File size: 7,673 Bytes
0838417 | 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 | from __future__ import annotations
import argparse
from pathlib import Path
import sys
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
import numpy as np
from scipy.stats import gaussian_kde
SCRIPT_ROOT = Path(__file__).resolve().parents[1]
if str(SCRIPT_ROOT) not in sys.path:
sys.path.insert(0, str(SCRIPT_ROOT))
from common import RESULTS_MECHANISM_FIG_ROOT, RESULTS_MECHANISM_TRACE_EP10_ROOT, RESULTS_MECHANISM_TRACE_QUICK_ROOT
MAX_POINTS_PER_COMPONENT = 2000
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Replot 2-task P/Q/C UMAP figure from paper mechanism traces.")
parser.add_argument("--bundle", choices=["quick_seed7", "ep10_seed7"], default="ep10_seed7")
return parser.parse_args()
plt.rcParams.update(
{
"font.family": "DejaVu Sans",
"font.size": 10,
"axes.titlesize": 13.5,
"axes.titleweight": "semibold",
"axes.labelsize": 11.5,
"axes.labelcolor": "#334155",
"axes.edgecolor": "#cbd5e1",
"axes.linewidth": 1.0,
"xtick.color": "#64748b",
"ytick.color": "#64748b",
"legend.fontsize": 11,
"figure.facecolor": "white",
"axes.facecolor": "#fbfcfe",
"savefig.facecolor": "white",
"savefig.bbox": "tight",
"savefig.pad_inches": 0.05,
}
)
def _load_trace(task: str, bundle: str) -> dict[str, np.ndarray]:
traces_dir = (RESULTS_MECHANISM_TRACE_QUICK_ROOT if bundle == "quick_seed7" else RESULTS_MECHANISM_TRACE_EP10_ROOT) / "traces"
payload = np.load(traces_dir / f"{task}_seed7_teacher_forced.npz", allow_pickle=True)
return {
"valid_mask": np.asarray(payload["valid_mask"]),
"q": np.asarray(payload["q"]),
"p": np.asarray(payload["p"]),
"c": np.asarray(payload["c"]),
}
def _project(array: np.ndarray) -> np.ndarray:
try:
from umap import UMAP
return UMAP(n_neighbors=35, min_dist=0.12, random_state=0, n_components=2).fit_transform(array)
except Exception:
centered = array - np.mean(array, axis=0, keepdims=True)
_, _, vt = np.linalg.svd(centered, full_matrices=False)
basis = vt[:2].T
return centered @ basis
def _pad(values: np.ndarray, target_dim: int) -> np.ndarray:
if values.shape[1] >= target_dim:
return values[:, :target_dim]
return np.concatenate([values, np.zeros((values.shape[0], target_dim - values.shape[1]))], axis=1)
def _standardize(values: np.ndarray) -> np.ndarray:
mean = np.mean(values, axis=0, keepdims=True)
std = np.std(values, axis=0, keepdims=True)
return (values - mean) / np.clip(std, 1e-6, None)
def _prepare_component(values: np.ndarray, target_dim: int) -> np.ndarray:
standardized = _standardize(values)
padded = _pad(standardized, target_dim)
return padded / np.sqrt(float(values.shape[1]))
def _subsample_even(values: np.ndarray, count: int) -> np.ndarray:
if values.shape[0] <= count:
return values
indices = np.linspace(0, values.shape[0] - 1, count, dtype=np.int32)
return values[indices]
def _draw_density_contours(ax: plt.Axes, points: np.ndarray, color: str) -> None:
if points.shape[0] < 32:
return
x = points[:, 0]
y = points[:, 1]
try:
kde = gaussian_kde(np.vstack([x, y]))
except Exception:
return
x_lo, x_hi = np.quantile(x, [0.01, 0.99])
y_lo, y_hi = np.quantile(y, [0.01, 0.99])
x_pad = max(1e-3, 0.12 * (x_hi - x_lo))
y_pad = max(1e-3, 0.12 * (y_hi - y_lo))
xx, yy = np.meshgrid(
np.linspace(x_lo - x_pad, x_hi + x_pad, 160),
np.linspace(y_lo - y_pad, y_hi + y_pad, 160),
)
zz = kde(np.vstack([xx.ravel(), yy.ravel()])).reshape(xx.shape)
levels = np.quantile(zz[zz > 0], [0.72, 0.86, 0.95])
ax.contour(xx, yy, zz, levels=levels, colors=[color], linewidths=[0.9, 1.15, 1.45], alpha=0.7, zorder=3)
def _set_view_limits(ax: plt.Axes, embedding: np.ndarray) -> None:
x_lo, x_hi = np.quantile(embedding[:, 0], [0.01, 0.99])
y_lo, y_hi = np.quantile(embedding[:, 1], [0.01, 0.99])
x_pad = max(1e-3, 0.10 * (x_hi - x_lo))
y_pad = max(1e-3, 0.10 * (y_hi - y_lo))
ax.set_xlim(x_lo - x_pad, x_hi + x_pad)
ax.set_ylim(y_lo - y_pad, y_hi + y_pad)
def _save_pair(fig: plt.Figure, stem: Path) -> None:
stem.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(stem.with_suffix(".pdf"))
fig.savefig(stem.with_suffix(".png"))
def main() -> int:
args = parse_args()
fig, axes = plt.subplots(1, 2, figsize=(6.6, 2.9), sharex=False, sharey=False, dpi=220)
colors = {"q": "#2b6cb0", "p": "#dd6b20", "c": "#2f855a"}
titles = {"finger_spin": "Finger Spin", "cheetah_run": "Cheetah Run"}
legend_handles = [
Line2D([0], [0], marker="o", linestyle="", markersize=6.5, markerfacecolor=colors[key], markeredgecolor="white", markeredgewidth=0.6, label=fr"${key}$")
for key in ("q", "p", "c")
]
for ax, task in zip(axes, ["finger_spin", "cheetah_run"]):
trace = _load_trace(task, args.bundle)
mask = trace["valid_mask"].reshape(-1)
q = trace["q"].reshape(-1, trace["q"].shape[-1])[mask]
p = trace["p"].reshape(-1, trace["p"].shape[-1])[mask]
c = trace["c"].reshape(-1, trace["c"].shape[-1])[mask]
count = min(MAX_POINTS_PER_COMPONENT, q.shape[0], p.shape[0], c.shape[0])
q = _subsample_even(q, count)
p = _subsample_even(p, count)
c = _subsample_even(c, count)
target_dim = max(q.shape[1], p.shape[1], c.shape[1])
stacked = np.concatenate(
[
_prepare_component(q, target_dim),
_prepare_component(p, target_dim),
_prepare_component(c, target_dim),
],
axis=0,
)
labels = (["q"] * count) + (["p"] * count) + (["c"] * count)
embedding = _project(stacked)
for key in ("q", "p", "c"):
selected = np.array([label == key for label in labels])
points = embedding[selected]
ax.scatter(
points[:, 0],
points[:, 1],
s=4,
alpha=0.20,
color=colors[key],
label=fr"${key}$",
linewidths=0.0,
edgecolors="none",
rasterized=True,
zorder=2,
)
_draw_density_contours(ax, points, colors[key])
_set_view_limits(ax, embedding)
ax.set_title(titles[task], loc="left", pad=8, color="#0f172a")
ax.set_xticks([])
ax.set_yticks([])
ax.set_xlabel("Embedding 1", labelpad=8)
ax.set_ylabel("Embedding 2", labelpad=8)
for spine in ("left", "bottom"):
ax.spines[spine].set_color("#cbd5e1")
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
fig.legend(legend_handles, ["q", "p", "c"], loc="upper center", ncol=3, frameon=False, bbox_to_anchor=(0.5, 1.02), handletextpad=0.35, columnspacing=1.2)
fig.tight_layout(pad=0.7, w_pad=1.2, rect=(0, 0, 1, 0.92))
out_stem = RESULTS_MECHANISM_FIG_ROOT / "pqc"
_save_pair(fig, out_stem)
plt.close(fig)
print(out_stem.with_suffix(".pdf"))
print(out_stem.with_suffix(".png"))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|