class OnnxLengthRegurator(nn.Module):
def __init__(self, max_len=1400):
super().__init__()
self.max_len = max_len
# The maximum length of the make_pad_mask is the
# maximum value of the duration.
self.make_pad_mask = None #MakePadMask(max_seq_len, flip=False)
def LR(self, x, duration):
duration = duration.to(torch.int32)
# positions = (torch.arange(self.max_len).unsqueeze(-1).unsqueeze(-1) < duration.cumsum(1)).sum(2)
positions = (torch.arange(self.max_len).unsqueeze(-1).unsqueeze(-1).to(torch.int32) < duration.cumsum(1).to(torch.int32)).sum(2)
mask = positions.ne(0)
index = ((duration.size(1) - positions) * mask).T
x_l = [torch.index_select(x[i], 0, index[i]) for i in range(x.size(0))]
outputs = torch.stack(x_l).to(x.dtype)
outputs = outputs * mask.T.unsqueeze(-1)
# outputs torch.Size([2, 1000, 256]) batch, mel_len, hidden
return outputs, mask.sum(0)
def forward(self, x, duration):
output, mel_len = self.LR(x, duration)
return output #, mel_len
TTS中的length_regular(分箱算法)的静态实现形式
最新推荐文章于 2024-09-28 22:05:59 发布
该代码定义了一个名为OnnxLengthRegulator的神经网络模块,用于处理序列数据的长度调整。它根据给定的持续时间(duration)动态计算padding位置,使用cumsum方法和torch的比较操作来确定非零元素的索引。然后,它选择并堆叠序列中的有效部分,最后返回处理后的输出和mask的总和,用于指示有效长度。
摘要由CSDN通过智能技术生成