File size: 36,562 Bytes
d5ee97c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 |
# -*- coding: utf-8 -*-
# Copyright 2020 Minh Nguyen (@dathudeptrai)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Based Trainer."""
import abc
import logging
import os
import tensorflow as tf
from tqdm import tqdm
from tensorflow_tts.optimizers import GradientAccumulator
from tensorflow_tts.utils import utils
class BasedTrainer(metaclass=abc.ABCMeta):
"""Customized trainer module for all models."""
def __init__(self, steps, epochs, config):
self.steps = steps
self.epochs = epochs
self.config = config
self.finish_train = False
self.writer = tf.summary.create_file_writer(config["outdir"])
self.train_data_loader = None
self.eval_data_loader = None
self.train_metrics = None
self.eval_metrics = None
self.list_metrics_name = None
def init_train_eval_metrics(self, list_metrics_name):
"""Init train and eval metrics to save it to tensorboard."""
self.train_metrics = {}
self.eval_metrics = {}
for name in list_metrics_name:
self.train_metrics.update(
{name: tf.keras.metrics.Mean(name="train_" + name, dtype=tf.float32)}
)
self.eval_metrics.update(
{name: tf.keras.metrics.Mean(name="eval_" + name, dtype=tf.float32)}
)
def reset_states_train(self):
"""Reset train metrics after save it to tensorboard."""
for metric in self.train_metrics.keys():
self.train_metrics[metric].reset_states()
def reset_states_eval(self):
"""Reset eval metrics after save it to tensorboard."""
for metric in self.eval_metrics.keys():
self.eval_metrics[metric].reset_states()
def update_train_metrics(self, dict_metrics_losses):
for name, value in dict_metrics_losses.items():
self.train_metrics[name].update_state(value)
def update_eval_metrics(self, dict_metrics_losses):
for name, value in dict_metrics_losses.items():
self.eval_metrics[name].update_state(value)
def set_train_data_loader(self, train_dataset):
"""Set train data loader (MUST)."""
self.train_data_loader = train_dataset
def get_train_data_loader(self):
"""Get train data loader."""
return self.train_data_loader
def set_eval_data_loader(self, eval_dataset):
"""Set eval data loader (MUST)."""
self.eval_data_loader = eval_dataset
def get_eval_data_loader(self):
"""Get eval data loader."""
return self.eval_data_loader
@abc.abstractmethod
def compile(self):
pass
@abc.abstractmethod
def create_checkpoint_manager(self, saved_path=None, max_to_keep=10):
"""Create checkpoint management."""
pass
def run(self):
"""Run training."""
self.tqdm = tqdm(
initial=self.steps, total=self.config["train_max_steps"], desc="[train]"
)
while True:
self._train_epoch()
if self.finish_train:
break
self.tqdm.close()
logging.info("Finish training.")
@abc.abstractmethod
def save_checkpoint(self):
"""Save checkpoint."""
pass
@abc.abstractmethod
def load_checkpoint(self, pretrained_path):
"""Load checkpoint."""
pass
def _train_epoch(self):
"""Train model one epoch."""
for train_steps_per_epoch, batch in enumerate(self.train_data_loader, 1):
# one step training
self._train_step(batch)
# check interval
self._check_log_interval()
self._check_eval_interval()
self._check_save_interval()
# check wheter training is finished
if self.finish_train:
return
# update
self.epochs += 1
self.train_steps_per_epoch = train_steps_per_epoch
logging.info(
f"(Steps: {self.steps}) Finished {self.epochs} epoch training "
f"({self.train_steps_per_epoch} steps per epoch)."
)
@abc.abstractmethod
def _eval_epoch(self):
"""One epoch evaluation."""
pass
@abc.abstractmethod
def _train_step(self, batch):
"""One step training."""
pass
@abc.abstractmethod
def _check_log_interval(self):
"""Save log interval."""
pass
@abc.abstractmethod
def fit(self):
pass
def _check_eval_interval(self):
"""Evaluation interval step."""
if self.steps % self.config["eval_interval_steps"] == 0:
self._eval_epoch()
def _check_save_interval(self):
"""Save interval checkpoint."""
if self.steps % self.config["save_interval_steps"] == 0:
self.save_checkpoint()
logging.info(f"Successfully saved checkpoint @ {self.steps} steps.")
def generate_and_save_intermediate_result(self, batch):
"""Generate and save intermediate result."""
pass
def _write_to_tensorboard(self, list_metrics, stage="train"):
"""Write variables to tensorboard."""
with self.writer.as_default():
for key, value in list_metrics.items():
tf.summary.scalar(stage + "/" + key, value.result(), step=self.steps)
self.writer.flush()
class GanBasedTrainer(BasedTrainer):
"""Customized trainer module for GAN TTS training (MelGAN, GAN-TTS, ParallelWaveGAN)."""
def __init__(
self,
steps,
epochs,
config,
strategy,
is_generator_mixed_precision=False,
is_discriminator_mixed_precision=False,
):
"""Initialize trainer.
Args:
steps (int): Initial global steps.
epochs (int): Initial global epochs.
config (dict): Config dict loaded from yaml format configuration file.
"""
super().__init__(steps, epochs, config)
self._is_generator_mixed_precision = is_generator_mixed_precision
self._is_discriminator_mixed_precision = is_discriminator_mixed_precision
self._strategy = strategy
self._already_apply_input_signature = False
self._generator_gradient_accumulator = GradientAccumulator()
self._discriminator_gradient_accumulator = GradientAccumulator()
self._generator_gradient_accumulator.reset()
self._discriminator_gradient_accumulator.reset()
def init_train_eval_metrics(self, list_metrics_name):
with self._strategy.scope():
super().init_train_eval_metrics(list_metrics_name)
def get_n_gpus(self):
return self._strategy.num_replicas_in_sync
def _get_train_element_signature(self):
return self.train_data_loader.element_spec
def _get_eval_element_signature(self):
return self.eval_data_loader.element_spec
def set_gen_model(self, generator_model):
"""Set generator class model (MUST)."""
self._generator = generator_model
def get_gen_model(self):
"""Get generator model."""
return self._generator
def set_dis_model(self, discriminator_model):
"""Set discriminator class model (MUST)."""
self._discriminator = discriminator_model
def get_dis_model(self):
"""Get discriminator model."""
return self._discriminator
def set_gen_optimizer(self, generator_optimizer):
"""Set generator optimizer (MUST)."""
self._gen_optimizer = generator_optimizer
if self._is_generator_mixed_precision:
self._gen_optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(
self._gen_optimizer, "dynamic"
)
def get_gen_optimizer(self):
"""Get generator optimizer."""
return self._gen_optimizer
def set_dis_optimizer(self, discriminator_optimizer):
"""Set discriminator optimizer (MUST)."""
self._dis_optimizer = discriminator_optimizer
if self._is_discriminator_mixed_precision:
self._dis_optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(
self._dis_optimizer, "dynamic"
)
def get_dis_optimizer(self):
"""Get discriminator optimizer."""
return self._dis_optimizer
def compile(self, gen_model, dis_model, gen_optimizer, dis_optimizer):
self.set_gen_model(gen_model)
self.set_dis_model(dis_model)
self.set_gen_optimizer(gen_optimizer)
self.set_dis_optimizer(dis_optimizer)
def _train_step(self, batch):
if self._already_apply_input_signature is False:
train_element_signature = self._get_train_element_signature()
eval_element_signature = self._get_eval_element_signature()
self.one_step_forward = tf.function(
self._one_step_forward, input_signature=[train_element_signature]
)
self.one_step_evaluate = tf.function(
self._one_step_evaluate, input_signature=[eval_element_signature]
)
self.one_step_predict = tf.function(
self._one_step_predict, input_signature=[eval_element_signature]
)
self._already_apply_input_signature = True
# run one_step_forward
self.one_step_forward(batch)
# update counts
self.steps += 1
self.tqdm.update(1)
self._check_train_finish()
def _one_step_forward(self, batch):
per_replica_losses = self._strategy.run(
self._one_step_forward_per_replica, args=(batch,)
)
return self._strategy.reduce(
tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None
)
@abc.abstractmethod
def compute_per_example_generator_losses(self, batch, outputs):
"""Compute per example generator losses and return dict_metrics_losses
Note that all element of the loss MUST has a shape [batch_size] and
the keys of dict_metrics_losses MUST be in self.list_metrics_name.
Args:
batch: dictionary batch input return from dataloader
outputs: outputs of the model
Returns:
per_example_losses: per example losses for each GPU, shape [B]
dict_metrics_losses: dictionary loss.
"""
per_example_losses = 0.0
dict_metrics_losses = {}
return per_example_losses, dict_metrics_losses
@abc.abstractmethod
def compute_per_example_discriminator_losses(self, batch, gen_outputs):
"""Compute per example discriminator losses and return dict_metrics_losses
Note that all element of the loss MUST has a shape [batch_size] and
the keys of dict_metrics_losses MUST be in self.list_metrics_name.
Args:
batch: dictionary batch input return from dataloader
outputs: outputs of the model
Returns:
per_example_losses: per example losses for each GPU, shape [B]
dict_metrics_losses: dictionary loss.
"""
per_example_losses = 0.0
dict_metrics_losses = {}
return per_example_losses, dict_metrics_losses
def _calculate_generator_gradient_per_batch(self, batch):
outputs = self._generator(**batch, training=True)
(
per_example_losses,
dict_metrics_losses,
) = self.compute_per_example_generator_losses(batch, outputs)
per_replica_gen_losses = tf.nn.compute_average_loss(
per_example_losses,
global_batch_size=self.config["batch_size"]
* self.get_n_gpus()
* self.config["gradient_accumulation_steps"],
)
if self._is_generator_mixed_precision:
scaled_per_replica_gen_losses = self._gen_optimizer.get_scaled_loss(
per_replica_gen_losses
)
if self._is_generator_mixed_precision:
scaled_gradients = tf.gradients(
scaled_per_replica_gen_losses, self._generator.trainable_variables
)
gradients = self._gen_optimizer.get_unscaled_gradients(scaled_gradients)
else:
gradients = tf.gradients(
per_replica_gen_losses, self._generator.trainable_variables
)
# gradient accumulate for generator here
if self.config["gradient_accumulation_steps"] > 1:
self._generator_gradient_accumulator(gradients)
# accumulate loss into metrics
self.update_train_metrics(dict_metrics_losses)
if self.config["gradient_accumulation_steps"] == 1:
return gradients, per_replica_gen_losses
else:
return per_replica_gen_losses
def _calculate_discriminator_gradient_per_batch(self, batch):
(
per_example_losses,
dict_metrics_losses,
) = self.compute_per_example_discriminator_losses(
batch, self._generator(**batch, training=True)
)
per_replica_dis_losses = tf.nn.compute_average_loss(
per_example_losses,
global_batch_size=self.config["batch_size"]
* self.get_n_gpus()
* self.config["gradient_accumulation_steps"],
)
if self._is_discriminator_mixed_precision:
scaled_per_replica_dis_losses = self._dis_optimizer.get_scaled_loss(
per_replica_dis_losses
)
if self._is_discriminator_mixed_precision:
scaled_gradients = tf.gradients(
scaled_per_replica_dis_losses,
self._discriminator.trainable_variables,
)
gradients = self._dis_optimizer.get_unscaled_gradients(scaled_gradients)
else:
gradients = tf.gradients(
per_replica_dis_losses, self._discriminator.trainable_variables
)
# accumulate loss into metrics
self.update_train_metrics(dict_metrics_losses)
# gradient accumulate for discriminator here
if self.config["gradient_accumulation_steps"] > 1:
self._discriminator_gradient_accumulator(gradients)
if self.config["gradient_accumulation_steps"] == 1:
return gradients, per_replica_dis_losses
else:
return per_replica_dis_losses
def _one_step_forward_per_replica(self, batch):
per_replica_gen_losses = 0.0
per_replica_dis_losses = 0.0
if self.config["gradient_accumulation_steps"] == 1:
(
gradients,
per_replica_gen_losses,
) = self._calculate_generator_gradient_per_batch(batch)
self._gen_optimizer.apply_gradients(
zip(gradients, self._generator.trainable_variables)
)
else:
# gradient acummulation here.
for i in tf.range(self.config["gradient_accumulation_steps"]):
reduced_batch = {
k: v[
i
* self.config["batch_size"] : (i + 1)
* self.config["batch_size"]
]
for k, v in batch.items()
}
# run 1 step accumulate
reduced_batch_losses = self._calculate_generator_gradient_per_batch(
reduced_batch
)
# sum per_replica_losses
per_replica_gen_losses += reduced_batch_losses
gradients = self._generator_gradient_accumulator.gradients
self._gen_optimizer.apply_gradients(
zip(gradients, self._generator.trainable_variables)
)
self._generator_gradient_accumulator.reset()
# one step discriminator
# recompute y_hat after 1 step generator for discriminator training.
if self.steps >= self.config["discriminator_train_start_steps"]:
if self.config["gradient_accumulation_steps"] == 1:
(
gradients,
per_replica_dis_losses,
) = self._calculate_discriminator_gradient_per_batch(batch)
self._dis_optimizer.apply_gradients(
zip(gradients, self._discriminator.trainable_variables)
)
else:
# gradient acummulation here.
for i in tf.range(self.config["gradient_accumulation_steps"]):
reduced_batch = {
k: v[
i
* self.config["batch_size"] : (i + 1)
* self.config["batch_size"]
]
for k, v in batch.items()
}
# run 1 step accumulate
reduced_batch_losses = (
self._calculate_discriminator_gradient_per_batch(reduced_batch)
)
# sum per_replica_losses
per_replica_dis_losses += reduced_batch_losses
gradients = self._discriminator_gradient_accumulator.gradients
self._dis_optimizer.apply_gradients(
zip(gradients, self._discriminator.trainable_variables)
)
self._discriminator_gradient_accumulator.reset()
return per_replica_gen_losses + per_replica_dis_losses
def _eval_epoch(self):
"""Evaluate model one epoch."""
logging.info(f"(Steps: {self.steps}) Start evaluation.")
# calculate loss for each batch
for eval_steps_per_epoch, batch in enumerate(
tqdm(self.eval_data_loader, desc="[eval]"), 1
):
# eval one step
self.one_step_evaluate(batch)
if eval_steps_per_epoch <= self.config["num_save_intermediate_results"]:
# save intermedia
self.generate_and_save_intermediate_result(batch)
logging.info(
f"(Steps: {self.steps}) Finished evaluation "
f"({eval_steps_per_epoch} steps per epoch)."
)
# average loss
for key in self.eval_metrics.keys():
logging.info(
f"(Steps: {self.steps}) eval_{key} = {self.eval_metrics[key].result():.4f}."
)
# record
self._write_to_tensorboard(self.eval_metrics, stage="eval")
# reset
self.reset_states_eval()
def _one_step_evaluate_per_replica(self, batch):
################################################
# one step generator.
outputs = self._generator(**batch, training=False)
_, dict_metrics_losses = self.compute_per_example_generator_losses(
batch, outputs
)
# accumulate loss into metrics
self.update_eval_metrics(dict_metrics_losses)
################################################
# one step discriminator
if self.steps >= self.config["discriminator_train_start_steps"]:
_, dict_metrics_losses = self.compute_per_example_discriminator_losses(
batch, outputs
)
# accumulate loss into metrics
self.update_eval_metrics(dict_metrics_losses)
################################################
def _one_step_evaluate(self, batch):
self._strategy.run(self._one_step_evaluate_per_replica, args=(batch,))
def _one_step_predict_per_replica(self, batch):
outputs = self._generator(**batch, training=False)
return outputs
def _one_step_predict(self, batch):
outputs = self._strategy.run(self._one_step_predict_per_replica, args=(batch,))
return outputs
@abc.abstractmethod
def generate_and_save_intermediate_result(self, batch):
return
def create_checkpoint_manager(self, saved_path=None, max_to_keep=10):
"""Create checkpoint management."""
if saved_path is None:
saved_path = self.config["outdir"] + "/checkpoints/"
os.makedirs(saved_path, exist_ok=True)
self.saved_path = saved_path
self.ckpt = tf.train.Checkpoint(
steps=tf.Variable(1),
epochs=tf.Variable(1),
gen_optimizer=self.get_gen_optimizer(),
dis_optimizer=self.get_dis_optimizer(),
)
self.ckp_manager = tf.train.CheckpointManager(
self.ckpt, saved_path, max_to_keep=max_to_keep
)
def save_checkpoint(self):
"""Save checkpoint."""
self.ckpt.steps.assign(self.steps)
self.ckpt.epochs.assign(self.epochs)
self.ckp_manager.save(checkpoint_number=self.steps)
utils.save_weights(
self._generator,
self.saved_path + "generator-{}.h5".format(self.steps)
)
utils.save_weights(
self._discriminator,
self.saved_path + "discriminator-{}.h5".format(self.steps)
)
def load_checkpoint(self, pretrained_path):
"""Load checkpoint."""
self.ckpt.restore(pretrained_path)
self.steps = self.ckpt.steps.numpy()
self.epochs = self.ckpt.epochs.numpy()
self._gen_optimizer = self.ckpt.gen_optimizer
# re-assign iterations (global steps) for gen_optimizer.
self._gen_optimizer.iterations.assign(tf.cast(self.steps, tf.int64))
# re-assign iterations (global steps) for dis_optimizer.
try:
discriminator_train_start_steps = self.config[
"discriminator_train_start_steps"
]
discriminator_train_start_steps = tf.math.maximum(
0, self.steps - discriminator_train_start_steps
)
except Exception:
discriminator_train_start_steps = self.steps
self._dis_optimizer = self.ckpt.dis_optimizer
self._dis_optimizer.iterations.assign(
tf.cast(discriminator_train_start_steps, tf.int64)
)
# load weights.
utils.load_weights(
self._generator,
self.saved_path + "generator-{}.h5".format(self.steps)
)
utils.load_weights(
self._discriminator,
self.saved_path + "discriminator-{}.h5".format(self.steps)
)
def _check_train_finish(self):
"""Check training finished."""
if self.steps >= self.config["train_max_steps"]:
self.finish_train = True
if (
self.steps != 0
and self.steps == self.config["discriminator_train_start_steps"]
):
self.finish_train = True
logging.info(
f"Finished training only generator at {self.steps}steps, pls resume and continue training."
)
def _check_log_interval(self):
"""Log to tensorboard."""
if self.steps % self.config["log_interval_steps"] == 0:
for metric_name in self.list_metrics_name:
logging.info(
f"(Step: {self.steps}) train_{metric_name} = {self.train_metrics[metric_name].result():.4f}."
)
self._write_to_tensorboard(self.train_metrics, stage="train")
# reset
self.reset_states_train()
def fit(self, train_data_loader, valid_data_loader, saved_path, resume=None):
self.set_train_data_loader(train_data_loader)
self.set_eval_data_loader(valid_data_loader)
self.train_data_loader = self._strategy.experimental_distribute_dataset(
self.train_data_loader
)
self.eval_data_loader = self._strategy.experimental_distribute_dataset(
self.eval_data_loader
)
with self._strategy.scope():
self.create_checkpoint_manager(saved_path=saved_path, max_to_keep=10000)
if len(resume) > 1:
self.load_checkpoint(resume)
logging.info(f"Successfully resumed from {resume}.")
self.run()
class Seq2SeqBasedTrainer(BasedTrainer, metaclass=abc.ABCMeta):
"""Customized trainer module for Seq2Seq TTS training (Tacotron, FastSpeech)."""
def __init__(
self, steps, epochs, config, strategy, is_mixed_precision=False,
):
"""Initialize trainer.
Args:
steps (int): Initial global steps.
epochs (int): Initial global epochs.
config (dict): Config dict loaded from yaml format configuration file.
strategy (tf.distribute): Strategy for distributed training.
is_mixed_precision (bool): Use mixed_precision training or not.
"""
super().__init__(steps, epochs, config)
self._is_mixed_precision = is_mixed_precision
self._strategy = strategy
self._model = None
self._optimizer = None
self._trainable_variables = None
# check if we already apply input_signature for train_step.
self._already_apply_input_signature = False
# create gradient accumulator
self._gradient_accumulator = GradientAccumulator()
self._gradient_accumulator.reset()
def init_train_eval_metrics(self, list_metrics_name):
with self._strategy.scope():
super().init_train_eval_metrics(list_metrics_name)
def set_model(self, model):
"""Set generator class model (MUST)."""
self._model = model
def get_model(self):
"""Get generator model."""
return self._model
def set_optimizer(self, optimizer):
"""Set optimizer (MUST)."""
self._optimizer = optimizer
if self._is_mixed_precision:
self._optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(
self._optimizer, "dynamic"
)
def get_optimizer(self):
"""Get optimizer."""
return self._optimizer
def get_n_gpus(self):
return self._strategy.num_replicas_in_sync
def compile(self, model, optimizer):
self.set_model(model)
self.set_optimizer(optimizer)
self._trainable_variables = self._train_vars()
def _train_vars(self):
if self.config["var_train_expr"]:
list_train_var = self.config["var_train_expr"].split("|")
return [
v
for v in self._model.trainable_variables
if self._check_string_exist(list_train_var, v.name)
]
return self._model.trainable_variables
def _check_string_exist(self, list_string, inp_string):
for string in list_string:
if string in inp_string:
return True
return False
def _get_train_element_signature(self):
return self.train_data_loader.element_spec
def _get_eval_element_signature(self):
return self.eval_data_loader.element_spec
def _train_step(self, batch):
if self._already_apply_input_signature is False:
train_element_signature = self._get_train_element_signature()
eval_element_signature = self._get_eval_element_signature()
self.one_step_forward = tf.function(
self._one_step_forward, input_signature=[train_element_signature]
)
self.one_step_evaluate = tf.function(
self._one_step_evaluate, input_signature=[eval_element_signature]
)
self.one_step_predict = tf.function(
self._one_step_predict, input_signature=[eval_element_signature]
)
self._already_apply_input_signature = True
# run one_step_forward
self.one_step_forward(batch)
# update counts
self.steps += 1
self.tqdm.update(1)
self._check_train_finish()
def _one_step_forward(self, batch):
per_replica_losses = self._strategy.run(
self._one_step_forward_per_replica, args=(batch,)
)
return self._strategy.reduce(
tf.distribute.ReduceOp.SUM, per_replica_losses, axis=None
)
def _calculate_gradient_per_batch(self, batch):
outputs = self._model(**batch, training=True)
per_example_losses, dict_metrics_losses = self.compute_per_example_losses(
batch, outputs
)
per_replica_losses = tf.nn.compute_average_loss(
per_example_losses,
global_batch_size=self.config["batch_size"]
* self.get_n_gpus()
* self.config["gradient_accumulation_steps"],
)
if self._is_mixed_precision:
scaled_per_replica_losses = self._optimizer.get_scaled_loss(
per_replica_losses
)
if self._is_mixed_precision:
scaled_gradients = tf.gradients(
scaled_per_replica_losses, self._trainable_variables
)
gradients = self._optimizer.get_unscaled_gradients(scaled_gradients)
else:
gradients = tf.gradients(per_replica_losses, self._trainable_variables)
# gradient accumulate here
if self.config["gradient_accumulation_steps"] > 1:
self._gradient_accumulator(gradients)
# accumulate loss into metrics
self.update_train_metrics(dict_metrics_losses)
if self.config["gradient_accumulation_steps"] == 1:
return gradients, per_replica_losses
else:
return per_replica_losses
def _one_step_forward_per_replica(self, batch):
if self.config["gradient_accumulation_steps"] == 1:
gradients, per_replica_losses = self._calculate_gradient_per_batch(batch)
self._optimizer.apply_gradients(
zip(gradients, self._trainable_variables), 1.0
)
else:
# gradient acummulation here.
per_replica_losses = 0.0
for i in tf.range(self.config["gradient_accumulation_steps"]):
reduced_batch = {
k: v[
i
* self.config["batch_size"] : (i + 1)
* self.config["batch_size"]
]
for k, v in batch.items()
}
# run 1 step accumulate
reduced_batch_losses = self._calculate_gradient_per_batch(reduced_batch)
# sum per_replica_losses
per_replica_losses += reduced_batch_losses
gradients = self._gradient_accumulator.gradients
self._optimizer.apply_gradients(
zip(gradients, self._trainable_variables), 1.0
)
self._gradient_accumulator.reset()
return per_replica_losses
@abc.abstractmethod
def compute_per_example_losses(self, batch, outputs):
"""Compute per example losses and return dict_metrics_losses
Note that all element of the loss MUST has a shape [batch_size] and
the keys of dict_metrics_losses MUST be in self.list_metrics_name.
Args:
batch: dictionary batch input return from dataloader
outputs: outputs of the model
Returns:
per_example_losses: per example losses for each GPU, shape [B]
dict_metrics_losses: dictionary loss.
"""
per_example_losses = 0.0
dict_metrics_losses = {}
return per_example_losses, dict_metrics_losses
def _eval_epoch(self):
"""Evaluate model one epoch."""
logging.info(f"(Steps: {self.steps}) Start evaluation.")
# calculate loss for each batch
for eval_steps_per_epoch, batch in enumerate(
tqdm(self.eval_data_loader, desc="[eval]"), 1
):
# eval one step
self.one_step_evaluate(batch)
if eval_steps_per_epoch <= self.config["num_save_intermediate_results"]:
# save intermedia
self.generate_and_save_intermediate_result(batch)
logging.info(
f"(Steps: {self.steps}) Finished evaluation "
f"({eval_steps_per_epoch} steps per epoch)."
)
# average loss
for key in self.eval_metrics.keys():
logging.info(
f"(Steps: {self.steps}) eval_{key} = {self.eval_metrics[key].result():.4f}."
)
# record
self._write_to_tensorboard(self.eval_metrics, stage="eval")
# reset
self.reset_states_eval()
def _one_step_evaluate_per_replica(self, batch):
outputs = self._model(**batch, training=False)
_, dict_metrics_losses = self.compute_per_example_losses(batch, outputs)
self.update_eval_metrics(dict_metrics_losses)
def _one_step_evaluate(self, batch):
self._strategy.run(self._one_step_evaluate_per_replica, args=(batch,))
def _one_step_predict_per_replica(self, batch):
outputs = self._model(**batch, training=False)
return outputs
def _one_step_predict(self, batch):
outputs = self._strategy.run(self._one_step_predict_per_replica, args=(batch,))
return outputs
@abc.abstractmethod
def generate_and_save_intermediate_result(self, batch):
return
def create_checkpoint_manager(self, saved_path=None, max_to_keep=10):
"""Create checkpoint management."""
if saved_path is None:
saved_path = self.config["outdir"] + "/checkpoints/"
os.makedirs(saved_path, exist_ok=True)
self.saved_path = saved_path
self.ckpt = tf.train.Checkpoint(
steps=tf.Variable(1), epochs=tf.Variable(1), optimizer=self.get_optimizer()
)
self.ckp_manager = tf.train.CheckpointManager(
self.ckpt, saved_path, max_to_keep=max_to_keep
)
def save_checkpoint(self):
"""Save checkpoint."""
self.ckpt.steps.assign(self.steps)
self.ckpt.epochs.assign(self.epochs)
self.ckp_manager.save(checkpoint_number=self.steps)
utils.save_weights(
self._model,
self.saved_path + "model-{}.h5".format(self.steps)
)
def load_checkpoint(self, pretrained_path):
"""Load checkpoint."""
self.ckpt.restore(pretrained_path)
self.steps = self.ckpt.steps.numpy()
self.epochs = self.ckpt.epochs.numpy()
self._optimizer = self.ckpt.optimizer
# re-assign iterations (global steps) for optimizer.
self._optimizer.iterations.assign(tf.cast(self.steps, tf.int64))
# load weights.
utils.load_weights(
self._model,
self.saved_path + "model-{}.h5".format(self.steps)
)
def _check_train_finish(self):
"""Check training finished."""
if self.steps >= self.config["train_max_steps"]:
self.finish_train = True
def _check_log_interval(self):
"""Log to tensorboard."""
if self.steps % self.config["log_interval_steps"] == 0:
for metric_name in self.list_metrics_name:
logging.info(
f"(Step: {self.steps}) train_{metric_name} = {self.train_metrics[metric_name].result():.4f}."
)
self._write_to_tensorboard(self.train_metrics, stage="train")
# reset
self.reset_states_train()
def fit(self, train_data_loader, valid_data_loader, saved_path, resume=None):
self.set_train_data_loader(train_data_loader)
self.set_eval_data_loader(valid_data_loader)
self.train_data_loader = self._strategy.experimental_distribute_dataset(
self.train_data_loader
)
self.eval_data_loader = self._strategy.experimental_distribute_dataset(
self.eval_data_loader
)
with self._strategy.scope():
self.create_checkpoint_manager(saved_path=saved_path, max_to_keep=10000)
if len(resume) > 1:
self.load_checkpoint(resume)
logging.info(f"Successfully resumed from {resume}.")
self.run()
|