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

chnage gmm activations

This commit is contained in:
Eren Golge 2019-11-09 02:33:37 +01:00
parent 015f7780f4
commit 6f3dd1b6ae
1 changed files with 8 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
layers/common_layers.py
View File

@ -116,7 +116,7 @@ class GravesAttention(nn.Module):
self._mask_value = 0.0 self._mask_value = 0.0
self.K = K self.K = K
# self.attention_alignment = 0.05 # self.attention_alignment = 0.05
self.epsilon = 1e-5 self.eps = 1e-5
self.J = None self.J = None
self.N_a = nn.Sequential( self.N_a = nn.Sequential(
nn.Linear(query_dim, query_dim, bias=True), nn.Linear(query_dim, query_dim, bias=True),
@ -157,26 +157,25 @@ class GravesAttention(nn.Module):
k_t = gbk_t[:, 2, :] k_t = gbk_t[:, 2, :]
# attention GMM parameters # attention GMM parameters
# g_t = torch.softmax(g_t, dim=-1) + self.epsilon # distribution weight inv_sig_t = torch.exp(-torch.clamp(b_t, min=-7, max=9)) # variance
# sig_t = torch.exp(b_t) + self.epsilon # variance
# mu_t = self.mu_prev + self.attention_alignment * torch.exp(k_t) # mean
sig_t = torch.pow(torch.nn.functional.softplus(b_t), 2)
mu_t = self.mu_prev + torch.nn.functional.softplus(k_t) mu_t = self.mu_prev + torch.nn.functional.softplus(k_t)
# TODO try sigmoid here g_t = torch.softmax(g_t, dim=-1) * inv_sig_t + self.eps
g_t = (torch.softmax(g_t, dim=-1) / sig_t)
# each B x K x T_in # each B x K x T_in
g_t = g_t.unsqueeze(2).expand(g_t.size(0), g_t = g_t.unsqueeze(2).expand(g_t.size(0),
g_t.size(1), g_t.size(1),
inputs.size(1)) inputs.size(1))
sig_t = sig_t.unsqueeze(2).expand_as(g_t) inv_sig_t = inv_sig_t.unsqueeze(2).expand_as(g_t)
mu_t_ = mu_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)] j = self.J[:g_t.size(0), :, :inputs.size(1)]
# attention weights # attention weights
phi_t = g_t * torch.exp(-0.5 * sig_t * (mu_t_ - j)**2) phi_t = g_t * torch.exp(-0.5 * inv_sig_t * (mu_t_ - j)**2)
alpha_t = self.COEF * torch.sum(phi_t, 1) alpha_t = self.COEF * torch.sum(phi_t, 1)
if alpha_t.max() > 1e+2:
breakpoint()
# apply masking # apply masking
if mask is not None: if mask is not None:
alpha_t.data.masked_fill_(~mask, self._mask_value) alpha_t.data.masked_fill_(~mask, self._mask_value)