code-HFD / benchmark /evaluate.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
fb0074a verified
Raw History Blame Contribute Delete
14.1 kB
"""Pack HFD motion segments for the unchanged OmniHM Text/Speech evaluator.
Also score the complete idle/listen/motion episode and each GT-defined phase.
Physics is evaluated with the same FK, contact thresholds and jerk definition.
"""
import argparse
import json
import os
import subprocess
import sys
from pathlib import Path
from types import SimpleNamespace
from common import WORK, OMG, OMNI, EVALUATOR, PEER, qpos_from_absolute, sha256, write_json
from session_data import flattened_rows
import numpy as np
def verify_scores(cohort, generated, output):
"""Verify the complete score bundle against the frozen cohort, not file presence."""
def read(path):
return json.loads(Path(path).read_text())
def require(ok, message):
if not ok:raise ValueError(message)
manifest=read(cohort/'manifest.json');identity=read(generated/'identity.json')
protocol=read(output/'protocol.json');native=output/'motion'
rows=flattened_rows(manifest['rows']);n=len(rows)
ids=np.asarray([r['sample_id'] for r in rows])
require(identity['cohort_sha256']==sha256(cohort/'manifest.json'),'Generated cohort differs')
require(all(protocol.get(k)==v for k,v in identity.items()),'Metric generation identity differs')
require(protocol['count']==n and protocol['session_count']==len(manifest['rows']),'Metric counts differ')
state=read(native/'run_state.json')
require(state.get('status')=='complete' and state.get('kind')=='benchmark'
and state.get('paper_row_complete') is True,'Native scoring is incomplete')
raw=read(native/'metrics.json');paper=read(native/'paper_metrics.json')
require(raw['protocol']['num_samples']==n,'Native sample count differs')
require(paper.get('complete') is True and len(paper['values'])==14
and np.isfinite(np.asarray(paper['values'],float)).all(),'Incomplete/nonfinite paper row')
sys.path[:0]=[str(OMNI),str(OMG/'src')]
from metrics import build_paper_payload
require(paper==build_paper_payload('text',raw),'Displayed metrics differ from raw native results')
tracker=raw['tracker_executed']
require(tracker.get('enabled') is True and tracker.get('paper_eligible') is True
and tracker['num_samples']==n,'Missing/noncanonical/partial tracking')
for key in ('generated_retrieval_diagnostics','reference_retrieval_diagnostics'):
require(raw['generated_target']['text'][key]['retrieval_num_samples']==n//32*32,
'Retrieval tail count differs')
arrays={
'sample_metrics.npz':('contact_sliding_speed','body_jerk_mean'),
'tracker_failure_flags.npz':('mpjpe','g_mpjpe','e_vel','e_acc','fall','joint_limit'),
'embeddings.npz':('reference_embeddings','generated_embeddings','text_embeddings',
'distribution_reference_embeddings','distribution_generated_embeddings')}
for name,required in arrays.items():
with np.load(native/name,allow_pickle=False) as z:
require(np.array_equal(z['sample_id'],ids),f'{name}: sample identities/order differ')
require(set(required)<=set(z.files),f'{name}: missing metric arrays')
require(all(z[k].dtype.kind in 'biufc' for k in required),f'{name}: nonnumeric metric array')
for key in z.files:
value=z[key]
if value.dtype.kind in 'biufc':
require(value.ndim>=1 and len(value)==n and np.isfinite(value).all(),
f'{name}: invalid array {key}')
with np.load(output/'inputs/conditions.npz',allow_pickle=False) as z:
require(np.array_equal(z['sample_id'],ids),'Packed input identities differ')
require(z['captions'].tolist()==[r['caption'] for r in rows],'Final-command captions differ')
phase=read(output/'phase_physics.json')
require(phase['summary']['all']['sources']==len(manifest['rows']),'Incomplete full-session physics')
all_rows=[r for r in phase['per_segment'] if r['phase']=='all']
require(len(all_rows)==len(manifest['rows']) and {r['sample_id'] for r in all_rows}
=={r['sample_id'] for r in manifest['rows']},'Full-session physics identities differ')
for row in phase['per_segment']:
require(all(np.isfinite(v) for k in ('generated','gt') for v in row[k].values()),
'Nonfinite phase physics')
require(set(protocol['source_hashes'])=={str(p) for p in (OMNI/'evaluate.py',OMNI/'metrics.py',EVALUATOR)},
'Missing native metric/encoder provenance')
for path,expected in protocol['source_hashes'].items():
require(sha256(path)==expected,'Metric implementation/encoder changed: '+path)
adapter=read(native/'hfd_adapter.json')
require(adapter['adapter_sha256']==sha256(WORK/'native_eval.py'),'Native adapter changed')
receipt=dict(complete=True,session_count=len(manifest['rows']),motion_segments=n,
retrieval_count=n//32*32,cohort_sha256=identity['cohort_sha256'],
paper_metrics_sha256=sha256(native/'paper_metrics.json'),
scope='Completion/identity/finite-value check; no extra generation or score changes')
write_json(output/'result_audit.json',receipt)
return receipt
def pack(rows, cohort, generated, destination):
sys.path.insert(0,str(OMG/'src'))
from omg.robots.g1.constants import G1_JOINT_NAMES
pred=[];gt=[]
for row in rows:
s,e=row['score_start'],row['score_end']
with np.load(generated/(row.get('parent_sample_id',row['sample_id'])+'.npz')) as z:p=z['absolute']
with np.load(cohort/row['reference_file']) as z:g=z['absolute']
if len(p)<e or len(g)<e:raise ValueError('Truncated motion segment')
pred.append(qpos_from_absolute(p[s:e]));gt.append(qpos_from_absolute(g[s:e]))
lengths=np.array([len(x) for x in gt],np.int32)
if [len(x) for x in pred]!=lengths.tolist():raise ValueError('Mismatched timing')
n=int(lengths.max())
common=dict(fps=np.float32(30),valid=np.arange(n)[None]<lengths[:,None],lengths=lengths,
sample_id=np.array([r['sample_id'] for r in rows]),joint_names=np.array(G1_JOINT_NAMES),
captions=np.array([r['caption'] for r in rows]),dataset=np.array(['hfd-val']*len(rows)),
dataset_index=np.array([r['dataset_index'] for r in rows],np.int32))
destination.mkdir(parents=True,exist_ok=True)
for name,values in [('generated',pred),('reference',gt)]:
padded=np.stack([np.concatenate([x,np.repeat(x[-1:],n-len(x),axis=0)]) for x in values])
np.savez_compressed(destination/f'{name}.npz',qpos_36=padded,**common)
np.savez_compressed(destination/'conditions.npz',sample_id=common['sample_id'],
captions=np.array([r['caption'] for r in rows]),dataset=np.array(['hfd-val']*len(rows)),
dataset_index=np.array([r['dataset_index'] for r in rows],np.int32))
return n
def phase_physics(rows,cohort,generated,output):
import torch
sys.path.insert(0,str(OMG/'src'))
from omg.generation.metrics.physical import physical_qpos_metrics
from omg.robots.g1.kinematics import G1Kinematics
rep=SimpleNamespace(kinematics=G1Kinematics(OMG/'assets/robots/g1/g1_kinematics.json').cuda().eval())
keys=['contact_sliding_speed','body_jerk_mean']
records=[]
for row in rows:
with np.load(generated/(row['sample_id']+'.npz')) as z:p=z['absolute']
with np.load(cohort/row['reference_file']) as z:g=z['absolute'];labels=z['controls']
# Tail shorter than one emission is not silently padded or time-warped.
n=min(len(p),len(g))
qpred=qpos_from_absolute(p[:n]);qgt=qpos_from_absolute(g[:n])
time=np.arange(len(qpred))/30
states=labels[np.minimum(np.rint(time*50).astype(int),n-1)]
for phase,label in [('all',None),('idle',2),('listen',0),('motion',1)]:
mask=np.ones(len(qpred),bool) if label is None else states==label
# Split discontinuities: never differentiate across distinct phase intervals.
boundaries=np.flatnonzero(np.diff(np.r_[False,mask,False].astype(int))).reshape(-1,2)
for start,end in boundaries:
if end-start<4:continue
record=dict(sample_id=row['sample_id'],phase=phase,frames=int(end-start),
start_frame=int(start),end_frame=int(end),fps=30)
for name,q in [('generated',qpred),('gt',qgt)]:
x=torch.from_numpy(q[start:end][None]).cuda()
with torch.inference_mode():
value=physical_qpos_metrics(qpos_36=x,representation=rep,fps=30.,
valid=torch.ones(x.shape[:2],device=x.device,dtype=torch.bool),
contact_height_threshold=.12,contact_penetration_tolerance=.02)
record[name]={k:float(value[k].cpu()) for k in keys}
records.append(record)
# First aggregate segments per source, then give each source equal weight.
summary={}
for phase in ['all','idle','listen','motion']:
subset=[r for r in records if r['phase']==phase]
ids=sorted(set(r['sample_id'] for r in subset))
summary[phase]=dict(sources=len(ids),frames=sum(r['frames'] for r in subset))
for name in ['generated','gt']:
summary[phase][name]={}
for key in keys:
values=[]
for sid in ids:
same=[r for r in subset if r['sample_id']==sid]
values.append(np.average([r[name][key] for r in same],weights=[r['frames'] for r in same]))
summary[phase][name][key]=dict(mean=float(np.mean(values)),std=float(np.std(values))) if values else None
write_json(output,dict(fps=30,units=dict(contact_sliding_speed='m/s',body_jerk_mean='m/s^3'),
phase_source='fixed GT timeline; never predicted state',
aggregation='frame-weighted phase segments within source, then equal source mean; ddof=0',
note='All includes transitions. Short <4-frame phase segments excluded only from phase summaries.',
summary=summary,per_segment=records))
def main():
p=argparse.ArgumentParser()
p.add_argument('--cohort',type=Path,default=WORK/'outputs/sessions2048')
p.add_argument('--generated',type=Path,default=WORK/'outputs/session_ckpt_1800000/generated')
p.add_argument('--output',type=Path,default=WORK/'outputs/session_ckpt_1800000/eval')
p.add_argument('--tracker',choices=['run','skip'],default='run')
p.add_argument('--verify-only',action='store_true',help='Check a finished score bundle without re-running inference/evaluation')
a=p.parse_args()
a.cohort=a.cohort.resolve();a.generated=a.generated.resolve();a.output=a.output.resolve()
if a.verify_only:
print(json.dumps(verify_scores(a.cohort,a.generated,a.output),ensure_ascii=False))
return
a.output.mkdir(parents=True,exist_ok=True)
manifest=json.loads((a.cohort/'manifest.json').read_text())
identity=json.loads((a.generated/'identity.json').read_text())
if identity['cohort_sha256']!=sha256(a.cohort/'manifest.json'):raise ValueError('Wrong cohort')
rows=manifest['rows']
for row in rows:
meta=json.loads((a.generated/(row['sample_id']+'.json')).read_text())
if any(meta.get(k)!=v for k,v in identity.items()):raise ValueError('Mixed generation identity')
packdir=a.output/'inputs'
segments=flattened_rows(rows)
frames=pack(segments,a.cohort,a.generated,packdir)
write_json(a.output/'scored_segments.json',[{k:r[k] for k in ('sample_id','parent_sample_id','turn_index','caption','task','score_start','score_end') if k in r} for r in segments])
phase_physics(rows,a.cohort,a.generated,a.output/'phase_physics.json')
options={
'task':'text','generated-path':packdir/'generated.npz','reference-path':packdir/'reference.npz',
'condition-path':packdir/'conditions.npz','output-dir':a.output/'motion',
'omg-root':OMG,'expected-frames':frames,'expected-fps':30,'device':'cuda','batch-size':32,
'validate-only':0,'kinematics-path':OMG/'assets/robots/g1/g1_kinematics.json',
'contact-height-threshold':.12,'contact-penetration-tolerance':.02,
'evaluator-checkpoint':EVALUATOR,'retrieval-evaluator-checkpoint':EVALUATOR,
'motion-key':'body_pos_local','t5-model':OMG/'models/t5-3b-local',
'text-cache-dir':a.output/'text_cache','text-batch-size':16,'seed':0,
'beat-direction':'music_to_motion','beat-threshold':.5,'sigma-frames':3.,
'min-motion-beat-distance-seconds':.25,'tracker-mode':a.tracker,
'tracker-artifact':a.output/'unused.npz',
'holomotion-onnx':OMG/'models/HoloMotion/HoloMotion_v1.2/holomotion_v1.2_motion_tracking_model/exported/model.onnx',
'tracker-robot-xml':OMG/'assets/holomotion/g1_29dof/scene_29dof.xml',
'tracker-target-fps':50,'tracker-providers':'TensorrtExecutionProvider,CUDAExecutionProvider',
'tracker-control-substeps':10,'tracker-action-clip':10,'tracker-steps':0}
command=[sys.executable,'-B',str(WORK/'native_eval.py')]
for k,v in options.items():command.extend(['--'+k,str(v)])
write_json(a.output/'protocol.json',dict(**identity,conditioning='speech, not text',
reference_table='Table 3 / Table 6, same motion metrics; not music metrics',
session_count=len(rows),count=len(segments),small_sample_diagnostic=len(rows)<2048,direct_paper_comparison=False,
retrieval_count=len(segments)//32*32,retrieval_tail_omitted=len(segments)%32,
aggregation='Equal eligible source-occurrence weight; full-session physics separately. Retrieval uses complete batch32 only.',
source_hashes={str(x):sha256(x) for x in [OMNI/'evaluate.py',OMNI/'metrics.py',EVALUATOR]},
command=command,scope='motion-only tracking/distribution/retrieval; full and phase physics separately'))
subprocess.run(command,cwd=OMG,check=True,env={**os.environ,'PYTHONDONTWRITEBYTECODE':'1'})
if a.tracker=='run':verify_scores(a.cohort,a.generated,a.output)
if __name__=='__main__':main()