File size: 12,635 Bytes
4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 6011afd 4956e51 | 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 | # tools/make_fewshot_configs.py
#
# Create few-shot variants of existing configs by moving K images from OOD -> TRAIN.
# Works with density-aware configs (ood_same_density / ood_diff_density) if present.
#
# Output config name: <base_cfg>__fs<K>
# Output splits written:
# - train
# - id_test
# - ood_test
# - ood_same_density (if present in base)
# - ood_diff_density (if present in base)
#
# Memory-safe:
# - Never builds giant Python lists of rows
# - Streams master parquet via pyarrow.dataset + record batches
# - Writes parquet shards with lazy-open writers + safe rotation
from __future__ import annotations
import argparse
import json
import os
import random
import zlib
from dataclasses import dataclass
from pathlib import Path
from typing import Dict, Iterable, List, Optional, Set, Tuple
import pyarrow as pa
import pyarrow.compute as pc
import pyarrow.dataset as ds
import pyarrow.parquet as pq
def _stable_seed(base_seed: int, cfg_name: str, k: int) -> int:
# deterministic across runs/machines
h = zlib.adler32(cfg_name.encode("utf-8")) & 0xFFFFFFFF
return (base_seed * 1_000_003 + h * 97 + k * 7919) & 0xFFFFFFFF
def _list_parquet_files(cfg_dir: Path, split: str) -> List[Path]:
return sorted(cfg_dir.glob(f"{split}-*.parquet"))
def _read_filenames_from_split(cfg_dir: Path, split: str) -> Set[str]:
files = _list_parquet_files(cfg_dir, split)
if not files:
return set()
d = ds.dataset([str(p) for p in files], format="parquet")
# only pull filename column to keep memory low
col = d.to_table(columns=["filename"]).column("filename")
return set(col.to_pylist())
def _sample_k_from_set(items: Set[str], k: int, seed: int) -> List[str]:
if k <= 0:
return []
if len(items) < k:
raise ValueError(f"Not enough items to sample k={k}. pool={len(items)}")
rng = random.Random(seed)
items_list = sorted(items) # stable ordering
return rng.sample(items_list, k)
@dataclass
class _SplitWriter:
out_dir: Path
split_name: str
rows_per_shard: int
compression: str = "zstd"
# internal state
shard_idx: int = 0
rows_in_shard: int = 0
total_rows: int = 0
schema: Optional[pa.Schema] = None
writer: Optional[pq.ParquetWriter] = None
def _open(self, schema: pa.Schema) -> None:
self.schema = schema
out_path = self.out_dir / f"{self.split_name}-{self.shard_idx:05d}.parquet"
self.writer = pq.ParquetWriter(out_path, schema, compression=self.compression)
self.rows_in_shard = 0
def _close(self) -> None:
if self.writer is not None:
self.writer.close()
self.writer = None
def write_batch(self, batch: pa.RecordBatch) -> None:
if batch is None or batch.num_rows == 0:
return
if self.schema is None:
self._open(batch.schema)
# rotate before writing if already full
if self.rows_in_shard >= self.rows_per_shard:
self._close()
self.shard_idx += 1
# lazy-open on next write, but we can open now (schema known)
self._open(self.schema)
offset = 0
while offset < batch.num_rows:
remaining = self.rows_per_shard - self.rows_in_shard
take = min(remaining, batch.num_rows - offset)
sub = batch.slice(offset, take)
if self.writer is None:
# can happen after rotation
if self.schema is None:
self.schema = sub.schema
self._open(self.schema)
# write
self.writer.write_table(pa.Table.from_batches([sub]))
self.rows_in_shard += sub.num_rows
self.total_rows += sub.num_rows
offset += take
# rotate if filled exactly
if self.rows_in_shard >= self.rows_per_shard:
self._close()
self.shard_idx += 1
# keep lazy-open until next write
def finalize(self) -> int:
self._close()
return self.total_rows
def _stream_master_and_write_splits(
master_dir: Path,
out_cfg_dir: Path,
split_to_files: Dict[str, Set[str]],
batch_rows: int,
compression: str = "zstd",
) -> Dict[str, int]:
"""
Single-pass scan over master parquet. Routes each row into 1..N splits
depending on membership of filename in split_to_files[split].
Writes parquet shards per split under out_cfg_dir.
Returns counts per split.
"""
master = ds.dataset(str(master_dir), format="parquet")
writers: Dict[str, _SplitWriter] = {}
# Build Arrow arrays for membership checks per batch:
# We'll compute masks via pc.is_in(filename_col, value_set=array_of_values)
# For large sets this could be heavy; our sets are moderate and batch_rows is controlled.
split_value_arrays: Dict[str, pa.Array] = {}
for split, files in split_to_files.items():
if not files:
continue
split_value_arrays[split] = pa.array(list(files), type=pa.string())
# Ensure output dir exists
out_cfg_dir.mkdir(parents=True, exist_ok=True)
# scan with a reasonable batch size
scanner = master.scanner(batch_size=batch_rows)
for rb in scanner.to_batches():
# must have filename column for routing
if "filename" not in rb.schema.names:
raise RuntimeError("Master parquet must contain a 'filename' column.")
fn = rb.column(rb.schema.get_field_index("filename"))
for split, values_arr in split_value_arrays.items():
mask = pc.is_in(fn, value_set=values_arr)
# If no matches, skip
if pc.sum(mask).as_py() == 0:
continue
filtered = rb.filter(mask)
if split not in writers:
split_out = out_cfg_dir
writers[split] = _SplitWriter(
out_dir=split_out,
split_name=split,
rows_per_shard=ROWS_PER_SHARD,
compression=compression,
)
writers[split].write_batch(filtered)
counts: Dict[str, int] = {}
for split, w in writers.items():
counts[split] = w.finalize()
return counts
def materialize_fewshot_config(
repo_root: Path,
base_cfg: str,
k: int,
seed: int,
rows_per_shard: int,
scan_batch_size: int,
) -> Dict[str, object]:
"""
Creates <base_cfg>__fs<k> under data/configs by moving k images
from base ood_train into train.
ood_test is NEVER modified and stays identical across the base config
and all few-shot variants, ensuring consistent OOD evaluation.
Few-shot images are drawn from a single deterministic ordering of ood_train
(seeded by base_cfg name only, not k), so that
fs1_images ⊂ fs10_images ⊂ fs100_images ⊂ fsall_images.
Pass k=-1 (or k >= len(ood_train)) to move all ood_train images (fsall).
"""
data_dir = repo_root / "data"
cfgs_dir = data_dir / "configs"
master_dir = data_dir / "master"
base_dir = cfgs_dir / base_cfg
if not base_dir.exists():
raise FileNotFoundError(f"Base config not found: {base_dir}")
out_cfg = f"{base_cfg}__fs{k}" if k >= 0 else f"{base_cfg}__fsall"
out_dir = cfgs_dir / out_cfg
if out_dir.exists():
return {"status": "skip_exists", "out_cfg": out_cfg}
# Load filename sets from base config
train_files = _read_filenames_from_split(base_dir, "train")
id_test_files = _read_filenames_from_split(base_dir, "id_test")
ood_train_files = _read_filenames_from_split(base_dir, "ood_train")
ood_test_files = _read_filenames_from_split(base_dir, "ood_test") # never changes
if not ood_train_files:
raise RuntimeError(
f"Base config '{base_cfg}' has no ood_train split. "
"Regenerate base configs with the updated make_configs.py."
)
# Build a deterministic ordering seeded only on (seed, base_cfg) so that
# all k values draw progressive prefixes of the same list.
ordering_seed = _stable_seed(seed, base_cfg, 0) # k=0 → ordering never depends on actual k
ordered_ood_train = sorted(ood_train_files) # stable base ordering
rng = random.Random(ordering_seed)
rng.shuffle(ordered_ood_train)
# Determine how many to move
k_actual = len(ordered_ood_train) if k < 0 else min(k, len(ordered_ood_train))
chosen = ordered_ood_train[:k_actual]
chosen_set = set(chosen)
# Move chosen from ood_train -> train; ood_test is completely untouched
new_train = set(train_files) | chosen_set
new_id_test = set(id_test_files)
new_ood_train = set(ood_train_files) - chosen_set
new_ood_test = set(ood_test_files) # unchanged
# Build split membership map (skip empty splits)
split_to_files: Dict[str, Set[str]] = {
"train": new_train,
"id_test": new_id_test,
"ood_test": new_ood_test,
}
if new_ood_train:
split_to_files["ood_train"] = new_ood_train
# Write protocol/manifest
out_dir.mkdir(parents=True, exist_ok=True)
proto = {
"base_config": base_cfg,
"fewshot_k": k_actual,
"seed": seed,
"ordering_seed": ordering_seed,
"moved_from_ood_train_to_train": sorted(chosen),
"counts": {
"train": len(new_train),
"id_test": len(new_id_test),
"ood_train": len(new_ood_train),
"ood_test": len(new_ood_test),
},
}
(out_dir / "protocol.json").write_text(json.dumps(proto, indent=2))
# Stream master once and write all splits
# Interpret scan_batch_size as a multiplier; enforce a reasonable minimum.
batch_rows = max(2048, int(scan_batch_size) * 2048)
global ROWS_PER_SHARD
ROWS_PER_SHARD = rows_per_shard
counts_written = _stream_master_and_write_splits(
master_dir=master_dir,
out_cfg_dir=out_dir,
split_to_files=split_to_files,
batch_rows=batch_rows,
compression="zstd",
)
proto["counts_written"] = counts_written
(out_dir / "protocol.json").write_text(json.dumps(proto, indent=2))
return {"status": "ok", "out_cfg": out_cfg, "counts_written": counts_written}
def parse_fewshot_ks(s: str) -> List[int]:
"""Parse comma-separated k values. Use 'all' or -1 to indicate fsall (all ood_train)."""
s = s.strip()
if not s:
return []
parts = [p.strip() for p in s.split(",")]
out: List[int] = []
for p in parts:
if not p:
continue
if p.lower() == "all":
out.append(-1)
else:
out.append(int(p))
return out
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--repo_root", required=True, help="path to hf_repo root (contains data/ and tools/)")
ap.add_argument(
"--base_configs",
nargs="+",
required=True,
help="base config names (dirs under data/configs)",
)
ap.add_argument("--fewshot_ks", required=True, help="comma-separated list, e.g. 10,50,100")
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--rows_per_shard", type=int, default=256)
ap.add_argument("--scan_batch_size", type=int, default=8, help="controls streaming batch rows (multiplier)")
args = ap.parse_args()
repo_root = Path(args.repo_root).resolve()
if not (repo_root / "data" / "configs").exists():
raise FileNotFoundError(f"Expected data/configs under {repo_root}")
ks = parse_fewshot_ks(args.fewshot_ks)
if not ks:
raise ValueError("fewshot_ks is empty. Example: --fewshot_ks 10,50,100")
# Ensure master exists
master_dir = repo_root / "data" / "master"
if not master_dir.exists():
raise FileNotFoundError(f"Master dir missing: {master_dir}")
for base_cfg in args.base_configs:
for k in ks:
out = materialize_fewshot_config(
repo_root=repo_root,
base_cfg=base_cfg,
k=k,
seed=args.seed,
rows_per_shard=args.rows_per_shard,
scan_batch_size=args.scan_batch_size,
)
if out["status"] == "skip_exists":
print(f"Skip (exists): {out['out_cfg']}")
else:
print(f"Wrote: {out['out_cfg']} counts={out.get('counts_written')}")
print("Done.")
if __name__ == "__main__":
# global used inside streaming writer constructor to avoid passing around everywhere
ROWS_PER_SHARD = 256
main()
|