Handle dynamic shape in apply_reshape (#36)

* Handle dynamic shape in apply_reshape

onnxconverter_common/onnx_ops.py

@@ -548,7 +548,7 @@ def apply_relu(scope, input_name, output_name, container, operator_name=None):
 
 
 def apply_reshape(scope, input_name, output_name, container, operator_name=None, desired_shape=None):
-    if len(list(i for i in desired_shape if i is not None and i < 0)) > 1:
+    if not isinstance(desired_shape, str) and len(list(i for i in desired_shape if i is not None and i < 0)) > 1:
         raise ValueError('There can only be one -1 in the targeted shape of a Reshape but got %s' % desired_shape)
 
     name = _create_name_or_use_existing_one(scope, 'Reshape', operator_name)
@@ -557,9 +557,11 @@ def apply_reshape(scope, input_name, output_name, container, operator_name=None,
         container.add_node('Reshape', input_name, output_name, op_version=1, name=name, shape=desired_shape,
                            consumed_inputs=[0])
     else:
-        # The shape attribute of Reshape becomes a tensor input, so we create one tensor to store that attribute.
-        desired_shape_name = scope.get_unique_variable_name('shape_tensor')
-        container.add_initializer(desired_shape_name, onnx_proto.TensorProto.INT64, [len(desired_shape)], desired_shape)
+        if isinstance(desired_shape, str):
+            desired_shape_name = desired_shape
+        else:
+            desired_shape_name = scope.get_unique_variable_name('shape_tensor')
+            container.add_initializer(desired_shape_name, onnx_proto.TensorProto.INT64, [len(desired_shape)], desired_shape)
 
         # Create ONNX Reshape operator
         if isinstance(input_name, list):