train_tacotron2.py

# -*- 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.
"""Train Tacotron2."""
import tensorflow as tf

physical_devices = tf.config.list_physical_devices("GPU")
for i in range(len(physical_devices)):
    tf.config.experimental.set_memory_growth(physical_devices[i], True)

import sys

sys.path.append(".")

import argparse
import logging
import os

import numpy as np
import yaml
from tqdm import tqdm

import tensorflow_tts
from examples.tacotron2.tacotron_dataset import CharactorMelDataset
from tensorflow_tts.configs.tacotron2 import Tacotron2Config
from tensorflow_tts.models import TFTacotron2
from tensorflow_tts.optimizers import AdamWeightDecay, WarmUp
from tensorflow_tts.trainers import Seq2SeqBasedTrainer
from tensorflow_tts.utils import calculate_2d_loss, calculate_3d_loss, return_strategy


class Tacotron2Trainer(Seq2SeqBasedTrainer):
    """Tacotron2 Trainer class based on Seq2SeqBasedTrainer."""

    def __init__(
        self,
        config,
        strategy,
        steps=0,
        epochs=0,
        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.
            is_mixed_precision (bool): Use mixed precision or not.

        """
        super(Tacotron2Trainer, self).__init__(
            steps=steps,
            epochs=epochs,
            config=config,
            strategy=strategy,
            is_mixed_precision=is_mixed_precision,
        )
        # define metrics to aggregates data and use tf.summary logs them
        self.list_metrics_name = [
            "stop_token_loss",
            "mel_loss_before",
            "mel_loss_after",
            "guided_attention_loss",
        ]
        self.init_train_eval_metrics(self.list_metrics_name)
        self.reset_states_train()
        self.reset_states_eval()

        self.config = config

    def compile(self, model, optimizer):
        super().compile(model, optimizer)
        self.binary_crossentropy = tf.keras.losses.BinaryCrossentropy(
            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
        )
        self.mse = tf.keras.losses.MeanSquaredError(
            reduction=tf.keras.losses.Reduction.NONE
        )
        self.mae = tf.keras.losses.MeanAbsoluteError(
            reduction=tf.keras.losses.Reduction.NONE
        )

    def _train_step(self, batch):
        """Here we re-define _train_step because apply input_signature make
        the training progress slower on my experiment. Note that input_signature
        is apply on based_trainer by default.
        """
        if self._already_apply_input_signature is False:
            self.one_step_forward = tf.function(
                self._one_step_forward, experimental_relax_shapes=True
            )
            self.one_step_evaluate = tf.function(
                self._one_step_evaluate, experimental_relax_shapes=True
            )
            self.one_step_predict = tf.function(
                self._one_step_predict, experimental_relax_shapes=True
            )
            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_evaluate_per_replica(self, batch):
        """One step evaluate per GPU

        Tacotron-2 used teacher-forcing when training and evaluation.
        So we need pass `training=True` for inference step.

        """
        outputs = self._model(**batch, training=True)
        _, dict_metrics_losses = self.compute_per_example_losses(batch, outputs)

        self.update_eval_metrics(dict_metrics_losses)

    def _one_step_predict_per_replica(self, batch):
        """One step predict per GPU

        Tacotron-2 used teacher-forcing when training and evaluation.
        So we need pass `training=True` for inference step.

        """
        outputs = self._model(**batch, training=True)
        return outputs

    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.
        """
        (
            decoder_output,
            post_mel_outputs,
            stop_token_predictions,
            alignment_historys,
        ) = outputs

        mel_loss_before = calculate_3d_loss(
            batch["mel_gts"], decoder_output, loss_fn=self.mae
        )
        mel_loss_after = calculate_3d_loss(
            batch["mel_gts"], post_mel_outputs, loss_fn=self.mae
        )

        # calculate stop_loss
        max_mel_length = (
            tf.reduce_max(batch["mel_lengths"])
            if self.config["use_fixed_shapes"] is False
            else [self.config["max_mel_length"]]
        )
        stop_gts = tf.expand_dims(
            tf.range(tf.reduce_max(max_mel_length), dtype=tf.int32), 0
        )  # [1, max_len]
        stop_gts = tf.tile(
            stop_gts, [tf.shape(batch["mel_lengths"])[0], 1]
        )  # [B, max_len]
        stop_gts = tf.cast(
            tf.math.greater_equal(stop_gts, tf.expand_dims(batch["mel_lengths"], 1)),
            tf.float32,
        )

        stop_token_loss = calculate_2d_loss(
            stop_gts, stop_token_predictions, loss_fn=self.binary_crossentropy
        )

        # calculate guided attention loss.
        attention_masks = tf.cast(
            tf.math.not_equal(batch["g_attentions"], -1.0), tf.float32
        )
        loss_att = tf.reduce_sum(
            tf.abs(alignment_historys * batch["g_attentions"]) * attention_masks,
            axis=[1, 2],
        )
        loss_att /= tf.reduce_sum(attention_masks, axis=[1, 2])

        per_example_losses = (
            stop_token_loss + mel_loss_before + mel_loss_after + loss_att
        )

        dict_metrics_losses = {
            "stop_token_loss": stop_token_loss,
            "mel_loss_before": mel_loss_before,
            "mel_loss_after": mel_loss_after,
            "guided_attention_loss": loss_att,
        }

        return per_example_losses, dict_metrics_losses

    def generate_and_save_intermediate_result(self, batch):
        """Generate and save intermediate result."""
        import matplotlib.pyplot as plt

        # predict with tf.function for faster.
        outputs = self.one_step_predict(batch)
        (
            decoder_output,
            mel_outputs,
            stop_token_predictions,
            alignment_historys,
        ) = outputs
        mel_gts = batch["mel_gts"]
        utt_ids = batch["utt_ids"]

        # convert to tensor.
        # here we just take a sample at first replica.
        try:
            mels_before = decoder_output.values[0].numpy()
            mels_after = mel_outputs.values[0].numpy()
            mel_gts = mel_gts.values[0].numpy()
            alignment_historys = alignment_historys.values[0].numpy()
            utt_ids = utt_ids.values[0].numpy()
        except Exception:
            mels_before = decoder_output.numpy()
            mels_after = mel_outputs.numpy()
            mel_gts = mel_gts.numpy()
            alignment_historys = alignment_historys.numpy()
            utt_ids = utt_ids.numpy()

        # check directory
        dirname = os.path.join(self.config["outdir"], f"predictions/{self.steps}steps")
        if not os.path.exists(dirname):
            os.makedirs(dirname)

        for idx, (mel_gt, mel_before, mel_after, alignment_history) in enumerate(
            zip(mel_gts, mels_before, mels_after, alignment_historys), 0
        ):
            mel_gt = tf.reshape(mel_gt, (-1, 80)).numpy()  # [length, 80]
            mel_before = tf.reshape(mel_before, (-1, 80)).numpy()  # [length, 80]
            mel_after = tf.reshape(mel_after, (-1, 80)).numpy()  # [length, 80]

            # plot figure and save it
            utt_id = utt_ids[idx]
            figname = os.path.join(dirname, f"{utt_id}.png")
            fig = plt.figure(figsize=(10, 8))
            ax1 = fig.add_subplot(311)
            ax2 = fig.add_subplot(312)
            ax3 = fig.add_subplot(313)
            im = ax1.imshow(np.rot90(mel_gt), aspect="auto", interpolation="none")
            ax1.set_title("Target Mel-Spectrogram")
            fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax1)
            ax2.set_title(f"Predicted Mel-before-Spectrogram @ {self.steps} steps")
            im = ax2.imshow(np.rot90(mel_before), aspect="auto", interpolation="none")
            fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax2)
            ax3.set_title(f"Predicted Mel-after-Spectrogram @ {self.steps} steps")
            im = ax3.imshow(np.rot90(mel_after), aspect="auto", interpolation="none")
            fig.colorbar(mappable=im, shrink=0.65, orientation="horizontal", ax=ax3)
            plt.tight_layout()
            plt.savefig(figname)
            plt.close()

            # plot alignment
            figname = os.path.join(dirname, f"{idx}_alignment.png")
            fig = plt.figure(figsize=(8, 6))
            ax = fig.add_subplot(111)
            ax.set_title(f"Alignment @ {self.steps} steps")
            im = ax.imshow(
                alignment_history, aspect="auto", origin="lower", interpolation="none"
            )
            fig.colorbar(im, ax=ax)
            xlabel = "Decoder timestep"
            plt.xlabel(xlabel)
            plt.ylabel("Encoder timestep")
            plt.tight_layout()
            plt.savefig(figname)
            plt.close()


def main():
    """Run training process."""
    parser = argparse.ArgumentParser(
        description="Train FastSpeech (See detail in tensorflow_tts/bin/train-fastspeech.py)"
    )
    parser.add_argument(
        "--train-dir",
        default=None,
        type=str,
        help="directory including training data. ",
    )
    parser.add_argument(
        "--dev-dir",
        default=None,
        type=str,
        help="directory including development data. ",
    )
    parser.add_argument(
        "--use-norm", default=1, type=int, help="usr norm-mels for train or raw."
    )
    parser.add_argument(
        "--outdir", type=str, required=True, help="directory to save checkpoints."
    )
    parser.add_argument(
        "--config", type=str, required=True, help="yaml format configuration file."
    )
    parser.add_argument(
        "--resume",
        default="",
        type=str,
        nargs="?",
        help='checkpoint file path to resume training. (default="")',
    )
    parser.add_argument(
        "--verbose",
        type=int,
        default=1,
        help="logging level. higher is more logging. (default=1)",
    )
    parser.add_argument(
        "--mixed_precision",
        default=0,
        type=int,
        help="using mixed precision for generator or not.",
    )
    parser.add_argument(
        "--pretrained",
        default="",
        type=str,
        nargs="?",
        help="pretrained weights .h5 file to load weights from. Auto-skips non-matching layers",
    )
    parser.add_argument(
        "--use-fal",
        default=0,
        type=int,
        help="Use forced alignment guided attention loss or regular",
    )
    args = parser.parse_args()

    # return strategy
    STRATEGY = return_strategy()

    # set mixed precision config
    if args.mixed_precision == 1:
        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})

    args.mixed_precision = bool(args.mixed_precision)
    args.use_norm = bool(args.use_norm)
    args.use_fal = bool(args.use_fal)

    # set logger
    if args.verbose > 1:
        logging.basicConfig(
            level=logging.DEBUG,
            stream=sys.stdout,
            format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
        )
    elif args.verbose > 0:
        logging.basicConfig(
            level=logging.INFO,
            stream=sys.stdout,
            format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
        )
    else:
        logging.basicConfig(
            level=logging.WARN,
            stream=sys.stdout,
            format="%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
        )
        logging.warning("Skip DEBUG/INFO messages")

    # check directory existence
    if not os.path.exists(args.outdir):
        os.makedirs(args.outdir)

    # check arguments
    if args.train_dir is None:
        raise ValueError("Please specify --train-dir")
    if args.dev_dir is None:
        raise ValueError("Please specify --valid-dir")

    # load and save config
    with open(args.config) as f:
        config = yaml.load(f, Loader=yaml.Loader)
    config.update(vars(args))
    config["version"] = tensorflow_tts.__version__

    # get dataset
    if config["remove_short_samples"]:
        mel_length_threshold = config["mel_length_threshold"]
    else:
        mel_length_threshold = 0

    if config["format"] == "npy":
        charactor_query = "*-ids.npy"
        mel_query = "*-raw-feats.npy" if args.use_norm is False else "*-norm-feats.npy"
        align_query = "*-alignment.npy" if args.use_fal is True else ""
        charactor_load_fn = np.load
        mel_load_fn = np.load
    else:
        raise ValueError("Only npy are supported.")

    train_dataset = CharactorMelDataset(
        dataset=config["tacotron2_params"]["dataset"],
        root_dir=args.train_dir,
        charactor_query=charactor_query,
        mel_query=mel_query,
        charactor_load_fn=charactor_load_fn,
        mel_load_fn=mel_load_fn,
        mel_length_threshold=mel_length_threshold,
        reduction_factor=config["tacotron2_params"]["reduction_factor"],
        use_fixed_shapes=config["use_fixed_shapes"],
        align_query=align_query,
    )

    # update max_mel_length and max_char_length to config
    config.update({"max_mel_length": int(train_dataset.max_mel_length)})
    config.update({"max_char_length": int(train_dataset.max_char_length)})

    with open(os.path.join(args.outdir, "config.yml"), "w") as f:
        yaml.dump(config, f, Dumper=yaml.Dumper)
    for key, value in config.items():
        logging.info(f"{key} = {value}")

    train_dataset = train_dataset.create(
        is_shuffle=config["is_shuffle"],
        allow_cache=config["allow_cache"],
        batch_size=config["batch_size"]
        * STRATEGY.num_replicas_in_sync
        * config["gradient_accumulation_steps"],
    )

    valid_dataset = CharactorMelDataset(
        dataset=config["tacotron2_params"]["dataset"],
        root_dir=args.dev_dir,
        charactor_query=charactor_query,
        mel_query=mel_query,
        charactor_load_fn=charactor_load_fn,
        mel_load_fn=mel_load_fn,
        mel_length_threshold=mel_length_threshold,
        reduction_factor=config["tacotron2_params"]["reduction_factor"],
        use_fixed_shapes=False,  # don't need apply fixed shape for evaluation.
        align_query=align_query,
    ).create(
        is_shuffle=config["is_shuffle"],
        allow_cache=config["allow_cache"],
        batch_size=config["batch_size"] * STRATEGY.num_replicas_in_sync,
    )

    # define trainer
    trainer = Tacotron2Trainer(
        config=config,
        strategy=STRATEGY,
        steps=0,
        epochs=0,
        is_mixed_precision=args.mixed_precision,
    )

    with STRATEGY.scope():
        # define model.
        tacotron_config = Tacotron2Config(**config["tacotron2_params"])
        tacotron2 = TFTacotron2(config=tacotron_config, name="tacotron2")
        tacotron2._build()
        tacotron2.summary()

        if len(args.pretrained) > 1:
            tacotron2.load_weights(args.pretrained, by_name=True, skip_mismatch=True)
            logging.info(
                f"Successfully loaded pretrained weight from {args.pretrained}."
            )

        # AdamW for tacotron2
        learning_rate_fn = tf.keras.optimizers.schedules.PolynomialDecay(
            initial_learning_rate=config["optimizer_params"]["initial_learning_rate"],
            decay_steps=config["optimizer_params"]["decay_steps"],
            end_learning_rate=config["optimizer_params"]["end_learning_rate"],
        )

        learning_rate_fn = WarmUp(
            initial_learning_rate=config["optimizer_params"]["initial_learning_rate"],
            decay_schedule_fn=learning_rate_fn,
            warmup_steps=int(
                config["train_max_steps"]
                * config["optimizer_params"]["warmup_proportion"]
            ),
        )

        optimizer = AdamWeightDecay(
            learning_rate=learning_rate_fn,
            weight_decay_rate=config["optimizer_params"]["weight_decay"],
            beta_1=0.9,
            beta_2=0.98,
            epsilon=1e-6,
            exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"],
        )

        _ = optimizer.iterations

    # compile trainer
    trainer.compile(model=tacotron2, optimizer=optimizer)

    # start training
    try:
        trainer.fit(
            train_dataset,
            valid_dataset,
            saved_path=os.path.join(config["outdir"], "checkpoints/"),
            resume=args.resume,
        )
    except KeyboardInterrupt:
        trainer.save_checkpoint()
        logging.info(f"Successfully saved checkpoint @ {trainer.steps}steps.")


if __name__ == "__main__":
    main()