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

simpler gmm attention implementaiton

pull/10/head
root 2020-01-09 15:56:09 +01:00
parent f2b6d00c45
commit 5e148038be
2 changed files with 7 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
config.json
View File

@ -109,7 +109,7 @@
[ [
{ {
"name": "ljspeech", "name": "ljspeech",
"path": "/data5/ro/shared/data/keithito/LJSpeech-1.1/", "path": "/root/LJSpeech-1.1/",
// "path": "/home/erogol/Data/LJSpeech-1.1", // "path": "/home/erogol/Data/LJSpeech-1.1",
"meta_file_train": "metadata_train.csv", "meta_file_train": "metadata_train.csv",
"meta_file_val": "metadata_val.csv" "meta_file_val": "metadata_val.csv"

15
layers/common_layers.py
View File

@ -132,7 +132,7 @@ class GravesAttention(nn.Module):
def init_states(self, inputs): def init_states(self, inputs):
if self.J is None or inputs.shape[1] > self.J.shape[-1]: if self.J is None or inputs.shape[1] > self.J.shape[-1]:
self.J = torch.arange(0, inputs.shape[1]).to(inputs.device).expand([inputs.shape[0], self.K, inputs.shape[1]]) self.J = torch.arange(0, inputs.shape[1]+1).to(inputs.device) + 0.5
self.attention_weights = torch.zeros(inputs.shape[0], inputs.shape[1]).to(inputs.device) self.attention_weights = torch.zeros(inputs.shape[0], inputs.shape[1]).to(inputs.device)
self.mu_prev = torch.zeros(inputs.shape[0], self.K).to(inputs.device) self.mu_prev = torch.zeros(inputs.shape[0], self.K).to(inputs.device)
@ -164,17 +164,14 @@ class GravesAttention(nn.Module):
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
# each B x K x T_in j = self.J[:inputs.size(1)+1]
g_t = g_t.unsqueeze(2).expand(g_t.size(0),
g_t.size(1),
inputs.size(1))
sig_t = sig_t.unsqueeze(2).expand_as(g_t)
mu_t_ = mu_t.unsqueeze(2).expand_as(g_t)
j = self.J[:g_t.size(0), :, :inputs.size(1)]
# attention weights # attention weights
phi_t = g_t * torch.exp(-0.5 * (mu_t_ - j)**2 / (sig_t**2)) phi_t = g_t.unsqueeze(-1) * torch.exp(-0.5 * (mu_t.unsqueeze(-1) - j)**2 / (sig_t.unsqueeze(-1)**2))
# discritize attention weights
alpha_t = self.COEF * torch.sum(phi_t, 1) alpha_t = self.COEF * torch.sum(phi_t, 1)
alpha_t = alpha_t[:, 1:] - alpha_t[:, :-1]
# apply masking # apply masking
if mask is not None: if mask is not None: