Fix Transpose + Pad handler, for Keras app MobilenetV2 model

Author: jignparm

tests/run_pretrained_models.yaml

@@ -399,6 +399,18 @@ keras_resnet50:
   model: ResNet50
   model_type: keras
   input_get: get_ramp
+  inputs:
+    "input_1:0": [1, 224, 224, 3]
+  outputs:
+    - Identity:0
+
+keras_mobilenet_v2:
+  tf_min_version: 2.1
+  disabled: false
+  url: module://tensorflow.keras.applications.mobilenet_v2/MobileNetV2
+  model: MobileNetV2
+  model_type: keras
+  input_get: get_ramp
   inputs:
     "input_1:0": [1, 224, 224, 3]
   outputs:

tf2onnx/graph.py

@@ -504,6 +504,13 @@ def make_const(self, name, np_val, skip_conversion=False, raw=True):
         self.set_dtype(name, utils.map_numpy_to_onnx_dtype(np_val.dtype))
         return node
 
+    def copy_const(self, node, name=None):
+        """Copy a const node, using name is specified"""
+        # TODO: support attr copy starting opset 12
+        if name is None:
+            name = utils.make_name(node.name)
+        return self.make_const(name, node.get_tensor_value(as_list=False))
+
     def make_node(self, op_type, inputs, attr=None, output_count=1, outputs=None, skip_conversion=True,
                   op_name_scope=None, name=None, shapes=None, dtypes=None, domain=constants.ONNX_DOMAIN,
                   infer_shape_dtype=True):

tf2onnx/optimizer/transpose_optimizer.py

@@ -602,14 +602,19 @@ def _pad_handler(self, trans, node):
             new_pads = [pads[0], pads[3], pads[1], pads[2], pads[4], pads[7], pads[5], pads[6]]
             node.set_attr("pads", new_pads)
             return self._switch_transpose_and_node(node, trans)
-        if node.inputs[1].is_const():
-            if node.inputs[1].data_format in ["NHWC", "unkown"]:
-                pads = node.inputs[1].get_tensor_value()
+
+        input1 = node.inputs[1]
+        if input1.is_const():
+            if input1.data_format in ["NHWC", "unkown"]:
+                if not self._nodes_has_single_consumer_node([input1]):
+                    input1 = self.g.copy_const(input1)
+                    self.input[1] = input1.output[0]
+                pads = input1.get_tensor_value()
                 # NHWC->NCHW
                 new_pads = np.array([pads[0], pads[3], pads[1], pads[2], pads[4], pads[7], pads[5], pads[6]],
                                     dtype=np.int64)
-                node.inputs[1].set_tensor_value(new_pads)
-                node.inputs[1].data_format = "NCHW"
+                input1.set_tensor_value(new_pads)
+                input1.data_format = "NCHW"
             return self._switch_transpose_and_node(node, trans)
         return False