Spaces:
Sleeping
Sleeping
File size: 21,655 Bytes
1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 236f287 1832e16 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 |
import json
import random
from concurrent.futures import ThreadPoolExecutor
from datetime import datetime
import librosa
import pandas as pd
import numpy as np
from audio import (
assign_outputs_to_refs_by_corr,
loudness_normalize,
make_union_voiced_mask,
)
from config import *
from distortions import apply_adv_distortions, apply_distortions
from metrics import (
compute_pm,
compute_ps,
diffusion_map_torch,
pm_ci_components_full,
ps_ci_components_full,
)
from models import embed_batch, load_model
from utils import *
def compute_mapss_measures(
models,
mixtures,
*,
systems=None,
algos=None,
experiment_id=None,
layer=DEFAULT_LAYER,
add_ci=DEFAULT_ADD_CI,
alpha=DEFAULT_ALPHA,
seed=42,
on_missing="skip",
verbose=False,
max_gpus=None,
):
gpu_distributor = GPUWorkDistributor(max_gpus)
ngpu = get_gpu_count(max_gpus)
if on_missing not in {"skip", "error"}:
raise ValueError("on_missing must be 'skip' or 'error'.")
torch.manual_seed(seed)
random.seed(seed)
np.random.seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
canon_mix = canonicalize_mixtures(mixtures, systems=systems)
mixture_entries = []
for m in canon_mix:
entries = []
for i, refp in enumerate(m.refs):
sid = m.speaker_ids[i]
entries.append(
{"id": sid, "ref": Path(refp), "mixture": m.mixture_id, "outs": {}}
)
mixture_entries.append(entries)
for m, mix_entries in zip(canon_mix, mixture_entries):
for algo, out_list in (m.systems or {}).items():
mapping = assign_outputs_to_refs_by_corr(
[e["ref"] for e in mix_entries], out_list
)
for idx, e in enumerate(mix_entries):
j = mapping[idx]
if j is not None:
e["outs"][algo] = out_list[j]
if algos is None:
algos_to_run = sorted(
{algo for algo in canon_mix[0].systems.keys()} if canon_mix and canon_mix[0].systems else []
)
else:
algos_to_run = list(algos)
exp_id = experiment_id or datetime.now().strftime("%Y%m%d_%H%M%S")
exp_root = os.path.join(RESULTS_ROOT, f"experiment_{exp_id}")
os.makedirs(exp_root, exist_ok=True)
params = {
"models": models,
"layer": layer,
"add_ci": add_ci,
"alpha": alpha,
"seed": seed,
"batch_size": BATCH_SIZE,
"ngpu": ngpu,
"max_gpus": max_gpus,
}
with open(os.path.join(exp_root, "params.json"), "w") as f:
json.dump(params, f, indent=2)
canon_struct = [
{
"mixture_id": m.mixture_id,
"references": [str(p) for p in m.refs],
"systems": {
a: [str(p) for p in outs] for a, outs in (m.systems or {}).items()
},
"speaker_ids": m.speaker_ids,
}
for m in canon_mix
]
with open(os.path.join(exp_root, "manifest_canonical.json"), "w") as f:
json.dump(canon_struct, f, indent=2)
print(f"Starting experiment {exp_id} with {ngpu} GPUs")
print(f"Results will be saved to: {exp_root}")
clear_gpu_memory()
get_gpu_memory_info(verbose)
flat_entries = [e for mix in mixture_entries for e in mix]
all_refs = {}
if verbose:
print("Loading reference signals...")
for e in flat_entries:
wav, _ = librosa.load(str(e["ref"]), sr=SR)
all_refs[e["id"]] = torch.from_numpy(loudness_normalize(wav))
if verbose:
print("Computing voiced masks...")
win = int(ENERGY_WIN_MS * SR / 1000)
hop = int(ENERGY_HOP_MS * SR / 1000)
voiced_mask_mix = []
total_frames_per_mix = [] # Store total frames for each mixture
for i, mix in enumerate(mixture_entries):
if verbose:
print(f" Computing mask for mixture {i + 1}/{len(mixture_entries)}")
if ngpu > 0:
with torch.cuda.device(0):
refs_for_mix = [all_refs[e["id"]].cuda() for e in mix]
mask = make_union_voiced_mask(refs_for_mix, win, hop)
voiced_mask_mix.append(mask.cpu())
total_frames_per_mix.append(mask.shape[0])
# Explicitly delete GPU tensors
for ref in refs_for_mix:
del ref
torch.cuda.empty_cache()
else:
refs_for_mix = [all_refs[e["id"]].cpu() for e in mix]
mask = make_union_voiced_mask(refs_for_mix, win, hop)
voiced_mask_mix.append(mask.cpu())
total_frames_per_mix.append(mask.shape[0])
ordered_speakers = [e["id"] for e in flat_entries]
# Initialize storage for all mixtures and algorithms
all_mixture_results = {} # mixture_id -> {algo -> {model -> data}}
for mix_idx, (mix_canon, mix_entries) in enumerate(zip(canon_mix, mixture_entries)):
mixture_id = mix_canon.mixture_id
all_mixture_results[mixture_id] = {}
# Get total frames for this mixture
total_frames = total_frames_per_mix[mix_idx]
# Get speakers for this mixture
mixture_speakers = [e["id"] for e in mix_entries]
for algo_idx, algo in enumerate(algos_to_run):
if verbose:
print(f"\nProcessing Mixture {mixture_id}, Algorithm {algo_idx + 1}/{len(algos_to_run)}: {algo}")
# Remove the old algo_dir creation here - we don't need these empty folders anymore
all_outs = {}
missing = []
for e in mix_entries:
assigned_path = e.get("outs", {}).get(algo)
if assigned_path is None:
missing.append((e["mixture"], e["id"]))
continue
wav, _ = librosa.load(str(assigned_path), sr=SR)
all_outs[e["id"]] = torch.from_numpy(loudness_normalize(wav))
if missing:
msg = f"[{algo}] missing outputs for {len(missing)} speaker(s) in mixture {mixture_id}"
if on_missing == "error":
raise FileNotFoundError(msg)
else:
if verbose:
warnings.warn(msg + " Skipping those speakers.")
if not all_outs:
if verbose:
warnings.warn(f"[{algo}] No outputs for mixture {mixture_id}. Skipping.")
continue
# Initialize storage for this algorithm
if algo not in all_mixture_results[mixture_id]:
all_mixture_results[mixture_id][algo] = {}
# Initialize frame-wise storage with NaN for all frames
ps_frames = {m: {s: [np.nan] * total_frames for s in mixture_speakers} for m in models}
pm_frames = {m: {s: [np.nan] * total_frames for s in mixture_speakers} for m in models}
ps_bias_frames = {m: {s: [np.nan] * total_frames for s in mixture_speakers} for m in models}
ps_prob_frames = {m: {s: [np.nan] * total_frames for s in mixture_speakers} for m in models}
pm_bias_frames = {m: {s: [np.nan] * total_frames for s in mixture_speakers} for m in models}
pm_prob_frames = {m: {s: [np.nan] * total_frames for s in mixture_speakers} for m in models}
for model_idx, mname in enumerate(models):
if verbose:
print(f" Processing Model {model_idx + 1}/{len(models)}: {mname}")
for metric_type in ["PS", "PM"]:
clear_gpu_memory()
gc.collect()
model_wrapper, layer_eff = load_model(mname, layer, max_gpus)
get_gpu_memory_info(verbose)
# Process only this mixture
speakers_this_mix = [e for e in mix_entries if e["id"] in all_outs]
if not speakers_this_mix:
continue
if verbose:
print(f" Processing {metric_type} for mixture {mixture_id}")
all_signals_mix = []
all_masks_mix = []
all_labels_mix = []
for e in speakers_this_mix:
s = e["id"]
if metric_type == "PS":
dists = [
loudness_normalize(d)
for d in apply_distortions(all_refs[s].numpy(), "all")
]
else:
dists = [
loudness_normalize(d)
for d in apply_adv_distortions(
all_refs[s].numpy(), "all"
)
]
sigs = [all_refs[s].numpy(), all_outs[s].numpy()] + dists
lbls = ["ref", "out"] + [f"d{i}" for i in range(len(dists))]
masks = [voiced_mask_mix[mix_idx]] * len(sigs)
all_signals_mix.extend(sigs)
all_masks_mix.extend(masks)
all_labels_mix.extend([f"{s}-{l}" for l in lbls])
try:
# Process in smaller batches
batch_size = min(2, BATCH_SIZE)
embeddings_list = []
for i in range(0, len(all_signals_mix), batch_size):
batch_sigs = all_signals_mix[i:i + batch_size]
batch_masks = all_masks_mix[i:i + batch_size]
batch_embs = embed_batch(
batch_sigs,
batch_masks,
model_wrapper,
layer_eff,
use_mlm=False,
)
if batch_embs.numel() > 0:
embeddings_list.append(batch_embs.cpu())
torch.cuda.empty_cache()
if embeddings_list:
embeddings = torch.cat(embeddings_list, dim=0)
E, L, D = embeddings.shape
if L == 0:
if verbose:
print(
f" WARNING: mixture {mixture_id} produced 0 frames after masking; skipping.")
continue
# Get valid frame indices for this mixture
mask = voiced_mask_mix[mix_idx]
valid_frame_indices = torch.where(mask)[0].tolist()
if verbose:
print(f" Computing {metric_type} scores for {mname}...")
# Process frames with their stored embeddings and labels
with ThreadPoolExecutor(
max_workers=min(2, ngpu if ngpu > 0 else 1)
) as executor:
def process_frame(f, frame_idx, embeddings_mix, labels_mix):
try:
frame_emb = embeddings_mix[:, f, :].detach().cpu().numpy()
if add_ci:
coords_d, coords_c, eigvals, k_sub_gauss = (
gpu_distributor.execute_on_gpu(
diffusion_map_torch,
frame_emb,
labels_mix,
alpha=alpha,
eig_solver="full",
return_eigs=True,
return_complement=True,
return_cval=add_ci,
)
)
else:
coords_d = gpu_distributor.execute_on_gpu(
diffusion_map_torch,
frame_emb,
labels_mix,
alpha=alpha,
eig_solver="full",
return_eigs=False,
return_complement=False,
return_cval=False,
)
coords_c = None
eigvals = None
k_sub_gauss = 1
if metric_type == "PS":
score = compute_ps(
coords_d, labels_mix, max_gpus
)
bias = prob = None
if add_ci:
bias, prob = ps_ci_components_full(
coords_d,
coords_c,
eigvals,
labels_mix,
delta=DEFAULT_DELTA_CI,
)
return frame_idx, "PS", score, bias, prob
else:
score = compute_pm(
coords_d, labels_mix, "gamma", max_gpus
)
bias = prob = None
if add_ci:
bias, prob = pm_ci_components_full(
coords_d,
coords_c,
eigvals,
labels_mix,
delta=DEFAULT_DELTA_CI,
K=k_sub_gauss,
)
return frame_idx, "PM", score, bias, prob
except Exception as ex:
if verbose:
print(f" ERROR frame {frame_idx}: {ex}")
return None
futures = [
executor.submit(process_frame, f, valid_frame_indices[f], embeddings,
all_labels_mix)
for f in range(L)
]
for fut in futures:
result = fut.result()
if result is None:
continue
frame_idx, metric, score, bias, prob = result
if metric == "PS":
for sp in score:
if sp in mixture_speakers:
ps_frames[mname][sp][frame_idx] = score[sp]
if add_ci and bias is not None:
ps_bias_frames[mname][sp][frame_idx] = bias[sp]
ps_prob_frames[mname][sp][frame_idx] = prob[sp]
else:
for sp in score:
if sp in mixture_speakers:
pm_frames[mname][sp][frame_idx] = score[sp]
if add_ci and bias is not None:
pm_bias_frames[mname][sp][frame_idx] = bias[sp]
pm_prob_frames[mname][sp][frame_idx] = prob[sp]
except Exception as ex:
if verbose:
print(f" ERROR processing mixture {mixture_id}: {ex}")
continue
finally:
# Always clean up after processing a mixture
del all_signals_mix, all_masks_mix
if 'embeddings_list' in locals():
del embeddings_list
clear_gpu_memory()
gc.collect()
del model_wrapper
clear_gpu_memory()
gc.collect()
# Store results for this mixture and algorithm
all_mixture_results[mixture_id][algo][mname] = {
'ps_frames': ps_frames[mname],
'pm_frames': pm_frames[mname],
'ps_bias_frames': ps_bias_frames[mname] if add_ci else None,
'ps_prob_frames': ps_prob_frames[mname] if add_ci else None,
'pm_bias_frames': pm_bias_frames[mname] if add_ci else None,
'pm_prob_frames': pm_prob_frames[mname] if add_ci else None,
'total_frames': total_frames
}
# Save results for this mixture after processing all algorithms
if verbose:
print(f" Saving results for mixture {mixture_id}...")
# Create timestamps in milliseconds - using lowercase hop
timestamps_ms = [i * hop * 1000 / SR for i in range(total_frames)]
for model in models:
# Prepare PS data
ps_data = {'timestamp_ms': timestamps_ms}
pm_data = {'timestamp_ms': timestamps_ms}
ci_data = {'timestamp_ms': timestamps_ms} if add_ci else None
# Combine data from all algorithms for this mixture
for algo in algos_to_run:
if algo not in all_mixture_results[mixture_id]:
continue
if model not in all_mixture_results[mixture_id][algo]:
continue
model_data = all_mixture_results[mixture_id][algo][model]
# Add PS data
for speaker in mixture_speakers:
col_name = f"{algo}_{speaker}"
ps_data[col_name] = model_data['ps_frames'][speaker]
pm_data[col_name] = model_data['pm_frames'][speaker]
if add_ci and ci_data is not None:
ci_data[f"{algo}_{speaker}_ps_bias"] = model_data['ps_bias_frames'][speaker]
ci_data[f"{algo}_{speaker}_ps_prob"] = model_data['ps_prob_frames'][speaker]
ci_data[f"{algo}_{speaker}_pm_bias"] = model_data['pm_bias_frames'][speaker]
ci_data[f"{algo}_{speaker}_pm_prob"] = model_data['pm_prob_frames'][speaker]
# Save CSV files for this mixture
mixture_dir = os.path.join(exp_root, mixture_id)
os.makedirs(mixture_dir, exist_ok=True)
pd.DataFrame(ps_data).to_csv(
os.path.join(mixture_dir, f"ps_scores_{model}.csv"),
index=False
)
pd.DataFrame(pm_data).to_csv(
os.path.join(mixture_dir, f"pm_scores_{model}.csv"),
index=False
)
if add_ci and ci_data is not None:
pd.DataFrame(ci_data).to_csv(
os.path.join(mixture_dir, f"ci_{model}.csv"),
index=False
)
del all_outs
clear_gpu_memory()
gc.collect()
print(f"\nEXPERIMENT COMPLETED")
print(f"Results saved to: {exp_root}")
del all_refs, voiced_mask_mix
# Import and call the cleanup function
from models import cleanup_all_models
cleanup_all_models()
clear_gpu_memory()
get_gpu_memory_info(verbose)
gc.collect()
return exp_root |