Upsample has scales in attributes up until opset 8

Author: NikolasMarkou

tests/test_optimizers.py

@@ -1177,26 +1177,28 @@ def test_cast_back_to_back_non_const_mixed_types(self):
         self.run_and_compare(["res", "res2", "res3"], {"u": np.random.randn(1, 2, 3).astype(np.float32)}, model_proto,
                              "Cast", 5)
 
+    @check_opset_max_version(8, "until opset 8 scales is in attributes")
     def test_upsample_all_ones_removed(self):
+        shape = (1, 1, 32, 32)
         node1 = helper.make_node(
-            "Upsample",
-            ["X"],
-            ["Y"],
+            op_type="Upsample",
+            inputs=["X"],
+            outputs=["Y"],
             scales=[1., 1., 1., 1.],
             name="upsample1")
 
         graph = helper.make_graph(
             [node1],
             "test_upsample_all_ones",
-            [helper.make_tensor_value_info("X", TensorProto.FLOAT, (1, 32, 32, 1))],
-            [helper.make_tensor_value_info("Y", TensorProto.FLOAT, (1, 32, 32, 1))],
+            [helper.make_tensor_value_info("X", TensorProto.FLOAT, shape)],
+            [helper.make_tensor_value_info("Y", TensorProto.FLOAT, shape)],
         )
 
         model_proto = self.make_model(graph, producer_name="onnx-tests")
 
         self.run_and_compare(
             ["Y"],
-            {"X": np.random.randn(1, 32, 32, 1).astype(np.float32)},
+            {"X": np.random.randn(*shape).astype(np.float32)},
             model_proto,
             "Upsample",
             0)

tf2onnx/optimizer/upsample_optimizer.py

@@ -3,6 +3,7 @@
 """
 
 from __future__ import unicode_literals
+from onnx import helper
 
 from .optimizer_base import GraphOptimizerBase
 
@@ -14,21 +15,24 @@ class UpsampleOptimizer(GraphOptimizerBase):
 
     def __init__(self):  # pylint: disable=useless-super-delegation
         super(UpsampleOptimizer, self).__init__()
+        self._g = None
 
     def _optimize(self, graph):
         return self._apply_optimization(
             graph,
             self._optimize_at_current_graph_level)
 
     def _optimize_at_current_graph_level(self, graph):
+        self._g = graph
         # replace upsample node with all ones in scale with identity node
-        for n in graph.get_nodes():
+        for n in self._g.get_nodes():
             if n.type == "Upsample":
-                scales = n.get_attr_value("scales")
-                if all([s == 1 for s in scales]):
-                    n.type = "Identity"
-                    if len(n.input) > 0:
-                        n.input = [n.input[0]]
-                    self.logger.debug("replacing " + n.name +
-                                      " with Identity operation")
-        return graph
+                # upsample in opset <=8 has scales in attributes
+                if self._g.opset <= 8:
+                    scales = n.get_attr_value("scales")
+                    if scales and all([float(s) == 1. for s in scales]):
+                        n.type = "Identity"
+                        self.logger.debug("replacing " + n.name +
+                                          " with Identity operation ")
+                # upsample in opset > 8 has scales in input[1]
+        return self._g