File size: 9,110 Bytes
63a1291 | 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 | """Build the PixelModel v3 train/eval data from MS-COCO.
Same source as v1: `sayakpaul/coco-30-val-2014` (MS-COCO val2014 caption/image
pairs). The set is split *deterministically by image hash* into a train subset
and an eval subset so the two are provably disjoint and reproducible - exactly
the regression-safe protocol v1 used.
Outputs (into --out, default ../pm-work):
coco_train.npz images (uint8, N x S x S x 3) + token ids (int32, N x T)
coco_eval.npz eval images + raw caption strings (for FID/CLIP)
vocab.json word-level tokenizer vocabulary (built from TRAIN only)
split_manifest.json hashes + counts, so disjointness is auditable
Run on the training machine (needs `datasets` + network):
python fetch_coco_subset.py --out ../pm-work
"""
from __future__ import annotations
import argparse
import hashlib
import io
import json
import os
import time
import numpy as np
from PIL import Image
from model import build_vocab, encode_caption, PAD_ID
def center_square_resize(img: Image.Image, size: int) -> np.ndarray:
"""Center-crop to a square, resize to size x size, return uint8 HWC RGB."""
img = img.convert("RGB")
w, h = img.size
s = min(w, h)
left = (w - s) // 2
top = (h - s) // 2
img = img.crop((left, top, left + s, top + s)).resize((size, size), Image.BICUBIC)
return np.asarray(img, dtype=np.uint8)
def image_hash(img: Image.Image) -> str:
"""Stable content hash of an image, independent of dataset row order."""
buf = io.BytesIO()
img.convert("RGB").resize((64, 64), Image.BICUBIC).save(buf, format="PNG")
return hashlib.md5(buf.getvalue()).hexdigest()
def caption_of(example) -> str:
"""COCO rows expose captions under a few different keys across mirrors."""
for key in ("caption", "captions", "text", "sentences", "annotations_captions"):
if key in example and example[key]:
v = example[key]
if isinstance(v, (list, tuple)):
return str(v[0])
return str(v)
raise KeyError(f"no caption field found; row keys = {list(example.keys())}")
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--out", default="../pm-work", help="output directory")
ap.add_argument("--dataset", default="sayakpaul/coco-30-val-2014",
help="HF dataset id (MS-COCO caption/image pairs)")
ap.add_argument("--split", default="train", help="HF split to pull from")
ap.add_argument("--n-train", type=int, default=20000)
ap.add_argument("--n-eval", type=int, default=5000)
ap.add_argument("--store-size", type=int, default=144,
help="square size images are stored at (must be >= train crop)")
ap.add_argument("--vocab-size", type=int, default=8192, help="base model, incl. PAD+UNK")
ap.add_argument("--vocab-size-xl", type=int, default=16384, help="XL variant vocab cap")
ap.add_argument("--min-freq", type=int, default=2)
ap.add_argument("--max-tokens", type=int, default=20)
ap.add_argument("--eval-hash-mod", type=int, default=6,
help="1/eval-hash-mod of images are reserved for eval")
ap.add_argument("--streaming", action="store_true",
help="stream the split instead of downloading it first. "
"Faster to start, but a long collection loop can die "
"to a dropped connection; off by default.")
ap.add_argument("--load-retries", type=int, default=6)
ap.add_argument("--max-reconnects", type=int, default=8,
help="resume attempts if iteration dies mid-collection")
ap.add_argument("--seed", type=int, default=0)
args = ap.parse_args()
os.makedirs(args.out, exist_ok=True)
from datasets import load_dataset
def open_dataset():
"""Load the split, retrying the network part. Default is NON-streaming:
the shards are downloaded once (HF retries/resumes those internally) and
iteration is then purely local, so a long collection loop cannot die
halfway to a dropped HTTP connection."""
last = None
for t in range(args.load_retries):
try:
return load_dataset(args.dataset, split=args.split,
streaming=args.streaming)
except Exception as e:
last = e
wait = 5 * (t + 1)
print(f"[data] load attempt {t + 1}/{args.load_retries} failed "
f"({type(e).__name__}: {e}); retrying in {wait}s", flush=True)
time.sleep(wait)
raise RuntimeError(f"could not load {args.dataset}: {last}")
train_imgs, train_caps_raw, train_hashes = [], [], []
eval_imgs, eval_caps, eval_hashes = [], [], []
seen = set()
def need_more():
return len(train_imgs) < args.n_train or len(eval_imgs) < args.n_eval
last_report = 0
reconnects = 0
print(f"[data] loading {args.dataset} split={args.split} "
f"(streaming={args.streaming})", flush=True)
while need_more():
ds = open_dataset()
try:
for ex in ds:
if not need_more():
break
try:
img = ex.get("image") or ex.get("img")
if img is None:
continue
cap = caption_of(ex)
h = image_hash(img)
if h in seen:
continue
seen.add(h)
is_eval = (int(h, 16) % args.eval_hash_mod == 0)
if is_eval:
if len(eval_imgs) >= args.n_eval:
continue
eval_imgs.append(center_square_resize(img, 256))
eval_caps.append(cap)
eval_hashes.append(h)
else:
if len(train_imgs) >= args.n_train:
continue
train_imgs.append(center_square_resize(img, args.store_size))
train_caps_raw.append(cap)
train_hashes.append(h)
except Exception:
continue
total = len(train_imgs) + len(eval_imgs)
if total - last_report >= 1000:
last_report = total
print(f"[data] train={len(train_imgs)} eval={len(eval_imgs)}", flush=True)
break
except Exception as e:
reconnects += 1
print(f"[data] iteration died ({type(e).__name__}: {e}); "
f"reconnect {reconnects}/{args.max_reconnects} with "
f"train={len(train_imgs)} eval={len(eval_imgs)}", flush=True)
if reconnects >= args.max_reconnects:
print("[data] giving up on reconnects; proceeding with what we have",
flush=True)
break
time.sleep(5)
assert set(train_hashes).isdisjoint(set(eval_hashes)), "train/eval overlap!"
print(f"[data] collected train={len(train_imgs)} eval={len(eval_imgs)} "
f"(disjoint: {set(train_hashes).isdisjoint(set(eval_hashes))})")
vocab = build_vocab(train_caps_raw, args.vocab_size, args.min_freq)
with open(os.path.join(args.out, "vocab.json"), "w") as f:
json.dump(vocab, f)
vocab_xl = build_vocab(train_caps_raw, args.vocab_size_xl, args.min_freq)
with open(os.path.join(args.out, "vocab_xl.json"), "w") as f:
json.dump(vocab_xl, f)
print(f"[data] vocab size = {len(vocab)} (cap {args.vocab_size}), "
f"vocab_xl = {len(vocab_xl)} (cap {args.vocab_size_xl})")
train_tokens = np.stack(
[encode_caption(c, vocab, args.max_tokens) for c in train_caps_raw])
np.savez_compressed(
os.path.join(args.out, "coco_train.npz"),
images=np.stack(train_imgs),
tokens=train_tokens,
captions=np.array(train_caps_raw, dtype=object),
store_size=args.store_size,
max_tokens=args.max_tokens,
)
np.savez_compressed(
os.path.join(args.out, "coco_eval.npz"),
images=np.stack(eval_imgs),
captions=np.array(eval_caps, dtype=object),
)
with open(os.path.join(args.out, "split_manifest.json"), "w") as f:
json.dump({
"dataset": args.dataset,
"n_train": len(train_imgs),
"n_eval": len(eval_imgs),
"eval_hash_mod": args.eval_hash_mod,
"vocab_size": len(vocab),
"store_size": args.store_size,
"max_tokens": args.max_tokens,
"disjoint": bool(set(train_hashes).isdisjoint(set(eval_hashes))),
"train_hash_sample": train_hashes[:8],
"eval_hash_sample": eval_hashes[:8],
}, f, indent=2)
mb = os.path.getsize(os.path.join(args.out, "coco_train.npz")) / 1e6
print(f"[data] wrote coco_train.npz ({mb:.1f} MB), coco_eval.npz, vocab.json")
print(f"[data] done -> {args.out}")
if __name__ == "__main__":
main()
|