DL-Art-School/codes/scripts/audio/gen/use_discrete_vocoder.py

51 lines
2.6 KiB
Python
Raw Normal View History

import argparse
import torchaudio
from data.audio.unsupervised_audio_dataset import load_audio
2021-12-17 03:47:37 +00:00
from scripts.audio.gen.speech_synthesis_utils import do_spectrogram_diffusion, \
load_discrete_vocoder_diffuser, wav_to_mel, convert_mel_to_codes
from utils.audio import plot_spectrogram
from utils.util import load_model_from_config
def roundtrip_vocoding(dvae, vocoder, diffuser, clip, cond=None, plot_spec=False):
clip = clip.unsqueeze(0)
if cond is None:
cond = clip
else:
cond = cond.unsqueeze(0)
mel = wav_to_mel(clip)
if plot_spec:
plot_spectrogram(mel[0].cpu())
codes = convert_mel_to_codes(dvae, mel)
2022-01-20 18:28:50 +00:00
return do_spectrogram_diffusion(vocoder, dvae, diffuser, codes, cond, spectrogram_compression_factor=256, plt_spec=plot_spec)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
2022-01-20 18:28:50 +00:00
parser.add_argument('-opt', type=str, help='Path to options YAML file used to train the diffusion model',
default='X:\\dlas\\experiments\\train_diffusion_vocoder_22k_level.yml')
parser.add_argument('-diffusion_model_name', type=str, help='Name of the diffusion model in opt.', default='generator')
2022-01-20 18:28:50 +00:00
parser.add_argument('-diffusion_model_path', type=str, help='Diffusion model checkpoint to load.', default='X:\\dlas\\experiments\\train_diffusion_vocoder_22k_level\\models\\2500_generator.pth')
parser.add_argument('-dvae_model_name', type=str, help='Name of the DVAE model in opt.', default='dvae')
2022-01-20 18:28:50 +00:00
parser.add_argument('-input_file', type=str, help='Path to the input audio file.', default='Y:\\clips\\books1\\3_dchha04 Romancing The Tribes\\00036.wav')
parser.add_argument('-cond', type=str, help='Path to the conditioning input audio file.', default='Y:\\clips\\books1\\3042_18_Holden__000000000\\00037.wav')
args = parser.parse_args()
print("Loading DVAE..")
dvae = load_model_from_config(args.opt, args.dvae_model_name)
print("Loading Diffusion Model..")
diffusion = load_model_from_config(args.opt, args.diffusion_model_name, also_load_savepoint=False, load_path=args.diffusion_model_path)
print("Loading data..")
diffuser = load_discrete_vocoder_diffuser()
inp = load_audio(args.input_file, 22050).cuda()
2021-12-17 03:47:37 +00:00
cond = inp if args.cond is None else load_audio(args.cond, 22050)
if cond.shape[-1] > 44100+10000:
cond = cond[:,10000:54100]
2022-01-20 18:28:50 +00:00
cond = cond.cuda()
print("Performing inference..")
roundtripped = roundtrip_vocoding(dvae, diffusion, diffuser, inp, cond).cpu()
2022-01-20 18:28:50 +00:00
torchaudio.save('roundtrip_vocoded_output.wav', roundtripped.squeeze(0), 22050)