File size: 6,058 Bytes
84591c3 07c6451 b97e423 e2cf5a4 b97e423 84591c3 07c6451 84591c3 07c6451 84591c3 b97e423 84591c3 | 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 | """Show a stimulus (image, prompt, segmentation) or a whole minimal pair.
python scripts/visualize.py S46 # one stimulus
python scripts/visualize.py X-REL-CAT-LEFT-DOG-RICH # every member of a pair, side by side
python scripts/visualize.py S46 --out s46.png # save instead of opening a window
python scripts/visualize.py --list # list the stimuli and the pairs
Objects named in the prompt are outlined in green, unmentioned ones in orange, negated ones
("with no lamp") in purple; background regions (sky, wall, floor...) are labelled in grey.
Run it from a local copy of the dataset (git clone, or huggingface_hub.snapshot_download), or point
--root at one. Needs: pip install matplotlib pillow numpy
"""
import argparse
import json
import sys
import textwrap
from pathlib import Path
import numpy as np
from PIL import Image
STATUS = {"mentioned": "#2e9e44", "unmentioned": "#e07b00", "negated": "#8e44ad", "background": "#7f7f7f"}
def load(root):
rows = [json.loads(l) for l in open(root / "data" / "metadata.jsonl", encoding="utf-8") if l.strip()]
pairs = [json.loads(l) for l in open(root / "pairs" / "pairs.jsonl", encoding="utf-8") if l.strip()]
return {r["sample_id"]: r for r in rows}, {p["pair_id"]: p for p in pairs}
def segment_ids(root, row):
seg = np.asarray(Image.open(root / row["segmentation"]).convert("RGB"), dtype=np.uint32)
return seg[..., 0] + 256 * seg[..., 1] + 256 ** 2 * seg[..., 2]
def status(obj, stuff_ids):
if obj["negated"]:
return "negated"
if obj["mentioned"]:
return "mentioned"
return "background" if obj["segment_id"] in stuff_ids else "unmentioned"
def draw(ax_img, ax_seg, root, row, role=None):
import matplotlib.patches as patches
image = np.asarray(Image.open(root / "data" / row["file_name"]).convert("RGB"))
title = f'{row["sample_id"]}{f" ({role})" if role else ""}: "{row["prompt"]}"'
ax_img.imshow(image)
ax_img.set_title(textwrap.fill(title, 75), fontsize=10)
if row.get("control_type"):
sub = f'control [{row["control_type"]}]'
else:
sub = f'{row["case_id"]} {row["subcategory"]}' + (f' [{row["case_variant"]}]' if row.get("case_variant") else "")
ax_img.set_xlabel(textwrap.fill(f"{sub}. Unmentioned: {row['unmentioned_content']}", 80), fontsize=8)
for ax in (ax_img, ax_seg):
ax.set_xticks([])
ax.set_yticks([])
if not row.get("segmentation"): # generated images: no segmentation yet, show the human description
ax_seg.axis("off")
text = row.get("description") or "(no segmentation)"
if row.get("generation_prompt"):
text += f'\n\nGeneration prompt ({row["image_source"]}, seed {row.get("generation_seed")}): {row["generation_prompt"]}'
ax_seg.text(0, 1, textwrap.fill(text, 70, replace_whitespace=False), va="top", fontsize=8, wrap=True)
return
ids = segment_ids(root, row)
info = json.load(open((root / row["segmentation"]).with_suffix(".json"), encoding="utf-8"))["segments_info"]
stuff_ids = {s["id"] for s in info if not s["isthing"]}
rng = np.random.default_rng(0)
overlay = image.astype(float) * 0.35
for obj in row["objects"]:
m = ids == obj["segment_id"]
overlay[m] += rng.integers(60, 255, 3) * 0.65
ax_seg.imshow(overlay.clip(0, 255).astype(np.uint8))
ax_seg.set_title("segmentation", fontsize=10)
for obj in row["objects"]:
st = status(obj, stuff_ids)
x, y, w, h = obj["bbox"]
if st != "background":
ax_seg.add_patch(patches.Rectangle((x, y), w, h, fill=False, edgecolor=STATUS[st], linewidth=1.8))
ax_seg.text(x + 3, y + 14, obj["label"], fontsize=7, color="white",
bbox=dict(facecolor=STATUS[st], edgecolor="none", pad=1.2, alpha=0.9))
def main():
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("id", nargs="?", help="a sample_id (S46) or a pair_id (P-C2, X-REL-...)")
p.add_argument("--root", type=Path, default=Path(__file__).resolve().parent.parent, help="dataset folder")
p.add_argument("--out", type=Path, help="save the figure to this file instead of showing it")
p.add_argument("--list", action="store_true", help="list the stimuli and the pairs")
args = p.parse_args()
samples, pairs = load(args.root)
if args.list or not args.id:
for s in samples.values():
print(f'{s["sample_id"]:5} {s["case_id"] or "ctl":3} "{s["prompt"]}"')
print()
for pr in pairs.values():
print(f'{pr["pair_id"]:30} {pr["axis"]:6} {pr["family_expected"]:11} {", ".join(m["sample_id"] for m in pr["members"])}')
return
import matplotlib
if args.out:
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.lines import Line2D
if args.id in samples:
members, title = [(samples[args.id], None)], None
elif args.id in pairs:
pr = pairs[args.id]
members = [(samples[m["sample_id"]], m["role"]) for m in pr["members"]]
title = f'{pr["pair_id"]}: same {"prompt" if pr["axis"] == "image" else "image"}, ' \
f'expected {pr["family_expected"]}'
else:
sys.exit(f"unknown id {args.id!r} (try --list)")
fig, axes = plt.subplots(len(members), 2, figsize=(11, 3.9 * len(members) + 0.4), squeeze=False)
for (row, role), (ax_img, ax_seg) in zip(members, axes):
draw(ax_img, ax_seg, args.root, row, role)
legend = [Line2D([0], [0], color=c, lw=4, label=k) for k, c in STATUS.items()]
fig.legend(handles=legend, loc="lower center", ncol=4, fontsize=8, frameon=False)
if title:
fig.suptitle(title, fontsize=11)
fig.tight_layout(rect=(0, 0.06 / len(members), 1, 1 - 0.04 / len(members) if title else 1))
if args.out:
fig.savefig(args.out, dpi=110)
print(f"saved {args.out}")
else:
plt.show()
if __name__ == "__main__":
main()
|