Medix-AI/coqui-tts
Medix-AI
/
coqui-tts
mirror of https://github.com/coqui-ai/TTS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

graves v2

This commit is contained in:
root 2020-01-15 01:53:27 +01:00 committed by erogol
parent cf7d968f57
commit 72817438db
2 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
config.json
View File

@ -1,6 +1,6 @@
{ {
"model": "Tacotron2", // one of the model in models/ "model": "Tacotron2", // one of the model in models/
"run_name": "ljspeech-graves", "run_name": "ljspeech-gravesv2",
"run_description": "tacotron2 wuth graves attention", "run_description": "tacotron2 wuth graves attention",
// AUDIO PARAMETERS // AUDIO PARAMETERS

8
layers/common_layers.py
View File

@ -118,7 +118,7 @@ class GravesAttention(nn.Module):
def __init__(self, query_dim, K): def __init__(self, query_dim, K):
super(GravesAttention, self).__init__() super(GravesAttention, self).__init__()
self._mask_value = 0.0 self._mask_value = 1e-8
self.K = K self.K = K
# self.attention_alignment = 0.05 # self.attention_alignment = 0.05
self.eps = 1e-5 self.eps = 1e-5
@ -165,12 +165,14 @@ class GravesAttention(nn.Module):
# attention GMM parameters # attention GMM parameters
sig_t = torch.nn.functional.softplus(b_t) + self.eps sig_t = torch.nn.functional.softplus(b_t) + self.eps
mu_t = self.mu_prev + torch.nn.functional.softplus(k_t) mu_t = self.mu_prev + torch.nn.functional.softplus(k_t)
g_t = torch.softmax(g_t, dim=-1) / sig_t + self.eps g_t = torch.softmax(g_t, dim=-1) / sig_t + self.eps
j = self.J[:inputs.size(1)+1] j = self.J[:inputs.size(1)+1]
# attention weights # attention weights
phi_t = g_t.unsqueeze(-1) * (1 / (1 + torch.exp((mu_t.unsqueeze(-1) - j) / sig_t.unsqueeze(-1)))) phi_t = g_t.unsqueeze(-1) * (1 / (1 + torch.sigmoid((mu_t.unsqueeze(-1) - j) / sig_t.unsqueeze(-1))))
# discritize attention weights # discritize attention weights
alpha_t = torch.sum(phi_t, 1) alpha_t = torch.sum(phi_t, 1)
@ -182,8 +184,6 @@ class GravesAttention(nn.Module):
alpha_t.data.masked_fill_(~mask, self._mask_value) alpha_t.data.masked_fill_(~mask, self._mask_value)
context = torch.bmm(alpha_t.unsqueeze(1), inputs).squeeze(1) context = torch.bmm(alpha_t.unsqueeze(1), inputs).squeeze(1)
# for better visualization
# self.attention_weights = torch.clamp(alpha_t, min=0)
self.attention_weights = alpha_t self.attention_weights = alpha_t
self.mu_prev = mu_t self.mu_prev = mu_t
return context return context