optimize(infer): move attentions into rvc

rvc/norms.py

@@ -0,0 +1,18 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, channels: int, eps: float = 1e-5):
+        super(LayerNorm, self).__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x: torch.Tensor):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)