File size: 12,893 Bytes
571a68d | 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 | """
Full skeleton audit: naming + structural features for all 79 skeletons.
Checks:
1. Canonical name quality (empty, duplicate, semantic sense)
2. Side tag correctness (L/R balance, missed sides)
3. Symmetry pair validity (matched pairs have similar bone lengths/depths)
4. Topology structure (tree depth, branching factor, connected)
5. Geodesic distance matrix sanity
6. Rest pose geometry (bone lengths, height, spread)
7. Cross-dataset consistency for shared canonical names
"""
import sys
import os
from pathlib import Path
import numpy as np
from collections import Counter, defaultdict
project_root = Path(__file__).parent.parent
sys.path.insert(0, str(project_root))
from src.data.skeleton_graph import SkeletonGraph
RESULT_DIR = project_root / 'logs' / 'data_fix_20260318'
# Severity levels
ERRORS = [] # Must fix
WARNINGS = [] # Should review
INFO = [] # FYI
def log_error(dataset, msg):
ERRORS.append(f'[ERROR] {dataset}: {msg}')
def log_warn(dataset, msg):
WARNINGS.append(f'[WARN] {dataset}: {msg}')
def log_info(dataset, msg):
INFO.append(f'[INFO] {dataset}: {msg}')
def audit_skeleton(dataset_id: str, skel_data: dict, motion_sample: dict = None):
"""Full audit of one skeleton."""
sg = SkeletonGraph.from_dict(skel_data)
J = sg.num_joints
canon = [str(n) for n in skel_data.get('canonical_names', [])]
raw = [str(n) for n in skel_data['joint_names']]
results = {'dataset': dataset_id, 'num_joints': J, 'issues': []}
# ===== 1. Canonical Name Quality =====
if not canon:
log_error(dataset_id, 'No canonical_names field in skeleton.npz')
results['issues'].append('no_canonical')
else:
# Empty names
empty = [i for i, c in enumerate(canon) if not c.strip()]
if empty:
log_error(dataset_id, f'{len(empty)} empty canonical names at indices {empty[:5]}')
# Duplicate canonical names (within same skeleton)
dupes = {c: [i for i, x in enumerate(canon) if x == c] for c in set(canon)}
dupes = {c: idxs for c, idxs in dupes.items() if len(idxs) > 1}
if dupes:
for c, idxs in list(dupes.items())[:3]:
raw_names = [raw[i] for i in idxs]
log_warn(dataset_id, f'Duplicate canonical "{c}" for raw: {raw_names}')
# Names that are just numbers or single chars
bad_names = [c for c in canon if len(c) <= 2 and not c.isalpha()]
if bad_names:
log_warn(dataset_id, f'Very short canonical names: {bad_names[:5]}')
# Prefix residue (bip01, bn_, jt_, mixamorig still present)
prefix_residue = [c for c in canon if any(p in c.lower() for p in ['bip01', 'bn_', 'jt_', 'mixamorig', 'npc_'])]
if prefix_residue:
log_warn(dataset_id, f'Prefix residue in canonical: {prefix_residue[:5]}')
# ===== 2. Side Tag Correctness =====
n_left = sum(1 for t in sg.side_tags if t == 'left')
n_right = sum(1 for t in sg.side_tags if t == 'right')
n_center = sum(1 for t in sg.side_tags if t == 'center')
if n_left != n_right and J > 5: # small skeletons may be asymmetric
log_warn(dataset_id, f'L/R imbalance: L={n_left} R={n_right} C={n_center}')
# Check if canonical names agree with side tags
if canon:
for i, (c, t) in enumerate(zip(canon, sg.side_tags)):
if 'left' in c and t != 'left':
log_warn(dataset_id, f'Joint {i} "{raw[i]}" canonical="{c}" but side_tag="{t}"')
elif 'right' in c and t != 'right':
log_warn(dataset_id, f'Joint {i} "{raw[i]}" canonical="{c}" but side_tag="{t}"')
# ===== 3. Symmetry Pair Validation =====
for i, j in sg.symmetry_pairs:
# Check bone lengths match (should be ~equal for symmetric joints)
bl_i = sg.bone_lengths[i]
bl_j = sg.bone_lengths[j]
if bl_i > 0.01 and bl_j > 0.01:
ratio = max(bl_i, bl_j) / min(bl_i, bl_j)
if ratio > 1.5:
log_warn(dataset_id, f'Sym pair ({raw[i]}, {raw[j]}) bone length mismatch: {bl_i:.3f} vs {bl_j:.3f}')
# Check depths match
if sg.depths[i] != sg.depths[j]:
log_warn(dataset_id, f'Sym pair ({raw[i]}, {raw[j]}) depth mismatch: {sg.depths[i]} vs {sg.depths[j]}')
# ===== 4. Topology Structure =====
max_depth = sg.depths.max()
max_degree = sg.degrees.max()
leaf_count = sum(1 for d in sg.degrees if d == 0)
root_count = sum(1 for p in sg.parent_indices if p == -1)
if root_count != 1:
log_error(dataset_id, f'Expected 1 root, found {root_count}')
if max_depth > 20:
log_warn(dataset_id, f'Very deep tree: max_depth={max_depth}')
# Check connectivity (all joints reachable from root)
reachable = set()
queue = [i for i, p in enumerate(sg.parent_indices) if p == -1]
while queue:
curr = queue.pop(0)
reachable.add(curr)
for j in range(J):
if sg.parent_indices[j] == curr and j not in reachable:
queue.append(j)
if len(reachable) != J:
log_error(dataset_id, f'Disconnected: only {len(reachable)}/{J} joints reachable from root')
# ===== 5. Geodesic Distance Sanity =====
geo = sg.geodesic_dist
if geo.shape != (J, J):
log_error(dataset_id, f'Geodesic distance shape mismatch: {geo.shape} vs ({J},{J})')
else:
# Should be symmetric
asym = np.abs(geo - geo.T).max()
if asym > 0.01:
log_error(dataset_id, f'Geodesic distance not symmetric: max asymmetry={asym}')
# Diagonal should be 0
diag_max = np.diag(geo).max()
if diag_max > 0.01:
log_error(dataset_id, f'Geodesic distance diagonal non-zero: max={diag_max}')
# Max distance should be reasonable
max_geo = geo.max()
if max_geo > J:
log_warn(dataset_id, f'Geodesic max={max_geo} exceeds num_joints={J}')
# No unreachable pairs (distance should be < J+1)
unreachable = (geo >= J + 1).sum()
if unreachable > 0:
log_error(dataset_id, f'{unreachable} unreachable joint pairs in geodesic matrix')
# ===== 6. Rest Pose Geometry =====
offsets = sg.rest_offsets
bone_lengths = sg.bone_lengths
# Zero-length bones (excluding root)
zero_bones = sum(1 for i, bl in enumerate(bone_lengths) if bl < 1e-6 and sg.parent_indices[i] >= 0)
if zero_bones > 0:
log_info(dataset_id, f'{zero_bones} zero-length bones')
# Very long bones (> 10x median)
nonzero_bl = bone_lengths[bone_lengths > 1e-6]
if len(nonzero_bl) > 0:
median_bl = np.median(nonzero_bl)
long_bones = [(raw[i], bl) for i, bl in enumerate(bone_lengths)
if bl > 10 * median_bl and sg.parent_indices[i] >= 0]
if long_bones:
log_warn(dataset_id, f'Unusually long bones (>10x median={median_bl:.4f}): {long_bones[:3]}')
# ===== 7. Motion Data Spot Check =====
if motion_sample is not None:
T = int(motion_sample['num_frames'])
jp = motion_sample['joint_positions'][:T]
lp = motion_sample['local_positions'][:T]
vel = motion_sample['velocities'][:T]
# Check joint count matches
if jp.shape[1] != J:
log_error(dataset_id, f'Motion joints={jp.shape[1]} != skeleton joints={J}')
# Check for NaN/Inf
for key in ['joint_positions', 'local_positions', 'velocities']:
arr = motion_sample[key][:T]
if np.any(np.isnan(arr)):
log_error(dataset_id, f'NaN in motion {key}')
if np.any(np.isinf(arr)):
log_error(dataset_id, f'Inf in motion {key}')
# Height sanity
height = jp[0, :, 1].max() - jp[0, :, 1].min()
if height < 0.05:
log_warn(dataset_id, f'Very small body height: {height:.4f}m')
elif height > 10:
log_warn(dataset_id, f'Very large body height: {height:.2f}m (scale issue?)')
# Velocity sanity
vel_max = np.linalg.norm(vel, axis=-1).max()
if vel_max > 50:
log_warn(dataset_id, f'Very high velocity: max={vel_max:.1f} m/s')
# Collect summary
results['n_left'] = n_left
results['n_right'] = n_right
results['n_center'] = n_center
results['n_sym'] = len(sg.symmetry_pairs)
results['max_depth'] = int(max_depth)
results['max_degree'] = int(max_degree)
results['leaf_count'] = leaf_count
results['n_canonical'] = len(canon)
results['n_dupe_canonical'] = sum(len(v) - 1 for v in dupes.values()) if canon and dupes else 0
return results
def main():
all_results = []
# Audit human datasets
human_datasets = ['humanml3d', 'lafan1', '100style', 'bandai_namco', 'cmu_mocap', 'mixamo']
for ds in human_datasets:
ds_path = project_root / 'data' / 'processed' / ds
skel_data = dict(np.load(ds_path / 'skeleton.npz', allow_pickle=True))
# Load a motion sample
motions_dir = ds_path / 'motions'
files = sorted(os.listdir(motions_dir))
motion_sample = dict(np.load(motions_dir / files[0], allow_pickle=True)) if files else None
r = audit_skeleton(ds, skel_data, motion_sample)
all_results.append(r)
# Audit Zoo species
zoo_path = project_root / 'data' / 'processed' / 'truebones_zoo'
skel_dir = zoo_path / 'skeletons'
motions_dir = zoo_path / 'motions'
# Build species → motion mapping
species_motions = {}
for f in sorted(os.listdir(motions_dir))[:200]:
d = dict(np.load(motions_dir / f, allow_pickle=True))
sp = str(d.get('species', ''))
if sp and sp not in species_motions:
species_motions[sp] = d
for skel_file in sorted(skel_dir.glob('*.npz')):
species = skel_file.stem
skel_data = dict(np.load(skel_file, allow_pickle=True))
motion_sample = species_motions.get(species)
r = audit_skeleton(f'zoo/{species}', skel_data, motion_sample)
all_results.append(r)
# ===== Cross-dataset canonical consistency =====
# For human datasets, check that same canonical name → similar tree depth
canonical_depths = defaultdict(list)
for ds in human_datasets:
skel_data = dict(np.load(project_root / 'data' / 'processed' / ds / 'skeleton.npz', allow_pickle=True))
sg = SkeletonGraph.from_dict(skel_data)
canon = [str(n) for n in skel_data.get('canonical_names', [])]
for c, d in zip(canon, sg.depths):
canonical_depths[c].append((ds, int(d)))
for c, entries in canonical_depths.items():
depths = [d for _, d in entries]
if len(set(depths)) > 1 and max(depths) - min(depths) > 2:
sources = [(ds, d) for ds, d in entries]
log_warn('cross-dataset', f'Canonical "{c}" has depth variance: {sources}')
# ===== Write Report =====
report_path = RESULT_DIR / 'skeleton_audit_report.md'
with open(report_path, 'w') as f:
f.write('# Skeleton Audit Report\n\n')
f.write(f'**Date**: 2026-03-18\n')
f.write(f'**Datasets**: {len(human_datasets)} human + {len(list(skel_dir.glob("*.npz")))} zoo species\n\n')
# Summary table
f.write('## Summary\n\n')
f.write(f'| Dataset | J | L | R | C | Sym | Depth | Leaves | Canon | Dupes |\n')
f.write(f'|---------|:-:|:-:|:-:|:-:|:---:|:-----:|:------:|:-----:|:-----:|\n')
for r in all_results:
f.write(f'| {r["dataset"]:20s} | {r["num_joints"]:3d} | {r["n_left"]:2d} | {r["n_right"]:2d} | '
f'{r["n_center"]:2d} | {r["n_sym"]:2d} | {r["max_depth"]:2d} | {r["leaf_count"]:2d} | '
f'{r["n_canonical"]:3d} | {r["n_dupe_canonical"]:2d} |\n')
# Issues
f.write(f'\n## Errors ({len(ERRORS)})\n\n')
if ERRORS:
for e in ERRORS:
f.write(f'- {e}\n')
else:
f.write('None.\n')
f.write(f'\n## Warnings ({len(WARNINGS)})\n\n')
if WARNINGS:
for w in WARNINGS:
f.write(f'- {w}\n')
else:
f.write('None.\n')
f.write(f'\n## Info ({len(INFO)})\n\n')
if INFO:
for i in INFO:
f.write(f'- {i}\n')
else:
f.write('None.\n')
print(f'Audit complete: {len(ERRORS)} errors, {len(WARNINGS)} warnings, {len(INFO)} info')
print(f'Report: {report_path}')
# Print to console too
if ERRORS:
print(f'\n=== ERRORS ({len(ERRORS)}) ===')
for e in ERRORS:
print(f' {e}')
if WARNINGS:
print(f'\n=== WARNINGS ({len(WARNINGS)}) ===')
for w in WARNINGS[:30]:
print(f' {w}')
if len(WARNINGS) > 30:
print(f' ... and {len(WARNINGS) - 30} more')
if __name__ == '__main__':
main()
|