Spaces:
Running
on
A10G
Running
on
A10G
File size: 36,331 Bytes
bdab1da |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 |
import json
import logging
import math
import os
import time
from contextlib import suppress
import numpy as np
import torch
import torch.nn.functional as F
try:
import wandb
except ImportError:
wandb = None
from open_clip import ClipLoss, gather_features
from .distributed import is_master
from .zero_shot import zero_shot_eval
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def unwrap_model(model):
if hasattr(model, "module"):
return model.module
else:
return model
def train_one_epoch(
model, data, epoch, optimizer, scaler, scheduler, args, tb_writer=None
):
device = torch.device(args.device)
autocast = torch.cuda.amp.autocast if args.precision == "amp" else suppress
model.train()
loss = ClipLoss(
local_loss=args.local_loss,
gather_with_grad=args.gather_with_grad,
cache_labels=True,
rank=args.rank,
world_size=args.world_size,
use_horovod=args.horovod,
mlp_loss=args.clap_mlploss,
weight_loss_kappa=args.kappa,
)
dataloader, sampler = data["train"].dataloader, data["train"].sampler
if args.distributed and sampler is not None:
sampler.set_epoch(epoch)
num_batches_per_epoch = dataloader.num_batches
sample_digits = math.ceil(math.log(dataloader.num_samples + 1, 10))
# for toy dataset
if args.dataset_type == "toy":
dataloader.dataset.generate_queue()
loss_m = AverageMeter()
batch_time_m = AverageMeter()
data_time_m = AverageMeter()
end = time.time()
for i, batch in enumerate(dataloader):
# logging.info(f"batch {i} of {num_batches_per_epoch}")
step = num_batches_per_epoch * epoch + i
if isinstance(scheduler, dict):
for s in scheduler.values():
s(step)
else:
scheduler(step)
audios = batch # contains mel_spec, wavform, and longer list
texts = batch["text"]
# audios = audios.to(device=device, non_blocking=True)
# texts = texts.to(device=device, non_blocking=True)
data_time_m.update(time.time() - end)
if isinstance(optimizer, dict):
for o_ in optimizer.values():
o_.zero_grad()
else:
optimizer.zero_grad()
with autocast():
(
audio_features,
text_features,
audio_features_mlp,
text_features_mlp,
logit_scale_a,
logit_scale_t,
) = model(audios, texts, device)
if args.clap_mlploss:
total_loss = loss(
audio_features=audio_features,
text_features=text_features,
logit_scale_a=logit_scale_a,
logit_scale_t=logit_scale_t,
audio_features_mlp=audio_features_mlp,
text_features_mlp=text_features_mlp,
)
else:
total_loss = loss(
audio_features=audio_features,
text_features=text_features,
logit_scale_a=logit_scale_a,
)
if isinstance(optimizer, dict):
if scaler is not None:
scaler.scale(total_loss).backward()
for o_ in optimizer.values():
if args.horovod:
o_.synchronize()
scaler.unscale_(o_)
with o_.skip_synchronize():
scaler.step(o_)
else:
scaler.step(o_)
scaler.update()
else:
total_loss.backward()
for o_ in optimizer.values():
o_.step()
else:
if scaler is not None:
scaler.scale(total_loss).backward()
if args.horovod:
optimizer.synchronize()
scaler.unscale_(optimizer)
with optimizer.skip_synchronize():
scaler.step(optimizer)
else:
scaler.step(optimizer)
scaler.update()
else:
total_loss.backward()
optimizer.step()
# Note: we clamp to 4.6052 = ln(100), as in the original paper.
with torch.no_grad():
unwrap_model(model).logit_scale_a.clamp_(0, math.log(100))
if args.clap_mlploss:
unwrap_model(model).logit_scale_t.clamp_(0, math.log(100))
batch_time_m.update(time.time() - end)
end = time.time()
batch_count = i + 1
if is_master(args) and (i % 100 == 0 or batch_count == num_batches_per_epoch):
if isinstance(audios, dict):
batch_size = len(audios["waveform"])
else:
batch_size = len(audios)
num_samples = batch_count * batch_size * args.world_size
samples_per_epoch = dataloader.num_samples
percent_complete = 100.0 * batch_count / num_batches_per_epoch
# NOTE loss is coarsely sampled, just master node and per log update
loss_m.update(total_loss.item(), batch_size)
logit_scale_scalar_a = logit_scale_a.item()
logit_scale_scalar_t = logit_scale_t.item()
if isinstance(optimizer, dict):
if args.clap_mlploss:
logging.info(
f"Train Epoch: {epoch} [{num_samples:>{sample_digits}}/{samples_per_epoch} ({percent_complete:.0f}%)] "
f"Loss: {loss_m.val:#.5g} ({loss_m.avg:#.4g}) "
f"Data (t): {data_time_m.avg:.3f} "
f"Batch (t): {batch_time_m.avg:.3f} "
f"LR: {[o_.param_groups[0]['lr'] for o_ in optimizer.values()]} "
f"Logit Scale Audio: {logit_scale_scalar_a:.3f}"
f"Logit Scale Text: {logit_scale_scalar_t:.3f}"
)
log_data = {
"loss": loss_m.val,
"data_time": data_time_m.val,
"batch_time": batch_time_m.val,
"scale_audio": logit_scale_scalar_a,
"scale_text": logit_scale_scalar_t,
"lr": [o_.param_groups[0]["lr"] for o_ in optimizer.values()],
}
else:
logging.info(
f"Train Epoch: {epoch} [{num_samples:>{sample_digits}}/{samples_per_epoch} ({percent_complete:.0f}%)] "
f"Loss: {loss_m.val:#.5g} ({loss_m.avg:#.4g}) "
f"Data (t): {data_time_m.avg:.3f} "
f"Batch (t): {batch_time_m.avg:.3f} "
f"LR: {[o_.param_groups[0]['lr'] for o_ in optimizer.values()]} "
f"Logit Scale Audio: {logit_scale_scalar_a:.3f}"
)
log_data = {
"loss": loss_m.val,
"data_time": data_time_m.val,
"batch_time": batch_time_m.val,
"scale_audio": logit_scale_scalar_a,
"lr": [o_.param_groups[0]["lr"] for o_ in optimizer.values()],
}
else:
if args.clap_mlploss:
logging.info(
f"Train Epoch: {epoch} [{num_samples:>{sample_digits}}/{samples_per_epoch} ({percent_complete:.0f}%)] "
f"Loss: {loss_m.val:#.5g} ({loss_m.avg:#.4g}) "
f"Data (t): {data_time_m.avg:.3f} "
f"Batch (t): {batch_time_m.avg:.3f} "
f"LR: {optimizer.param_groups[0]['lr']:5f} "
f"Logit Scale Audio: {logit_scale_scalar_a:.3f}"
f"Logit Scale Text: {logit_scale_scalar_t:.3f}"
)
# Save train loss / etc. Using non avg meter values as loggers have their own smoothing
log_data = {
"loss": loss_m.val,
"data_time": data_time_m.val,
"batch_time": batch_time_m.val,
"scale_audio": logit_scale_scalar_a,
"scale_text": logit_scale_scalar_t,
"lr": optimizer.param_groups[0]["lr"],
}
else:
logging.info(
f"Train Epoch: {epoch} [{num_samples:>{sample_digits}}/{samples_per_epoch} ({percent_complete:.0f}%)] "
f"Loss: {loss_m.val:#.5g} ({loss_m.avg:#.4g}) "
f"Data (t): {data_time_m.avg:.3f} "
f"Batch (t): {batch_time_m.avg:.3f} "
f"LR: {optimizer.param_groups[0]['lr']:5f} "
f"Logit Scale Audio: {logit_scale_scalar_a:.3f}"
)
# Save train loss / etc. Using non avg meter values as loggers have their own smoothing
log_data = {
"loss": loss_m.val,
"data_time": data_time_m.val,
"batch_time": batch_time_m.val,
"scale_audio": logit_scale_scalar_a,
"lr": optimizer.param_groups[0]["lr"],
}
for name, val in log_data.items():
name = "train/" + name
if tb_writer is not None:
tb_writer.add_scalar(name, val, step)
if args.wandb:
assert wandb is not None, "Please install wandb."
wandb.log({name: val, "step": step})
# resetting batch / data time meters per log window
batch_time_m.reset()
data_time_m.reset()
# end for
def evaluate(model, data, epoch, args, tb_writer=None):
metrics = {}
if not args.parallel_eval:
if not is_master(args):
return metrics
device = torch.device(args.device)
model.eval()
# CHANGE
# zero_shot_metrics = zero_shot_eval(model, data, epoch, args)
# metrics.update(zero_shot_metrics)
if is_master(args):
print("Evaluating...")
autocast = torch.cuda.amp.autocast if args.precision == "amp" else suppress
if args.val_dataset_names == ["Clotho", "audiocaps"]:
# if only clotho and audiocaps are used, then we will use a different evaluation function.
# This is because in the Clotho and audiocaps valid and test set, there are 5 text for 1 audio.
if args.parallel_eval:
# (yusong): just a hack here. Don't use parallel eval when evaluating only clotho and audiocaps.
raise NotImplementedError(
"Parallel evaluation not supported for eval only Clotho and audiocaps."
)
val_metrics_per_dataset = evaluate_clotho_audiocaps(
model, data, epoch, args, autocast, device, tb_writer
)
for m in val_metrics_per_dataset.values():
metrics.update(m)
if "epoch" not in metrics.keys():
metrics.update({"epoch": epoch})
metrics = select_top_metric_clotho_audiocaps(
metrics, val_metrics_per_dataset, args
)
elif "val" in data and (
args.val_frequency
and ((epoch % args.val_frequency) == 0 or epoch == args.epochs)
):
dataloader = data["val"].dataloader
num_samples = 0
samples_per_val = dataloader.num_samples
# FIXME this does not scale past small eval datasets
# all_audio_features @ all_text_features will blow up memory and compute very quickly
eval_info = {}
if args.clap_mlploss:
eval_info["all"] = {
"cumulative_loss": 0.0,
"num_samples": 0,
"all_audio_features": [],
"all_text_features": [],
"all_audio_features_mlp": [],
"all_text_features_mlp": [],
} # cumulative_loss = 0.0
else:
eval_info["all"] = {
"cumulative_loss": 0.0,
"num_samples": 0,
"all_audio_features": [],
"all_text_features": [],
} # cumu
# all_audio_features, all_text_features, all_audio_features_mlp, all_text_features_mlp = [], [], [], []
with torch.no_grad():
for i, batch in enumerate(dataloader):
audios = batch # contains mel_spec, wavform, and longer list
texts = batch["text"]
# audios = audios.to(device=device, non_blocking=True)
all_names = list(
set(["-".join(b.split("/")[-3:-1]) for b in batch["__url__"]])
)
for name in all_names:
if name not in eval_info.keys():
if args.clap_mlploss:
eval_info[name] = {
"cumulative_loss": 0.0,
"num_samples": 0,
"all_audio_features": [],
"all_text_features": [],
"all_audio_features_mlp": [],
"all_text_features_mlp": [],
}
else:
eval_info[name] = {
"cumulative_loss": 0.0,
"num_samples": 0,
"all_audio_features": [],
"all_text_features": [],
}
with autocast():
(
audio_features,
text_features,
audio_features_mlp,
text_features_mlp,
logit_scale_a,
logit_scale_t,
) = model(audios, texts, device)
if args.parallel_eval:
# multi-GPU eval
if args.clap_mlploss:
(
audio_features,
text_features,
audio_features_mlp,
text_features_mlp,
) = gather_features(
audio_features=audio_features,
text_features=text_features,
audio_features_mlp=audio_features_mlp,
text_features_mlp=text_features_mlp,
local_loss=False,
gather_with_grad=False,
rank=args.rank,
world_size=args.world_size,
use_horovod=args.horovod,
mlp_loss=args.clap_mlploss,
)
else:
(audio_features, text_features,) = gather_features(
audio_features=audio_features,
text_features=text_features,
local_loss=False,
gather_with_grad=False,
rank=args.rank,
world_size=args.world_size,
use_horovod=args.horovod,
mlp_loss=args.clap_mlploss,
)
if is_master(args):
num_samples += audio_features.shape[0]
for n in [*all_names, "all"]:
if n == "all":
eval_info[n]["all_audio_features"].append(
audio_features.cpu()
)
eval_info[n]["all_text_features"].append(
text_features.cpu()
)
if args.clap_mlploss:
eval_info[n]["all_audio_features_mlp"].append(
audio_features_mlp.cpu()
)
eval_info[n]["all_text_features_mlp"].append(
text_features_mlp.cpu()
)
else:
idx = np.where(
np.array(
[
"-".join(b.split("/")[-3:-1])
for b in batch["__url__"]
]
)
== n
)[0]
eval_info[n]["all_audio_features"].append(
audio_features.cpu().index_select(
0, torch.tensor(idx).long()
)
)
eval_info[n]["all_text_features"].append(
text_features.cpu().index_select(
0, torch.tensor(idx).long()
)
)
if args.clap_mlploss:
eval_info[n]["all_audio_features_mlp"].append(
audio_features_mlp.cpu().index_select(
0, torch.tensor(idx).long()
)
)
eval_info[n]["all_text_features_mlp"].append(
text_features_mlp.cpu().index_select(
0, torch.tensor(idx).long()
)
)
# print(f'eval step {i}') # (yusong): for debug
# cumulative_loss += total_loss * batch_size
# num_samples += batch_size
if is_master(args) and (i % 100) == 0: # and i != 0:
logging.info(
f"Eval Epoch: {epoch} [{num_samples} / {samples_per_val}]"
)
if is_master(args):
val_metrics_per_dataset = {}
for n in eval_info.keys():
if args.clap_mlploss:
metrics_single_dataset = get_metrics(
audio_features=torch.cat(
eval_info[n]["all_audio_features"]
),
text_features=torch.cat(eval_info[n]["all_text_features"]),
logit_scale_a=logit_scale_a.cpu(),
audio_features_mlp=torch.cat(
eval_info[n]["all_audio_features_mlp"]
),
text_features_mlp=torch.cat(
eval_info[n]["all_text_features_mlp"]
),
logit_scale_t=logit_scale_t.cpu(),
mlp_loss=args.clap_mlploss,
)
else:
metrics_single_dataset = get_metrics(
audio_features=torch.cat(
eval_info[n]["all_audio_features"]
),
text_features=torch.cat(eval_info[n]["all_text_features"]),
logit_scale_a=logit_scale_a.cpu(),
mlp_loss=args.clap_mlploss,
)
val_metrics_per_dataset[n] = {
n + "/" + k: v for k, v in metrics_single_dataset.items()
}
metrics.update(val_metrics_per_dataset[n])
if "epoch" not in metrics.keys():
metrics.update({"epoch": epoch})
if is_master(args):
if not metrics:
return metrics
logging.info(
f"Eval Epoch: {epoch} "
+ "\n".join(
[
"\t".join([f"{k}: {round(v, 4):.4f}" for k, v in m.items()])
for m in val_metrics_per_dataset.values()
]
)
)
if args.save_logs:
for name, val in metrics.items():
if tb_writer is not None:
tb_writer.add_scalar(f"val/{name}", val, epoch)
with open(os.path.join(args.checkpoint_path, "results.jsonl"), "a+") as f:
f.write(json.dumps(metrics))
f.write("\n")
if args.wandb:
assert wandb is not None, "Please install wandb."
for name, val in metrics.items():
wandb.log({f"val/{name}": val, "epoch": epoch})
return metrics
else:
return metrics
def get_metrics(
audio_features,
text_features,
logit_scale_a,
audio_features_mlp=None,
text_features_mlp=None,
logit_scale_t=None,
mlp_loss=False,
):
metrics = {}
if mlp_loss:
# Set up audio to text & text to audio similary matrice
a_logits_per_audio = (
(logit_scale_a * audio_features @ text_features_mlp.t()).detach().cpu()
)
a_logits_per_text = a_logits_per_audio.t().detach().cpu()
t_logits_per_audio = (
(logit_scale_t * audio_features_mlp @ text_features.t()).detach().cpu()
)
t_logits_per_text = t_logits_per_audio.t().detach().cpu()
labels = torch.arange(audio_features.shape[0]).long()
# Change the loss from two terms into four terms with 2x2 combined CE loss
total_loss = (
F.cross_entropy(a_logits_per_audio, labels)
+ F.cross_entropy(a_logits_per_text, labels)
+ F.cross_entropy(t_logits_per_audio, labels)
+ F.cross_entropy(t_logits_per_text, labels)
) / 4
metrics[f"cumulative_loss"] = total_loss.item()
metrics[f"num_samples"] = audio_features.shape[0]
logits = {
"audio_to_text": (a_logits_per_audio + t_logits_per_audio) / 2,
"text_to_audio": (a_logits_per_text + t_logits_per_text) / 2,
}
ground_truth = torch.arange(len(text_features)).view(-1, 1)
else:
# print("text_features", text_features)
# print("text_features.shape", text_features.shape)
logits_per_audio = (
(logit_scale_a * audio_features @ text_features.t()).detach().cpu()
)
logits_per_text = logits_per_audio.t().detach().cpu()
labels = torch.arange(audio_features.shape[0]).long()
# Change the loss from two terms into four terms with 2x2 combined CE loss
total_loss = (
F.cross_entropy(logits_per_audio, labels)
+ F.cross_entropy(logits_per_text, labels)
) / 2
metrics[f"cumulative_loss"] = total_loss.item()
metrics[f"num_samples"] = audio_features.shape[0]
logits = {"audio_to_text": logits_per_audio, "text_to_audio": logits_per_text}
ground_truth = torch.arange(len(text_features)).view(-1, 1)
for name, logit in logits.items():
ranking = torch.argsort(logit, descending=True)
preds = torch.where(ranking == ground_truth)[
1
] # (yusong) this line is slow because it uses single thread
preds = preds.detach().cpu().numpy()
metrics[f"{name}_mean_rank"] = preds.mean() + 1
metrics[f"{name}_median_rank"] = np.floor(np.median(preds)) + 1
for k in [1, 5, 10]:
metrics[f"{name}_R@{k}"] = np.mean(preds < k)
# map@10
metrics[f"{name}_mAP@10"] = np.mean(np.where(preds < 10, 1 / (preds + 1), 0.0))
return metrics
def evaluate_clotho_audiocaps(
model, data, epoch, args, autocast, device, tb_writer=None
):
"""
Adapted from https://github.com/XinhaoMei/audio-text_retrieval/blob/main/tools/utils.py.
1. for text-to-audio retrieval, do 5 times and average the results
2. for R@1, R@5, R@10 in audio-to-text retrieval, take the best rank among 5 text
3. for map@10 in audio-to-text retrieval:
3.1: sort the rank of 5 text
3.2: exclude the rank >=10 (0-index)
3.3: compute the map regarding the remaining ranks: np.mean(np.arange(1, len(ranks)+1) / ranks).
(3.3) That is, take the top ranks of 5 text that is < 10, and assign the descending number as ground truth.
(3.3) E.g.: the ground truth of first rank of the 5 text should be 1, the second rank should be 2, etc.
"""
# TODO: (yusong) only support single GPU evaluation and only support non-mlp case for now.
dataloader = data["val"].dataloader
with torch.no_grad():
eval_info = {}
for i, batch in enumerate(dataloader):
audios = batch # contains mel_spec, wavform, and longer list
# each item in the list has 5 texts
if args.tmodel == "transformer":
from open_clip import tokenize
texts = [tokenize(t) for t in batch["full_text"]]
texts = torch.cat(texts)
else:
from .data import tokenizer
texts = [
tokenizer(t) for t in batch["full_text"]
] # 5 texts for each audio
texts = {
k: torch.cat([t[k] for t in texts]) for k in texts[0].keys()
} # 5 x batch
# audios = audios.to(device=device, non_blocking=True)
all_names = list(
set(["-".join(b.split("/")[-3:-1]) for b in batch["__url__"]])
)
for name in all_names:
if name not in eval_info.keys():
# we will not use mlp outputs even if args.clap_mlploss=True
eval_info[name] = {
"cumulative_loss": 0.0,
"num_samples": 0,
"all_audio_features": [],
"all_text_features": [],
}
with autocast():
audio_features = model(audios, None, device)
text_features = model(None, texts, device)
audio_features = F.normalize(audio_features, dim=-1)
text_features = F.normalize(text_features, dim=-1)
all_names = list(
set(["-".join(b.split("/")[-3:-1]) for b in batch["__url__"]])
)
for n in all_names:
idx = np.where(
np.array(
["-".join(b.split("/")[-3:-1]) for b in batch["__url__"]]
)
== n
)[0]
eval_info[n]["all_audio_features"].append(
audio_features.cpu().index_select(0, torch.tensor(idx).long())
)
# (yusong) please double-check. This is for selecting 5 text features at once.
# because idx is a list of indices in size of num_samples,
# and text_features is a tensor of size (5*num_samples, dim)
# so we need to select 5 consecutive indices at once for a single index in idx.
eval_info[n]["all_text_features"].append(
text_features.cpu()
.reshape([-1, 5, text_features.shape[1]])
.index_select(0, torch.tensor(idx).long())
.reshape([-1, text_features.shape[1]])
)
val_metrics_all = {}
for n in eval_info.keys():
logit_scale_a, logit_scale_t = model(None, None, device)
logit_scale_a = logit_scale_a.cpu()
audio_features = torch.cat(eval_info[n]["all_audio_features"], dim=0)
text_features = torch.cat(eval_info[n]["all_text_features"], dim=0)
logits_per_audio = (
(logit_scale_a * audio_features @ text_features.t()).detach().cpu()
)
logits_per_text = logits_per_audio.t().detach().cpu()
# logits_per_audio shape: [num_samples, num_samples*5]
# logits_per_text shape: [num_samples*5, num_samples]
logging.info(
f"dataset {n}, logits_per_audio shape: {logits_per_audio.shape}, "
f"logits_per_text shape: {logits_per_text.shape}"
)
metrics = {}
num_samples = audio_features.shape[0]
metrics[f"num_samples"] = num_samples
# (yusong) the following code is very important, please double-check:
# logits_per_audio.reshape(num_samples, num_samples, 5)[:, :, d]
# logits_per_text.reshape(num_samples, 5, num_samples)[:, d, :]
# Those two are retrieving one of the 5 text for each audio.
labels = torch.arange(audio_features.shape[0]).long()
audio_to_text_loss = [
F.cross_entropy(
logits_per_audio.reshape(num_samples, num_samples, 5)[:, :, d],
labels,
)
for d in range(5)
]
text_to_audio_loss = [
F.cross_entropy(
logits_per_text.reshape(num_samples, 5, num_samples)[:, d, :],
labels,
)
for d in range(5)
]
total_loss = (np.mean(audio_to_text_loss) + np.mean(text_to_audio_loss)) / 2
metrics[f"cumulative_loss"] = total_loss.item()
# text to audio: do 5 times
pred_text = []
for d in range(5):
logit = logits_per_text.reshape(num_samples, 5, num_samples)[:, d, :]
ground_truth = torch.arange(len(logit)).view(-1, 1)
ranking = torch.argsort(
logit, descending=True
) # [num_samples, num_samples]
preds = torch.where(ranking == ground_truth)[1]
pred_text.append(preds.detach().cpu().numpy())
pred_text_concat = np.concatenate(pred_text, axis=0) # [5*num_samples]
metrics[f"text_to_audio_mean_rank"] = pred_text_concat.mean() + 1
metrics[f"text_to_audio_median_rank"] = (
np.floor(np.median(pred_text_concat)) + 1
)
for k in [1, 5, 10]:
metrics[f"text_to_audio_R@{k}"] = np.mean(pred_text_concat < k)
# map@10
metrics[f"text_to_audio_mAP@10"] = np.mean(
np.where(pred_text_concat < 10, 1 / (pred_text_concat + 1), 0.0)
)
# audio to text: take the best result
# for audio to text map 10, sort and assign descending ground truth.
# see https://github.com/XinhaoMei/audio-text_retrieval/blob/main/tools/utils.py#L103
# map@10
map_all = []
pred_audio_all = []
for d in range(num_samples):
# logits_per_audio: [num_samples, num_samples*5]
logit_single = logits_per_audio[d, :] # [5*num_samples]
# Ground-truth index: [d*5, d*5+1, d*5+2, d*5+3, d*5+4]
ranking = torch.argsort(
logit_single, descending=True
) # [5*num_samples]
# ranking: the index of first match, second match, ...
ground_truth = torch.arange(d * 5, d * 5 + 5)[None]
all_pred = torch.where(
torch.stack([ranking] * 5) == ground_truth.view(-1, 1)
)[1]
min_pred = torch.min(all_pred)
pred_audio_all.append(min_pred.detach().cpu().numpy())
all_pred_filter = all_pred[all_pred < 10].detach().cpu().numpy()
# /5 because we have 5 text, so it means for the text rank >=10 we count as 0.
map_single = (
np.sum(
(np.arange(1, len(all_pred_filter) + 1) / (all_pred_filter + 1))
)
/ 5
)
map_all.append(map_single)
metrics[f"audio_to_text_mAP@10"] = np.mean(map_all)
for k in [1, 5, 10]:
metrics[f"audio_to_text_R@{k}"] = np.mean(np.array(pred_audio_all) < k)
val_metrics_all[n] = {n + "/" + k: v for k, v in metrics.items()}
return val_metrics_all
def calculate_selection_performance_clotho_audiocaps(val_metrics_per_dataset):
"""
Calculate performance for Clotho+AudioCaps for model selection.
"""
selection_performance_all = []
for n in val_metrics_per_dataset.keys():
selection_performance = (
val_metrics_per_dataset[n][f"{n}/audio_to_text_mAP@10"]
+ val_metrics_per_dataset[n][f"{n}/text_to_audio_mAP@10"]
) / 2
selection_performance_all.append(selection_performance)
return np.mean(selection_performance_all)
def select_top_metric_clotho_audiocaps(metrics, val_metrics_per_dataset, args):
# val_metrics_per_dataset: dict, key: dataset name, value: dict, key: metric name, value: metric value
# metrics: dict, key: metric name, value: metric value
# Hack: use args to save the top performance
if not hasattr(args, "top_selection_performance"):
selection_performance = calculate_selection_performance_clotho_audiocaps(
val_metrics_per_dataset
)
# TODO: write the if and else together
metric_update = {}
for n in val_metrics_per_dataset.keys():
for k in val_metrics_per_dataset[n].keys():
metric_update[
k.split("/")[0] + "-top" + "/" + k.split("/")[1]
] = val_metrics_per_dataset[n][k]
metric_update["top_selection_performance"] = selection_performance
metric_update["top-selection-epoch"] = metrics["epoch"]
metrics.update(metric_update)
args.top_metric = metric_update
args.top_selection_performance = selection_performance
else:
selection_performance_new = calculate_selection_performance_clotho_audiocaps(
val_metrics_per_dataset
)
selection_performance_old = args.top_selection_performance
if selection_performance_new > selection_performance_old:
metric_update = {}
for n in val_metrics_per_dataset.keys():
for k in val_metrics_per_dataset[n].keys():
metric_update[
k.split("/")[0] + "-top" + "/" + k.split("/")[1]
] = val_metrics_per_dataset[n][k]
metric_update["top_selection_performance"] = selection_performance_new
metric_update["top-selection-epoch"] = metrics["epoch"]
metrics.update(metric_update)
args.top_metric = metric_update
args.top_selection_performance = selection_performance_new
else:
metrics.update(args.top_metric)
return metrics
|