activevision's picture
Initial release v0.4.0 — ActiveVision benchmark (85 instances, 17 tasks)
f69e256 verified
raw
history blame
30.4 kB
"""Line Intersections (turtle-based, anchored endpoints).
Each curve has two preset endpoints placed on the perimeter of the canvas.
A straight "leader" segment connects each endpoint to an interior point
(~150px inward), keeping the perimeter band free of crossings. The middle
body of the curve is a turtle-style walk through waypoints distributed
across a generous interior zone — tangled, but not concentrated at the
very center.
Only one of the two endpoints of each curve is labeled (A, B, C, ...).
The model starts tracing from the labeled endpoint and follows the
queried curve through the interior tangle. At every crossing it passes
through, it records the label of the OTHER curve involved (for a
self-crossing, that label is the queried curve's own letter). The
answer is the ordered sequence of labels.
"""
from __future__ import annotations
import argparse
import json
import math
import os
import random
import string
import sys
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Tuple
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
from tqdm import tqdm
LINE_COLOR = "#2f2f2f"
LABEL_COLORS = [
"#c23030",
"#1a6dba",
"#2d8e2d",
"#c47a18",
"#7a35a0",
]
# ── Perimeter anchors ──────────────────────────────────────────────
def _perimeter_point(s: float, width: int, height: int, margin: int
) -> Tuple[np.ndarray, np.ndarray]:
"""Map s ∈ [0, 4) to (position, inward_unit_vector) on rect perimeter."""
side = int(s) % 4
t = s - int(s)
if side == 0: # top edge, left -> right
x = margin + (width - 2 * margin) * t
y = margin
inward = np.array([0.0, 1.0])
elif side == 1: # right edge, top -> bottom
x = width - margin
y = margin + (height - 2 * margin) * t
inward = np.array([-1.0, 0.0])
elif side == 2: # bottom edge, right -> left
x = width - margin - (width - 2 * margin) * t
y = height - margin
inward = np.array([0.0, -1.0])
else: # left edge, bottom -> top
x = margin
y = height - margin - (height - 2 * margin) * t
inward = np.array([1.0, 0.0])
return np.array([x, y]), inward
def sample_perimeter_anchors(
rng: random.Random,
num_anchors: int,
width: int,
height: int,
margin: int,
min_gap: float | None = None,
jitter_deg: float = 10.0,
max_attempts: int = 300,
) -> List[Tuple[np.ndarray, np.ndarray]]:
"""Sample `num_anchors` positions around the canvas perimeter, enforcing
a minimum gap (in perimeter units, 1 unit = one side)."""
if min_gap is None:
# 50% of perfectly even spacing
min_gap = (4.0 / num_anchors) * 0.5
for _ in range(max_attempts):
positions = sorted(rng.uniform(0.0, 4.0) for _ in range(num_anchors))
ok = True
for i in range(num_anchors):
gap = (positions[(i + 1) % num_anchors] - positions[i]) % 4.0
if gap < min_gap:
ok = False
break
if ok:
anchors = []
for s in positions:
pt, inward = _perimeter_point(s, width, height, margin)
# small angular jitter of the inward direction
ang = math.atan2(inward[1], inward[0])
ang += math.radians(rng.uniform(-jitter_deg, jitter_deg))
inward = np.array([math.cos(ang), math.sin(ang)])
anchors.append((pt, inward))
return anchors
return None
# ── Turtle body with anchored start/end ────────────────────────────
def _angle_diff(current: float, target: float) -> float:
d = target - current
return (d + math.pi) % (2 * math.pi) - math.pi
def build_anchored_curve(
rng: random.Random,
width: int,
height: int,
start_pos: np.ndarray,
start_inward: np.ndarray,
end_pos: np.ndarray,
end_inward: np.ndarray,
leader_length: float = 150.0,
step_size: float = 3.3,
max_turn: float = 0.028,
drift_rate: float = 0.007,
num_waypoints: int = 2,
interior_margin: int = 260,
max_steps: int = 1800,
) -> np.ndarray:
"""Single-turtle curve from start_pos to a point near end_pos.
Phase-dependent turning:
- Entry leader (first `leader_length` of path): max_turn = 0 (straight).
- Body (interior): waypoint steering within max_turn clamp.
- Exit leader (within `leader_length` of end_pos perimeter plane):
head-alignment only (no waypoint steering), rotating toward
-end_inward at up to max_turn/step. Then drives straight.
Because the entire curve is a single turtle walk, there are no splice
discontinuities. The curve terminates when it crosses the perimeter plane
at end_pos; the final point is wherever it actually crosses (close to
end_pos when head-alignment locks in time).
"""
pts: List[np.ndarray] = [start_pos.copy()]
x, y = float(start_pos[0]), float(start_pos[1])
theta = math.atan2(start_inward[1], start_inward[0])
# Waypoints in interior, ordered by progress from start→end.
raw_wps: List[np.ndarray] = []
for _ in range(num_waypoints):
wx = rng.uniform(interior_margin, width - interior_margin)
wy = rng.uniform(interior_margin, height - interior_margin)
raw_wps.append(np.array([wx, wy]))
direction = end_pos - start_pos
dir_norm = direction / (np.linalg.norm(direction) + 1e-9)
raw_wps.sort(key=lambda p: float(np.dot(p - start_pos, dir_norm)))
# Funnel waypoint deep on the exit axis pulls the turtle onto the axis
# before the exit leader. Virtual lookahead past the exit then keeps
# heading aligned with -end_inward as we approach end_pos.
funnel = end_pos + end_inward * 400.0
lookahead = end_pos + (-end_inward) * 240.0
waypoints: List[np.ndarray] = raw_wps + [funnel, lookahead]
exit_theta = math.atan2(-end_inward[1], -end_inward[0])
entry_steps = max(2, int(leader_length / step_size))
wp_idx = 0
drift = 0.0
# Perp to end_inward (for lateral-offset measurement along perimeter edge).
perp = np.array([-end_inward[1], end_inward[0]])
for step_count in range(max_steps):
# Signed perpendicular distance from turtle to end_pos perimeter plane.
# end_inward points from perimeter INTO canvas, so on interior side
# dot(pos - end_pos, end_inward) > 0; crossing the plane → <= 0.
sd = ((x - end_pos[0]) * end_inward[0]
+ (y - end_pos[1]) * end_inward[1])
# Lateral offset along the perimeter edge from end_pos.
lat = abs((x - end_pos[0]) * perp[0] + (y - end_pos[1]) * perp[1])
# Termination: turtle has reached/crossed the perimeter.
if sd <= 0.0:
break
if step_count < entry_steps:
# Entry leader: no turn (straight).
turn = 0.0
elif sd < leader_length and lat < 40.0:
# Exit leader: pure head-alignment, no positional steering.
head_err = _angle_diff(theta, exit_theta)
turn = max(-max_turn, min(max_turn, head_err))
else:
# Body: waypoint steering.
is_final = (wp_idx == len(waypoints) - 1)
if is_final:
wx, wy = waypoints[-1]
pull = 0.18
else:
wx, wy = waypoints[wp_idx]
steer_dist = math.hypot(wx - x, wy - y)
# reach must exceed turning diameter (2 * step / max_turn)
if steer_dist < 140.0:
wp_idx += 1
continue
pull = 0.09
drift += rng.gauss(0, drift_rate)
drift = max(-max_turn * 0.7, min(max_turn * 0.7, drift))
turn = drift + rng.gauss(0, max_turn * 0.12)
target_ang = math.atan2(wy - y, wx - x)
steer = _angle_diff(theta, target_ang)
turn += steer * pull
turn = max(-max_turn, min(max_turn, turn))
theta += turn
x += step_size * math.cos(theta)
y += step_size * math.sin(theta)
pts.append(np.array([x, y]))
return np.array(pts)
# ── Intersection detection ─────────────────────────────────────────
def _segments_cross(p0, p1, q0, q1) -> bool:
eps = 1e-8
o1 = (p1[0]-p0[0])*(q0[1]-p0[1]) - (p1[1]-p0[1])*(q0[0]-p0[0])
o2 = (p1[0]-p0[0])*(q1[1]-p0[1]) - (p1[1]-p0[1])*(q1[0]-p0[0])
o3 = (q1[0]-q0[0])*(p0[1]-q0[1]) - (q1[1]-q0[1])*(p0[0]-q0[0])
o4 = (q1[0]-q0[0])*(p1[1]-q0[1]) - (q1[1]-q0[1])*(p1[0]-q0[0])
return ((o1 > eps and o2 < -eps) or (o1 < -eps and o2 > eps)) and \
((o3 > eps and o4 < -eps) or (o3 < -eps and o4 > eps))
def _crossing_angle(p0, p1, q0, q1) -> float:
d1 = p1 - p0
d2 = q1 - q0
cos_a = np.dot(d1, d2) / (np.linalg.norm(d1) * np.linalg.norm(d2) + 1e-12)
return math.degrees(math.acos(min(1.0, abs(cos_a))))
def _find_crossings(
poly_a: np.ndarray,
poly_b: np.ndarray,
same_curve: bool = False,
min_seg_gap: int = 10,
):
"""Returns (count, worst_angle, details).
`details` is a list of dicts, one per crossing, with keys:
i, j — segment indices on poly_a and poly_b
ti, tj — crossing parameters ∈ (0, 1) along each segment
px, py — exact crossing point
angle — crossing angle in degrees
For same_curve=True, i < j is enforced so each crossing appears once.
"""
na = len(poly_a) - 1
nb = len(poly_b) - 1
a_min_x = np.minimum(poly_a[:-1, 0], poly_a[1:, 0])
a_max_x = np.maximum(poly_a[:-1, 0], poly_a[1:, 0])
a_min_y = np.minimum(poly_a[:-1, 1], poly_a[1:, 1])
a_max_y = np.maximum(poly_a[:-1, 1], poly_a[1:, 1])
b_min_x = np.minimum(poly_b[:-1, 0], poly_b[1:, 0])
b_max_x = np.maximum(poly_b[:-1, 0], poly_b[1:, 0])
b_min_y = np.minimum(poly_b[:-1, 1], poly_b[1:, 1])
b_max_y = np.maximum(poly_b[:-1, 1], poly_b[1:, 1])
all_min_x = np.concatenate([a_min_x, b_min_x])
all_max_x = np.concatenate([a_max_x, b_max_x])
cell_size = max(np.median(all_max_x - all_min_x), 1.0) * 3
grid_a = defaultdict(list)
for i in range(na):
cx0 = int(a_min_x[i] / cell_size); cx1 = int(a_max_x[i] / cell_size)
cy0 = int(a_min_y[i] / cell_size); cy1 = int(a_max_y[i] / cell_size)
for gx in range(cx0, cx1 + 1):
for gy in range(cy0, cy1 + 1):
grid_a[(gx, gy)].append(i)
grid_b = defaultdict(list)
for j in range(nb):
cx0 = int(b_min_x[j] / cell_size); cx1 = int(b_max_x[j] / cell_size)
cy0 = int(b_min_y[j] / cell_size); cy1 = int(b_max_y[j] / cell_size)
for gx in range(cx0, cx1 + 1):
for gy in range(cy0, cy1 + 1):
grid_b[(gx, gy)].append(j)
checked = set()
worst_angle = 180.0
details: List[Dict] = []
for cell_key in grid_a:
if cell_key not in grid_b:
continue
for i in grid_a[cell_key]:
for j in grid_b[cell_key]:
if same_curve:
ii, jj = min(i, j), max(i, j)
if jj - ii < min_seg_gap:
continue
key = (ii, jj)
else:
key = (i, j)
if key in checked:
continue
checked.add(key)
if a_max_x[i] < b_min_x[j] or b_max_x[j] < a_min_x[i] or \
a_max_y[i] < b_min_y[j] or b_max_y[j] < a_min_y[i]:
continue
if same_curve:
si, sj = key # enforce i < j for self-crossings
p0, p1 = poly_a[si], poly_a[si + 1]
q0, q1 = poly_b[sj], poly_b[sj + 1]
else:
si, sj = i, j
p0, p1 = poly_a[si], poly_a[si + 1]
q0, q1 = poly_b[sj], poly_b[sj + 1]
if not _segments_cross(p0, p1, q0, q1):
continue
d1 = p1 - p0
d2 = q1 - q0
denom = d1[0] * d2[1] - d1[1] * d2[0]
if abs(denom) < 1e-12:
continue
ti = ((q0[0] - p0[0]) * d2[1] - (q0[1] - p0[1]) * d2[0]) / denom
tj = ((q0[0] - p0[0]) * d1[1] - (q0[1] - p0[1]) * d1[0]) / denom
px = float(p0[0] + ti * d1[0])
py = float(p0[1] + ti * d1[1])
angle = _crossing_angle(p0, p1, q0, q1)
worst_angle = min(worst_angle, angle)
details.append({
"i": int(si), "j": int(sj),
"ti": float(ti), "tj": float(tj),
"px": px, "py": py, "angle": float(angle),
})
return len(details), worst_angle, details
def _details_to_points(details: List[Dict]) -> np.ndarray:
if not details:
return np.zeros((0, 2))
return np.array([[d["px"], d["py"]] for d in details])
def _curves_too_close(
poly_a: np.ndarray,
poly_b: np.ndarray,
same_curve: bool = False,
min_dist: float = 7.0,
sample_step: int = 3,
self_index_gap: int = 50,
known_crossings: np.ndarray | None = None,
crossing_exclude_radius: float = 30.0,
) -> bool:
"""Flag if ANY sample point on A comes within `min_dist` of B without
being attributable to a real crossing. Samples within
`crossing_exclude_radius` of any point in `known_crossings` are ignored
(that's the crossing region itself, where near-zero distance is expected).
"""
nb = len(poly_b) - 1
if nb <= 0 or len(poly_a) == 0:
return False
b_pts = poly_b[:-1]
b_vecs = poly_b[1:] - poly_b[:-1]
b_lens_sq = np.maximum((b_vecs ** 2).sum(axis=1), 1e-12)
has_crossings = known_crossings is not None and len(known_crossings) > 0
for idx in range(0, len(poly_a), sample_step):
px, py = poly_a[idx]
p = np.array([px, py])
dp = p - b_pts
t = (dp * b_vecs).sum(axis=1) / b_lens_sq
t = np.clip(t, 0.0, 1.0)
proj = b_pts + t[:, None] * b_vecs
dists = np.sqrt(((p - proj) ** 2).sum(axis=1))
if same_curve:
mask = np.abs(np.arange(nb) - idx) < self_index_gap
dists[mask] = 9999.0
if dists.min() < min_dist:
if has_crossings:
cross_dists = np.sqrt(((p - known_crossings) ** 2).sum(axis=1))
if cross_dists.min() < crossing_exclude_radius:
continue # close approach explained by a real crossing
return True
return False
# ── Instance sampling ──────────────────────────────────────────────
def sample_instance(
rng: random.Random,
width: int,
height: int,
min_lines: int = 4,
max_lines: int = 4,
min_total_crossings: int = 12,
max_total_crossings: int = 60,
perimeter_margin: int = 70,
leader_length: float = 150.0,
interior_margin: int = 260,
max_attempts: int = 2500,
) -> Dict | None:
labels = list(string.ascii_uppercase)
for _ in range(max_attempts):
num_lines = rng.randint(min_lines, max_lines)
anchors = sample_perimeter_anchors(
rng, num_lines * 2, width, height,
margin=perimeter_margin,
)
if anchors is None:
continue
# Pair anchor i with anchor i+num_lines: opposite-side pairing to
# force curves through the interior.
pairs = [(anchors[i], anchors[i + num_lines]) for i in range(num_lines)]
rng.shuffle(pairs)
polylines = []
for (sp, sd), (ep, ed) in pairs:
poly = build_anchored_curve(
rng, width, height,
start_pos=sp, start_inward=sd,
end_pos=ep, end_inward=ed,
leader_length=leader_length,
step_size=rng.uniform(3.0, 3.8),
max_turn=rng.uniform(0.034, 0.044),
drift_rate=rng.uniform(0.010, 0.016),
num_waypoints=rng.randint(4, 6),
interior_margin=interior_margin,
)
polylines.append(poly)
total = 0
global_min_angle = 180.0
self_details: Dict[int, List[Dict]] = {}
pair_details: Dict[Tuple[int, int], List[Dict]] = {}
for a in range(num_lines):
sc, ang, det = _find_crossings(polylines[a], polylines[a],
same_curve=True)
self_details[a] = det
total += sc
global_min_angle = min(global_min_angle, ang)
for b in range(a + 1, num_lines):
cc, ang, det = _find_crossings(polylines[a], polylines[b])
pair_details[(a, b)] = det
total += cc
global_min_angle = min(global_min_angle, ang)
if total < min_total_crossings or total > max_total_crossings:
continue
if global_min_angle < 45.0:
continue
# 1% of the canvas's short side keeps curves visibly separated
# at any resolution while still letting the sampler converge at
# higher difficulties.
min_dist_px = 0.01 * float(min(width, height))
too_close = False
for a in range(num_lines):
if _curves_too_close(polylines[a], polylines[a], same_curve=True,
min_dist=min_dist_px,
known_crossings=_details_to_points(self_details[a])):
too_close = True; break
for b in range(a + 1, num_lines):
if _curves_too_close(polylines[a], polylines[b],
min_dist=min_dist_px,
known_crossings=_details_to_points(pair_details[(a, b)])):
too_close = True; break
if too_close:
break
if too_close:
continue
# Reject if any two crossing points are too close to each other —
# that would let multiple intersections be mistaken for one.
all_cross_pts = []
for a in range(num_lines):
pts = _details_to_points(self_details[a])
if len(pts) > 0:
all_cross_pts.append(pts)
for b in range(a + 1, num_lines):
pts = _details_to_points(pair_details[(a, b)])
if len(pts) > 0:
all_cross_pts.append(pts)
if all_cross_pts:
stacked = np.concatenate(all_cross_pts, axis=0)
if len(stacked) >= 2:
diffs = stacked[:, None, :] - stacked[None, :, :]
dmat = np.sqrt((diffs ** 2).sum(axis=-1))
np.fill_diagonal(dmat, np.inf)
if dmat.min() < 20.0:
continue
# Pick label positions: one of the two endpoints per curve.
# "label_at_start" picks the first endpoint (index 0) when True.
label_starts = [rng.random() < 0.5 for _ in range(num_lines)]
# Per-curve sequence length: each self-crossing is hit twice during
# the trace (entering + leaving), each pair-crossing once.
seq_lens = []
for q in range(num_lines):
sl = 2 * len(self_details[q])
for b in range(num_lines):
if b == q:
continue
key = (min(q, b), max(q, b))
sl += len(pair_details[key])
seq_lens.append(sl)
# Prefer a query line whose sequence length is near ~10.
TARGET = 10
candidates = sorted(
range(num_lines), key=lambda i: abs(seq_lens[i] - TARGET)
)
query_line = candidates[0]
seq_len = seq_lens[query_line]
if seq_len < 8 or seq_len > 13:
continue
query_label = labels[query_line]
line_labels = labels[:num_lines]
# Build the ordered crossing sequence along the query curve.
query_poly = polylines[query_line]
qdiffs = np.diff(query_poly, axis=0)
qseg_lens = np.sqrt((qdiffs ** 2).sum(axis=1))
qcum = np.concatenate([[0.0], np.cumsum(qseg_lens)])
entries: List[Tuple[float, str]] = []
# Self-crossings contribute two entries (both passes), labeled Q.
for d in self_details[query_line]:
pos_a = qcum[d["i"]] + d["ti"] * qseg_lens[d["i"]]
pos_b = qcum[d["j"]] + d["tj"] * qseg_lens[d["j"]]
entries.append((pos_a, query_label))
entries.append((pos_b, query_label))
# Pair-crossings contribute one entry, labeled with the other curve.
for other in range(num_lines):
if other == query_line:
continue
if query_line < other:
dets = pair_details[(query_line, other)]
for d in dets:
pos = qcum[d["i"]] + d["ti"] * qseg_lens[d["i"]]
entries.append((pos, labels[other]))
else:
dets = pair_details[(other, query_line)]
for d in dets:
# query curve is poly_b in this pair → use j, tj
pos = qcum[d["j"]] + d["tj"] * qseg_lens[d["j"]]
entries.append((pos, labels[other]))
# Orientation: labeled endpoint at poly[0] → forward arc order;
# labeled endpoint at poly[-1] → reverse arc order.
trace_forward = label_starts[query_line]
entries.sort(key=lambda e: e[0] if trace_forward else -e[0])
answer_sequence = [label for _, label in entries]
answer = ", ".join(answer_sequence)
question = (
f"The image shows {num_lines} curves. Each curve has one labeled "
f"endpoint on the perimeter (labels: {', '.join(line_labels)}). "
f"Starting from the labeled endpoint of curve {query_label}, follow "
f"curve {query_label} through the interior tangle. At every "
f"crossing you pass through, record the label of the OTHER curve "
f"involved — when curve {query_label} crosses itself, record "
f"{query_label}. Report the ordered sequence of labels, separated "
f"by commas (for example: A, C, B, C, A). "
f"Provide your final answer enclosed in <answer>...</answer> tags."
)
return {
"width": width,
"height": height,
"num_lines": num_lines,
"query_line": query_line,
"query_label": query_label,
"line_labels": line_labels,
"total_crossings": total,
"sequence_length": seq_len,
"polylines": polylines,
"label_starts": label_starts,
"question": question,
"answer": answer,
}
return None
# ── Rendering ──────────────────────────────────────────────────────
def label_anchor_for_endpoint(
polyline: np.ndarray,
at_start: bool,
width: int,
height: int,
offset: float = 46.0,
tangent_span: int = 10,
) -> Tuple[float, float]:
"""Offset a label perpendicular to the leader direction at a labeled
curve endpoint. Prefer the side that keeps the label inside the canvas
and farther from the polyline."""
if at_start:
pt = polyline[0]
tan = polyline[min(tangent_span, len(polyline) - 1)] - polyline[0]
else:
pt = polyline[-1]
tan = polyline[-1] - polyline[max(-tangent_span - 1, -len(polyline))]
tnorm = math.hypot(float(tan[0]), float(tan[1]))
if tnorm < 1e-8:
ux, uy = 0.0, -1.0
else:
ux = float(tan[0]) / tnorm
uy = float(tan[1]) / tnorm
perps = [(-uy, ux), (uy, -ux)]
best = None
best_score = -1e9
for px, py in perps:
cx = float(pt[0]) + px * offset
cy = float(pt[1]) + py * offset
score = 0.0
if 30 < cx < width - 30 and 30 < cy < height - 30:
score += 200.0
diffs = polyline - np.array([cx, cy])
dists = np.sqrt((diffs ** 2).sum(axis=1))
score += float(dists.min())
if score > best_score:
best_score = score
best = (cx, cy)
cx, cy = best
return (
float(np.clip(cx, 28, width - 28)),
float(np.clip(cy, 28, height - 28)),
)
def render_instance(out_path: Path, record: Dict, noise_seed: int) -> None:
width = int(record["width"])
height = int(record["height"])
polylines = record["polylines"]
num_lines = record["num_lines"]
line_labels = record["line_labels"]
label_starts = record["label_starts"]
fig = plt.figure(figsize=(width / 100, height / 100), dpi=100)
ax = fig.add_axes([0, 0, 1, 1])
ax.set_xlim(0, width)
ax.set_ylim(height, 0)
ax.axis("off")
ax.set_facecolor("#f3efe8")
nrng = np.random.default_rng(noise_seed)
noise = nrng.normal(0.0, 1.0, size=(height, width))
noise = (noise - noise.min()) / max(noise.max() - noise.min(), 1e-6)
ax.imshow(noise, cmap="Greys", alpha=0.06, extent=(0, width, height, 0),
interpolation="bilinear")
for i in range(num_lines):
poly = polylines[i]
ax.plot(
poly[:, 0], poly[:, 1],
color=LINE_COLOR, linewidth=2.8,
solid_capstyle="round", solid_joinstyle="round",
zorder=2.0 + i * 0.05,
)
for i in range(num_lines):
poly = polylines[i]
labeled_pt = poly[0] if label_starts[i] else poly[-1]
lc = LABEL_COLORS[i % len(LABEL_COLORS)]
lx, ly = label_anchor_for_endpoint(
poly, at_start=label_starts[i], width=width, height=height,
)
# Dashed connector from endpoint to offset label
ax.plot([labeled_pt[0], lx], [labeled_pt[1], ly],
color="#9b7b56", linewidth=1.0, alpha=0.7,
linestyle=(0, (2.0, 3.2)), zorder=3.8)
# Dot at the actual endpoint
ax.scatter([labeled_pt[0]], [labeled_pt[1]], s=110,
facecolors=lc, edgecolors="white", linewidths=1.8, zorder=4.0)
# Label at offset position
ax.text(lx, ly, line_labels[i],
fontsize=17, fontweight="bold", color=lc,
ha="center", va="center", zorder=5.0,
bbox=dict(boxstyle="round,pad=0.2", facecolor="#f3efe8",
edgecolor=lc, alpha=0.95, linewidth=1.2))
fig.savefig(out_path, dpi=100, bbox_inches="tight", pad_inches=0)
plt.close(fig)
# ── Main ───────────────────────────────────────────────────────────
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--output-root", type=Path, required=True)
parser.add_argument("--count", type=int, default=20)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--width", type=int, default=1024)
parser.add_argument("--height", type=int, default=1024)
parser.add_argument("--min-lines", type=int, default=3)
parser.add_argument("--max-lines", type=int, default=5)
parser.add_argument("--min-crossings", type=int, default=3)
parser.add_argument("--max-crossings", type=int, default=25)
parser.add_argument("--difficulty", type=int, default=5,
help="Integer difficulty >=0; scales line and crossing counts.")
parser.add_argument("--workers", type=int, default=8)
args = parser.parse_args()
d = max(0, int(args.difficulty))
# num_lines grows as 3 + d//3 (no cap). num_crossings ∈ [3, 5 + 2d].
# Fixed 45° minimum crossing angle (handled inside sample_instance).
args.min_lines = 3 + d // 3
args.max_lines = 3 + d // 3
args.min_crossings = 10
args.max_crossings = 10 + 2 * d
# Canvas scaling: area ∝ N_d / N_0 with N_d = 5 + 2d.
N_d = 5 + 2 * d
N_0 = 5
s = math.sqrt(max(1.0, N_d / N_0))
args.width = int(round(args.width * s))
args.height = int(round(args.height * s))
out_root = args.output_root
img_dir = out_root / "images"
img_dir.mkdir(parents=True, exist_ok=True)
sys.path.insert(0, str(Path(__file__).resolve().parents[3]))
from _sample_pool import parallel_sample_records # noqa: E402
def _attempt(rng):
return sample_instance(
rng, args.width, args.height,
min_lines=args.min_lines, max_lines=args.max_lines,
min_total_crossings=args.min_crossings,
max_total_crossings=args.max_crossings,
max_attempts=50,
)
records_raw = parallel_sample_records(
_attempt, count=args.count, workers=args.workers,
seed_base=args.seed,
)
rng_render = random.Random(args.seed ^ 0xA5A5)
records = []
for idx, record in enumerate(records_raw):
name = f"line_intersections_{idx:05d}.png"
ns = rng_render.randint(0, 10**9)
render_instance(img_dir / name, record, noise_seed=ns)
record.pop("polylines")
record["image"] = f"images/{name}"
records.append(record)
print(f" {len(records)}/{args.count} valid samples (workers={args.workers})")
with (out_root / "annotations.jsonl").open("w") as fh:
for r in records:
fh.write(json.dumps(r) + "\n")
data_json = {
"task": "line_intersections",
"category": "sequential_traversal",
"count": len(records),
"items": [
{"image": r["image"], "question": r["question"], "answer": r["answer"]}
for r in records
],
}
(out_root / "data.json").write_text(json.dumps(data_json, indent=2))
print(f"Saved {len(records)} items to {out_root}")
if __name__ == "__main__":
main()