diff --git a/mace/python/tools/converter_tool/caffe_converter.py b/mace/python/tools/converter_tool/caffe_converter.py
index 3ff2fa231e2b461481a41c966b65831a78c955bc..56e8b645089758a23d7d70017e9032b3a6273432 100644
--- a/mace/python/tools/converter_tool/caffe_converter.py
+++ b/mace/python/tools/converter_tool/caffe_converter.py
@@ -187,6 +187,7 @@ class CaffeConverter(base_converter.ConverterInterface):
             'BatchNorm': self.convert_folded_batchnorm,
             'Crop': self.convert_crop,
             'Scale': self.convert_scale,
+            'ShuffleChannel': self.convert_channel_shuffle,
         }
         self._option = option
         self._mace_net_def = mace_pb2.NetDef()
@@ -656,3 +657,14 @@ class CaffeConverter(base_converter.ConverterInterface):
 
             ConverterUtil.add_data_format_arg(biasadd_op,
                                               DataFormat.NCHW)
+
+    def convert_channel_shuffle(self, caffe_op):
+        op = self.convert_general_op(caffe_op)
+        param = caffe_op.layer.shuffle_channel_param
+        op.type = MaceOp.ChannelShuffle.name
+
+        group_arg = op.arg.add()
+        group_arg.name = MaceKeyword.mace_group_str
+        group_arg.i = 1
+        if param.HasField('group'):
+            group_arg.i = param.group
diff --git a/mace/python/tools/converter_tool/shape_inference.py b/mace/python/tools/converter_tool/shape_inference.py
index da6541f499459532e9ead7b277cb1192a9b3fdd1..aeb19022badc324855d82820a61d5b9a2a7f1cb1 100644
--- a/mace/python/tools/converter_tool/shape_inference.py
+++ b/mace/python/tools/converter_tool/shape_inference.py
@@ -48,6 +48,7 @@ class ShapeInference(object):
             MaceOp.FullyConnected.name: self.infer_shape_fully_connected,
             MaceOp.Crop.name: self.infer_shape_crop,
             MaceOp.BiasAdd.name: self.infer_shape_general,
+            MaceOp.ChannelShuffle.name: self.infer_shape_channel_shuffle,
         }
 
         self._net = net
@@ -220,3 +221,7 @@ class ShapeInference(object):
         mace_check(len(op.input) == 2, "crop layer needs two inputs")
         output_shape = self._output_shape_cache[op.input[1]]
         self.add_output_shape(op, [output_shape])
+
+    def infer_shape_channel_shuffle(self, op):
+        output_shape = self._output_shape_cache[op.input[0]]
+        self.add_output_shape(op, [output_shape])
diff --git a/third_party/caffe/caffe.proto b/third_party/caffe/caffe.proto
index cec617b99b8a5cde9e93e2bb14be0cab21794908..54ccf20ca2378f7e15881930333f0014c1923b63 100644
--- a/third_party/caffe/caffe.proto
+++ b/third_party/caffe/caffe.proto
@@ -421,6 +421,7 @@ message LayerParameter {
   optional ThresholdParameter threshold_param = 128;
   optional TileParameter tile_param = 138;
   optional WindowDataParameter window_data_param = 129;
+  optional ShuffleChannelParameter shuffle_channel_param = 164;
 }
 
 // Message that stores parameters used to apply transformation
@@ -1439,3 +1440,7 @@ message PReLUParameter {
   // Whether or not slope parameters are shared across channels.
   optional bool channel_shared = 2 [default = false];
 }
+
+message ShuffleChannelParameter {
+  optional uint32 group = 1[default = 1]; // The number of group
+}