if iteration % 5 == 0:
if hp.loss == "BEGAN":
logger.log_training(iteration = iteration, loss_gan = loss_gan,
loss_dis = loss_dis, loss_cycle = loss_cycle, k = k, convergence = convergence)
if (iteration % 50 == 0):
save_checkpoint(m, opt, iteration, f'checkpoint/{args.checkpoint_path}/gen')
save_checkpoint(dis_high, opt_dis, iteration, f'checkpoint/{args.checkpoint_path}/dis')
idx = random.randint(0, fake_singing.size(0) - 1)
#mel = (mel * std) +mean
#z = (z * std) + mean
real_audio = melblock.inverse(singing).detach().cpu().numpy()
fake_audio = melblock.inverse(fake_singing).detach().cpu().numpy()
real_speech_audio = vocoder_speech.inverse(speech).detach().cpu().numpy()
#mel = (mel -mean)/ std
#z = (z - mean ) / std
"""
logger work like this:
logger only accept image, audio ,scalars type.
and the type of them is :
scalars : int
image : tensor
audio : ndarray
More details can see logger/logger.py
"""
loss_sp2sing = loss_sp2sing.mean()
OptimStep([(m, opt, loss_sp2sing, False)], 3)
if (iteration % 50 == 0):
logger.log_training(iteration=iteration, loss_sing2sp=loss_sp2sing)
save_checkpoint(m, opt, iteration,
f'checkpoint/{args.checkpoint_path}/gen')
idx = 0
fake_song_padded = torch.clamp(fake_song_padded, -10000, 1.8)
singing_melgan = fake_song_padded[idx:idx + 1]
singing_real_melgan = song_padded[idx:idx + 1]
if args.feat_type == "mel":
fake_singing_audio = vocoder.inverse(
singing_melgan).detach().cpu().numpy()[0]
real_singing_audio = vocoder.inverse(
singing_real_melgan).detach().cpu().numpy()[0]
else:
fake_singing_audio = gl_rec(
singing_melgan[0].detach().cpu().numpy())
real_singing_audio = gl_rec(
singing_real_melgan[0].detach().cpu().numpy())
logger.log_validation(
iteration=iteration,
mel_train_singing=("image", plot_spectrogram_to_numpy(),
song_padded[idx]),
mel_train_singing_generate=("image", plot_spectrogram_to_numpy(),
fake_song_padded[idx]),