syn_utils.py

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
import math
import os
import re
from pathlib import Path
from typing import Any
from typing import Dict
from typing import List
from typing import Optional

import numpy as np
import onnxruntime as ort
import paddle
from paddle import inference
from paddle import jit
from paddle.io import DataLoader
from paddle.static import InputSpec
from yacs.config import CfgNode

from paddlespeech.t2s.datasets.am_batch_fn import *
from paddlespeech.t2s.datasets.data_table import DataTable
from paddlespeech.t2s.datasets.vocoder_batch_fn import Clip_static
from paddlespeech.t2s.frontend import English
from paddlespeech.t2s.frontend.canton_frontend import CantonFrontend
from paddlespeech.t2s.frontend.mix_frontend import MixFrontend
from paddlespeech.t2s.frontend.zh_frontend import Frontend
from paddlespeech.t2s.modules.normalizer import ZScore
from paddlespeech.utils.dynamic_import import dynamic_import

# remove [W:onnxruntime: xxx] from ort
ort.set_default_logger_severity(3)

model_alias = {
    # acoustic model
    "speedyspeech":
    "paddlespeech.t2s.models.speedyspeech:SpeedySpeech",
    "speedyspeech_inference":
    "paddlespeech.t2s.models.speedyspeech:SpeedySpeechInference",
    "fastspeech2":
    "paddlespeech.t2s.models.fastspeech2:FastSpeech2",
    "fastspeech2_inference":
    "paddlespeech.t2s.models.fastspeech2:FastSpeech2Inference",
    "tacotron2":
    "paddlespeech.t2s.models.tacotron2:Tacotron2",
    "tacotron2_inference":
    "paddlespeech.t2s.models.tacotron2:Tacotron2Inference",
    # voc
    "pwgan":
    "paddlespeech.t2s.models.parallel_wavegan:PWGGenerator",
    "pwgan_inference":
    "paddlespeech.t2s.models.parallel_wavegan:PWGInference",
    "mb_melgan":
    "paddlespeech.t2s.models.melgan:MelGANGenerator",
    "mb_melgan_inference":
    "paddlespeech.t2s.models.melgan:MelGANInference",
    "style_melgan":
    "paddlespeech.t2s.models.melgan:StyleMelGANGenerator",
    "style_melgan_inference":
    "paddlespeech.t2s.models.melgan:StyleMelGANInference",
    "hifigan":
    "paddlespeech.t2s.models.hifigan:HiFiGANGenerator",
    "hifigan_inference":
    "paddlespeech.t2s.models.hifigan:HiFiGANInference",
    "wavernn":
    "paddlespeech.t2s.models.wavernn:WaveRNN",
    "wavernn_inference":
    "paddlespeech.t2s.models.wavernn:WaveRNNInference",
    "erniesat":
    "paddlespeech.t2s.models.ernie_sat:ErnieSAT",
    "erniesat_inference":
    "paddlespeech.t2s.models.ernie_sat:ErnieSATInference",
}


def denorm(data, mean, std):
    return data * std + mean


def norm(data, mean, std):
    return (data - mean) / std


def get_chunks(data, block_size: int, pad_size: int):
    data_len = data.shape[1]
    chunks = []
    n = math.ceil(data_len / block_size)
    for i in range(n):
        start = max(0, i * block_size - pad_size)
        end = min((i + 1) * block_size + pad_size, data_len)
        chunks.append(data[:, start:end, :])
    return chunks


# input
def get_sentences(text_file: Optional[os.PathLike], lang: str='zh'):
    # construct dataset for evaluation
    sentences = []
    with open(text_file, 'rt', encoding='utf-8') as f:
        for line in f:
            if line.strip() != "":
                items = re.split(r"\s+", line.strip(), 1)
                utt_id = items[0]
                if lang in {'zh', 'canton'}:
                    sentence = "".join(items[1:])
                elif lang == 'en':
                    sentence = " ".join(items[1:])
                elif lang == 'mix':
                    sentence = " ".join(items[1:])
            sentences.append((utt_id, sentence))
    return sentences


# am only
def get_test_dataset(test_metadata: List[Dict[str, Any]],
                     am: str,
                     speaker_dict: Optional[os.PathLike]=None,
                     voice_cloning: bool=False):
    # model: {model_name}_{dataset}
    am_name = am[:am.rindex('_')]
    am_dataset = am[am.rindex('_') + 1:]
    converters = {}
    if am_name == 'fastspeech2':
        fields = ["utt_id", "text"]
        if am_dataset in {"aishell3", "vctk", "mix",
                          "canton"} and speaker_dict is not None:
            print("multiple speaker fastspeech2!")
            fields += ["spk_id"]
        elif voice_cloning:
            print("voice cloning!")
            fields += ["spk_emb"]
        else:
            print("single speaker fastspeech2!")
    elif am_name == 'speedyspeech':
        fields = ["utt_id", "phones", "tones"]
    elif am_name == 'tacotron2':
        fields = ["utt_id", "text"]
        if voice_cloning:
            print("voice cloning!")
            fields += ["spk_emb"]
    elif am_name == 'erniesat':
        fields = [
            "utt_id", "text", "text_lengths", "speech", "speech_lengths",
            "align_start", "align_end"
        ]
        converters = {"speech": np.load}
    else:
        print("wrong am, please input right am!!!")

    test_dataset = DataTable(
        data=test_metadata, fields=fields, converters=converters)
    return test_dataset


# am and voc, for PTQ_static
def get_dev_dataloader(dev_metadata: List[Dict[str, Any]],
                       am: str,
                       batch_size: int=1,
                       speaker_dict: Optional[os.PathLike]=None,
                       voice_cloning: bool=False,
                       n_shift: int=300,
                       batch_max_steps: int=16200,
                       shuffle: bool=True):
    # model: {model_name}_{dataset}
    am_name = am[:am.rindex('_')]
    am_dataset = am[am.rindex('_') + 1:]
    converters = {}
    if am_name == 'fastspeech2':
        fields = ["utt_id", "text"]
        if am_dataset in {"aishell3", "vctk", "mix",
                          "canton"} and speaker_dict is not None:
            print("multiple speaker fastspeech2!")
            collate_fn = fastspeech2_multi_spk_batch_fn_static
            fields += ["spk_id"]
        elif voice_cloning:
            print("voice cloning!")
            collate_fn = fastspeech2_multi_spk_batch_fn_static
            fields += ["spk_emb"]
        else:
            print("single speaker fastspeech2!")
            collate_fn = fastspeech2_single_spk_batch_fn_static
    elif am_name == 'speedyspeech':
        fields = ["utt_id", "phones", "tones"]
        if am_dataset in {"aishell3", "vctk",
                          "mix"} and speaker_dict is not None:
            print("multiple speaker speedyspeech!")
            collate_fn = speedyspeech_multi_spk_batch_fn_static
            fields += ["spk_id"]
        else:
            print("single speaker speedyspeech!")
            collate_fn = speedyspeech_single_spk_batch_fn_static
        fields = ["utt_id", "phones", "tones"]
    elif am_name == 'tacotron2':
        fields = ["utt_id", "text"]
        if voice_cloning:
            print("voice cloning!")
            collate_fn = tacotron2_multi_spk_batch_fn_static
            fields += ["spk_emb"]
        else:
            print("single speaker tacotron2!")
            collate_fn = tacotron2_single_spk_batch_fn_static
    else:
        print("voc dataloader")

    # am
    if am_name not in {'pwgan', 'mb_melgan', 'hifigan'}:
        dev_dataset = DataTable(
            data=dev_metadata,
            fields=fields,
            converters=converters, )

        dev_dataloader = DataLoader(
            dev_dataset,
            shuffle=shuffle,
            drop_last=False,
            batch_size=batch_size,
            collate_fn=collate_fn)
    # vocoder
    else:
        # pwgan: batch_max_steps: 25500 aux_context_window: 2
        # mb_melgan: batch_max_steps: 16200 aux_context_window 0
        # hifigan: batch_max_steps: 8400 aux_context_window 0
        aux_context_window = 0
        if am_name == 'pwgan':
            aux_context_window = 2

        train_batch_fn = Clip_static(
            batch_max_steps=batch_max_steps,
            hop_size=n_shift,
            aux_context_window=aux_context_window)
        dev_dataset = DataTable(
            data=dev_metadata,
            fields=["wave", "feats"],
            converters={
                "wave": np.load,
                "feats": np.load,
            }, )

        dev_dataloader = DataLoader(
            dev_dataset,
            shuffle=shuffle,
            drop_last=False,
            batch_size=batch_size,
            collate_fn=train_batch_fn)

    return dev_dataloader


# frontend
def get_frontend(lang: str='zh',
                 phones_dict: Optional[os.PathLike]=None,
                 tones_dict: Optional[os.PathLike]=None,
                 use_rhy=False):
    if lang == 'zh':
        frontend = Frontend(
            phone_vocab_path=phones_dict,
            tone_vocab_path=tones_dict,
            use_rhy=use_rhy)
    elif lang == 'canton':
        frontend = CantonFrontend(phone_vocab_path=phones_dict)
    elif lang == 'en':
        frontend = English(phone_vocab_path=phones_dict)
    elif lang == 'mix':
        frontend = MixFrontend(
            phone_vocab_path=phones_dict, tone_vocab_path=tones_dict)
    else:
        print("wrong lang!")
    return frontend


def run_frontend(frontend: object,
                 text: str,
                 merge_sentences: bool=False,
                 get_tone_ids: bool=False,
                 lang: str='zh',
                 to_tensor: bool=True):
    outs = dict()
    if lang == 'zh':
        input_ids = {}
        if text.strip() != "" and re.match(r".*?<speak>.*?</speak>.*", text,
                                           re.DOTALL):
            input_ids = frontend.get_input_ids_ssml(
                text,
                merge_sentences=merge_sentences,
                get_tone_ids=get_tone_ids,
                to_tensor=to_tensor)
        else:
            input_ids = frontend.get_input_ids(
                text,
                merge_sentences=merge_sentences,
                get_tone_ids=get_tone_ids,
                to_tensor=to_tensor)
        phone_ids = input_ids["phone_ids"]
        if get_tone_ids:
            tone_ids = input_ids["tone_ids"]
            outs.update({'tone_ids': tone_ids})
    elif lang == 'canton':
        input_ids = frontend.get_input_ids(
            text, merge_sentences=merge_sentences, to_tensor=to_tensor)
        phone_ids = input_ids["phone_ids"]
    elif lang == 'en':
        input_ids = frontend.get_input_ids(
            text, merge_sentences=merge_sentences, to_tensor=to_tensor)
        phone_ids = input_ids["phone_ids"]
    elif lang == 'mix':
        input_ids = frontend.get_input_ids(
            text, merge_sentences=merge_sentences, to_tensor=to_tensor)
        phone_ids = input_ids["phone_ids"]
    else:
        print("lang should in {'zh', 'en', 'mix', 'canton'}!")
    outs.update({'phone_ids': phone_ids})
    return outs


# dygraph
def get_am_inference(am: str='fastspeech2_csmsc',
                     am_config: CfgNode=None,
                     am_ckpt: Optional[os.PathLike]=None,
                     am_stat: Optional[os.PathLike]=None,
                     phones_dict: Optional[os.PathLike]=None,
                     tones_dict: Optional[os.PathLike]=None,
                     speaker_dict: Optional[os.PathLike]=None,
                     return_am: bool=False):
    with open(phones_dict, 'rt', encoding='utf-8') as f:
        phn_id = [line.strip().split() for line in f.readlines()]
    vocab_size = len(phn_id)
    tone_size = None
    if tones_dict is not None:
        with open(tones_dict, 'rt', encoding='utf-8') as f:
            tone_id = [line.strip().split() for line in f.readlines()]
        tone_size = len(tone_id)
    spk_num = None
    if speaker_dict is not None:
        with open(speaker_dict, 'rt', encoding='utf-8') as f:
            spk_id = [line.strip().split() for line in f.readlines()]
        spk_num = len(spk_id)
    odim = am_config.n_mels
    # model: {model_name}_{dataset}
    am_name = am[:am.rindex('_')]
    am_dataset = am[am.rindex('_') + 1:]
    am_class = dynamic_import(am_name, model_alias)
    am_inference_class = dynamic_import(am_name + '_inference', model_alias)
    if am_name == 'fastspeech2':
        am = am_class(
            idim=vocab_size, odim=odim, spk_num=spk_num, **am_config["model"])
    elif am_name == 'speedyspeech':
        am = am_class(
            vocab_size=vocab_size,
            tone_size=tone_size,
            spk_num=spk_num,
            **am_config["model"])
    elif am_name == 'tacotron2':
        am = am_class(idim=vocab_size, odim=odim, **am_config["model"])
    elif am_name == 'erniesat':
        am = am_class(idim=vocab_size, odim=odim, **am_config["model"])
    else:
        print("wrong am, please input right am!!!")

    am.set_state_dict(paddle.load(am_ckpt)["main_params"])
    am.eval()
    am_mu, am_std = np.load(am_stat)
    am_mu = paddle.to_tensor(am_mu)
    am_std = paddle.to_tensor(am_std)
    am_normalizer = ZScore(am_mu, am_std)
    am_inference = am_inference_class(am_normalizer, am)
    am_inference.eval()
    if return_am:
        return am_inference, am
    else:
        return am_inference


def get_voc_inference(
        voc: str='pwgan_csmsc',
        voc_config: Optional[os.PathLike]=None,
        voc_ckpt: Optional[os.PathLike]=None,
        voc_stat: Optional[os.PathLike]=None, ):
    # model: {model_name}_{dataset}
    voc_name = voc[:voc.rindex('_')]
    voc_class = dynamic_import(voc_name, model_alias)
    voc_inference_class = dynamic_import(voc_name + '_inference', model_alias)
    if voc_name != 'wavernn':
        voc = voc_class(**voc_config["generator_params"])
        voc.set_state_dict(paddle.load(voc_ckpt)["generator_params"])
        voc.remove_weight_norm()
        voc.eval()
    else:
        voc = voc_class(**voc_config["model"])
        voc.set_state_dict(paddle.load(voc_ckpt)["main_params"])
        voc.eval()

    voc_mu, voc_std = np.load(voc_stat)
    voc_mu = paddle.to_tensor(voc_mu)
    voc_std = paddle.to_tensor(voc_std)
    voc_normalizer = ZScore(voc_mu, voc_std)
    voc_inference = voc_inference_class(voc_normalizer, voc)
    voc_inference.eval()
    return voc_inference


# dygraph to static graph
def am_to_static(am_inference,
                 am: str='fastspeech2_csmsc',
                 inference_dir=Optional[os.PathLike],
                 speaker_dict: Optional[os.PathLike]=None):
    # model: {model_name}_{dataset}
    am_name = am[:am.rindex('_')]
    am_dataset = am[am.rindex('_') + 1:]
    if am_name == 'fastspeech2':
        if am_dataset in {"aishell3", "vctk", "mix",
                          "canton"} and speaker_dict is not None:
            am_inference = jit.to_static(
                am_inference,
                input_spec=[
                    InputSpec([-1], dtype=paddle.int64),
                    InputSpec([1], dtype=paddle.int64),
                ], )
        else:
            am_inference = jit.to_static(
                am_inference, input_spec=[InputSpec([-1], dtype=paddle.int64)])

    elif am_name == 'speedyspeech':
        if am_dataset in {"aishell3", "vctk", "mix",
                          "canton"} and speaker_dict is not None:
            am_inference = jit.to_static(
                am_inference,
                input_spec=[
                    InputSpec([-1], dtype=paddle.int64),  # text
                    InputSpec([-1], dtype=paddle.int64),  # tone
                    InputSpec([1], dtype=paddle.int64),  # spk_id
                    None  # duration
                ])
        else:
            am_inference = jit.to_static(
                am_inference,
                input_spec=[
                    InputSpec([-1], dtype=paddle.int64),
                    InputSpec([-1], dtype=paddle.int64)
                ])

    elif am_name == 'tacotron2':
        am_inference = jit.to_static(
            am_inference, input_spec=[InputSpec([-1], dtype=paddle.int64)])
    elif am_name == 'vits':
        if am_dataset in {"aishell3", "vctk"} and speaker_dict is not None:
            am_inference = jit.to_static(
                am_inference,
                input_spec=[
                    InputSpec([-1], dtype=paddle.int64),
                    InputSpec([1], dtype=paddle.int64),
                ])
        else:
            am_inference = jit.to_static(
                am_inference, input_spec=[InputSpec([-1], dtype=paddle.int64)])
    jit.save(am_inference, os.path.join(inference_dir, am))
    am_inference = jit.load(os.path.join(inference_dir, am))
    return am_inference


def voc_to_static(voc_inference,
                  voc: str='pwgan_csmsc',
                  inference_dir=Optional[os.PathLike]):
    voc_inference = jit.to_static(
        voc_inference, input_spec=[
            InputSpec([-1, 80], dtype=paddle.float32),
        ])
    jit.save(voc_inference, os.path.join(inference_dir, voc))
    voc_inference = jit.load(os.path.join(inference_dir, voc))
    return voc_inference


# inference
def get_predictor(
        model_dir: Optional[os.PathLike]=None,
        model_file: Optional[os.PathLike]=None,
        params_file: Optional[os.PathLike]=None,
        device: str='cpu',
        # for gpu
        use_trt: bool=False,
        # for trt
        use_dynamic_shape: bool=True,
        min_subgraph_size: int=5,
        # for cpu
        cpu_threads: int=1,
        use_mkldnn: bool=False,
        # for trt or mkldnn
        precision: int="fp32"):
    """
    Args:
        model_dir (os.PathLike): root path of model.pdmodel and model.pdiparams.
        model_file (os.PathLike): name of model_file.
        params_file (os.PathLike): name of params_file.
        device (str): Choose the device you want to run, it can be: cpu/gpu, default is cpu.
        use_trt (bool): whether to use TensorRT or not in GPU.
        use_dynamic_shape (bool): use dynamic shape or not in TensorRT.
        use_mkldnn (bool): whether to use MKLDNN or not in CPU.
        cpu_threads (int): num of thread when use CPU.
        precision (str): mode of running (fp32/fp16/bf16/int8).  
    """
    rerun_flag = False
    if device != "gpu" and use_trt:
        raise ValueError(
            "Predict by TensorRT mode: {}, expect device=='gpu', but device == {}".
            format(precision, device))

    config = inference.Config(
        str(Path(model_dir) / model_file), str(Path(model_dir) / params_file))
    config.enable_memory_optim()
    config.switch_ir_optim(True)
    if device == "gpu":
        config.enable_use_gpu(100, 0)
    else:
        config.disable_gpu()
        config.set_cpu_math_library_num_threads(cpu_threads)
        if use_mkldnn:
            # fp32
            config.enable_mkldnn()
            if precision == "int8":
                config.enable_mkldnn_int8({
                    "conv2d_transpose", "conv2d", "depthwise_conv2d", "pool2d",
                    "transpose2", "elementwise_mul"
                })
                # config.enable_mkldnn_int8()
            elif precision in {"fp16", "bf16"}:
                config.enable_mkldnn_bfloat16()
            print("MKLDNN with {}".format(precision))
    if use_trt:
        if precision == "bf16":
            print("paddle trt does not support bf16, switching to fp16.")
            precision = "fp16"
        precision_map = {
            "int8": inference.Config.Precision.Int8,
            "fp32": inference.Config.Precision.Float32,
            "fp16": inference.Config.Precision.Half,
        }
        assert precision in precision_map.keys()
        pdtxt_name = model_file.split(".")[0] + "_" + precision + ".txt"
        if use_dynamic_shape:
            dynamic_shape_file = os.path.join(model_dir, pdtxt_name)
            if os.path.exists(dynamic_shape_file):
                config.enable_tuned_tensorrt_dynamic_shape(dynamic_shape_file,
                                                           True)
                # for fastspeech2
                config.exp_disable_tensorrt_ops(["reshape2"])
                print("trt set dynamic shape done!")
            else:
                # In order to avoid memory overflow when collecting dynamic shapes, it is changed to use CPU.
                config.disable_gpu()
                config.set_cpu_math_library_num_threads(10)
                config.collect_shape_range_info(dynamic_shape_file)
                print("Start collect dynamic shape...")
                rerun_flag = True

        if not rerun_flag:
            print("Tensor RT with {}".format(precision))
            config.enable_tensorrt_engine(
                workspace_size=1 << 30,
                max_batch_size=1,
                min_subgraph_size=min_subgraph_size,
                precision_mode=precision_map[precision],
                use_static=True,
                use_calib_mode=False, )

    predictor = inference.create_predictor(config)
    return predictor


def get_am_output(
        input: str,
        am_predictor: paddle.nn.Layer,
        am: str,
        frontend: object,
        lang: str='zh',
        merge_sentences: bool=True,
        speaker_dict: Optional[os.PathLike]=None,
        spk_id: int=0, ):
    am_name = am[:am.rindex('_')]
    am_dataset = am[am.rindex('_') + 1:]
    am_input_names = am_predictor.get_input_names()
    get_spk_id = False
    get_tone_ids = False
    if am_name == 'speedyspeech':
        get_tone_ids = True
    if am_dataset in {"aishell3", "vctk", "mix", "canton"} and speaker_dict:
        get_spk_id = True
        spk_id = np.array([spk_id])

    frontend_dict = run_frontend(
        frontend=frontend,
        text=input,
        merge_sentences=merge_sentences,
        get_tone_ids=get_tone_ids,
        lang=lang)

    if get_tone_ids:
        tone_ids = frontend_dict['tone_ids']
        tones = tone_ids[0].numpy()
        tones_handle = am_predictor.get_input_handle(am_input_names[1])
        tones_handle.reshape(tones.shape)
        tones_handle.copy_from_cpu(tones)
    if get_spk_id:
        spk_id_handle = am_predictor.get_input_handle(am_input_names[1])
        spk_id_handle.reshape(spk_id.shape)
        spk_id_handle.copy_from_cpu(spk_id)
    phone_ids = frontend_dict['phone_ids']
    phones = phone_ids[0].numpy()
    phones_handle = am_predictor.get_input_handle(am_input_names[0])
    phones_handle.reshape(phones.shape)
    phones_handle.copy_from_cpu(phones)

    am_predictor.run()
    am_output_names = am_predictor.get_output_names()
    am_output_handle = am_predictor.get_output_handle(am_output_names[0])
    am_output_data = am_output_handle.copy_to_cpu()
    return am_output_data


def get_voc_output(voc_predictor, input):
    voc_input_names = voc_predictor.get_input_names()
    mel_handle = voc_predictor.get_input_handle(voc_input_names[0])
    mel_handle.reshape(input.shape)
    mel_handle.copy_from_cpu(input)

    voc_predictor.run()
    voc_output_names = voc_predictor.get_output_names()
    voc_output_handle = voc_predictor.get_output_handle(voc_output_names[0])
    wav = voc_output_handle.copy_to_cpu()
    return wav


def get_am_sublayer_output(am_sublayer_predictor, input):
    am_sublayer_input_names = am_sublayer_predictor.get_input_names()
    input_handle = am_sublayer_predictor.get_input_handle(
        am_sublayer_input_names[0])
    input_handle.reshape(input.shape)
    input_handle.copy_from_cpu(input)

    am_sublayer_predictor.run()
    am_sublayer_names = am_sublayer_predictor.get_output_names()
    am_sublayer_handle = am_sublayer_predictor.get_output_handle(
        am_sublayer_names[0])
    am_sublayer_output = am_sublayer_handle.copy_to_cpu()
    return am_sublayer_output


def get_streaming_am_output(input: str,
                            am_encoder_infer_predictor,
                            am_decoder_predictor,
                            am_postnet_predictor,
                            frontend,
                            lang: str='zh',
                            merge_sentences: bool=True):
    get_tone_ids = False
    frontend_dict = run_frontend(
        frontend=frontend,
        text=input,
        merge_sentences=merge_sentences,
        get_tone_ids=get_tone_ids,
        lang=lang)
    phone_ids = frontend_dict['phone_ids']
    phones = phone_ids[0].numpy()
    am_encoder_infer_output = get_am_sublayer_output(
        am_encoder_infer_predictor, input=phones)

    am_decoder_output = get_am_sublayer_output(
        am_decoder_predictor, input=am_encoder_infer_output)

    am_postnet_output = get_am_sublayer_output(
        am_postnet_predictor, input=np.transpose(am_decoder_output, (0, 2, 1)))
    am_output_data = am_decoder_output + np.transpose(am_postnet_output,
                                                      (0, 2, 1))
    normalized_mel = am_output_data[0]
    return normalized_mel


# onnx
def get_sess(model_path: Optional[os.PathLike],
             device: str='cpu',
             cpu_threads: int=1,
             use_trt: bool=False):
    sess_options = ort.SessionOptions()
    sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
    sess_options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
    if 'gpu' in device.lower():
        device_id = int(device.split(':')[1]) if len(
            device.split(':')) == 2 else 0
        # fastspeech2/mb_melgan can't use trt now!
        if use_trt:
            provider_name = 'TensorrtExecutionProvider'
        else:
            provider_name = 'CUDAExecutionProvider'
        providers = [(provider_name, {'device_id': device_id})]
    elif device.lower() == 'cpu':
        providers = ['CPUExecutionProvider']
    sess_options.intra_op_num_threads = cpu_threads
    sess = ort.InferenceSession(
        model_path, providers=providers, sess_options=sess_options)
    return sess