Datasets:
File size: 16,966 Bytes
347d9a3 | 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 | """Per-hall-family FOA rendering on a selected allocated GPU.
RLR acoustic propagation remains the pinned SoundSpaces native implementation.
GPU selection is explicit; sources/receivers use their actual scene positions.
Only temporary kernels are discarded. Metadata, SIR/SNR and acceptance reports
are retained. Noise emission gains are calibrated by the static job and shared
unchanged by the dynamic job.
"""
from __future__ import annotations
import argparse
import fcntl
import gzip
import json
import math
import os
from pathlib import Path
import shutil
import tempfile
import time
import numpy as np
import soundfile as sf
from dynamic_demo.render import Renderer, make_dry, two_time_convolution
from dynamic_dataset.runtime import ensure_compute
from .common import *
from .noise import waveform as noise_waveform
from .plan import audit_geometry
class NativeRenderer(Renderer):
"""Use the existing audited kernel conversion with explicit GPU selection."""
def __init__(self, root, recipe, cache, gpu):
import quaternion
import habitat_sim
self.hs, self.root, self.recipe, self.cache, self.position = habitat_sim, root, recipe, cache, position
cache.mkdir(parents=True, exist_ok=True)
self.records = {}
context = dict(renderer=recipe['renderer'], foa_mapping=recipe['foa_mapping'],
room_assets_sha256=recipe['room_assets_sha256'])
path = cache/'cache_context.json'
if path.exists() and read(path) != context:
raise ValueError('Recovery cache belongs to another acoustic recipe')
write(path, context)
folder = root/'assets/room'
for name, digest in recipe['room_assets_sha256'].items():
if sha(folder/name) != digest:
raise ValueError('Frozen room asset checksum mismatch')
config = habitat_sim.SimulatorConfiguration()
config.scene_id = str(folder/'stage.stage_config.json')
config.scene_dataset_config_file = str(folder/'dataset.scene_dataset_config.json')
config.load_semantic_mesh = True
config.enable_physics = False
config.gpu_device_id = gpu
if hasattr(config, 'random_seed'):
config.random_seed = 20260826
self.sim = habitat_sim.Simulator(habitat_sim.Configuration(config, [habitat_sim.agent.AgentConfiguration()]))
if not self.sim.semantic_scene:
raise ValueError('Acoustic semantic scene was not loaded')
spec = habitat_sim.AudioSensorSpec()
spec.uuid = 'audio_sensor'
spec.channelLayout.channelType = habitat_sim.sensor.RLRAudioPropagationChannelLayoutType.Ambisonics
spec.channelLayout.channelCount = 4
spec.position = [0., 0., 0.]
for key, value in recipe['renderer'].items():
if hasattr(spec.acousticsConfig, key):
setattr(spec.acousticsConfig, key, value)
self.sim.add_sensor(spec)
self.sensor = self.sim.get_agent(0)._sensors['audio_sensor']
self.sensor.setAudioMaterialsJSON(str(folder/'materials.json'))
def kernel(self, source, receiver, te, tr):
h = super().kernel(source, receiver, te, tr)
# Peak-packet DRR is a stated approximation, and only labelled for LOS.
a, b = position(source, te), position(receiver, tr)
key = __import__('dynamic_demo.common', fromlist=['uid']).uid(
self.recipe['room_assets_sha256'], *[round(float(x), 7) for x in np.r_[a, b, receiver['yaw_deg']]])
record = self.records[key]
if 'early_to_late_energy_ratio_db' not in record:
energy = h[:, 0].astype(float)**2
peak = round(record['direct_delay_s']*SR)
left, right = max(0, peak-8), min(len(h), peak+40)
early = float(energy[left:right].sum())
late = max(float(energy.sum())-early, 1e-30)
record.update(early_to_late_energy_ratio_db=10*math.log10(max(early, 1e-30)/late),
drr_proxy_valid=record['los'], drr_proxy_window_s=[-.0005, .0025])
return h
def source_mask(source, samples):
mask = np.zeros(samples, bool)
for event in source['events']:
for a, b in event['activity_samples']:
mask[max(0, a):min(samples, b)] = True
return mask
def fixed_source(spec, ordinal=0):
return dict(entity_id='N'+str(ordinal+1).zfill(2), entity_type='noise',
initial_xyz_m=spec['source_positions_world_m'][ordinal], displacement_m=[0., 0., 0.],
motion_start_s=0., motion_duration_s=1., yaw_deg=0)
def noise_component(root, episode, receiver, engine):
spec = episode['noise']
count = episode['duration_samples']
# Supply a prehistory so steady noise is already reverberant at scene zero.
pre = math.ceil(engine.recipe['renderer']['irTime']*SR)
dry = noise_waveform(spec, count+pre, root)
dry = np.r_[dry, np.zeros(pre+1, np.float32)]
rn = sorted(set([0., pre/SR] + [round(t+pre/SR, 9) for t in rir_nodes(receiver)]))
result = np.zeros((count, 4), np.float64)
for ordinal in range(len(spec['source_positions_world_m'])):
source = fixed_source(spec, ordinal)
rendered = two_time_convolution(dry, [0.], rn,
lambda te, tr: engine.kernel(source, receiver, 0., max(0., tr-pre/SR)),
max_tail_s=engine.recipe['renderer']['irTime']+.001)
result += rendered[pre:pre+count]
return result.astype(np.float32)
def calibration(root, group, engine, dry_cache):
path = root/'private/noise_calibration'/(group['group_id']+'.json')
if path.exists():
return read(path)
baseline = group['static_episodes'][0]
source, receiver = baseline['humans'][0], baseline['robots'][0]
dry = make_dry(root, baseline)[0][source['entity_id']]
target = two_time_convolution(dry, [0.], [0.], lambda te, tr: engine.kernel(source, receiver, te, tr),
max_tail_s=group['renderer']['irTime']+.001)
noise = noise_component(root, baseline, receiver, engine)
mask = source_mask(source, len(dry))
signal_power = float(np.mean(target[mask, 0].astype(float)**2))
noise_power = float(np.mean(noise[mask, 0].astype(float)**2))
if signal_power <= 0 or noise_power <= 0:
raise ValueError('Invalid anchor signal/noise power')
result = dict(group_id=group['group_id'], reference_mode='static', reference_robot_id='R01',
reference_source_id='H01', channel='W', mask='target emitted-speech activity intervals',
signal_power=signal_power, unit_noise_power=noise_power,
gains={str(int(level)): math.sqrt(signal_power/(noise_power*10**(level/10))) for level in [20., 10., 0.]},
shared_across_all_robots_and_modes=True)
write(path, result)
return result
def render_group(root, group, mode, gpu, cache_dir):
# Shared-filesystem advisory locks prevent duplicate writers when a second
# allocated node helps an existing queue. The lock inode stays in place;
# the OS releases ownership if a stalled native process is killed.
root,cache_dir=Path(root),Path(cache_dir)
lock_path=root/mode/'reports/locks'/(group['group_id']+'.lock')
lock_path.parent.mkdir(parents=True,exist_ok=True)
cache_dir.mkdir(parents=True,exist_ok=True)
with lock_path.open('a+') as lock:
while True:
try:
fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
break
except BlockingIOError:
write(cache_dir/'progress.json',dict(wall_time=time.time(),waiting_for='group_render_lock'))
time.sleep(5)
try:
return _render_group(root,group,mode,gpu,cache_dir)
finally:
fcntl.flock(lock,fcntl.LOCK_UN)
def _render_group(root, group, mode, gpu, cache_dir):
result_path = root/mode/'reports/groups'/(group['group_id']+'.json')
if result_path.exists():
result = read(result_path)
if not result['passed']:
raise ValueError('Previous group report did not pass')
for recording in result['recordings']:
if sha(root/mode/recording['path']) != recording['sha256']:
raise ValueError('Completed audio changed; cannot silently resume')
return result
ensure_compute()
started = time.monotonic()
cache_dir = Path(cache_dir)
engine_root = cache_dir/'engine'
(engine_root/'assets').mkdir(parents=True, exist_ok=True)
link = engine_root/'assets/room'
if not link.exists():
link.symlink_to(root/'shared/rooms'/group['room_id']/'assets/room', target_is_directory=True)
engine = NativeRenderer(engine_root, group, cache_dir/'rirs', gpu)
original_kernel = engine.kernel
def progress_kernel(source, receiver, te, tr):
write(cache_dir/'progress.json', dict(wall_time=time.time(), source_id=source['entity_id'],
receiver_id=receiver['entity_id'], source_time_s=te, receiver_time_s=tr))
return original_kernel(source, receiver, te, tr)
engine.kernel = progress_kernel
components = {}
dry_cache = {}
records, measurements, checks = [], [], []
try:
calibration_result = None
if group['family_type'] == 'noise':
if mode == 'static':
calibration_result = calibration(root, group, engine, dry_cache)
else:
path = root/'private/noise_calibration'/(group['group_id']+'.json')
for _ in range(360):
if path.exists():
break
write(cache_dir/'progress.json',dict(wall_time=time.time(),waiting_for='static_noise_calibration'))
time.sleep(5)
if not path.exists():
raise TimeoutError('Static-job shared noise calibration is missing')
calibration_result = read(path)
for episode in group[mode+'_episodes']:
if audit_geometry(episode, group['room'], group['family_type'] == 'count') is None:
raise ValueError('Frozen episode failed independent geometry validation')
dry, _ = make_dry(root, episode)
for receiver in episode['robots']:
rid = receiver['entity_id']
stems = {}
for source in episode['humans'] + episode['playback_sources']:
sid = source['entity_id']
if sid == rid:
continue
fingerprint = uid('stem', source, receiver, episode['duration_samples'])
if fingerprint not in components:
components[fingerprint] = two_time_convolution(dry[sid], rir_nodes(source), rir_nodes(receiver),
lambda te, tr, source=source: engine.kernel(source, receiver, te, tr),
max_tail_s=group['renderer']['irTime']+.001)
stems[sid] = components[fingerprint]
noise = np.zeros((episode['duration_samples'], 4), np.float32)
if episode['noise']['enabled']:
noise_key = uid('noise', episode['noise']['seed'], episode['noise'].get('recording'),
receiver, episode['duration_samples'])
if noise_key not in components:
components[noise_key] = noise_component(root, episode, receiver, engine)
gain = calibration_result['gains'][str(int(episode['noise']['nominal_snr_db']))]
noise = components[noise_key]*gain
speech = sum((w.astype(np.float64) for w in stems.values()), np.zeros_like(noise, dtype=np.float64))
wave = (speech+noise).astype(np.float32)
if not np.isfinite(wave).all() or not np.any(wave):
raise ValueError('Non-finite or silent FOA recording')
path = root/mode/'public/audio/test'/episode['episode_id']/(rid+'_foa.wav')
path.parent.mkdir(parents=True, exist_ok=True)
temporary = path.with_suffix('.tmp.wav')
sf.write(temporary, wave, SR, subtype='FLOAT')
temporary.replace(path)
readback, rate = sf.read(path, dtype='float32')
np.testing.assert_array_equal(readback, wave)
if rate != SR:
raise ValueError('Stored FOA sample-rate mismatch')
record = dict(recording_id=uid(episode['episode_id'], rid, 'recording'),
episode_id=episode['episode_id'], scene_id=episode['scene_id'], episode_group_id=group['group_id'],
receiver_robot_id=rid, path=str(path.relative_to(root/mode)), sha256=sha(path),
num_samples=len(wave), sample_rate_hz=SR, num_channels=4, channels=['W','Y','Z','X'],
ordering='ACN', normalization='SN3D', subtype='FLOAT', global_start_sample=0,
common_final_gain=1., raw_peak=float(np.max(np.abs(wave))))
records.append(record)
checks.append(dict(episode_id=episode['episode_id'], receiver_id=rid,
own_playback_excluded=True, source_ids=list(stems), noise_included=episode['noise']['enabled'],
all_finite=True, source_stem_reconstruction_max_error=float(np.max(np.abs(wave.astype(float)-(speech+noise)))),
tail_energy_fraction=float(np.sum(wave[-8000:].astype(float)**2)/np.sum(wave.astype(float)**2))))
for human in episode['humans']:
mask = source_mask(human, len(wave))
target = stems[human['entity_id']][:, 0].astype(float)
interference = speech[:, 0]-target
sp = float(np.mean(target[mask]**2))
ip = float(np.mean(interference[mask]**2))
npower = float(np.mean(noise[mask, 0].astype(float)**2))
ratio = lambda denominator: 10*math.log10(sp/denominator) if sp > 0 and denominator > 1e-30 else None
visibility = [dict(time_s=s/SR, visible=__import__('dynamic_demo.render', fromlist=['visible']).visible(
position(human, s/SR), position(receiver, s/SR), group['room'])) for s in grid(episode)]
measurements.append(dict(episode_id=episode['episode_id'], receiver_id=rid, source_id=human['entity_id'],
mask='target emitted-speech activity intervals; W-channel received energy',
target_power=sp, competing_speech_power=ip, environmental_noise_power=npower,
sir_db=ratio(ip), snr_db=ratio(npower), sinr_db=ratio(ip+npower),
visibility=visibility, nominal_snr_db=episode['noise'].get('nominal_snr_db')))
print('EPISODE RENDERED', mode, group['index'], episode['branch'], len(engine.records), 'RIRs', flush=True)
rir_log = root/mode/'private/rir_audits'/(group['group_id']+'.jsonl.gz')
rir_log.parent.mkdir(parents=True, exist_ok=True)
with gzip.open(rir_log, 'wt') as output:
for record in sorted(engine.records.values(), key=lambda r:r['rir_id']):
output.write(json.dumps(record, separators=(',', ':'), allow_nan=False)+'\n')
result = dict(passed=True, group_id=group['group_id'], index=group['index'], mode=mode,
scene_id=group['scene_id'], room_id=group['room_id'], family_type=group['family_type'],
episode_ids=[e['episode_id'] for e in group[mode+'_episodes']], recordings=records,
acoustic_measurements=measurements, audio_checks=checks, rir_audit_path=str(rir_log.relative_to(root/mode)),
rir_audit_sha256=sha(rir_log), rir_count=len(engine.records), raw_peak=max(r['raw_peak'] for r in records),
final_normalization_applied=False, gpu_device_id=gpu, wall_seconds=time.monotonic()-started,
label_step_s=.5, source_rir_step_s=.2, receiver_rir_step_s=.2,
human_listening_audit_completed=False, model_inference_performed=False)
write(result_path, result)
print('GROUP COMPLETE', mode, group['index'], len(records), 'FOA', round(result['wall_seconds'], 1), 's', flush=True)
return result
finally:
engine.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--root', type=Path, required=True)
parser.add_argument('--mode', choices=['static','dynamic'], required=True)
parser.add_argument('--index', type=int, required=True)
parser.add_argument('--gpu', type=int, required=True)
parser.add_argument('--cache-dir', type=Path, required=True)
a = parser.parse_args()
group = read(a.root/'private/plans'/f'{a.index:04d}.json')
render_group(a.root.resolve(), group, a.mode, a.gpu, a.cache_dir)
|