Fix bug in ConditionalInput: cat along the feature dim, not the batch…

… dim (NVIDIA#7785)

Signed-off-by: anferico <[email protected]>

nemo/collections/tts/modules/submodules.py

@@ -493,7 +493,7 @@ def __init__(self, hidden_dim, condition_dim, condition_types=[]):
     def forward(self, inputs, conditioning=None):
         """
         Args:
-            inputs (torch.tensor): B x T x C tensor.
+            inputs (torch.tensor): B x T x H tensor.
             conditioning (torch.tensor): B x 1 x C conditioning embedding.
         """
         if len(self.condition_types) > 0:
@@ -510,7 +510,7 @@ def forward(self, inputs, conditioning=None):
 
             if "concat" in self.condition_types:
                 conditioning = conditioning.repeat(1, inputs.shape[1], 1)
-                inputs = torch.cat([inputs, conditioning])
+                inputs = torch.cat([inputs, conditioning], dim=-1)
                 inputs = self.concat_proj(inputs)
 
         return inputs