Datasets:
Tasks:
Visual Question Answering
Formats:
parquet
Size:
1K - 10K
Tags:
multimodal
vision-language-model
fine-grained-perception
self-distillation
on-policy-distillation
paired-images
License:
File size: 16,550 Bytes
498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a e01c86c 498228a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 | """
Construct the candidate paired-view pool for VisAlloc training.
Usage:
python scripts/prepare_data.py --data-dir ./cache/visalloc
This script:
1. Downloads the Vision-OPD-6K dataset from HuggingFace
(train.jsonl, student images, teacher crops)
2. Generates the negative teacher views: a red box drawn at a displaced
position (IoU < 0.1 vs the ground-truth box), using the exact recovered
positive crop window size, within-crop box, Pillow stroke, and 2x resize.
The official positive is reconstructed pixel-exactly before generation
3. Converts train.jsonl to train.parquet and attaches the `neg_bbox_images`
column for all source candidates; this script does not run paired-teacher
filtering or select the final 5,795 training samples
Outputs (under --data-dir):
images/ official student images (red-box full image)
teacher_images/ official positive teacher crops
teacher_neg/ generated negative teacher crops
results.json per-sample generation metadata (boxes, IoU)
train.parquet candidate pool with local image paths, before sample filtering
"""
from __future__ import annotations
import argparse
import hashlib
import itertools
import json
import os
import random
import subprocess
import sys
import time
from multiprocessing import Pool
from typing import Any
import cv2
import datasets
import numpy as np
from PIL import Image, ImageDraw
FRAME_WIDTH = 5 # verified against every official teacher image
NEG_IOU_LIMIT = 0.1
UPSTREAM_REVISION = "eb5c1c2e7b9a7b6a619efe4161c7369c71bf8af4"
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Prepare the VisAlloc candidate paired-view pool.")
parser.add_argument("--data-dir", default="./cache/visalloc", help="Output directory")
parser.add_argument("--hf-repo", default="yuanqianhao/Vision-OPD-6K", help="HuggingFace dataset repo")
parser.add_argument("--revision", default=UPSTREAM_REVISION, help="Pinned upstream dataset revision")
parser.add_argument("--skip-download", action="store_true", help="Skip downloading, only preprocess")
parser.add_argument("--nproc", type=int, default=16, help="Worker processes for negative-view generation")
return parser.parse_args()
# ---------------------------------------------------------------- download --
def download_dataset(repo_id: str, data_dir: str, revision: str = UPSTREAM_REVISION) -> None:
print(f"Downloading dataset from {repo_id} ...")
from huggingface_hub import snapshot_download
snapshot_download(repo_id, repo_type="dataset", revision=revision, local_dir=data_dir)
# Extract every shipped image bundle (multi-part tars are cat-joined).
for sub, prefix in (("images", "images.tar.gz"),
("teacher_images", "teacher_images.tar.gz"),
("original_images", "original_images.tar.gz")):
d = os.path.join(data_dir, sub)
if not os.path.isdir(d):
print(f"Warning: {d} not present in the download", file=sys.stderr)
continue
tars = sorted(f for f in os.listdir(d) if f.startswith(prefix))
if not tars:
continue
print(f"Extracting {sub} ...")
cat_process = subprocess.Popen(["cat", *tars], cwd=d, stdout=subprocess.PIPE)
try:
subprocess.run(["tar", "-xf", "-", "-C", "."], cwd=d, stdin=cat_process.stdout, check=True)
finally:
if cat_process.stdout is not None:
cat_process.stdout.close()
cat_returncode = cat_process.wait()
if cat_returncode != 0:
raise subprocess.CalledProcessError(cat_process.returncode, cat_process.args)
for f in tars:
os.remove(os.path.join(d, f))
print("Image extraction complete.")
# ------------------------------------------------------- negative views --
def iou(a, b):
ix0, iy0 = max(a[0], b[0]), max(a[1], b[1])
ix1, iy1 = min(a[2], b[2]), min(a[3], b[3])
inter = max(0, ix1 - ix0) * max(0, iy1 - iy0)
ua = (a[2] - a[0]) * (a[3] - a[1]) + (b[2] - b[0]) * (b[3] - b[1]) - inter
return inter / ua if ua else 0.0
def project_bands(values):
indices = np.flatnonzero(values >= values.max() * 0.65)
return [group for group in np.split(indices, np.flatnonzero(np.diff(indices) > 1) + 1) if len(group)]
def recover_geometry(original, official, gt):
w, h = official.size
if w % 2 or h % 2:
raise ValueError("Official teacher dimensions must match the verified 2x pipeline")
cw, ch = w // 2, h // 2
pixels = np.array(official)
red = (pixels[:, :, 0] > 245) & (pixels[:, :, 1] < 25) & (pixels[:, :, 2] < 25)
xb, yb = project_bands(red.sum(0)), project_bands(red.sum(1))
if len(xb) < 2 or len(yb) < 2:
return recover_difficult(original, official, gt)
box = [int(xb[0][0]) // 2, int(yb[0][0]) // 2, int(xb[-1][-1]) // 2, int(yb[-1][-1]) // 2]
if xb[-1][-1] == w - 1:
box[2] = cw
if yb[-1][-1] == h - 1:
box[3] = ch
x = max(0, min(original.width - cw, int((gt[0] + gt[2] - cw) / 2)))
y = max(0, min(original.height - ch, int((gt[1] + gt[3] - ch) / 2)))
small = np.array(official.resize((cw, ch), Image.Resampling.BOX))
strong = (
(small[:, :, 0].astype(int) > small[:, :, 1] * 1.5)
& (small[:, :, 0].astype(int) > small[:, :, 2] * 1.5)
& (small[:, :, 0] > 150)
)
valid = 1 - cv2.dilate(strong.astype("uint8"), np.ones((7, 7), np.uint8))
sx, sy = max(0, x - 4), max(0, y - 4)
ex, ey = min(original.width, x + cw + 5), min(original.height, y + ch + 5)
if valid.sum() == 0:
return recover_difficult(original, official, gt)
scores = cv2.matchTemplate(
np.array(original.crop((sx, sy, ex, ey))).astype("float32"),
small.astype("float32"),
cv2.TM_SQDIFF,
mask=np.repeat(valid[:, :, None], 3, axis=2),
)
_, _, loc, _ = cv2.minMaxLoc(scores)
ox, oy = sx + loc[0], sy + loc[1]
source_crop = original.crop((ox, oy, ox + cw, oy + ch))
def compare(candidate):
drawn = source_crop.copy()
ImageDraw.Draw(drawn).rectangle(candidate, outline=(255, 0, 0), width=FRAME_WIDTH)
return np.array(drawn.resize((w, h), Image.Resampling.LANCZOS))
rebuilt = compare(box)
if not np.array_equal(rebuilt, pixels):
# Natural red content can move a threshold edge by one pixel.
choices = [[v, v - 1, v + 1] for v in box]
for candidate in itertools.product(*choices):
if np.array_equal(compare(candidate), pixels):
box = list(candidate)
break
else:
return recover_difficult(original, official, gt)
return {
"positive_crop_window": [ox, oy, ox + cw, oy + ch],
"frame_box_in_crop": box,
"source_crop_size": [cw, ch],
"output_size": [w, h],
"resize_scale": [2, 2],
"frame_width": FRAME_WIDTH,
"reconstruction_exact": True,
}
def recover_difficult(original, official, gt):
w, h = official.size
cw, ch = w // 2, h // 2
pixels = np.array(official)
x = max(0, min(original.width - cw, int((gt[0] + gt[2] - cw) / 2)))
y = max(0, min(original.height - ch, int((gt[1] + gt[3] - ch) / 2)))
offsets = [0, 1, -1, 2, -2]
for ox, oy in itertools.product(
[x + v for v in offsets if 0 <= x + v <= original.width - cw],
[y + v for v in offsets if 0 <= y + v <= original.height - ch],
):
crop = original.crop((ox, oy, ox + cw, oy + ch))
bw, bh = gt[2] - gt[0], gt[3] - gt[1]
estimate = [
int(gt[0] - bw * 0.05) - ox,
int(gt[1] - bh * 0.05) - oy,
int(gt[2] + bw * 0.05) - ox,
int(gt[3] + bh * 0.05) - oy,
]
options = [[v + d for d in [0, 1, -1, 2, -2]] for v in estimate]
for box in itertools.product(*options):
if box[2] < box[0] or box[3] < box[1]:
continue
drawn = crop.copy()
ImageDraw.Draw(drawn).rectangle(box, outline="red", width=FRAME_WIDTH)
if np.array_equal(np.array(drawn.resize((w, h), Image.Resampling.LANCZOS)), pixels):
return {
"positive_crop_window": [ox, oy, ox + cw, oy + ch],
"frame_box_in_crop": list(box),
"source_crop_size": [cw, ch],
"output_size": [w, h],
"resize_scale": [2, 2],
"frame_width": FRAME_WIDTH,
"reconstruction_exact": True,
}
raise ValueError("No exact transform found for difficult sample")
_ROWS = None
_DATA_DIR = None
def _init_worker(rows, data_dir):
"""Pool initializer: safe under both fork and spawn start methods."""
global _ROWS, _DATA_DIR
_ROWS = rows
_DATA_DIR = data_dir
def render_teacher(original, window, geometry):
crop = original.crop(window)
if list(crop.size) != geometry["source_crop_size"]:
raise ValueError("Crop size differs from the official positive template")
ImageDraw.Draw(crop).rectangle(geometry["frame_box_in_crop"], outline=(255, 0, 0), width=FRAME_WIDTH)
return crop.resize(tuple(geometry["output_size"]), Image.Resampling.LANCZOS)
def sample_negative_window(gt, positive_window, image_size, rng):
"""Translate the whole official crop, preserving every within-crop coordinate."""
W, H = image_size
px, py, x2, y2 = positive_window
cw, ch = x2 - px, y2 - py
candidates = ((rng.randint(0, W - cw), rng.randint(0, H - ch)) for _ in range(2000))
corners = itertools.product((0, W - cw), (0, H - ch))
for x, y in itertools.chain(candidates, corners):
negative = [gt[0] + x - px, gt[1] + y - py, gt[2] + x - px, gt[3] + y - py]
if iou(negative, gt) < NEG_IOU_LIMIT:
return [x, y, x + cw, y + ch], negative
raise ValueError("No geometry-preserving negative window satisfies IoU < 0.1")
def _worker(i):
h = _ROWS[i]
try:
data_dir = _DATA_DIR
src_tpos = os.path.join(data_dir, h["teacher_images"][0])
with Image.open(os.path.join(data_dir, h["original_images"][0])) as image:
original = image.convert("RGB")
with Image.open(src_tpos) as image:
official = image.convert("RGB")
gt = h["bbox"]
geometry = recover_geometry(original, official, gt)
positive_window = geometry["positive_crop_window"]
reconstructed = render_teacher(original, positive_window, geometry)
if not np.array_equal(np.asarray(reconstructed), np.asarray(official)):
raise ValueError("Official positive reconstruction is not pixel-exact")
negative_window, negative_bbox = sample_negative_window(
gt, positive_window, original.size, random.Random(9000 + i)
)
negative = render_teacher(original, negative_window, geometry)
path = os.path.join(data_dir, "teacher_neg", f"{i:06d}.png")
negative.save(path, compress_level=1)
frame = Image.new("L", tuple(geometry["source_crop_size"]), 0)
ImageDraw.Draw(frame).rectangle(geometry["frame_box_in_crop"], outline=255, width=FRAME_WIDTH)
frame = frame.resize(official.size, Image.Resampling.LANCZOS)
with open(src_tpos, "rb") as stream:
positive_sha256 = hashlib.sha256(stream.read()).hexdigest()
with open(path, "rb") as stream:
negative_sha256 = hashlib.sha256(stream.read()).hexdigest()
return {
"idx": i,
"ok": True,
"gt_bbox": list(gt),
"neg_bbox": negative_bbox,
"iou": iou(negative_bbox, gt),
"crop_size": list(official.size),
"negative_crop_window": negative_window,
"original_size": list(original.size),
"target_bbox_in_crop": [
gt[0] - positive_window[0],
gt[1] - positive_window[1],
gt[2] - positive_window[0],
gt[3] - positive_window[1],
],
"positive_sha256": positive_sha256,
"negative_sha256": negative_sha256,
"shared_frame_mask_sha256": hashlib.sha256(frame.tobytes()).hexdigest(),
"generation_version": "2.0.0",
"resampling": "PIL.LANCZOS",
"renderer": "PIL.ImageDraw.rectangle",
**geometry,
}
except Exception as e:
return {"idx": i, "ok": False, "reason": f"exc:{type(e).__name__}:{str(e)[:180]}"}
def generate_negative_views(data_dir: str, nproc: int) -> list:
with open(os.path.join(data_dir, "train.jsonl"), encoding="utf-8") as f:
rows = [json.loads(line) for line in f]
os.makedirs(os.path.join(data_dir, "teacher_neg"), exist_ok=True)
t0 = time.time()
results = []
with Pool(nproc, initializer=_init_worker, initargs=(rows, data_dir)) as pool:
for n, r in enumerate(pool.imap_unordered(_worker, range(len(rows)), chunksize=16), 1):
results.append(r)
if n % 500 == 0:
ok = sum(1 for x in results if x.get("ok"))
print(f"[{n}/{len(rows)}] ok={ok} elapsed={time.time() - t0:.0f}s", flush=True)
results.sort(key=lambda r: r["idx"])
with open(os.path.join(data_dir, "results.json"), "w", encoding="utf-8") as f:
json.dump(results, f, ensure_ascii=False, indent=1)
ok = sum(1 for r in results if r.get("ok"))
success_rate = 100 * ok / len(results) if results else 0.0
print(f"Negative views: {ok}/{len(results)} ({success_rate:.1f}%) "
f"in {(time.time() - t0) / 60:.1f} min")
if ok != len(results):
raise RuntimeError("Negative generation incomplete; inspect results.json before using the dataset")
return results
# ---------------------------------------------------------------- parquet --
def clean_question(problem: str) -> str:
text = (problem or "").replace("<image>", "").strip()
hint = ("Only focus on the objects inside the red bounding box in the image "
"to answer this question.")
text = text.replace(f"\n\n{hint}", "").replace(hint, "")
return text.strip()
def build_record(item: dict[str, Any], data_dir: str, neg_path: str | None) -> dict[str, Any]:
record = {
"data_source": "zwz_rl_vqa_bbox_teacher",
"prompt": [{"role": "user", "content": item["problem"]}],
"images": [{"path": os.path.join(data_dir, item["images"][0])}],
"bbox_images": [{"path": os.path.join(data_dir, item["teacher_images"][0])}],
"ability": "visual_question_answering",
"reward_model": {"style": "none", "ground_truth": item.get("answer", "")},
"extra_info": {
"answer": item.get("answer", ""),
"question": clean_question(item.get("problem", "")),
"source_extra_info": item.get("extra_info", {}),
},
}
record["neg_bbox_images"] = [{"path": neg_path}] if neg_path is not None else []
return record
def convert_to_parquet(data_dir: str, results: list) -> None:
jsonl_path = os.path.join(data_dir, "train.jsonl")
if not os.path.exists(jsonl_path):
print(f"Error: {jsonl_path} not found", file=sys.stderr)
sys.exit(1)
by_idx_ok = {r["idx"]: r.get("ok", False) for r in results}
print("Converting train.jsonl to train.parquet ...")
records = []
with open(jsonl_path, encoding="utf-8") as f:
for i, line in enumerate(f):
item = json.loads(line)
neg = os.path.join(data_dir, "teacher_neg", f"{i:06d}.png") if by_idx_ok.get(i) else None
records.append(build_record(item, data_dir, neg))
dataset = datasets.Dataset.from_list(records)
output_path = os.path.join(data_dir, "train.parquet")
dataset.to_parquet(output_path)
n_neg = sum(1 for r in records if "neg_bbox_images" in r)
print(f"Saved {len(records)} records ({n_neg} with negative views) to {output_path}")
def main() -> None:
args = parse_args()
data_dir = os.path.abspath(args.data_dir)
os.makedirs(data_dir, exist_ok=True)
if not args.skip_download:
download_dataset(args.hf_repo, data_dir, args.revision)
results = generate_negative_views(data_dir, args.nproc)
convert_to_parquet(data_dir, results)
print(f"\nCandidate pool preparation complete. Apply sample selection before training: {data_dir}/train.parquet")
if __name__ == "__main__":
main()
|