diff --git a/mace/dsp/ops.h b/mace/dsp/ops.h
index c686482c92e7f25489a5034963b8eb0fe064d19a..f8e5e7f60da5ee7d965de33ace1083801e6884b5 100644
--- a/mace/dsp/ops.h
+++ b/mace/dsp/ops.h
@@ -150,7 +150,6 @@ DEF_OP_WREF(QuantizedSigmoid_8)
 DEF_OP_WREF(QuantizedTanh_8)
 DEF_OP_WREF(QuantizedSoftmax_8)
 DEF_OP_WREF(QuantizedLRN_8)
-DEF_OP_WREF(Quantizedpad2d_frame_8p)
 DEF_OP_WREF(QuantizedSub_8p8to32)
 DEF_OP_WREF(QuantizedMaximum_8)
 DEF_OP_WREF(QuantizedMinimum_8)
diff --git a/mace/python/tools/BUILD b/mace/python/tools/BUILD
index b2c7ae9e285303ff1c8dbebfa60d3d988b15739f..5a048c378e9937728d334545c2b444abc57fc3f9 100644
--- a/mace/python/tools/BUILD
+++ b/mace/python/tools/BUILD
@@ -1,6 +1,9 @@
 py_library(
     name = "tf_converter_lib",
-    srcs = ["tf_converter_lib.py", "tf_dsp_converter_lib.py"],
+    srcs = [
+        "tf_converter_lib.py",
+        "tf_dsp_converter_lib.py",
+        "tf_graph_util.py"],
     srcs_version = "PY2AND3",
     deps = [
         "//mace/proto:mace_py",
diff --git a/mace/python/tools/dsp_ops.py b/mace/python/tools/dsp_ops.py
index dec73b6b55a4d5c24d7d2e7da01271ba60529070..8874b0b33a3a42fa01a07bed0fb2b601dfad2b69 100644
--- a/mace/python/tools/dsp_ops.py
+++ b/mace/python/tools/dsp_ops.py
@@ -47,13 +47,14 @@ class DspOps(object):
       'Split': 'Split_f',
       'Transpose': 'Transpose_f',
       'Concat': 'Concat_f',
+      'AddN': 'AddN_f',
     }
   def has_op(self, tf_op):
     return tf_op in self.dsp_ops
 
   def map_nn_op(self, tf_op):
     if tf_op not in self.dsp_ops:
-      raise Exception('Could not map nn op')
+      raise Exception('Could not map nn op for: ', tf_op)
     return self.dsp_ops[tf_op]
 
 
diff --git a/mace/python/tools/tf_dsp_converter_lib.py b/mace/python/tools/tf_dsp_converter_lib.py
index e2639123b559a9b7aec1ce6f39cd14219ea8ef23..28cb19991e9a8ccc22ae0d39ad8777e58dd1aab3 100644
--- a/mace/python/tools/tf_dsp_converter_lib.py
+++ b/mace/python/tools/tf_dsp_converter_lib.py
@@ -4,6 +4,7 @@ import tensorflow as tf
 import numpy as np
 from operator import mul
 from dsp_ops import DspOps
+from mace.python.tools import tf_graph_util
 
 padding_mode = {
   'NA': 0,
@@ -162,6 +163,12 @@ def add_output_node(net_def, output_node):
   node_input.output_port = 0
 
 def convert_to_mace_pb(input_graph_def, input_dim, output_node):
+  """
+    nnlib does not have batch norm, so use tensorflow optimizer to fold
+     batch norm with convolution. The fold optimization reorders ops, so
+     we sort ops first by topology.
+  """
+  input_graph_def = tf_graph_util.sort_graph(input_graph_def)
   inputs = input_dim.split(';')
   input_shape = {}
   for input in inputs:
diff --git a/mace/python/tools/tf_graph_util.py b/mace/python/tools/tf_graph_util.py
new file mode 100644
index 0000000000000000000000000000000000000000..3b899c49907901b6ce11c322ee8aa9aeeae10ce7
--- /dev/null
+++ b/mace/python/tools/tf_graph_util.py
@@ -0,0 +1,26 @@
+import tensorflow as tf
+from collections import OrderedDict
+
+def sort_graph_node(node, nodes_map, ordered_nodes_map):
+    if node.name not in ordered_nodes_map:
+        for input_tensor_name in node.input:
+            input_node_name = input_tensor_name.split(':')[
+                0] if ':' in input_tensor_name else input_tensor_name
+            if input_node_name not in nodes_map or input_node_name in ordered_nodes_map:
+                continue
+
+            input_node = nodes_map[input_node_name]
+            sort_graph_node(input_node, nodes_map, ordered_nodes_map)
+            ordered_nodes_map[input_node_name] = input_node
+        ordered_nodes_map[node.name] = node
+
+def sort_graph(graph_def):
+    nodes_map = {}
+    ordered_nodes_map = OrderedDict()
+    for node in graph_def.node:
+        nodes_map[node.name] = node
+    for node in graph_def.node:
+        sort_graph_node(node, nodes_map, ordered_nodes_map)
+    sorted_graph = tf.GraphDef()
+    sorted_graph.node.extend([node for _, node in ordered_nodes_map.iteritems()])
+    return sorted_graph
\ No newline at end of file