RF backend make_output_tensor, TF-layers disallow transpose

Fix #1410

returnn/frontend/_backend.py

@@ -269,6 +269,19 @@ def transpose(tensor: Tensor, perm: Sequence[Union[Dim, int]], *, allow_int: boo
         out.raw_tensor = backend.transpose_raw(tensor.raw_tensor, perm_)
         return out
 
+    @staticmethod
+    def make_output_tensor(tensor: Tensor, dims: Sequence[Dim], *, name: str) -> Tensor:
+        """
+        :param tensor:
+        :param dims:
+        :param name:
+        :return: tensor with dims order like in dims
+        """
+        # noinspection PyProtectedMember
+        tensor = tensor._raw_backend.transpose(tensor, dims, allow_int=False)
+        tensor = tensor.copy(name=name)
+        return tensor
+
     @staticmethod
     def expand_dims_raw(raw_tensor: T, axis: int) -> T:
         """

returnn/frontend/run_ctx.py

@@ -237,8 +237,8 @@ def mark_as_output(self, tensor: Union[Tensor, Any], name: str, *, dims: Optiona
             # We try some reasonable defaults, specifically: BTF or BF.
             dims = _default_dim_order(tensor)
         assert set(dims) == set(tensor.dims), f"mark_as_output: tensor {tensor} does not have the dims {dims}"
-        tensor = tensor.copy_transpose(dims, allow_int=False)
-        tensor = tensor.copy(name=name)
+        # noinspection PyProtectedMember
+        tensor = tensor._raw_backend.make_output_tensor(tensor, dims, name=name)
         assert name not in self.outputs.data
         self.outputs.data[name] = tensor
 

returnn/tf/frontend_layers/_backend.py

@@ -180,8 +180,12 @@ def reshape_raw(raw_tensor: Layer, shape: Union[Sequence[Union[int, Layer]], Lay
     @staticmethod
     def transpose(tensor: Tensor, perm: Sequence[Union[Dim, int]], *, allow_int: bool = False) -> Tensor:
         """transpose"""
-        assert all(isinstance(d, Dim) for d in perm)  # axis as int not supported
-        return rfl.make_layer({"class": "transpose", "from": tensor, "perm": perm}, name="transpose")
+        raise Exception("TF-layers backend: order of dims is irrelevant")
+
+    @staticmethod
+    def make_output_tensor(tensor: Tensor, dims: Sequence[Dim], *, name: str) -> Tensor:
+        """only func where we have explicitly defined dim order in the output"""
+        return rfl.make_layer({"class": "transpose", "from": tensor, "perm": dims}, name=name)
 
     @staticmethod
     def copy(tensor: Tensor) -> Tensor: