Fix batched inputs longer than target_length (clips mixed) and keep the last full segment

modeling_dasheng.py

@@ -381,14 +381,15 @@ class AudioTransformerMAE_Encoder(nn.Module):
                 else:
                     splits = torch.stack(splits[:-1], dim=0)
             else:
-                splits = torch.stack(splits[:-1], dim=0)
+                splits = torch.stack(splits, dim=0)
             n_splits = len(splits)
             x = torch.flatten(splits, 0, 1)  # spl b c f t-> (spl b) c f t
         else:
             n_splits = 1
 
         x = self.forward_features(x)
-        x = torch.reshape(x, (x.shape[0] // n_splits, -1, x.shape[-1]))
+        # (spl b) t c -> b (spl t) c: each clip's own segments, in time order
+        x = x.reshape(n_splits, -1, *x.shape[1:]).transpose(0, 1).flatten(1, 2)
         if padding_start:
             x = x[:,:padding_start, :]
         return x