How to finefune hifigan vocoder for mixertts

[pivolan]

got this error:

 data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/usr/local/lib/python3.8/dist-packages/nemo/collections/tts/data/tts_dataset.py", line 1099, in __getitem__
    start = random.randint(0, mel.shape[1] - frames - 1)
  File "/usr/lib/python3.8/random.py", line 248, in randint
    return self.randrange(a, b+1)
  File "/usr/lib/python3.8/random.py", line 226, in randrange
    raise ValueError("empty range for randrange() (%d, %d, %d)" % (istart, istop, width))
ValueError: empty range for randrange() (0, -178, -178)


Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

tried to use mel_generator from examples for fastpitch little bit modify it,

import argparse
import json
from pathlib import Path

import numpy as np
import soundfile as sf
import torch
from joblib import Parallel, delayed
from tqdm import tqdm

from nemo.collections.tts.models import MixerTTSModel
from nemo.collections.tts.parts.utils.tts_dataset_utils import (
    BetaBinomialInterpolator,
    beta_binomial_prior_distribution,
)
from nemo.utils import logging


def get_args():
    parser = argparse.ArgumentParser(
        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
        description="Generate mel spectrograms with pretrained FastPitch model, and create manifests for finetuning Hifigan.",
    )
    parser.add_argument(
        "--model",
        required=True,
        type=Path,
        help="Specify a full path of a fastpitch model checkpoint with the suffix of either .ckpt or .nemo.",
    )
    parser.add_argument(
        "--manifest",
        nargs="+",
        required=True,
        type=Path,
        help="Specify a full path of a JSON manifest. You could add multiple manifests.",
    )
    parser.add_argument(
        "--to",
        required=True,
        type=Path,
        help="Specify a full path of output root that would contain new manifests.",
    )
    parser.add_argument(
        "--num-workers",
        default=-1,
        type=int,
        help="Specify the max number of concurrently Python workers processes. "
             "If -1 all CPUs are used. If 1 no parallel computing is used.",
    )
    parser.add_argument("--cpu", action='store_true', default=False, help="Generate mel spectrograms using CPUs.")
    args = parser.parse_args()
    return args


def __load_wav(audio_file):
    with sf.SoundFile(audio_file, 'r') as f:
        samples = f.read(dtype='float32')
    return samples.transpose()


def __generate_mels(entry, spec_model, device, use_beta_binomial_interpolator, mel_root):
    # Generate a spectrograms (we need to use ground truth alignment for correct matching between audio and mels)
    audio = __load_wav(entry["audio_filepath"])
    audio = torch.from_numpy(audio).unsqueeze(0).to(device)
    audio_len = torch.tensor(audio.shape[1], dtype=torch.long, device=device).unsqueeze(0)

    speaker = None

    with torch.no_grad():
        if "normalized_text" in entry:
            text = spec_model.parse(entry["normalized_text"], normalize=False)
        else:
            text = spec_model.parse(entry['text'])

        text_len = torch.tensor(text.shape[-1], dtype=torch.long, device=device).unsqueeze(0)
        spect, spect_len = spec_model.preprocessor(input_signal=audio, length=audio_len)

        # Generate attention prior and spectrogram inputs for HiFi-GAN
        if use_beta_binomial_interpolator:
            beta_binomial_interpolator = BetaBinomialInterpolator()
            attn_prior = (
                torch.from_numpy(beta_binomial_interpolator(spect_len.item(), text_len.item()))
                .unsqueeze(0)
                .to(text.device)
            )
        else:
            attn_prior = (
                torch.from_numpy(beta_binomial_prior_distribution(text_len.item(), spect_len.item()))
                .unsqueeze(0)
                .to(text.device)
            )

        spectrogram = spec_model.forward(
            text=text, text_len=text_len, spect=spect, spect_len=spect_len, attn_prior=attn_prior
        )[0]
        spectrogram = np.transpose(spectrogram[0].to('cpu').numpy(), (1, -2))

        save_path = mel_root / f"{Path(entry['audio_filepath']).stem}.npy"
        np.save(save_path, spectrogram)
        entry["mel_filepath"] = str(save_path)

    return entry

if some discussion found info that I should transpose before save, but with what params? I tries anything, but it still does not work and throw same error

[redoctopus]

You should only need to transpose the spectrogram's frequency/time axes, as by the time you're calling it you've already gotten rid of the batch dimension. You shouldn't need to change any other parameters.