fix convert model bugs when convert dense_relative_transformer models

fairseq/modules/relative_multihead_attention.py

@@ -273,4 +273,4 @@ class RelativeMultiheadAttention(nn.Module):
     x_tz_matmul = torch.bmm(x_t, z).transpose(0, 1).view(batch_size_mul_head, length, -1)
     x_tz_matmul = torch.bmm(x_t, z).transpose(0, 1).view(batch_size_mul_head, length, -1)
 
 
     #assert xy_matmul.size() == x_tz_matmul.size()
     #assert xy_matmul.size() == x_tz_matmul.size()
-    return xy_matmul + x_tz_matmul
+    return xy_matmul + x_tz_matmul
\ No newline at end of file

scripts/convert_t2t_to_fairseq.py

@@ -473,7 +473,7 @@ def convert_settings(settings):
     args['arch'] = 'relative_transformer'
     args['arch'] = 'relative_transformer'
 
 
   if use_relative_position_representation and use_dense:
   if use_relative_position_representation and use_dense:
-    args['arch'] = 'relative_dense_transformer'
+    args['arch'] = 'dense_relative_transformer'
   return argparse.Namespace(**args)
   return argparse.Namespace(**args)