File size: 8,503 Bytes
a8eb6e5 | 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 | #!/usr/bin/env python3
"""
Download only the files needed for training (defined by filtered_index.json)
from HuggingFaceVLA/community_dataset_v3.
Rate-limit-aware greedy scheduler: downloads small files when we have rate limit
headroom, swaps to large files (videos) when approaching the limit to keep
bandwidth busy while the window recovers. Goal: never actually hit a 429.
"""
import argparse
import json
import os
import sys
import time
import threading
from collections import defaultdict, deque
from concurrent.futures import ThreadPoolExecutor, as_completed, Future
from huggingface_hub import hf_hub_download
RATE_LIMIT = 100 # 1000 actual / ~10 API calls per hf_hub_download
RATE_WINDOW = 300 # 5 minutes
class RateLimitTracker:
"""Sliding window request counter."""
def __init__(self):
self.lock = threading.Lock()
self.timestamps: deque[float] = deque()
def record(self):
now = time.time()
with self.lock:
self.timestamps.append(now)
self._prune(now)
def count(self) -> int:
now = time.time()
with self.lock:
self._prune(now)
return len(self.timestamps)
def headroom(self) -> int:
"""How many more requests we can make in this window."""
return max(0, RATE_LIMIT - self.count())
def wait_if_needed(self):
"""If we've exhausted the window, sleep until oldest request expires."""
while self.headroom() <= 0:
with self.lock:
if self.timestamps:
wait = RATE_WINDOW - (time.time() - self.timestamps[0]) + 1
if wait > 0:
print(f" Rate limit reached, waiting {wait:.0f}s...", flush=True)
# Release lock while sleeping
else:
wait = 0
if wait > 0:
time.sleep(wait)
def _prune(self, now):
cutoff = now - RATE_WINDOW
while self.timestamps and self.timestamps[0] < cutoff:
self.timestamps.popleft()
class FileQueue:
"""Thread-safe queue that serves small or large files on demand."""
def __init__(self, small_files: list[str], large_files: list[str]):
self.lock = threading.Lock()
self.small = deque(small_files)
self.large = deque(large_files)
self.total = len(small_files) + len(large_files)
def get(self, prefer_small: bool) -> str | None:
with self.lock:
if prefer_small and self.small:
return self.small.popleft()
elif self.large:
return self.large.popleft()
elif self.small:
return self.small.popleft()
return None
def remaining(self) -> int:
with self.lock:
return len(self.small) + len(self.large)
def small_remaining(self) -> int:
with self.lock:
return len(self.small)
def large_remaining(self) -> int:
with self.lock:
return len(self.large)
def build_file_lists(index_path: str, output_dir: str) -> tuple[list[str], list[str], int]:
"""Returns (small_files, large_files, skipped_count) from filtered_index.json.
Skips files already on disk."""
with open(index_path) as f:
index = json.load(f)
datasets = defaultdict(list)
for ep in index["episodes"]:
datasets[ep["dataset"]].append(ep["episode_index"])
small = []
large = []
skipped = 0
def add_if_missing(filepath, target_list):
nonlocal skipped
if os.path.exists(os.path.join(output_dir, filepath)):
skipped += 1
else:
target_list.append(filepath)
for dataset_name, episode_indices in datasets.items():
prefix = dataset_name
add_if_missing(f"{prefix}/meta/info.json", small)
add_if_missing(f"{prefix}/meta/tasks.jsonl", small)
add_if_missing(f"{prefix}/meta/episodes.jsonl", small)
for ep_idx in episode_indices:
ep_str = f"episode_{ep_idx:06d}"
add_if_missing(f"{prefix}/data/chunk-000/{ep_str}.parquet", small)
add_if_missing(f"{prefix}/videos/chunk-000/observation.images.image/{ep_str}.mp4", large)
add_if_missing(f"{prefix}/videos/chunk-000/observation.images.image2/{ep_str}.mp4", large)
return small, large, skipped
# Shared state
tracker = RateLimitTracker()
queue: FileQueue = None
stats_lock = threading.Lock()
downloaded = 0
total_bytes = 0
failed = []
start_time = 0
# When headroom drops below this, prefer large files
HEADROOM_THRESHOLD = 50
def worker(output_dir, token):
"""Worker loop: grab a file based on rate limit state, download it, repeat."""
global downloaded, total_bytes
while True:
headroom = tracker.headroom()
prefer_small = headroom > HEADROOM_THRESHOLD
filepath = queue.get(prefer_small)
if filepath is None:
return
for attempt in range(10):
tracker.wait_if_needed()
tracker.record()
try:
path = hf_hub_download(
repo_id="HuggingFaceVLA/community_dataset_v3",
repo_type="dataset",
filename=filepath,
local_dir=output_dir,
token=token,
)
size = os.path.getsize(path)
with stats_lock:
downloaded += 1
total_bytes += size
_maybe_log()
break
except Exception as e:
if "429" in str(e) and attempt < 9:
time.sleep(30 * (attempt + 1))
continue
with stats_lock:
failed.append((filepath, str(e)))
_maybe_log()
break
def _maybe_log():
"""Log progress every 100 files. Must be called with stats_lock held."""
total = downloaded + len(failed)
if total % 100 == 0 and total > 0:
elapsed = time.time() - start_time
rate = total / elapsed if elapsed > 0 else 0
mb_s = (total_bytes / 1024 / 1024) / elapsed if elapsed > 0 else 0
gb_done = total_bytes / 1024 / 1024 / 1024
headroom = tracker.headroom()
remaining = queue.remaining()
est_min = remaining / rate / 60 if rate > 0 else 0
print(f" [{total}/{queue.total}] {gb_done:.1f}GB, "
f"{mb_s:.0f} MB/s, {rate:.1f} files/s, "
f"headroom: {headroom}/{RATE_LIMIT}, "
f"queued: {queue.small_remaining()}s+{queue.large_remaining()}L, "
f"~{est_min:.0f}min left", flush=True)
def main():
global queue, start_time
parser = argparse.ArgumentParser(description="Download training subset from community_dataset_v3")
parser.add_argument("--index", type=str, default="filtered_index.json")
parser.add_argument("--output", type=str, default="/data/community_dataset_v3")
parser.add_argument("--token", type=str, default=os.environ.get("HF_TOKEN"))
parser.add_argument("--workers", type=int, default=8)
args = parser.parse_args()
if not args.token:
print("ERROR: Set HF_TOKEN or pass --token")
return
small, large, skipped = build_file_lists(args.index, args.output)
queue = FileQueue(small, large)
print(f"Files to download: {queue.total} ({skipped} already on disk, skipped)")
print(f" Small (metadata+parquets): {len(small)}")
print(f" Large (videos): {len(large)}")
print(f" Workers: {args.workers}")
print(f" Rate limit: {RATE_LIMIT}/{RATE_WINDOW}s, "
f"swap to large files at <{HEADROOM_THRESHOLD} headroom")
print()
start_time = time.time()
with ThreadPoolExecutor(max_workers=args.workers) as pool:
futures = [pool.submit(worker, args.output, args.token)
for _ in range(args.workers)]
for f in futures:
f.result()
elapsed = time.time() - start_time
gb_total = total_bytes / 1024 / 1024 / 1024
print(f"\nDone in {elapsed/60:.1f} min: {downloaded} files, "
f"{gb_total:.1f}GB, {len(failed)} failed")
if failed:
print("Failed files:")
for f, err in failed[:20]:
print(f" {f}: {err}")
if len(failed) > 20:
print(f" ... and {len(failed) - 20} more")
sys.exit(1)
if __name__ == "__main__":
main()
|