File size: 10,469 Bytes
773c429 a4a865e 773c429 a4a865e 773c429 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 | """Physical interpretability of the next-generation backbones.
Every architecture is probed on the same held-out trajectories, and each probe
is chosen to expose the *physics* the model claims to represent:
koopman eigen-spectrum of the linear latent operator -> memory time constants
gdn norm of the delta-rule fast-weight matrix -> "memory written at yield"
xlstm forget-gate value -> memory horizon vs strain
node ||dz/dt|| of the learned vector field -> where the dynamics change
rssm prior std of the stochastic latent -> model uncertainty vs strain
flow spread of K sampled curves -> predictive interval quality
Outputs land in ``results/tables/interpret_*.csv`` and
``results/figures/fig_interpret_*.png``.
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
from typing import Dict, List
import numpy as np
import pandas as pd
from .data import (PROC, TARGETS, Normalizer, PackedDataset, build_static_frame,
collate)
from .models import build_model
ROOT = Path(__file__).resolve().parents[2]
RUNS = ROOT / "results" / "runs"
TABLES = ROOT / "results" / "tables"
FIGS = ROOT / "results" / "figures"
EPS = np.linspace(0.0, 0.2, 401)
def _plt():
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
return plt
def load(tag: str, device="cpu"):
import torch
d = RUNS / tag
ck = torch.load(d / "model.pt", map_location=device, weights_only=False)
meta = json.loads((d / "metrics.json").read_text())
cfg = ck["cfg"]
m = build_model(cfg["model"], n_static=int(np.sum(ck["keep_mask"])),
n_out=len(TARGETS), d_model=cfg["d_model"],
n_layers=cfg["n_layers"], d_state=cfg["d_state"],
dropout=cfg["dropout"], coupling=cfg["coupling"]).to(device)
m.load_state_dict(ck["state"])
m.eval()
return m, cfg, meta, ck
def _loader(tag, meta, ck, device="cpu", batch=2):
import torch
mtab = pd.read_csv(PROC / "samples_meta.csv")
traj = pd.read_csv(PROC / "trajectories.csv")
names = ck["names"]
km = np.asarray(ck["keep_mask"], dtype=bool)
kinds = build_static_frame(mtab)[2]
norm = Normalizer.from_json(meta["normalizer"])
ids = [int(s) for s in meta["test_ids"]]
ds = PackedDataset(ids, mtab, traj, norm, names, kinds, km)
dl = torch.utils.data.DataLoader(ds, batch_size=batch, shuffle=False,
collate_fn=collate)
return dl, norm, ids
def _collect_trace(model, dl, device="cpu", per_sample=False) -> np.ndarray:
"""Run the model with recording enabled; returns (T,) or (T, C) mean trace.
Each block appends one entry per timestep (a scalar or a tuple of scalars),
already averaged over the batch. Entries of the same block therefore stack
to ``(T, C)``; blocks of one architecture share ``C``.
"""
import torch
blocks = list(getattr(model, "blocks", []))
for b in blocks:
if hasattr(b, "record"):
b.record = True
b.trace = []
if hasattr(model, "record"):
model.record = True
model.trace = []
rows = []
with torch.no_grad():
for st, x, _Y, _M, _s in dl:
model(st.to(device), x.to(device))
per = [np.asarray(getattr(b, "trace", [])) for b in blocks
if len(getattr(b, "trace", []))]
if not per and len(getattr(model, "trace", [])):
# single trace on the model itself (NODE / RSSM)
per = [np.asarray(model.trace)]
if per:
ps = [p if p.ndim >= 2 else p[:, None] for p in per] # (T, C)
L = min(p.shape[0] for p in ps)
rows.append(np.mean(np.stack([p[:L] for p in ps], 0), 0))
for b in blocks:
if hasattr(b, "trace"):
b.trace = []
if hasattr(model, "trace"):
model.trace = []
for b in blocks:
if hasattr(b, "record"):
b.record = False
if not rows:
return np.zeros((0, 401))
return np.mean(np.stack(rows, 0), 0) # average over batches -> (T,) / (T, C)
# ---------------------------------------------------------------------------
# individual probes
# ---------------------------------------------------------------------------
def koopman_spectrum(tag: str) -> pd.DataFrame:
m, cfg, meta, ck = load(tag)
r, th = m.spectrum()
r = r.detach().numpy()
th = th.detach().numpy()
d_eps = 0.2 / 400.0
tau = -d_eps / np.log(np.clip(r, 1e-6, 1 - 1e-9)) # strain units
df = pd.DataFrame({"mode": np.arange(len(r)), "radius": r,
"angle": th, "tau_strain": tau,
"period_strain": 2 * np.pi * d_eps / np.maximum(th, 1e-6)})
TABLES.mkdir(parents=True, exist_ok=True)
df.to_csv(TABLES / "interpret_koopman.csv", index=False)
return df
def trace_probe(tag: str, kind: str) -> pd.DataFrame:
m, cfg, meta, ck = load(tag)
dl, norm, ids = _loader(tag, meta, ck)
tr = _collect_trace(m, dl)
if tr.size == 0:
return pd.DataFrame()
if tr.ndim == 1:
tr = tr[:, None]
if kind == "gdn":
names = ["mem_norm", "decay_alpha", "write_beta"]
elif kind == "xlstm":
names = ["mem_norm", "forget_gate"]
elif kind == "node":
names = ["dz_dt_norm"]
elif kind == "rssm":
names = ["prior_std"]
else:
names = [f"c{i}" for i in range(tr.shape[1])]
names = names[:tr.shape[1]]
T = tr.shape[0]
df = pd.DataFrame(tr, columns=names)
df.insert(0, "strain", np.linspace(0, 0.2, T))
df.to_csv(TABLES / f"interpret_{kind}.csv", index=False)
return df
def flow_uncertainty(tag: str, k: int = 16) -> Dict[str, float]:
import torch
m, cfg, meta, ck = load(tag)
dl, norm, ids = _loader(tag, meta, ck, batch=8)
S, Y, M = [], [], []
with torch.no_grad():
for st, x, y, msk, _ in dl:
z = m.sample(st, x, k=k) # (k, b, T, 13)
S.append(z.numpy())
Y.append(y.numpy())
M.append(msk.numpy())
S = np.concatenate(S, 1) # (k, N, T, 13)
Y = np.concatenate(Y, 0)
M = np.concatenate(M, 0)
raw = norm.inverse_targets(S) # (k, N, T, 13)
raww = norm.inverse_targets(Y)
mu = raw.mean(0)
sd = raw.std(0)
j = 0
mask = M[..., j] > 0
if mask.sum() < 10:
return {}
lo, hi = mu[..., j] - 1.645 * sd[..., j], mu[..., j] + 1.645 * sd[..., j]
cov = float(((raww[..., j] >= lo) & (raww[..., j] <= hi))[mask].mean())
err = np.abs(mu[..., j] - raww[..., j])[mask]
width = (hi - lo)[mask]
out = {"coverage90": cov, "mean_width": float(width.mean()),
"mean_abs_err": float(err.mean()),
"sharpness_ratio": float(err.mean() / max(width.mean(), 1e-9)),
"n_samples": int(mask.sum())}
pd.DataFrame([out]).to_csv(TABLES / "interpret_flow.csv", index=False)
return out
# ---------------------------------------------------------------------------
# figures
# ---------------------------------------------------------------------------
def fig_memory(traces: Dict[str, pd.DataFrame], fname="fig_interpret_memory.png"):
plt = _plt()
panels = [("xlstm", "forget_gate", "xLSTM forget gate $f_t$"),
("gdn", "mem_norm", r"Gated DeltaNet $\|S_t\|_F$"),
("node", "dz_dt_norm", r"Neural ODE $\|dz/dt\|$"),
("rssm", "prior_std", r"RSSM prior std $\sigma_t$")]
fig, axes = plt.subplots(2, 2, figsize=(7.2, 4.4), sharex=True)
for ax, (k, col, title) in zip(axes.ravel(), panels):
df = traces.get(k)
if df is None or col not in df:
ax.axis("off")
continue
v = df[col].to_numpy(dtype=float)
eps = df["strain"].to_numpy(dtype=float)
# plot the raw quantity (each panel has its own units); the title states
# the elastic (eps<0.02) -> plastic (eps>0.15) means, the same windows
# used in the text, so figure and table never disagree.
ax.plot(eps, v, lw=1.2, color="C0")
ax.set_title(f"{title}\n{v[eps < 0.02].mean():.4g} → {v[eps > 0.15].mean():.4g}",
fontsize=7.5)
ax.axvline(0.02, color="0.6", ls=":", lw=0.8)
ax.grid(alpha=0.2)
for ax in axes[-1]:
ax.set_xlabel("true strain")
fig.tight_layout()
FIGS.mkdir(parents=True, exist_ok=True)
fig.savefig(FIGS / fname, bbox_inches="tight")
plt.close(fig)
return FIGS / fname
def fig_koopman(df: pd.DataFrame, fname="fig_interpret_koopman.png"):
plt = _plt()
fig, ax = plt.subplots(1, 2, figsize=(6.4, 2.6))
th = np.linspace(0, 2 * np.pi, 200)
ax[0].plot(np.cos(th), np.sin(th), color="0.8", lw=0.8)
ax[0].scatter(df.radius * np.cos(df.angle), df.radius * np.sin(df.angle),
c=df.tau_strain, cmap="viridis", s=18)
ax[0].set_aspect("equal")
ax[0].set_title("Koopman eigenvalues", fontsize=8)
ax[1].bar(np.arange(len(df)), df.tau_strain, color="C0")
ax[1].set_yscale("log")
ax[1].set_xlabel("mode")
ax[1].set_ylabel(r"memory time $\tau_\varepsilon$")
ax[1].set_title("decay time per mode", fontsize=8)
fig.tight_layout()
fig.savefig(FIGS / fname, bbox_inches="tight")
plt.close(fig)
return FIGS / fname
def main():
ap = argparse.ArgumentParser("pimamba.interpret")
ap.add_argument("--koopman", default=None)
ap.add_argument("--gdn", default=None)
ap.add_argument("--xlstm", default=None)
ap.add_argument("--node", default=None)
ap.add_argument("--rssm", default=None)
ap.add_argument("--flow", default=None)
ap.add_argument("--device", default="cpu")
a = ap.parse_args()
traces: Dict[str, pd.DataFrame] = {}
if a.koopman:
df = koopman_spectrum(a.koopman)
print(fig_koopman(df))
print(df.describe().to_string())
for kind, tag in [("gdn", a.gdn), ("xlstm", a.xlstm), ("node", a.node),
("rssm", a.rssm)]:
if tag:
df = trace_probe(tag, kind)
traces[kind] = df
print(f"{kind}: {df.shape}")
if traces:
print(fig_memory(traces))
if a.flow:
print(flow_uncertainty(a.flow))
if __name__ == "__main__":
main() |