B-DoPED / scripts /shard_utils.py
1Konny's picture
Initial release: B-DoPED dataset (library + scripts + rendered outputs)
40980f7
Raw
History Blame
3.79 kB
"""
shard_utils.py — pack rendered pose directories into plain (stored) tar shards.
Each shard `shard_{group:05d}.tar` holds a contiguous block of pose ids; in-tar layout is
`pose_{id:06d}/{modality}/{file}`. Tars are STORED (no compression) since PNG/NPZ are already
compressed. Per-shard sha256 + a `shards_index.csv` are emitted.
Used by render_smal_multiview.py (--shard) to pack rendered output into tar shards.
"""
import os
import csv
import tarfile
import hashlib
from collections import defaultdict
def sha256_file(path, buf=1 << 20):
h = hashlib.sha256()
with open(path, "rb") as f:
for chunk in iter(lambda: f.read(buf), b""):
h.update(chunk)
return h.hexdigest()
def pack_to_shards(items, shard_dir, shard_size, modalities=("rgb", "seg", "npz"),
index_csv=None, log=print):
"""
items : iterable of (pose_id:int, src_pose_dir:str). Need not be sorted.
shard_dir : output directory for shard_*.tar (+ index_csv if given).
shard_size : number of pose ids per shard group (group = pose_id // shard_size).
modalities : subdirectories under each pose dir to include (others, e.g. overlays, are skipped).
Returns the list of index rows (dicts).
"""
os.makedirs(shard_dir, exist_ok=True)
groups = defaultdict(list)
for pid, src in items:
groups[pid // shard_size].append((int(pid), src))
rows = []
for g in sorted(groups):
members = sorted(groups[g], key=lambda x: x[0])
shard_path = os.path.join(shard_dir, f"shard_{g:05d}.tar")
n_files = 0
with tarfile.open(shard_path, "w") as tar: # "w" = stored, no compression
for pid, src in members:
arc_pose = f"pose_{pid:06d}"
for mod in modalities:
md = os.path.join(src, mod)
if not os.path.isdir(md):
continue
for fn in sorted(os.listdir(md)):
tar.add(os.path.join(md, fn), arcname=f"{arc_pose}/{mod}/{fn}")
n_files += 1
sz = os.path.getsize(shard_path)
sha = sha256_file(shard_path)
rows.append({"shard": os.path.basename(shard_path),
"pose_start": members[0][0], "pose_end": members[-1][0],
"num_poses": len(members), "num_files": n_files,
"bytes": sz, "sha256": sha})
log(f"[shard] {os.path.basename(shard_path)} poses {members[0][0]}-{members[-1][0]} "
f"({len(members)}) {sz/1e9:.2f} GB {n_files} files")
if index_csv:
fields = ["shard", "pose_start", "pose_end", "num_poses", "num_files", "bytes", "sha256"]
existing = []
if os.path.exists(index_csv):
with open(index_csv, newline="") as f:
existing = [r for r in csv.DictReader(f)]
by_shard = {r["shard"]: r for r in existing}
for r in rows:
by_shard[r["shard"]] = {k: str(r[k]) for k in fields}
with open(index_csv, "w", newline="") as f:
w = csv.DictWriter(f, fieldnames=fields)
w.writeheader()
for s in sorted(by_shard):
w.writerow(by_shard[s])
log(f"[shard] wrote index {index_csv} ({len(by_shard)} shards)")
return rows
def modalities_from_flags(save_rgb=True, save_mask=True, save_npz=True,
save_keypoints=False, save_depth=False, save_canny=False):
mods = []
if save_rgb: mods.append("rgb")
if save_mask: mods.append("seg")
if save_npz: mods.append("npz")
if save_keypoints: mods.append("rgb_with_keypoints2d")
if save_depth: mods.append("depth")
if save_canny: mods.append("canny")
return tuple(mods)