Merge branch 'main' into ps

inference/tts/ps_flow.py

@@ -10,8 +10,9 @@ class PortaSpeechFlowInfer(BaseTTSInfer):
         ph_dict_size = len(self.ph_encoder)
         word_dict_size = len(self.word_encoder)
         model = PortaSpeechFlow(ph_dict_size, word_dict_size, self.hparams)
-        model.eval()
         load_ckpt(model, hparams['work_dir'], 'model')
+        model.post_flow.store_inverse()
+        model.eval()
         return model
 
     def forward_model(self, inp):

modules/tts/portaspeech/fvae.py

@@ -125,7 +125,7 @@ class FVAE(nn.Module):
             return z_q, loss_kl, z_p, m_q, logs_q
         else:
             latent_shape = [cond_sqz.shape[0], self.latent_size, cond_sqz.shape[2]]
-            z_p = self.prior_dist.sample(latent_shape).to(cond.device) * noise_scale
+            z_p = torch.randn(latent_shape).to(cond.device) * noise_scale
             if self.use_prior_flow:
                 z_p = self.prior_flow(z_p, 1, cond_sqz, reverse=True)
             return z_p

modules/tts/portaspeech/portaspeech_flow.py

@@ -70,6 +70,6 @@ class PortaSpeechFlow(PortaSpeech):
                 ret['postflow'] = None
         else:
             nonpadding = torch.ones_like(x_recon[:, :1, :])
-            z_post = prior_dist.sample(x_recon.shape).to(g.device) * self.hparams['noise_scale']
+            z_post = torch.randn(x_recon.shape).to(g.device) * self.hparams['noise_scale']
             x_recon, _ = self.post_flow(z_post, nonpadding, g, reverse=True)
             ret['mel_out'] = x_recon.transpose(1, 2)