Merge pull request #619 from zhijxu-MS/add_transpose_sub

add trans opt for sub

Author: zhijxu-MS

tests/test_optimizers.py

@@ -368,6 +368,32 @@ def _make_loop(external_inputs, outputs):
         self.run_transpose_compare(["Y"], {"array": np.random.randn(10, 3, 4, 5).astype(np.float32)},
                                    model_proto, remaining_transpose_num=0)
 
+    def test_trans_with_sub(self):
+        io_shape = [2, 3, 4, 5]
+        const_shapes = [[2, 4, 5, 3], [4, 5, 3], [5, 3], [3]]
+        for trans_is_first_input in [True, False]:
+            for const_shape in const_shapes:
+                node1 = helper.make_node("Transpose", ["X"], ["Y"], perm=[0, 2, 3, 1], name="trans_a")
+                const_tensor = helper.make_tensor(name='const', data_type=TensorProto.FLOAT, dims=const_shape,
+                                                  vals=np.random.randn(*const_shape).flatten().astype(np.float32))
+                node2 = helper.make_node("Constant", [], ["const"], value=const_tensor, name="const")
+                if trans_is_first_input:
+                    node3 = helper.make_node("Sub", ["Y", "const"], ["Z"], name="sub")
+                else:
+                    node3 = helper.make_node("Sub", ["const", "Y"], ["Z"], name="sub")
+
+                node4 = helper.make_node("Transpose", ["Z"], ["res"], perm=[0, 3, 1, 2], name="trans_b")
+                graph = helper.make_graph(
+                    [node1, node2, node3, node4],
+                    "test_trans_with_sub",
+                    [helper.make_tensor_value_info("X", TensorProto.FLOAT, io_shape)],
+                    [helper.make_tensor_value_info("res", TensorProto.FLOAT, io_shape)],
+                )
+
+                model_proto = helper.make_model(graph, producer_name="onnx-tests")
+                self.run_transpose_compare(["res"], {"X": np.random.randn(2, 3, 4, 5).astype(np.float32)},
+                                           model_proto, remaining_transpose_num=0)
+
     def test_trans_output_as_graph_outputs(self):
         """
         If transpose's output is graph's output, don't optimize it.

tf2onnx/optimizer/transpose_optimizer.py

@@ -186,6 +186,7 @@ def _initialize_handlers(self):
             "Slice": self._slice_handler,
             "Split": self._split_handler,
             "Squeeze": self._squeeze_handler,
+            "Sub": self._sub_handler,
             "Tanh": self._simple_through_handler,
             "Transpose": self._transpose_handler,
         }
@@ -522,6 +523,9 @@ def _calculate_new_attr(ori_perm, ori_squeeze_axes):
             return True
         return False
 
+    def _sub_handler(self, trans, node):
+        return self._handle_node_having_branches(node)
+
     def _pad_handler(self, trans, node):
         # [N-start, H-start, W-start, C-start, N-end, H-end,  W-end, C-end]
         pads = node.get_attr('pads').ints  # [x1_begin, x2_begin...x1_end, x2_end,...]