diff --git a/mace/python/tools/converter_tool/tensorflow_converter.py b/mace/python/tools/converter_tool/tensorflow_converter.py
index 2e373138be34d386c002234dda5dacc0bee7297e..f6f0877de9b6ecda45a3117d2b22241ad0306203 100644
--- a/mace/python/tools/converter_tool/tensorflow_converter.py
+++ b/mace/python/tools/converter_tool/tensorflow_converter.py
@@ -39,6 +39,7 @@ tf_dilations_str = 'dilations'
 tf_data_format_str = 'data_format'
 tf_kernel_str = 'ksize'
 tf_epsilon_str = 'epsilon'
+tf_is_training_str = 'is_training'
 tf_align_corners = 'align_corners'
 tf_block_size = 'block_size'
 tf_squeeze_dims = 'squeeze_dims'
@@ -516,6 +517,10 @@ class TensorflowConverter(base_converter.ConverterInterface):
         op = self.convert_general_op(tf_op)
         op.type = MaceOp.FoldedBatchNorm.name
 
+        is_training = tf_op.get_attr(tf_is_training_str)
+        assert is_training is False, 'Only support batch normalization ' \
+            'with is_training False, but got %s' % is_training
+
         gamma_value = tf_op.inputs[1].eval().astype(np.float32)
         beta_value = tf_op.inputs[2].eval().astype(np.float32)
         mean_value = tf_op.inputs[3].eval().astype(np.float32)