fix

paddlespeech/t2s/models/diffsinger/diffsinger.py

@@ -270,10 +270,10 @@ class DiffSinger(nn.Layer):
         mel_fs2 = mel_fs2.unsqueeze(0).transpose((0, 2, 1))
         cond_fs2 = self.fs2.encoder_infer(text, note, note_dur, is_slur)
         cond_fs2 = cond_fs2.transpose((0, 2, 1))
-        mel, _ = self.diffusion(mel_fs2, cond_fs2)
-        # noise = paddle.randn(mel_fs2.shape)
-        # mel = self.diffusion.inference(
-        #     noise=noise, cond=cond_fs2, ref_x=mel_fs2, num_inference_steps=100)
+        # mel, _ = self.diffusion(mel_fs2, cond_fs2)
+        noise = paddle.randn(mel_fs2.shape)
+        mel = self.diffusion.inference(
+            noise=noise, cond=cond_fs2, ref_x=mel_fs2, num_inference_steps=100)
         mel = mel.transpose((0, 2, 1))
         return mel[0]
 

paddlespeech/t2s/modules/diffusion.py

@@ -372,11 +372,9 @@ class GaussianDiffusion(nn.Layer):
                 The noises which is added to the input.
 
         """
-        # print("xxxxxxxxxxxxxxxx1: ", x, x.shape)
         x = x.transpose((0, 2, 1))
         x = self.norm_spec(x)
         x = x.transpose((0, 2, 1))
-        print("xxxxxxxxxxxxxxxx2: ", x, x.shape)
         noise_scheduler = self.noise_scheduler
 
         # Sample noise that we'll add to the mel-spectrograms
@@ -411,7 +409,7 @@ class GaussianDiffusion(nn.Layer):
                   num_inference_steps: Optional[int]=1000,
                   strength: Optional[float]=None,
                   scheduler_type: Optional[str]="ddpm",
-                  clip_noise: Optional[bool]=False,
+                  clip_noise: Optional[bool]=True,
                   clip_noise_range: Optional[Tuple[float, float]]=(-1, 1),
                   callback: Optional[Callable[[int, int, int, paddle.Tensor],
                                               None]]=None,