fairtalking-second-work / scripts /analysis /visualize_motivation.py
huahua123313's picture
Add files using upload-large-folder tool
eea47ad verified
Raw
History Blame Contribute Delete
9.13 kB
"""Visualize motivation evidence for CTA, given a feature .npz dumped by
scripts/analysis/dump_cta_features.py.
Produces three figures:
1) ``hist_<key>.pdf`` for key in {l_av, l_va, asym}
Per-class histogram + KDE. The asym histogram is THE figure that
directly demonstrates the detection signal.
2) ``scatter_lva_lav.pdf``
L_AV vs L_VA scatter plot. Shows whether fake samples collapse onto a
near-diagonal manifold (i.e. V is fully determined by A under the
generator), while real samples stay off-diagonal.
3) ``tsne_residual.pdf``
2-up: t-SNE of the V-prediction residual (left) and A-prediction
residual (right), colored by real/fake. Lets you see which residual
direction carries more class-separating information.
The script avoids any non-stdlib dependency beyond numpy / matplotlib /
sklearn (already required by the project for compute_extra_metrics.py).
Usage
-----
python3 scripts/analysis/visualize_motivation.py \\
--npz outputs/analysis/cta_features_oursval.npz \\
--out-dir outputs/analysis/figs_oursval
# subsample t-SNE to speed up (default 2000)
python3 scripts/analysis/visualize_motivation.py \\
--npz <...>.npz --out-dir <...>/figs --tsne-n 1000
"""
from __future__ import annotations
import argparse
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
def _palette():
return {0: "#2E86C1", 1: "#E74C3C"} # blue=real, red=fake
def _label_name(y: int) -> str:
return "real" if y == 0 else "fake"
# ---------------------------------------------------------------------------
# 1) Histograms
# ---------------------------------------------------------------------------
def plot_histogram(values, labels, out_path, title, xlabel, bins=60):
fig, ax = plt.subplots(figsize=(6.0, 4.0))
pal = _palette()
lo = float(np.percentile(values, 0.5))
hi = float(np.percentile(values, 99.5))
edges = np.linspace(lo, hi, bins + 1)
for y in (0, 1):
v = values[labels == y]
if len(v) == 0:
continue
ax.hist(v, bins=edges, density=True, alpha=0.55,
color=pal[y], edgecolor="white", linewidth=0.4,
label=f"{_label_name(y)} (n={len(v)}, mean={v.mean():.4f})")
ax.axvline(0.0, color="0.4", linestyle="--", linewidth=0.8, alpha=0.7)
ax.set_xlabel(xlabel)
ax.set_ylabel("density")
ax.set_title(title)
ax.legend(frameon=False, loc="best")
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
fig.tight_layout()
fig.savefig(out_path, dpi=200)
fig.savefig(out_path.with_suffix(".png"), dpi=200)
plt.close(fig)
print(f"[viz] wrote {out_path}")
# ---------------------------------------------------------------------------
# 2) L_AV vs L_VA scatter
# ---------------------------------------------------------------------------
def plot_scatter(l_av, l_va, labels, out_path, max_points_per_class=1500):
fig, ax = plt.subplots(figsize=(5.5, 5.0))
pal = _palette()
rng = np.random.default_rng(0)
for y in (1, 0):
mask = labels == y
x = l_av[mask]
z = l_va[mask]
if len(x) > max_points_per_class:
idx = rng.choice(len(x), size=max_points_per_class, replace=False)
x = x[idx]; z = z[idx]
ax.scatter(x, z, s=8, c=pal[y], alpha=0.45, edgecolors="none",
label=f"{_label_name(y)} (n={mask.sum()})")
lo = float(min(l_av.min(), l_va.min()))
hi = float(max(l_av.max(), l_va.max()))
pad = (hi - lo) * 0.05
ax.plot([lo - pad, hi + pad], [lo - pad, hi + pad],
color="0.3", linestyle="--", linewidth=0.8, alpha=0.8, label="y = x")
ax.set_xlim(lo - pad, hi + pad)
ax.set_ylim(lo - pad, hi + pad)
ax.set_xlabel(r"$L_{A \to V}$ (A->V prediction MSE)")
ax.set_ylabel(r"$L_{V \to A}$ (V->A prediction MSE)")
ax.set_title("Per-sample prediction-loss asymmetry")
ax.legend(frameon=False, loc="best")
ax.set_aspect("equal", adjustable="box")
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
fig.tight_layout()
fig.savefig(out_path, dpi=200)
fig.savefig(out_path.with_suffix(".png"), dpi=200)
plt.close(fig)
print(f"[viz] wrote {out_path}")
# ---------------------------------------------------------------------------
# 3) Residual t-SNE
# ---------------------------------------------------------------------------
def _tsne(x, n, seed=0):
if len(x) > n:
rng = np.random.default_rng(seed)
idx = rng.choice(len(x), size=n, replace=False)
x = x[idx]
else:
idx = np.arange(len(x))
try:
from sklearn.manifold import TSNE
except ImportError as e:
raise SystemExit(
"[viz] sklearn is required for t-SNE. Install: pip install scikit-learn"
) from e
tsne = TSNE(
n_components=2, perplexity=min(30, max(5, len(x) // 50)),
init="pca", learning_rate="auto", random_state=seed,
)
return tsne.fit_transform(x), idx
def plot_residual_tsne(r_av, r_va, labels, out_path, n=2000):
fig, axes = plt.subplots(1, 2, figsize=(11, 5))
pal = _palette()
for ax, x, name in [(axes[0], r_av, r"$r_{A \to V}$ (V token residual)"),
(axes[1], r_va, r"$r_{V \to A}$ (A token residual)")]:
emb, idx = _tsne(x, n=n)
y = labels[idx]
for cls in (1, 0):
m = y == cls
ax.scatter(emb[m, 0], emb[m, 1], s=10, c=pal[cls], alpha=0.55,
edgecolors="none", label=f"{_label_name(cls)} ({m.sum()})")
ax.set_title(name)
ax.set_xticks([]); ax.set_yticks([])
for s in ax.spines.values():
s.set_visible(False)
ax.legend(frameon=False, loc="best")
fig.suptitle("t-SNE of cross-modal prediction residuals", fontsize=11)
fig.tight_layout()
fig.savefig(out_path, dpi=200)
fig.savefig(out_path.with_suffix(".png"), dpi=200)
plt.close(fig)
print(f"[viz] wrote {out_path}")
# ---------------------------------------------------------------------------
def write_per_generator_summary(asym, labels, generators, out_path):
rows = [("group", "n", "asym_mean", "asym_median", "asym_std")]
rows.append(("real (all)", int((labels == 0).sum()),
float(asym[labels == 0].mean()),
float(np.median(asym[labels == 0])),
float(asym[labels == 0].std())))
fakes = labels == 1
if fakes.any():
for g in sorted(set(generators[fakes].tolist())):
sel = fakes & (generators == g)
v = asym[sel]
rows.append((f"fake/{g}", int(sel.sum()),
float(v.mean()), float(np.median(v)), float(v.std())))
rows.append(("fake (all)", int(fakes.sum()),
float(asym[fakes].mean()), float(np.median(asym[fakes])),
float(asym[fakes].std())))
with open(out_path, "w") as f:
f.write(",".join(map(str, rows[0])) + "\n")
for r in rows[1:]:
f.write(f"{r[0]},{r[1]},{r[2]:.6f},{r[3]:.6f},{r[4]:.6f}\n")
print(f"[viz] wrote {out_path}")
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument("--npz", required=True, help="Path to dump_cta_features.py output.")
p.add_argument("--out-dir", required=True, help="Directory for figures + summary csv.")
p.add_argument("--tsne-n", type=int, default=2000, help="t-SNE subsample size.")
args = p.parse_args()
npz_path = Path(args.npz).resolve()
out_dir = Path(args.out_dir).resolve()
out_dir.mkdir(parents=True, exist_ok=True)
print(f"[viz] loading {npz_path}")
data = np.load(npz_path, allow_pickle=True)
l_av = data["l_av"].astype(np.float64)
l_va = data["l_va"].astype(np.float64)
asym = data["asym"].astype(np.float64)
label = data["label"].astype(np.int64)
gen = data["generator"].astype(object)
r_av = data["r_av"].astype(np.float32)
r_va = data["r_va"].astype(np.float32)
meta = data["meta"].item() if "meta" in data.files else {}
print(f"[viz] n={len(l_av)} reals={(label==0).sum()} fakes={(label==1).sum()}")
if meta:
print(f"[viz] meta: {meta}")
plot_histogram(l_av, label, out_dir / "hist_l_av.pdf",
r"Distribution of $L_{A \to V}$", r"$L_{A \to V}$")
plot_histogram(l_va, label, out_dir / "hist_l_va.pdf",
r"Distribution of $L_{V \to A}$", r"$L_{V \to A}$")
plot_histogram(asym, label, out_dir / "hist_asym.pdf",
r"Asymmetry score $s = L_{V \to A} - L_{A \to V}$",
r"$s_\mathrm{asym}$")
plot_scatter(l_av, l_va, label, out_dir / "scatter_lva_lav.pdf")
plot_residual_tsne(r_av, r_va, label, out_dir / "tsne_residual.pdf",
n=args.tsne_n)
write_per_generator_summary(asym, label, gen,
out_dir / "per_generator_asym.csv")
print("[viz] DONE.")
return 0
if __name__ == "__main__":
raise SystemExit(main())