fix: the vocoder expects [B, K, T] tensors and this applies during tr…

…aining too
EveryVoiceTTS · Dec 10, 2024 · 9549082 · 9549082
1 parent 229645c
commit 9549082
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/fs2/model.py b/fs2/model.py
@@ -353,7 +353,7 @@ def _validation_global_step_0(self, batch, batch_idx) -> None:
             self.config.preprocessing.audio.output_sampling_rate,
         )
         if self.config.training.vocoder_path:
-            input_ = batch["mel"]
+            input_ = batch["mel"].transpose(1, 2)
             vocoder_ckpt = torch.load(
                 self.config.training.vocoder_path, map_location=input_.device
             )
@@ -431,7 +431,7 @@ def _validation_batch_idx_0(self, batch, batch_idx, output) -> None:
             )
 
         if self.config.training.vocoder_path:
-            input_ = output[self.output_key]
+            input_ = output[self.output_key].transpose(1, 2)
             vocoder_ckpt = torch.load(
                 self.config.training.vocoder_path, map_location=input_.device
             )