 measurement (b): WiFlow-STD fine-tuned on our fresh ESP32 paired dataset.
Dataset: ~/wiflow-std-bench/paired-20260610.jsonl -- 2,046 paired windows collected
2026-06-10 22:10-22:40 (ONE subject, ONE room, ONE ESP32 node, varied poses).
Per record: csi = flat float32 list, csi_shape, kp = 17 COCO [x, y] normalized [0,1]
camera coords, conf (MediaPipe mean confidence, all > 0.5 in this set), ts_start/ts_end.
Aligner: scripts/align-ground-truth.js, non-overlapping 20-frame windows (~0.42 s each).
Dataset findings (MEASURED on this file, 2026-06-10):
- csi_shape is HETEROGENEOUS, not uniformly [70, 20]: 1,347x [70,20], 284x [134,20],
243x [26,20], 130x [12,20], 42x [20,20]. The ESP32 stream emits mixed frame types
and the aligner stamps each window's subcarrier count from frame[0]
(extractCsiMatrix: nSc = window[0].subcarriers), zero-padding/truncating the rest.
Even native-70 windows contain ~20.4% internally zero-padded short frames
(subcarriers 40..69 all-zero for those frames).
- LAYOUT BUG: the aligner fills matrix[f * nSc + s] (frame-major) but declares
shape [nSc, nFrames]. The true layout is (frame, subcarrier); we reshape
(nFrames, nSc) and transpose. Confirmed by coherent per-frame zero-tails.
- Handling here (primary suite, "all2046"): every frame's subcarrier axis is
linearly resampled to 70 bins (np.interp over a normalized index domain;
identity for native-70 frames) so the pre-registered n=2,046 and split sizes
hold. Secondary suite ("native70") restricts to the 1,347 native [70,20]
windows (temporal 70/15/15 of those) as a homogeneity robustness check.
Pre-registered protocol (followed exactly):
1. TEMPORAL split (records are time-sorted; asserted): first 70% train (1,432),
next 15% val (307), last 15% test (307). No shuffling across time. Seed 42
for everything else.
2. Model: upstream WiFlow-STD trunk (WiFlowPoseModel) with a learned 1x1 Conv1d
projection 70->540 prepended, and K=17 via the parameter-free adaptive pool
(AdaptiveAvgPool2d((17, 1)) instead of (15, 1)) -- pretrained weights load
for any K. CSI normalization: divide by the TRAIN-split 99th-percentile
amplitude, clip to [0, 1] (documented in output JSON).
3. Three runs, <=60 epochs, early-stop patience 8 on val MPJPE, batch 32,
AdamW, fp32 (no autocast):
(i) pretrained-init: trunk init from upstream/test/best_pose_model.pth
(the measurement-(a) retrained checkpoint, ~96% PCK@20 on WiFlow data;
key remap att.->attention. / final_conv.->decoder. applied defensively
as in eval_repro.py -- a no-op for this checkpoint, which already uses
the new names). Discriminative lr: adapter 1e-4, trunk 1e-5.
(ii) scratch: same architecture, random init, all params lr 1e-4.
(iii) frozen-trunk: pretrained trunk frozen (requires_grad=False AND held in
.eval() so BatchNorm running stats cannot drift -- pure transfer probe);
only the 70->540 adapter trains, lr 1e-4.
4. Metrics on the temporal TEST split: torso-normalized PCK@10/20/30/40/50 and
MPJPE. Upstream utils/metrics.py calculate_pck(use_torso_norm=True) hardcodes
NECK_IDX/PELVIS_IDX = 2, 12 -- a 15-keypoint convention that is WRONG for our
17 COCO keypoints (2 = right_eye, 12 = right_hip). We therefore reimplement the
identical math (per-frame norm distance, clamp min 0.01, mean over all
keypoints x frames) with torso = ||l_shoulder(5) - l_hip(11)||.
Also reported: prediction std across test frames (constant-pose detector;
must be > 0) and the mean-pose-predictor baseline (train-split mean pose
evaluated on test -- the honesty bar).
Usage (on ruvultra):
nice -n 10 nohup ~/wiflow-std-bench/venv/bin/python train_measb.py > train_measb.log 2>&1 &
NOTE: deployed to ruvultra as a standalone single file, so it deliberately
inlines its helpers. The reference implementations (upstream import shim,
np.load mmap patch, key-remap loader, canonical evaluate loop) live in
benchmarks/wiflow-std/_bench_common.py — keep copies in sync.
"""
import json
import os
import random
import sys
import time
import numpy as np
import torch
import torch.nn as nn
BENCH = os.path.expanduser("~/wiflow-std-bench")
UPSTREAM = os.path.join(BENCH, "upstream")
MEASB = os.path.join(BENCH, "measb")
DATA = os.path.join(BENCH, "paired-20260610.jsonl")
CHECKPOINT = os.path.join(UPSTREAM, "test", "best_pose_model.pth")
sys.path.insert(0, UPSTREAM)
# Upstream defect (1): models/__init__.py imports a name tcn.py does not define.
# Register a stub package so the broken __init__ never executes (as eval_repro.py).
import types # noqa: E402
_models_pkg = types.ModuleType("models")
_models_pkg.__path__ = [os.path.join(UPSTREAM, "models")]
sys.modules["models"] = _models_pkg
from models.pose_model import WiFlowPoseModel # noqa: E402
SEED = 42
K = 17
N_SUBC = 70
TRUNK_IN = 540
BATCH = 32 # <= 64 per protocol (GPU shared with the efficiency sweep)
MAX_EPOCHS = 60
PATIENCE = 8
LR_ADAPTER = 1e-4
LR_TRUNK_FT = 1e-5 # 10x lower for the pretrained trunk vs the fresh adapter
L_SHOULDER, L_HIP = 5, 11
THRESHOLDS = (0.1, 0.2, 0.3, 0.4, 0.5)
def set_seed(seed=SEED):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def resample_subcarriers(frame_major, n_out=N_SUBC):
"""(nFrames, nSc) -> (nFrames, n_out) by per-frame linear interpolation.
Identity for nSc == n_out. Normalized index domain [0, 1] on both sides.
"""
nf, nsc = frame_major.shape
if nsc == n_out:
return frame_major
xi = np.linspace(0.0, 1.0, nsc)
xo = np.linspace(0.0, 1.0, n_out)
return np.stack([np.interp(xo, xi, frame_major[f]) for f in range(nf)]).astype(np.float32)
def load_dataset():
csi, kps, confs, ts, native70 = [], [], [], [], []
shape_counts = {}
with open(DATA) as f:
for line in f:
r = json.loads(line)
nsc, nf = r["csi_shape"]
shape_counts[f"{nsc}x{nf}"] = shape_counts.get(f"{nsc}x{nf}", 0) + 1
assert nf == 20, r["csi_shape"]
# Aligner layout bug: data is frame-major despite the declared
# [nSc, nFrames] shape -- reshape (nFrames, nSc), then resample the
# subcarrier axis to 70 and transpose to (70 subcarriers, 20 frames).
fm = np.asarray(r["csi"], dtype=np.float32).reshape(nf, nsc)
csi.append(resample_subcarriers(fm).T)
kp = np.asarray(r["kp"], dtype=np.float32)
assert kp.shape == (K, 2), kp.shape
kps.append(kp)
confs.append(r["conf"])
ts.append(r["ts_start"])
native70.append(nsc == N_SUBC)
assert all(ts[i] <= ts[i + 1] for i in range(len(ts) - 1)), "records not time-sorted"
return (np.stack(csi), np.stack(kps), np.asarray(confs, dtype=np.float32),
np.asarray(native70), shape_counts, ts[0], ts[-1])
def temporal_split(n):
n_train = int(round(n * 0.70))
n_val = int(round(n * 0.15))
return slice(0, n_train), slice(n_train, n_train + n_val), slice(n_train + n_val, n)
class AdaptedWiFlow(nn.Module):
"""1x1 Conv1d adapter 70->540 + upstream WiFlow-STD trunk with K=17 pool head."""
def __init__(self, k=K, dropout=0.5):
super().__init__()
self.adapter = nn.Conv1d(N_SUBC, TRUNK_IN, kernel_size=1)
nn.init.kaiming_normal_(self.adapter.weight, mode="fan_out", nonlinearity="relu")
nn.init.constant_(self.adapter.bias, 0)
self.trunk = WiFlowPoseModel(dropout=dropout)
# K=17 via the parameter-free adaptive pool: decoder emits [B, 2, 15, 20]
# spatial maps; pooling H->17 instead of 15 yields [B, 17, 2] with no new
# parameters, so the pretrained state_dict loads strict=True for any K.
self.trunk.avg_pool = nn.AdaptiveAvgPool2d((k, 1))
def forward(self, x):
return self.trunk(self.adapter(x))
def load_pretrained_trunk(trunk, path):
state = torch.load(path, map_location="cpu", weights_only=True)
# Defensive remap as in eval_repro.py (no-op for the retrained checkpoint).
renames = {"att.": "attention.", "final_conv.": "decoder."}
state = {next((new + k[len(old):] for old, new in renames.items()
if k.startswith(old)), k): v
for k, v in state.items()}
trunk.load_state_dict(state, strict=True)
def pck_torso(pred, target, thresholds=THRESHOLDS):
"""Upstream calculate_pck math, torso = l_shoulder(5)<->l_hip(11) for 17-kp COCO."""
norm = torch.sqrt(((target[:, L_SHOULDER] - target[:, L_HIP]) ** 2).sum(dim=1))
norm = torch.clamp(norm, min=0.01)
dist = torch.sqrt(((pred - target) ** 2).sum(dim=2)) / norm.unsqueeze(1)
return {f"pck@{int(t * 100)}": (dist <= t).float().mean().item() for t in thresholds}
def mpjpe(pred, target):
return torch.sqrt(((pred - target) ** 2).sum(dim=2)).mean().item()
@torch.no_grad()
def predict(model, x, batch=256):
model.eval()
return torch.cat([model(x[i:i + batch]) for i in range(0, len(x), batch)])
def eval_preds(pred, target):
out = pck_torso(pred, target)
out["mpjpe"] = mpjpe(pred, target)
# Constant-pose detector: std across test frames per coordinate, mean over
# the 17x2 coordinates. 0.0 == degenerate constant predictor.
out["pred_std"] = pred.std(dim=0).mean().item()
return out
def train_run(name, x_tr, y_tr, x_va, y_va, device, pretrained, freeze_trunk,
lr_trunk):
set_seed(SEED)
model = AdaptedWiFlow().to(device)
if pretrained:
load_pretrained_trunk(model.trunk, CHECKPOINT)
if freeze_trunk:
for p in model.trunk.parameters():
p.requires_grad = False
groups = [{"params": model.adapter.parameters(), "lr": LR_ADAPTER}]
else:
groups = [{"params": model.adapter.parameters(), "lr": LR_ADAPTER},
{"params": model.trunk.parameters(), "lr": lr_trunk}]
opt = torch.optim.AdamW(groups)
loss_fn = nn.MSELoss()
n = len(x_tr)
best_val, best_state, best_epoch, bad = float("inf"), None, -1, 0
history = []
t0 = time.time()
for epoch in range(MAX_EPOCHS):
model.train()
if freeze_trunk:
model.trunk.eval() # keep BatchNorm running stats fixed: pure transfer
perm = torch.randperm(n, device=device)
ep_loss = 0.0
for i in range(0, n, BATCH):
idx = perm[i:i + BATCH]
opt.zero_grad()
loss = loss_fn(model(x_tr[idx]), y_tr[idx])
loss.backward()
opt.step()
ep_loss += loss.item() * len(idx)
val_mpjpe = mpjpe(predict(model, x_va), y_va)
history.append({"epoch": epoch, "train_mse": ep_loss / n, "val_mpjpe": val_mpjpe})
marker = ""
if val_mpjpe < best_val:
best_val, best_epoch, bad = val_mpjpe, epoch, 0
best_state = {k: v.detach().cpu().clone() for k, v in model.state_dict().items()}
marker = " *"
else:
bad += 1
print(f"[{name}] epoch {epoch:02d} train_mse {ep_loss / n:.6f} "
f"val_mpjpe {val_mpjpe:.5f}{marker}", flush=True)
if bad >= PATIENCE:
print(f"[{name}] early stop at epoch {epoch} (best {best_epoch})", flush=True)
break
model.load_state_dict(best_state)
torch.save(best_state, os.path.join(MEASB, f"{name}_best.pth"))
return model, {"best_epoch": best_epoch, "best_val_mpjpe": best_val,
"epochs_run": len(history), "wall_seconds": round(time.time() - t0, 1),
"history": history}
def run_suite(tag, csi, kps, device):
"""Temporal 70/15/15 split, mean-pose baseline, three training runs."""
n = len(csi)
tr, va, te = temporal_split(n)
print(f"=== suite {tag}: n={n} train={tr.stop} val={va.stop - va.start} "
f"test={te.stop - te.start} ===", flush=True)
# CSI normalization constant from TRAIN split only.
train_p99 = float(np.percentile(csi[tr], 99))
train_max = float(csi[tr].max())
print(f"[{tag}] train p99={train_p99:.3f} max={train_max:.3f} -> /p99, clip [0,1]",
flush=True)
csi_n = np.clip(csi / train_p99, 0.0, 1.0).astype(np.float32)
x = torch.from_numpy(csi_n).to(device)
y = torch.from_numpy(kps).to(device)
x_tr, y_tr = x[tr], y[tr]
x_va, y_va = x[va], y[va]
x_te, y_te = x[te], y[te]
suite = {
"n_windows": n,
"split": {"n_train": int(tr.stop), "n_val": int(va.stop - va.start),
"n_test": int(te.stop - te.start)},
"csi_norm": {"method": "divide by train-split p99 amplitude, clip [0,1]",
"train_p99": train_p99, "train_max": train_max},
"runs": {},
}
# Honesty bar: mean-pose predictor fit on TRAIN, evaluated on TEST.
mean_pose = y_tr.mean(dim=0, keepdim=True).expand(len(y_te), -1, -1)
suite["mean_pose_baseline"] = eval_preds(mean_pose, y_te)
suite["mean_pose_baseline"]["note"] = "train-split mean pose; pred_std 0 by construction"
print(f"[{tag}] mean-pose baseline:", json.dumps(suite["mean_pose_baseline"]),
flush=True)
configs = [
("pretrained", dict(pretrained=True, freeze_trunk=False, lr_trunk=LR_TRUNK_FT)),
("scratch", dict(pretrained=False, freeze_trunk=False, lr_trunk=LR_ADAPTER)),
("frozen_trunk", dict(pretrained=True, freeze_trunk=True, lr_trunk=0.0)),
]
for name, cfg in configs:
print(f"=== run: {tag}/{name} {cfg} ===", flush=True)
model, train_info = train_run(f"{tag}_{name}", x_tr, y_tr, x_va, y_va,
device, **cfg)
test_metrics = eval_preds(predict(model, x_te), y_te)
n_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
suite["runs"][name] = {"config": cfg, "trainable_params": n_trainable,
"train": {k: v for k, v in train_info.items()
if k != "history"},
"history": train_info["history"],
"test": test_metrics}
print(f"[{tag}/{name}] TEST:", json.dumps(test_metrics), flush=True)
return suite
def main():
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"device {device}, torch {torch.__version__}", flush=True)
set_seed(SEED)
csi, kps, confs, native70, shape_counts, ts_first, ts_last = load_dataset()
print(f"shape distribution: {shape_counts}", flush=True)
results = {
"protocol": {
"dataset": DATA, "n_windows": len(csi),
"ts_first": ts_first, "ts_last": ts_last,
"conf_mean": float(confs.mean()), "conf_min": float(confs.min()),
"csi_shape_distribution": shape_counts,
"csi_layout_note": "aligner stores frame-major data under a transposed "
"[nSc, nFrames] shape label; corrected on load",
"csi_resample": "per-frame linear interp of subcarrier axis to 70 bins "
"(identity for native-70 frames); native-70 windows still "
"contain ~20.4% internally zero-padded short frames",
"split": "temporal 70/15/15 (no shuffle across time)",
"model": "1x1 Conv1d 70->540 adapter + WiFlowPoseModel trunk, "
"AdaptiveAvgPool2d((17,1)) head (parameter-free K=17)",
"checkpoint": CHECKPOINT,
"checkpoint_note": "measurement-(a) retrained checkpoint (~96% PCK@20 on "
"WiFlow data); att./final_conv. remap applied "
"defensively (no-op, already new-style keys)",
"optimizer": f"AdamW, adapter lr {LR_ADAPTER}, fine-tuned trunk lr "
f"{LR_TRUNK_FT} (10x lower), scratch all {LR_ADAPTER}",
"batch": BATCH, "max_epochs": MAX_EPOCHS, "patience": PATIENCE,
"precision": "fp32", "seed": SEED,
"pck": "torso-normalized, torso = ||l_shoulder(5) - l_hip(11)||, "
"clamp min 0.01, mean over keypoints x frames "
"(upstream math; upstream 2/12 indices are a 15-kp convention)",
},
# Primary: all 2,046 windows (pre-registered n), subcarrier axis resampled.
"all2046": None,
# Secondary robustness check: the 1,347 native [70,20] windows only.
"native70": None,
}
results["all2046"] = run_suite("all2046", csi, kps, device)
results["native70"] = run_suite("native70", csi[native70], kps[native70], device)
out = os.path.join(MEASB, "measurement_b.json")
with open(out, "w") as f:
json.dump(results, f, indent=2)
print(f"wrote {out}", flush=True)
if __name__ == "__main__":
main()