Support tensorflow 2.2

.azure-pipelines/linux-CI-keras-applications-nightly.yml

@@ -64,7 +64,7 @@ jobs:
         ONNX_PATH: onnx==1.6.0
         INSTALL_KERAS:
         UNINSTALL_KERAS: pip uninstall keras -y
-        INSTALL_TENSORFLOW: pip install tensorflow==2.1.0
+        INSTALL_TENSORFLOW: pip install tensorflow==2.2.0
         INSTALL_ORT: pip install -i https://test.pypi.org/simple/ ort-nightly
         INSTALL_KERAS_RESNET: pip install keras-resnet
         INSTALL_TRANSFORMERS: pip install transformers

.azure-pipelines/win32-CI-keras-applications-nightly.yml

@@ -64,7 +64,7 @@ jobs:
         ONNX_PATH: onnx==1.6.0
         INSTALL_KERAS:
         UNINSTALL_KERAS: pip uninstall keras -y
-        INSTALL_TENSORFLOW: pip install tensorflow==2.1.0
+        INSTALL_TENSORFLOW: pip install tensorflow==2.2.0
         INSTALL_ORT: pip install onnxruntime==1.1.1
         INSTALL_KERAS_RESNET: pip install keras-resnet
         INSTALL_TRANSFORMERS: pip install transformers

applications/nightly_build/test_keras_applications_v2.py

@@ -54,7 +54,6 @@ def test_InceptionV3(self):
         keras.backend.set_learning_phase(0)
         InceptionV3 = keras.applications.inception_v3.InceptionV3
         model = InceptionV3(include_top=True)
-        model.save('inception.h5')
         res = run_image(model, self.model_files, img_path, target_size=299, tf_v2=True)
         self.assertTrue(*res)
 

keras2onnx/parser.py

@@ -482,20 +482,28 @@ def _get_output_nodes(node_list, node):
     return [n_ for n_ in node_list if n_ not in nodes_has_children]  # need to keep the order.
 
 
+def _filter_out_input(node_name):
+    # tf.keras BN layer sometimes create a placeholder node 'scale' in tf2.x. 
+    # Given bn layer will be converted in a whole layer, it's fine to just filter this node out.
+    filter_out = re.match(r"batch_normalization_\d+\/scale$", node_name)
+    filter_out = filter_out or re.match(r"batch_normalization_\d+\/cond/input", node_name) # since tf 2.2
+    return filter_out
+
+
 def _advance_by_input(cur_node, layer_nodes, subgraph, inputs, graph_inputs, q_overall):
     for input_ in cur_node.inputs:
         predecessor = input_.op
         if is_placeholder_node(predecessor):
-            # tf.keras BN layer sometimes create a placeholder node 'scale' in tf2.x.
-            # Given bn layer will be converted in a whole layer, it's fine to just filter this node out.
-            if not re.match(r"batch_normalization_\d+\/scale$", predecessor.name):
+            if not _filter_out_input(predecessor.name):
                 inputs.add(predecessor)
                 graph_inputs.add(predecessor)
+                continue
         if predecessor in layer_nodes or len(layer_nodes) == 0:
             subgraph.append(predecessor)
         else:
-            inputs.add(predecessor)
-            q_overall.put_nowait(predecessor)
+            if not _filter_out_input(predecessor.name):
+                inputs.add(predecessor)
+                q_overall.put_nowait(predecessor)
 
 
 def _visit_nodelist(activated_keras_nodes, input_nodes, layer_key,
@@ -735,7 +743,15 @@ def parse_graph_modeless(topo, graph, target_opset, input_names, output_names, k
 
     for ts_i_ in input_tensors:
         var_type = _adjust_input_batch_size(infer_variable_type(ts_i_, target_opset))
-        str_value = ts_i_.name
+        if ts_i_.name.endswith(':0'):
+            str_value = ts_i_.name[:-2]
+            op = top_level.declare_local_operator(TYPES.Identity)
+            var0 = top_level.get_local_variable_or_declare_one(str_value, var_type)
+            var1 = top_level.get_local_variable_or_declare_one(ts_i_.name, var_type)
+            op.add_input(var0)
+            op.add_output(var1)
+        else:
+            str_value = ts_i_.name
         top_level.get_local_variable_or_declare_one(str_value, var_type)
         topo.raw_model.add_input_name(str_value)
 

tests/test_utils.py

@@ -137,7 +137,9 @@ def run_onnx_runtime(case_name, onnx_model, data, expected, model_files, rtol=1.
     if expected is None:
         return
 
-    if not isinstance(expected, list):
+    if isinstance(expected, tuple):
+        expected = list(expected)
+    elif not isinstance(expected, list):
         expected = [expected]
 
     res = all(np.allclose(expected[n_], actual[n_], rtol=rtol, atol=atol) for n_ in range(len(expected)))