PIVOT / scripts /figure1.py
pranamanam's picture
Upload 176 files
6fa9282 verified
Raw
History Blame Contribute Delete
4.07 kB
"""Draw the PIVOT overview with editable vector paths and embedded fonts."""
from pathlib import Path
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib import font_manager as fm
from matplotlib.patches import Ellipse, FancyArrowPatch, PathPatch, FancyBboxPatch
from matplotlib.path import Path as MPath
ROOT = Path(__file__).resolve().parents[1]
for f in (ROOT / "assets/fonts").glob("*.ttf"):
fm.fontManager.addfont(str(f))
plt.rcParams.update(
{
"font.family": "Ubuntu",
"mathtext.fontset": "cm",
"pdf.fonttype": 42,
"ps.fonttype": 42,
"font.size": 13,
}
)
blue = "#62AEDD"
navy = "#245E84"
red = "#9D2944"
ink = "#243746"
pale = "#E9F4FB"
gray = "#9EB0BD"
fig, ax = plt.subplots(figsize=(13.6, 4.8))
ax.set(xlim=(0, 13.6), ylim=(0, 4.8))
ax.axis("off")
def text(x, y, s, size=13, color=ink, **kw):
ax.text(
x,
y,
s,
fontsize=size,
color=color,
ha=kw.pop("ha", "center"),
va="center",
**kw,
)
def arrow(start, end, color=blue, rad=0, lw=2.2):
ax.add_patch(
FancyArrowPatch(
start,
end,
arrowstyle="-|>",
mutation_scale=16,
connectionstyle=f"arc3,rad={rad}",
color=color,
lw=lw,
)
)
def cloud(x, y, w, h, color, seed):
rng = np.random.default_rng(seed)
ax.add_patch(Ellipse((x, y), w, h, facecolor=color, alpha=0.10, edgecolor="none"))
z = rng.normal(size=(36, 2))
z = z[np.linalg.norm(z, axis=1) < 2]
ax.scatter(
x + z[:, 0] * w / 5,
y + z[:, 1] * h / 5,
s=rng.uniform(9, 24, len(z)),
color=color,
alpha=0.7,
linewidth=0,
)
cloud(1.45, 3.08, 2.1, 1.72, blue, 2)
text(1.45, 4.19, "Control population", 18, navy, weight="bold")
text(1.45, 2.02, r"$c_0\sim\rho_0$", 18)
# One broad map across the figure, with faint parallel sample trajectories.
for off in [-0.32, -0.12, 0.12, 0.32]:
verts = [(2.55, 3.1 + off), (4.3, 3.8 + off), (6.7, 3.9 + off), (8.75, 3.15 + off)]
ax.add_patch(
PathPatch(
MPath(verts, [MPath.MOVETO, MPath.CURVE4, MPath.CURVE4, MPath.CURVE4]),
fill=False,
edgecolor=blue,
alpha=0.20,
lw=10,
)
)
arrow((2.6, 3.1), (8.8, 3.14), blue, 0.20, 3)
text(5.6, 3.72, r"$X_\theta(0,1,c_0,e_u)$", 23)
text(5.6, 3.17, "Predict the response", 18, navy, weight="bold")
text(5.6, 4.38, r"$u=\{(g_j,o_j)\}_{j=1}^{M}\quad\longmapsto\quad e_u$", 17)
arrow((5.6, 4.12), (5.6, 3.92), gray, 0, 1.2)
cloud(9.4, 3.18, 1.75, 1.5, blue, 9)
text(9.4, 4.19, "Predicted cells", 18, navy, weight="bold")
cloud(11.97, 3.7, 1.65, 1.18, red, 3)
text(11.97, 4.56, "Target cells", 18, red, weight="bold")
text(12, 2.88, r"$c^\star\sim\rho^\star$", 18)
arrow((10.2, 3.23), (11.22, 3.63), red, -0.08, 1.6)
text(11, 2.57, "Endpoint reward", 17, red)
# Reward gradients return to the intervention coordinates.
arrow((11.3, 2.31), (5.22, 1.89), red, -0.17, 2.6)
text(7.95, 1.36, "Optimize intervention embeddings", 17, red)
text(3.05, 1.72, "Rank admissible interventions", 17, navy, weight="bold")
arrow((5.01, 2.03), (3.13, 2.25), navy, -0.10, 1.5)
text(3.45, 2.48, r"$e^{(L)}\;\longrightarrow\;u_{1:K}$", 18)
ax.add_patch(
FancyBboxPatch(
(0.18, 0.12),
13.22,
1.03,
boxstyle="round,pad=0.02,rounding_size=.08",
facecolor=pale,
edgecolor="none",
)
)
text(0.51, 0.78, "PIVOT", 18, navy, ha="left", weight="bold")
text(
7.8,
0.74,
r"$\hat c_1=X_\theta(0,1,c_0,e)\qquad g_e=J_eX_\theta^{\mathsf{T}}\nabla_{\hat c_1}r\qquad e^+=e+\gamma\,\dfrac{g_e}{\|g_e\|_2+\epsilon}$",
20,
)
text(
7.6,
0.29,
"Endpoint prediction, reward gradients, and admissible interventions.",
16,
)
fig.subplots_adjust(left=0, right=1, bottom=0, top=1)
for ext in ["pdf", "png", "svg"]:
fig.savefig(ROOT / "assets" / f"figure1.{ext}", dpi=250, facecolor="white")