code-HFD / benchmark /test_native_eval.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
fb0074a verified
Raw History Blame Contribute Delete
10.9 kB
"""Non-multiple-of-32 sessions must not disable native input integrity checks."""
import tempfile
import unittest
from unittest.mock import patch
import json
from pathlib import Path
import numpy as np
from common import WORK, OMG, OMNI, EVALUATOR, sha256, write_json
from native_eval import install_tail_policy, install_bounded_motion_conversion
class EntryPointTests(unittest.TestCase):
def test_provenance_does_not_prepopulate_native_output(self):
import native_eval
with tempfile.TemporaryDirectory() as tmp:
output=Path(tmp)/'new_output'
def native_main(*args, **kwargs):
self.assertFalse(output.exists())
output.mkdir()
raise SystemExit(0)
with patch.object(native_eval.sys,'argv',['native_eval.py','--output-dir',str(output)]), \
patch.object(native_eval,'install_tail_policy'), \
patch.object(native_eval,'install_bounded_motion_conversion'), \
patch.object(native_eval.runpy,'run_path',side_effect=native_main):
native_eval.main()
self.assertTrue((output/'hfd_adapter.json').is_file())
class BoundedFKTests(unittest.TestCase):
def test_native_fk_matches_for_all_sources_frames_and_motion_keys(self):
import torch
from omg.benchmarks.runners import common
from omg.robots.g1.kinematics import G1Kinematics
kin=G1Kinematics(str(OMG/'assets/robots/g1/g1_kinematics.json')).eval()
rng=torch.Generator().manual_seed(42)
qpos=torch.randn((33,9,36),generator=rng)*.1
qpos[...,2]+=.8
quat=torch.randn((33,9,4),generator=rng)
qpos[...,3:7]=quat/torch.linalg.vector_norm(quat,dim=-1,keepdim=True)
original=common._motion_for_evaluator
try:
install_bounded_motion_conversion()
with torch.inference_mode():
for key in ('qpos_36','body_pos_local','body_link_pos_local'):
expected=original(qpos,key,kinematics=kin)
actual=common._motion_for_evaluator(qpos,key,kinematics=kin)
torch.testing.assert_close(actual,expected,rtol=0,atol=0)
finally:
common._motion_for_evaluator=original
class ScoreBundleTests(unittest.TestCase):
def fixture(self, root):
import importlib.util
from metrics import build_paper_payload
from session_data import flattened_rows
spec=importlib.util.spec_from_file_location('hfd_score_check',WORK/'evaluate.py')
module=importlib.util.module_from_spec(spec);spec.loader.exec_module(module)
cohort,generated,output=(root/x for x in ('cohort','generated','eval'))
native=output/'motion';native.mkdir(parents=True)
original=[dict(sample_id=f's{i}',caption='stand',reference_frames=51) for i in range(32)]
write_json(cohort/'manifest.json',dict(rows=original))
rows=flattened_rows(original);ids=np.array([r['sample_id'] for r in rows])
identity=dict(step=1,cohort_sha256=sha256(cohort/'manifest.json'))
write_json(generated/'identity.json',identity)
write_json(output/'protocol.json',dict(**identity,count=32,session_count=32,
source_hashes={str(p):sha256(p) for p in (OMNI/'evaluate.py',OMNI/'metrics.py',EVALUATOR)}))
write_json(native/'run_state.json',dict(status='complete',kind='benchmark',paper_row_complete=True))
raw=dict(protocol=dict(num_samples=32),generated_target=dict(
text=dict(motion_fid=.1,matching_score=1.,r_precision=[.3,.5,.7],diversity_generated=1.2,
generated_retrieval_diagnostics=dict(retrieval_num_samples=32),
reference_retrieval_diagnostics=dict(retrieval_num_samples=32)),
physical=dict(contact_sliding_speed=dict(mean=.1),body_jerk_mean=dict(mean=20.))),
tracker_executed=dict(enabled=True,paper_eligible=True,num_samples=32,
metrics={k:dict(mean=1.) for k in ('mpjpe','g_mpjpe','e_vel','e_acc')},
failures=dict(fall_rate_percent=0.,joint_limit_rate_percent=0.)))
write_json(native/'metrics.json',raw)
write_json(native/'paper_metrics.json',build_paper_payload('text',raw))
np.savez(native/'sample_metrics.npz',sample_id=ids,
contact_sliding_speed=np.full(32,.1),body_jerk_mean=np.full(32,20.))
np.savez(native/'tracker_failure_flags.npz',sample_id=ids,
**{k:np.ones(32) for k in ('mpjpe','g_mpjpe','e_vel','e_acc')},
fall=np.zeros(32,bool),joint_limit=np.zeros(32,bool))
np.savez(native/'embeddings.npz',sample_id=ids,
**{k:np.ones((32,512),np.float32) for k in ('reference_embeddings','generated_embeddings',
'text_embeddings','distribution_reference_embeddings','distribution_generated_embeddings')})
(output/'inputs').mkdir()
np.savez(output/'inputs/conditions.npz',sample_id=ids,captions=np.array(['stand']*32))
write_json(output/'phase_physics.json',dict(summary=dict(all=dict(sources=32)),
per_segment=[dict(sample_id=r['sample_id'],phase='all',frames=31,
generated=dict(jerk=1.),gt=dict(jerk=1.)) for r in rows]))
write_json(native/'hfd_adapter.json',dict(adapter_sha256=sha256(WORK/'native_eval.py')))
return module.verify_scores,(cohort,generated,output)
def test_complete_bundle_and_incomplete_status(self):
with tempfile.TemporaryDirectory() as tmp:
verify,args=self.fixture(Path(tmp));out=args[-1]
self.assertTrue(verify(*args)['complete'])
state=out/'motion/run_state.json'
write_json(state,dict(status='running',kind='benchmark',paper_row_complete=True))
with self.assertRaisesRegex(ValueError,'incomplete'):verify(*args)
def test_mixed_order_nonfinite_and_inconsistent_table_rejected(self):
with tempfile.TemporaryDirectory() as tmp:
verify,args=self.fixture(Path(tmp));out=args[-1];native=out/'motion'
path=native/'embeddings.npz'
with np.load(path) as z:original={k:z[k] for k in z.files}
np.savez(path,**{**original,'sample_id':original['sample_id'][::-1]})
with self.assertRaisesRegex(ValueError,'identities/order'):verify(*args)
values=original['generated_embeddings'].copy();values[-1]=np.nan
np.savez(path,**{**original,'generated_embeddings':values})
with self.assertRaisesRegex(ValueError,'invalid array'):verify(*args)
np.savez(path,**original)
paper=json.loads((native/'paper_metrics.json').read_text());paper['values'][0]+=1
write_json(native/'paper_metrics.json',paper)
with self.assertRaisesRegex(ValueError,'Displayed metrics'):verify(*args)
class TailPolicyTests(unittest.TestCase):
@classmethod
def setUpClass(cls):
import metrics
from omg.benchmarks.runners import text
cls.metrics,cls.text=metrics,text
cls.original_loader=staticmethod(metrics.load_benchmark_inputs)
cls.original_summary=staticmethod(text._text_retrieval_summary)
install_tail_policy()
@classmethod
def tearDownClass(cls):
cls.metrics.load_benchmark_inputs=cls.original_loader
cls.text._text_retrieval_summary=cls.original_summary
def fixture(self, directory, count):
import importlib.util
spec=importlib.util.spec_from_file_location('hfd_tail_pack',WORK/'evaluate.py')
module=importlib.util.module_from_spec(spec);spec.loader.exec_module(module)
motion=np.zeros((51,36),np.float32);motion[:,31]=.8;motion[:,32]=1
np.savez(directory/'source.npz',absolute=motion)
rows=[dict(sample_id='source',score_start=0,score_end=51,caption='stand',
dataset_index=i,reference_file='source.npz') for i in range(count)]
# Generated file is shared here, but packed identities must be unique.
for i,row in enumerate(rows):
row['sample_id']=f'sample_{i}'
np.savez(directory/(row['sample_id']+'.npz'),absolute=motion)
module.pack(rows,directory,directory,directory/'packed')
return dict(task='text',generated_path=str(directory/'packed/generated.npz'),
reference_path=str(directory/'packed/reference.npz'),
condition_path=str(directory/'packed/conditions.npz'),
expected_frames=31,expected_fps=30)
def test_complete_batch_unchanged_and_all_33_retained(self):
with tempfile.TemporaryDirectory() as tmp:
root=Path(tmp);args=self.fixture(root,32)
old=self.original_loader(**args);new=self.metrics.load_benchmark_inputs(**args)
self.assertEqual(old.keys(),new.keys())
for key in old:np.testing.assert_equal(old[key],new[key])
args=self.fixture(root,33)
with self.assertRaisesRegex(ValueError,'batch-32'):self.original_loader(**args)
self.assertEqual(self.metrics.load_benchmark_inputs(**args)['num_samples'],33)
with self.assertRaisesRegex(ValueError,'batch-32'):
self.metrics.load_benchmark_inputs(**self.fixture(root,31))
def test_tail_still_validates_identity_caption_quaternion_and_lengths(self):
with tempfile.TemporaryDirectory() as tmp:
root=Path(tmp);args=self.fixture(root,33);path=Path(args['generated_path'])
with np.load(path) as z:original={k:z[k] for k in z.files}
for field in ('sample_id','captions','qpos_36','lengths'):
changed={k:v.copy() for k,v in original.items()}
if field in ('sample_id','captions'):changed[field][-1]='BAD'
elif field=='qpos_36':changed[field][-1,0,3:7]=0
else:changed[field][-1]-=1
np.savez(path,**changed)
with self.subTest(field=field),self.assertRaises(ValueError):
self.metrics.load_benchmark_inputs(**args)
def test_native_retrieval_formula_and_full_matching_preserved(self):
rng=np.random.default_rng(42)
motion=rng.normal(size=(33,16)).astype(np.float32);text=motion.copy();text[-1]+=5
expected=self.original_summary(motion,text,require_full_batches=False)
actual=self.text._text_retrieval_summary(motion,text,require_full_batches=True)
self.assertEqual(actual,expected)
self.assertEqual(actual['retrieval_num_samples'],32)
self.assertAlmostEqual(actual['matching_score'],float(np.linalg.norm(motion-text,axis=-1).mean()),places=6)
self.assertEqual(self.text._text_retrieval_summary(motion[:32],text[:32]),
self.original_summary(motion[:32],text[:32]))
if __name__=='__main__':unittest.main()