onnx_export.py

### from @nshmyrev's fork :  https://github.com/alphacep/MB-iSTFT-VITS2/blob/main/export.py

import librosa
import matplotlib.pyplot as plt

import os
import json
import math

import requests
import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import DataLoader

import commons
import utils
from data_utils import TextAudioLoader, TextAudioCollate, TextAudioSpeakerLoader, TextAudioSpeakerCollate
from models import SynthesizerTrn
from text.symbols import symbols

import numpy as np
from scipy.io.wavfile import write
import re
from scipy import signal


#- Variable section
PATH_TO_CONFIG = "./configs/PATH_TO_CONFIG.json" # path to config
PATH_TO_MODEL = "PATH_TO_G_XXXX.pth" # path to model
SPEAKER_ID = None # sid
SCALE_CONFIG = torch.FloatTensor([0.667, 1.0, 0.8]) # scales -> noise, noise_w, length
OPSET_VERSION = 15


hps = utils.get_hparams_from_file(PATH_TO_CONFIG)

if "use_mel_posterior_encoder" in hps.model.keys() and hps.model.use_mel_posterior_encoder == True:
    print("Using mel posterior encoder for VITS2")
    posterior_channels = 80  # vits2
    hps.data.use_mel_posterior_encoder = True
else:
    print("Using lin posterior encoder for VITS1")
    posterior_channels = hps.data.filter_length // 2 + 1
    hps.data.use_mel_posterior_encoder = False

net_g = SynthesizerTrn(
    len(symbols),
    posterior_channels,
    hps.train.segment_size // hps.data.hop_length,
    n_speakers=hps.data.n_speakers,
    is_onnx=True, # !
    **hps.model)

_ = utils.load_checkpoint(PATH_TO_MODEL, net_g, None)

num_symbols = net_g.n_vocab
num_speakers = net_g.n_speakers


def infer_forward(text, text_lengths, scales, sid=None):
    noise_scale = scales[0]
    length_scale = scales[1]
    noise_scale_w = scales[2]
    audio = net_g.infer(
            text,
            text_lengths,
            noise_scale=noise_scale,
            length_scale=length_scale,
            noise_scale_w=noise_scale_w,
            sid=sid,
    )[0].unsqueeze(1)

    return audio


with torch.no_grad():
    net_g.dec.remove_weight_norm()
    net_g.flow.remove_weight_norm() # Remove weightnorm in flows - a8d9f74
    net_g.forward = infer_forward

net_g.eval()

# dummy initialization
dmy_text = torch.randint(low=0, high=num_symbols, size=(1, 50), dtype=torch.long)
dmy_text_length = torch.LongTensor([dmy_text.size(1)])
dummy_input = (dmy_text, dmy_text_length, SCALE_CONFIG, SPEAKER_ID) # infer_forward()


# Export
torch.onnx.export(
        model=net_g,
        args=dummy_input,
        f="model.onnx",
        verbose=True,
        opset_version=OPSET_VERSION,
        input_names=["input", "input_lengths", "scales", "sid"],
        output_names=["output"],
        dynamic_axes={
            "input": {0: "batch_size", 1: "phonemes"},
            "input_lengths": {0: "batch_size"},
            "output": {0: "batch_size", 1: "time"},
        },
)