Is the dimension of the alignment of FastSpeech the same as that of tacotron2? about fastspeech HOT 5 CLOSED

alignment.py

from fastspeech.

I'm sorry to reply you so late.

from fastspeech.

您好,谢谢您的实现以及回复,对我很有帮助.
但关于alignment.py中的get_D函数我还有一些问题

def get_D(alignment):
D = np.array([0 for _ in range(np.shape(alignment)[1])])

for i in range(np.shape(alignment)[0]):
    max_index = alignment[i].tolist().index(alignment[i].max())
    D[max_index] = D[max_index] + 1

for i in range(np.shape(D)[0]):
    if D[i] > 0:
        D[i] = D[i] - 1

return D

其中为什么需要以下的步骤?

for i in range(np.shape(D)[0]):
if D[i] > 0:
D[i] = D[i] - 1

我加上这一步之后,sum(D)=S-T(S and T are the lengths of the ground-truth spectrograms and phonemes),我对原文的理解应该是sum(D)=S.
希望您有空可以帮忙解答一下疑惑,谢谢!

from fastspeech.

不好意思，现在才看到。
那段代码具体的意思是
首先alignment的第二维是字符的长度，
for i in range(np.shape(alignment)[0]):
max_index = alignment[i].tolist().index(alignment[i].max())
D[max_index] = D[max_index] + 1
为的是找到每个mel列中对应的最大的字符的index
D就是一个长度为字符数的计数器

for i in range(np.shape(D)[0]):
if D[i] > 0:
D[i] = D[i] - 1
上面这一步的目的是为了将所有计数减一，并防止0变成负一，因为在之后的代码中，计数为1的字符实际上会扩充成2个，所以需要减一。

from fastspeech.

好的,了解了,非常感谢!

from fastspeech.