def _prepare_sp(self, path):
sp = np.load(str(path))
sp = self._normalize(encode_sp(sp, sr=22050, mel_bins=36))
sp = self._crop(sp, upper_bound=128)
return sp
content_discriminator = ContentDiscriminator()
content_discriminator.to_gpu()
dis_con_opt = set_optimizer(content_discriminator, alpha=0.0001)
for epoch in range(epochs):
sum_gen_loss = 0
sum_dis_loss = 0
for batch in range(0, Ntrain, batchsize):
x_sp_box = []
y_sp_box = []
for _ in range(batchsize):
# sp loading -> mel conversion -> normalization -> crop
rnd_x = np.random.randint(x_len)
sp_x = np.load(x_path + x_list[rnd_x])
sp_x = normalize(encode_sp(sp_x, mel_bins=36))
rnd_y = np.random.randint(y_len)
sp_y = np.load(y_path + y_list[rnd_y])
sp_y = normalize(encode_sp(sp_y, mel_bins=36))
sp_x = crop(sp_x, upper_bound=128)
sp_y = crop(sp_y, upper_bound=128)
x_sp_box.append(sp_x[np.newaxis,:])
y_sp_box.append(sp_y[np.newaxis,:])
x = chainer.as_variable(xp.array(x_sp_box).astype(xp.float32))
y = chainer.as_variable(xp.array(y_sp_box).astype(xp.float32))
# Discriminator update
a_out, b_out = generator(x, y)
a_enc, _, a_fake, _, a_infer = a_out
serializers.load_npz('./generator_xy.model', generator_xy)
#serializers.load_npz('./generator_mtok.model', generator_xy)
x_sp_box = []
rnd = np.random.randint(x_len)
wav = x_wav_path + x_list[rnd]
#wav = './Dataset/jsut_ver1.1/basic5000/wav16/BASIC5000_3895.wav'
print(wav)
xx = load(wav, sampling_rate=16000)
f0, sp, ap = audio2af(xx)
f0_tmp = np.exp((np.ma.log(f0) - 5.35) / 0.24 * 0.27 + 5.67)
#f0_tmp = np.exp((np.ma.log(f0) - 6.20) / 0.29 * 0.27 + 5.67)
sp, mean, std = normalize(encode_sp(sp))
sp_y = np.zeros_like(sp, dtype=np.float64)
length = sp.shape[0]
for index in range(int(sp.shape[0] / 128) + 1):
if 128 * (index + 1) > length:
sp_tmp = sp[128 * index:length]
sp_tmp = np.pad(sp_tmp, ((0, 128 - sp_tmp.shape[0]), (0, 0)),
'constant',
constant_values=0)
else:
sp_tmp = sp[128 * index:128 * (index + 1)]
x_sp_box.append(sp_tmp[np.newaxis, :])