if args.load_checkpoint==True:
m, opt, iteration = load_checkpoint(f'checkpoint/{args.checkpoint_path}/gen', m, opt)
dis_high, opt_dis, iteration = load_checkpoint(f'checkpoint/{args.checkpoint_path}/dis', dis_high, opt_dis)
'''
###########################################################
In general, we preprocess data to npy, and put them in
specific folder. Dataloader load npy file.
But in this example, I show that how to transfrom audio
into stft, melspectrogram by torch.nn.module (MelSpectrogram).
###########################################################
'''
#melblock = MelSpectrogram(hp).cuda()
melblock = MelVocoder(path = "vocoder/melgan-neurips/scripts/logs/NUS")
vocoder_speech = torch.hub.load('descriptinc/melgan-neurips', 'load_melgan')
import random
while True:
speech = next(inf_iterator_tr_speech).cuda()
singing = next(inf_iterator_tr_sing).cuda()
scale = 1 + random.random()
speech_2x= F.interpolate(speech, scale_factor=scale, mode='nearest')
#mel = (mel-mean)/std
singing = singing[:,:,:min(speech_2x.size(2), singing.size(2))]
speech_2x = speech_2x[:,:,:min(speech_2x.size(2), singing.size(2))]
singing = F.pad(singing,(0,(singing.size(2)//8+1)*8 - singing.size(2)), 'reflect')
speech_2x = F.pad(speech_2x,(0,(speech_2x.size(2)//8+1)*8 - speech_2x.size(2)), 'reflect')
sys.path.append('../logger')
from logger import Logger
from logger_utils import prepare_directories_and_logger
from plotting_utils import plot_spectrogram_to_numpy
sys.path.append('../utils')
from optim_step import *
from save_and_load import save_checkpoint, load_checkpoint
from torch.nn import functional as F
sys.path.append('/home/ericwudayi/AiVocal/SkipVQVC/vocoder/melgan-neurips')
from mel2wav.interface import MelVocoder
#vocoder = torch.hub.load('descriptinc/melgan-neurips', 'load_melgan')
vocoder = MelVocoder(
path=
"/home/ericwudayi/AiVocal/ai_singing/vocoder/melgan-neurips/scripts/logs/NUS"
)
#vocoder = torch.hub.load('descriptinc/melgan-neurips', 'load_melgan')
def train_(args, model, opt, latent_loss_weight, criterion, loader, epochs,
inf_iterator_test, logger, iteration):
for epoch in range(epochs):
mse_sum = 0
mse_n = 0
for i, (audio, pitch) in enumerate(loader):
audio = audio.cuda().float()
dis_high = ymp.construct_model(f"model_config/{hp.config_dis}/1.yaml")
dis_high = dis_high.cuda()
opt_dis = optim.Adam(dis_high.parameters(), lr=1e-4)
iteration = 0
if args.load_checkpoint == True:
m, opt, iteration = load_checkpoint(
f'checkpoint/{args.checkpoint_path}/gen', m, opt)
##########################################################
'''
### Vocoder block ###
MelGan vocoder, vocoder/modules
'''
vocoder = MelVocoder(path="vocoder/melgan-neurips/scripts/logs/NUS")
#######################################################################
while True:
song_padded, read_padded, pitch_padded, read_real = \
next(inf_iterator_tr_speech)
song_padded, read_padded, pitch_padded = \
song_padded.float().cuda(), read_padded.float().cuda(), pitch_padded.long().cuda()
song_padded = song_padded[..., :song_padded.size(2) // 8 * 8]
read_padded = read_padded[..., :read_padded.size(2) // 8 * 8]
pitch_padded = pitch_padded[..., :read_padded.size(2) // 8 * 8]
factor = 16